Rational Optimization of Small Molecules for Alzheimer's Disease

Rational Optimization of Small Molecules for Alzheimer’s Disease Premortem Diagnosis

DISSERTATION

Presented in Partial Fulfillment of Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University

Katryna Cisek

Biophysics Graduate Program

The Ohio State University

2012

Dissertation Committee:

Jeffrey A. Kuret, PhD, Advisor

Christopher Hadad, PhD

Pui-Kai Tom Li, PhD

Michael Tweedle, PhD

Katryna Cisek

2012

ABSTRACT

Alzheimer’s disease is a debilitating, progressive neurodegenerative disorder that affects a large percentage of the elderly population. Currently, there is no definitive premortem diagnosis and no cure. Pathological brain tissue examination upon autopsy is used to identify the characteristic neurofibrillary tangles of tau protein and amyloid deposits that define this disease. These protein aggregates accumulate for many years before the onset of clinical symptoms; therefore, their in situ detection would be an invaluable tool for early premortem diagnosis. Because radiolabeled small molecules used for whole brain imaging, such as those used for PET imaging, have the advantage of tracking the spatiotemporal pattern of molecular targets, this approach has tremendous utility for neurodegenerative disorders. More specifically, the detection of tau-bearing neurofibrillary tangles is especially promising as the amount and spatiotemporal pattern of tau-aggregate deposition is the gold standard of postmortem disease assessment, as it correlates with loss of neurons. The main challenge in the development and optimization of a tau-selective imaging agent is the structure of these aggregates, which adopts a cross- ß-sheet of interdigitating monomers. Although multiple scaffold classes have been reported to bind cross-ß-sheet structure, their mechanism of binding and their ability to selectively bind different aggregates of varying protein composition are not well understood. There are no crystal or NMR structures that would reveal the atomic-level binding modes of this interaction. Most small molecule development studies focus on iterative structure-activity relationship modifications and testing of known scaffolds, such as the benzothiazole dye and commonly used tissue staining agent Thioflavin T. Even though quantitative structure activity relationship studies have been employed to investigate amyloid-binding

compounds, these retrospective studies have not elucidated any novel compound molecular properties that could explain the mechanism of interaction. Moreover, these studies have not rationalized the binding activity or selectivity of cross-ß-sheet-binding ligands in the context of computational binding models. The project described herein is a ligand-based quantitative structure activity relationship approach to identify descriptors of binding affinity and selectivity for two series of over fifty closely related benzothiazole derivatives and indolinones reported to displace Thioflavin T fluorescent probe from synthetic aggregates composed of tau and ß-amyloid peptide. This is a novel computational approach in that it seeks to elucidate binding affinity, selectivity as well as binding site density in the context of existing experimental and computational data. I developed a two-step regression analysis for the identification of two-dimensional “global” descriptors of ligand potency and selectivity, in conjunction with a three-dimensional analysis for the identification of volumetric “local” regions for the optimization of R-group substituents of potent and selective ligands. The resulting models were statistically robust and predictive and provided clear guidelines for ligand optimization. For the indolinone series, the model was successful in predicting novel potent analogues, which upon synthesis and testing were shown to be very promising candidates for further preclinical studies. The dissertation comprises of six chapters; the 1st chapter introduces the key challenges of Alzheimer’s disease premortem diagnosis, namely the tau-based imaging strategy for early and differential diagnosis, and the current status of the field, including the limitations of current diagnostic agents, quantitative structure-activity relationships and computational models of compounds binding aggregates. The 2nd chapter is a feasibility study confirming that computational models are capable of identifying not only potency-driving features of ligands, but also descriptors that account for compound selectivity. Chapter 3 explains the mechanism of polarizability within a neural background that was identified in Chapter 2 as an important feature for affinity and selectivity, as well as its implication for binding site density. The following two chapters are quantitative structure-affinity relationship analyses of two different compound libraries, the benzothiazole-aryls and indolinones. The studies interpret features

iii

identified as the most significant drivers of binding affinity and selectivity and provide rational approaches for optimizing the scaffolds. Chapter 6 concludes with limitations of the computational studies and the significance of polarizability for aggregation inhibitors.

ACKNOWLEDGEMENTS

First and foremost I would like to thank my advisor, Dr. Jeff Kuret, who has been an excellent advisor, teacher and mentor. His leadership has allowed me to grow as a scientist, to develop my critical thinking and writing skills, and most importantly to fearlessly pursue challenging projects. I feel very lucky to have been a student in his lab. Big thanks to all of my lab colleagues. Dr. Nicolette Honson started the project and taught me many experimental methods; she and Jordan Jensen generated a lot of data in the lab and were instrumental to the success of the project and many publications that followed. Thank you to Kelsey Schafer for TEM imaging, as well as experimental work on inhibitors where I was able to contribute computational results. I would also like to thank current lab members Kristin Funk for confocal images, and Grace Cooper for continuing the project. Thank you to former lab members Edward Chang, Sohee Kim, Swati Naphade, Vandana Kumari and In Hee Park who were wonderful to work with. I would like to thank my committee; Dr. Christopher Hadad, for sharing his expertise with computational methods, Dr. Michael Tweedle and Dr. Pui-Kai Tom Li for their collaboration on the project. Dr. Michael Zhu and Dr. Dennis McKay have generously provided the use of their plate readers and Dr. Michael Darby for synthesis. Very special thanks to my family, especially my parents, Anna and Roman Cisek, whose unconditional support was paramount to my success in graduate school, as well as my brothers, Andrew and Lukas, sister-in-law Agnes, and other family and friends who have encouraged me from the very beginning.

VITA

2005……………………………………………Bachelor of Science in Computer Science University of Northern Iowa

2005……………………………………………………..……Bachelor of Arts in German University of Northern Iowa

2011……………………………………..………..………Master of Science in Biophysics Ohio State University

2005-Present…………………………………………………Graduate Research Associate Ohio State University

PUBLICATIONS

Jensen, J., Cisek, K., and Kuret, J. (2012) Benzothiazole-aryl compounds as potential tau- directed imaging agents [In Preparation].

Cisek, K., Jensen, J., Funk, K., Bhasin, D., Li, PK, Darby, M., and Kuret, J. (2012) Indolinone compounds as potential tau-directed imaging agents [In Preparation].

Cisek K, Kuret J. QSAR-guided development of tau-directed imaging agents. (2012) Int J Alzheimers Dis. [In Preparation].

Cisek K, Jensen JR, Honson NS, Schafer KN, Cooper GL and Kuret J. (2012) Ligand electronic properties modulate tau filament binding site density [submitted].

Cisek K, Kuret J. (2012) QSAR studies for prediction of cross-β sheet aggregate binding affinity and selectivity. Bioorg Med Chem. 20(4):1434-41.

Kuret J, Jensen J, Cisek K. Benzothiazole and indolinone ligands for pre-mortem detection of aggregated tau proteins, provisional US patent filed July 27, 2011 [OSU Tech Licensing ID 2011-196].

Jensen JR, Cisek K, Funk KE, Naphade S, Schafer KN, Kuret J. (2011) Research towards tau imaging. J Alzheimers Dis. 26 Suppl 3:147-57.

Jensen JR, Cisek K, Honson NS, Kuret J. (2011) Ligand polarizability contributes to tau fibril binding affinity.Bioorg Med Chem. 19(17):5147-54.

Schafer KN, Murale DP, Kim K, Cisek K, Kuret J, Churchill DG. (2011) Structure- activity relationship of cyclic thiacarbocyanine tau aggregation inhibitors. Bioorg Med Chem Lett. 21(11):3273-6.

Kim S, Jensen JR, Cisek K, Funk KE, Naphade S, Schafer K, Kuret J. (2010) Imaging as a strategy for premortem diagnosis and staging of tauopathies, Curr Alzheimer Res. 7(3):230-4.

Fuh B, Sobo M, Cen L, Josiah D, Hutzen B, Cisek K, Bhasin D, Regan N, Lin L, Chan C, Caldas H, DeAngelis S, Li C, Li PK, Lin J. (2009) LLL-3 inhibits STAT3 activity, suppresses glioblastoma cell growth and prolongs survival in a mouse glioblastoma model, Br J Cancer. 100(1):106-12.

Bhasin D, Cisek K, Pandharkar T, Regan N, Li C, Pandit B, Lin J, Li PK. (2008) Design, synthesis, and studies of small molecule STAT3 inhibitors, Bioorg Med Chem Lett. 18(1):391-5.

FIELD OF STUDY

Major Field: Biophysics Graduate Program

vii

TABLE OF CONTENTS

Abstract……………………………………………………………………………………ii Acknowledgements………………………………....……………………………….…….v Vita…………………………………………………………………………………..……vi List of Tables…………………………………………………………………………….xii List of Figures……………………………………………………………………...…....xiii List of Abbreviations…………………………………………………………...…….…xiv Chapters 1. Introduction……………………………………………………………………………..1 1.1. Premortem diagnostic agents for Alzheimer’s disease .…...……..………...…….1 1.2. Tau-based imaging as a strategy for differential diagnosis.……..……….....……3 1.3. Structural complexity of the molecular target in disease state ……………..……4 1.4. Computational analysis of potent and selective aggregate-binding ligands ..……5 1.5. QSAR approaches to aggregate binding targets ………...………………….…....7 1.6. Computational modeling of ligand-aggregate binding interactions ………….….9 1.7. Summary……………………………………………………………...…………11 1.8. Figures…………………………………………………………………………...12 2. Qsar studies for prediction of cross- sheet aggregate binding affinity and selectivity………………..………………..……………………………………………...17 2.1. Introduction…………………………………………………………...…………17 2.2. Experimental procedures………………………………………………….…….19 2.2.1. Bioactivity Data………………………………….………………………..19 2.2.2. Chemical structures and calculation of molecular descriptors…..……..…19 2.2.3. Model generation………………………………………………………....20 2.2.4. Model validation ………...... ………………..…………………..….……21 2.2.5. Statistical methods …………...……………..…………………………….21

viii

2.3. Results…………………………………………………………………………...22 2.3.1. Experimental data, descriptor sets and workflow…………...……...…..…22 2.3.2. Descriptors for relative ThT displacement affinity…………………...... …22 2.3.3. Construction and validation of predictive MLR models………..…..…..…24 2.3.4. Descriptors for relative ThT displacement selectivity……………….…....26 2.4. Discussion……………………………………………………………………….27 2.5. Tables……...…………………………………………………………………….31 2.6. Figures………...………………………………………………………….……...39 3. The mechanism of polarizability-mediated ligand binding potential on tau fibrils…...44 3.1. Introduction……………………………………………………………….…..…44 3.2. Experimental procedures………………………………………………….…….46 3.2.1. Materials……………………………………………………………..……46 3.2.2. Protein preparation…………..……………………………………….……46 3.2.3. Filament length distributions………………...…………………..….…….47 3.2.4. Fluorescence displacement assays…………….…...……….……………..47 3.2.5. Radioligand binding assays…………...……….…...……….……………..47 3.2.6. Spectrophotometry…………………………….…...……….……………..48 3.2.7. Analytical methods……………...…………….…...……….……………..48 3.2.8. Computational chemistry……….…………….…....……….……………..49 3.3. Results ………………..…………………………………….……………….…..49 3.3.1. Tau filament populations for binding studies……………………………..49 3.3.2. Displacement probes…………………………………. …………...….…..50 3.3.3. Benzothiazole ligands interact differentially with tau aggregates……...…52 3.3.4. Electronic properties of tau aggregate binding ligands………….……...…53 3.4. Discussion……………………………………………………………………….54 3.5 Tables…………………………………………………………………………….58 3.6 Figures……………………………………………………………………….…...59 4. QSAR modeling of benzothiazole-aryls for tau-directed imaging agent development……………………………………………………………………………...66 4.1. Introduction…………………………………………………………………..…66

4.2. Methods………………..……………………………………………………..…67 4.2.1. Materials…………………………………………………………………..68 4.2.2. Fibrillization.………………………………………………………………68 4.2.3. Displacement assays…………………...……………………………...…..68 4.2.4. Computational procedures and QSAR model generation…….…….…….69 4.3. Results……………………………………………………………………..……69 4.3.1.Workflow of QSAR models for affinity and selectivity .………… ……...69 4.3.2. Classical PLR model for identification of most significant binding affinity descriptors………………………………………………..………..…………..…71 4.3.3. Classical MLR model for identification binding affinity and selectivity descriptors…………………………………………………………………….….73 4.3.4. 3D-QSAR model for affinity and selectivity in the context of 2D descriptors…………………………………………………………….……….....74 4.3.5. QSAR predictivity assessment using external validation………………...76 4.4. Discussion…………………………………………………………………...…..76 4.5. Tables……………………………………………………………………….…...80 4.6. Figures…………………………………………………………………………...90 5. Qsar modeling of 3-substituted indolinones for tau-directed imaging agent development……………………………………………………………………………...93 5.1. Introduction………………………………………………………………...……93 5.2. Experimental procedures……………………………………………………..…95 5.2.1. Materials…………………………………………………………………..95 5.2.2. Fibrillization.………………………………………………………………95 5.2.3. Displacement assays…………………...……………………………...…..95 5.2.4. Computational procedures and QSAR model generation……..…….…….95 5.3. Results……………………………………………………………….……..……95 5.3.1.Workflow of QSAR models for affinity and selectivity .…….…… ……...95 5.3.2. Classical PLR model for identification of most significant binding affinity descriptors……………………………………………..…..…………………..…96

5.3.3. Classical MLR model for identification binding affinity and selectivity descriptors…………………………………………………………………….….98 5.3.4. 3D-QSAR model for affinity and selectivity in the context of 2D descriptors……………………………………………………………………....100 5.3.5. QSAR predictivity assessment using external validation and novel analogue design…………………………………………………………………………...101 5.4. Discussion……………………………………………………………..…...…..102 5.5. Tables……………………………………………………………………….….104 5.6. Figures…………………………………………………………………..……...114 6. Conclusions and future directions…………………………………………….……...117 References………………………………………………………………………...…….121

LIST OF TABLES

Table 2.1. Compound structures and characteristics...……….………………….…… 31 Table 2.2. Compound MLR descriptors ………..…...………………………….…… 34 Table 2.3. PLR models and statistics for Aβ and insulin………..……………………..35

Table 2.4. Top-ranked PLR descriptors for Aβ40 and insulin pAC50 ………..……….36 Table 2.5. Correlation matrix…………………………………………..………………37 Table 2.6. MLR external validation…………………………………..………………. 38 Table 3.1. Compound structures and characteristics…………………..……………....58 Table 4.1. Compound structures and characteristics……...………….…..…………....80 Table 4.2. Compound MLR descriptors ...... ……....…………………………….…… 85 Table 4.3. Training set statistics for Tau and Aß……………………….………….…..86 Table 4.4. Test set statistics for Tau and Aß…………………………….………….….87

Table 4.5. Top-ranked PLR descriptors for tau and Aß40 pAC50 ………..…………...88 Table 4.6. Correlation matrix………………………………………………..…………89 Table 5.1 Compound structures and characteristics………………...…….………....104 Table 5.2. Compound MLR descriptors …………..……………..…………….…… 109 Table 5.3. Training set statistics for Tau and Aß …………………………..………...110 Table 5.4. Test set statistics for Tau and Aß………………………………..………...111

Table 5.5. Top-ranked PLR descriptors for tau and Aß42 pAC50………..…………..112 Table 5.6. Correlation matrix………………………………………………..………..113

xii

LIST OF FIGURES

Figure 1.1. AD pathology in human brain…………….……………………………….. 12 Figure 1.2. Diversity of cross-ß-sheet binding ligands………………………………….13 Figure 1.3. Tau-bearing fibrils in AD pathology. ………………………………………14 Figure 1.4. Ligand – protein binding models……………………………………….…. 15 Figure 1.5. Ligand selectivity for cross-ß-sheet structures…………..…………….…....16 Figure 2.1. QSAR modeling workflow …………………………………………………39 Figure 2.2. Correlation plots for MLR models …………………………………………40 Figure 2.3. Comparison of Aβ and insulin MLR equation coefficients. ………………41 Figure. 2.4. Structural features influencing ThT displacement activity ……….……….42 Figure 2.5. Reported binding modes for Pittsburgh Compound B ……….…………….43 Figure 3.1. Tau filament morphology and length distribution. …………………………59 Figure 3.2. Optical properties of probes and ligands. ………………………………...... 60 Figure 3.3. Probe interactions with filamentous 2N4R- and PHF-tau. …………………61 Figure 3.4. ThS displacement assays for tau filament binding………………………… 62 Figure 3.5. 125I-IMSB displacement assays for tau filament binding……………….…..63 Figure 3.6. Ligand electrostatic potential surfaces……………………………………...64 Figure 3.7. Ligand electronic difference density maps. ………………………………..65 Figure 4.1. QSAR modeling workflow……………………………………………….. 90 Figure 4.2. Pharmacophore alignment …………………………………………………91 Figure 4.3. 3D-QSAR contour maps for benzothiazoles.…….……………………….. 92 Figure 5.1. QSAR modeling workflow……………………………………………….114 Figure 5.2. Pharmacophore alignment………………………………………………...115 Figure 5.3. 3D-QSAR contour maps for indolinones……….…………….…………..116

xiii

LIST OF ABBREVIATIONS

2N4R, two N-terminal region four microtubule binding repeat region; A1-40, amyloid beta 40-mer peptide; A1-42, amyloid beta 42-mer peptide; AC50, activity (ligand) concentration for 50% probe displacement, AD, Alzheimer’s disease; BBB, blood-brain barrier; BTA, benzothiazole-aryl; CBD, cortical basal degeneration; CSF, cerebral spinal fluid; FTLD, frontotemporal lobar degeneration; HBA, hydrogen-bond acceptor; HBD, hydrogen-bond donor; LBD, Lewy Body dementia; MCI, mild cognitive impairment; MLR, multiple least-squares regression; PD, Parkinson’s disease; PET, positron emission tomography; PiD, Pick’s disease; PK/PD, pharmacokinetic/pharmacodynamic; PHF, paired helical filament; PLR, partial least-squares regression; PSP, progressive supranuclear palsy; SA, surface area; SAR, structure-activity relationship; SEE, standard error of the estimate; ThS, Thioflavin S; ThT, Thioflavin T; tPSA, topological polar surface area; QSAR, quantitative structure-activity relationships; Vol, volume;

xiv

CHAPTER 1

INTRODUCTION

1.1. Premortem diagnostic agents for Alzheimer’s disease

Alzheimer’s disease (AD) is a progressive dementia that leads to cognitive decline, memory loss and other behavioral and personality changes (Dubois, Feldman et al. 2007). It is the most prevalent form of neurodegeneration, affecting more than 5% of the over 65-year-old population, with prevalence expected to exceed 115 million persons worldwide by 2050 (Assoc 2012). In addition to the emotional and physical burdens of dementia, not only for the patient but the family and caregivers, the staggering costs for health care and hospice services are estimated to reach $200 billion, not accounting for unpaid caregivers, in the United States this year alone (Assoc 2012). Currently, there is no definitive and reliable pre-mortem diagnostic for AD; live patients are evaluated on the basis of their medical history and various physical and mental examinations (Dubois, Feldman et al. 2007). This type of diagnosis is usually only able to identify late stage disease after significant cognitive decline, and has poor sensitivity and specificity (80% and 70%, respectively), because the differential diagnosis is complicated by similar clinical presentation among dementias (Knopman, DeKosky et al. 2001). Moreover, treatments that slow the progression of cognitive decline and manage the behavioral symptoms are palliative at best and have only short-term effectiveness (Melnikova 2007). There is no cure for AD.

Definitive diagnosis of AD is obtained only at autopsy. This is because despite its dementing phenotype, AD is defined on the basis of characteristic lesions in the brain: extracellular amyloid plaques, composed of the Aβ peptide and intracellular neurofibrillary tangles (NFT), composed of the microtubule-associated protein tau (Figure 1.1A and 1.1B) (Berchtold and Cotman 1998). The deposition of lesions in AD follows a well-defined pathway, where the density and spatial distribution of the lesions indicates disease severity, cognitive decline and neurodegeneration (Braak and Braak 1991). This is assessed postmortem on the basis of brain tissue section staining with fluorescent molecules, such as Thioflavin T (ThT), and captured with confocal microscopy (Figure 1.1C). However, Aß plaque deposition alone poorly correlates with neuronal loss, and considerable plaque load has been identified in cognitively normal elderly populations (Braak and Braak 1991; Terry, Masliah et al. 1991; Masliah, Mallory et al. 1993). Conversely, the detection of tau-bearing NFTs has advantages over amyloid plaques because: 1) NFT formation precedes amyloid deposition in particular regions of the brain, such as the entorhinal cortex, decades before clinical symptoms appear, and may identify very early stages of disease or at risk populations (Duyckaerts and Hauw 1997; Morsch, Simon et al. 1999), 2) the spatiotemporal deposition of NFTs closely correlates with neurodegeneration and cognitive decline (Braak and Braak 1991; Ghoshal, Garcia-Sierra et al. 2002; Royall, Palmer et al. 2002), and 3) the gold standard of postmortem AD assessment, Braak staging, is based on NFT deposition (Braak and Braak 1991). Therefore, an NFT-selective diagnostic agent could have tremendous utility for tracking disease progression, identifying at risk populations, and monitoring utility of disease-modifying therapies. Diagnostic agents must meet the criteria for biomarkers established by expert consensus if they are to be approved for use in the clinic. The recently reevaluated AD biomarker criteria include identification of AD neuropathology, validation in confirmed cases, 80% specificity for accurately diagnosing AD (rate of true AD positives) and 80% sensitivity for accurately differentiating AD from other dementias (rate of false AD positives) in addition to reliability, reproducibility, non-invasiveness, cost-effectiveness and ease (1998; Albert, DeKosky et al. 2011). Because tau and Aβ peptide accumulate in

the cerebrospinal fluid (CSF) that occupies the subarachnoid space and the ventricles to protect and cushion the brain, CSF immunoassays have been used to detect the levels of tau and Aß peptides present in the lesions. However, CSF collection is an invasive lumbar puncture procedure and the assays have not consistently met the specificity and sensitivity criteria for AD cases vs. control (Shaw, Vanderstichele et al. 2009). In addition to lacking standardization, CSF immunoassays cannot track the spatiotemporal lesion distribution for AD staging and differential diagnosis (de Leon, DeSanti et al. 2004; Hampel, Mitchell et al. 2004). Combined with emerging therapies for AD, PET imaging is the only direct in situ modality for premortem diagnosis of AD which has the potential to change the standard of care for this deadly disease. 1.2. Tau-based imaging as a strategy for differential diagnosis Whole brain imaging methods rely on radioactive isotope decay to reveal the spatial distribution of biological targets with radiolabeled brain penetrant probes detectable by modern scanners (Baert 2008). For example, positron emission tomography (PET) utilizes probes containing a positron-emitting radioisotope. Upon collision of positrons with tissue electrons the annihilation of both particles results in release of coincident photons that can be captured in a detector ring (Ter-Pogossian, Phelps et al. 1975). To leverage this approach for AD diagnosis will require a brain-penetrant probe capable of specifically binding AD-related brain lesions (Nordberg, Rinne et al. 2010). To date, only one Aß-targeting 18F-radiolabeled PET agent, Amyvid, has been approved for use in the clinic (Hsiao, Huang et al. 2012), while others, such as 18F-Flutemetamol and 18F- Florbetaben (Klunk, Engler et al. 2004; Choi, Golding et al. 2009; Johnson, Jeppsson et al. 2009) are undergoing advanced clinical trials for premortem detection of AD. Although these PET agents generate high signal from binding amyloid, they have limited utility due to a high rate of false positives from nearly a third of normal controls and cognitively normal elderly populations bearing considerable plaque load (Braak and Braak 1991; Terry, Masliah et al. 1991; Masliah, Mallory et al. 1993) as well as false negative signal from certain cases of mild cognitive impairment (Kemppainen, Aalto et al. 2007; Rowe, Ng et al. 2007). Moreover, the poor correlation between amyloid deposition and neurodegeneration limits the feasibility of AD staging, as the levels of Aß

plateau during the course of disease (Engler, Forsberg et al. 2006; Jack, Lowe et al. 2009). While intended to specifically target Aß, these radiotracers do not selectively bind amyloid deposits, but tau-bearing NFTs as well. Therefore, in addition to detecting AD, the PET agents also generate positive signal for certain forms of tauopathies, such as frontotemporal lobar degeneration (FTLD), including progressive supranuclear palsy (PSP) (Goedert 2004), corticobasal degeneration (CBD) (Dickson 1999; Dickson, Bergeron et al. 2002) and Pick disease (PiD) (Dickson 2001) which present with NFT lesions. Consequently, the lack of specificity and selectivity of Aß imaging agents severely limits their utility for differential diagnosis. (Forman, Farmer et al. 2006; Knibb, Xuereb et al. 2006; Geser, Martinez-Lage et al. 2009; Rabinovici and Miller 2010) Although NFTs are also present in other dementias, such as FTLD-tau including PiD, PSP, and CBD, (Dickson 1999; Dickson 2001; Dickson, Bergeron et al. 2002; Goedert 2004) the regions and distribution of lesions are unique to each dementia (Cairns, Bigio et al. 2007). Whereas Pick bodies affect the dentate gyrus of the hippocampus, amygdala, and frontal and temporal neocortex (Dickson 2001), CBD is characterized by cerebral cortex and the basal ganglia, and PSP with cerebellum, brainstem, and limited cortical involvement (Dickson 1999; Boxer, Geschwind et al. 2006; Cairns, Bigio et al. 2007; Josephs, Whitwell et al. 2008). Therefore, a selective NFT-directed radiotracer signal could effectively diagnose different tauopathies with similar clinical presentation. Selective radiotracers are currently under development, most notably the NFT-directed 18F-THK523 (Fodero-Tavoletti, Okamura et al. 2011), which has shown favorable radiotracer kinetics (e.g. affinity, selectivity for tau) in a microPET study using the most aggressive tau mouse model with florid tangle pathology (rTg4510; nmol/g amounts of tau aggregates not published). However, by late stage disease the human brain accumulates only ~0.16 – 1 nmol/g wet tissue of insoluble tau aggregates (Greenberg and Davies 1990; Bramblett, Trojanowski et al. 1992). Therefore, further studies are needed in human cases to verify that the radioligand is capable of producing sufficient net PET signal over background in human patients with lower levels of tau aggregates. 1.3. Structural complexity of the molecular target in disease state

Although several scaffolds, such as benzothiazoles and imidazothiazoles (Figure 1.2) (Wischik, Harrington et al. 2004; Kudo, Susuki et al. 2006; Kemp, LJ et al. 2010; Maccioni-Baraona, Rojo et al. 2010) have been disclosed in patent literature as tau- binding agents for AD diagnostic agent development, none provide insight on mode of action or a rationale for tau selectivity. The difficulty in developing a probe for selectively targeting the tangles stems from the three-dimensional structure of the target in disease state. Normally, the tau protein is highly expressed in the axonal regions of neurons that can extend hundreds of centimeters, where its function is to stimulate, support and stabilize tubulin assembly into microtubules (Cleveland, Hwo et al. 1975). It is an unfolded monomeric peptide (55,000 to 68,000 kD) that associates with the microtubule based on tau isoform and phosphorylation state (Weingarten, Lockwood et al. 1975; Goedert, Crowther et al. 1991). Alternative splicing of the tau gene gives rise to six different isoforms in adult humans that differ in the number of positively-charged microtubule-binding domains they contain; three isoforms have three binding domains, and the other three isoforms have four binding domains (Goedert 1995). The isoform sequences vary between 352 to 441 amino acids, with the shortest isoform having m1, m3 and m4 binding domains and the longest isoform (2N4R) having m1, m2, m3 and m4, in addition to two 29 amino acid N-terminal inserts (Goedert, Crowther et al. 1991). However, in disease state, an unknown event causes tau to disassociate from the microtubule (Goedert, Spillantini et al. 1991; Iqbal and Grundkeiqbal 1991) and undergo a conformational change from a soluble unfolded monomer to an insoluble highly ordered aggregate termed paired helical filaments (PHF) (Iqbal, Alonso et al. 1994). PHF morphology has been elucidated using transmission electron microscopy (TEM) (Figure 1.3A). At present there is no crystal structure or NMR structural data available for a PHF or tau filament; structural data is available only for Aβ (Petkova, Buntkowsky et al. 2004). The Aβ fibrils have distinct morphologies that share common secondary and tertiary structures, but differ primarily in overall symmetry and quaternary structure (Ksiezak-Reding and Wall 2005). However, recent x-ray diffraction data has shown that tau filaments, both those extracted from human brain and those assembled in vitro from recombinant protein, share the cross-ß structure characteristic of Aβ: a parallel, in

register β-sheet conformation with a steric zipper spine extending out the filament axis composed of interdigitating side chains (Figure 1.3B) (Petkova, Yau et al. 2006; von Bergen, Barghorn et al. 2006). Instead of a three-dimensional binding site pocket that is the hallmark of globular proteins, such as enzymes, the parallel, in register ß-sheets present a surface with shallow grooves formed by the side chains of peptide residues (Petkova, Yau et al. 2006). Moreover, unlike the classic “lock and key” enzymes that have a one-to-one stoichiometry, filaments theoretically can have a high frequency of binding sites that depend on the amino acid composition of the fibrils (Groenning 2009). With multiple compounds capable of binding the filament surface at the same time, the interactions between small molecules and these sites share commonalities with adsorption phenomena, much like the process of dyeing cotton fibers (Timofei, Kurunczi et al. 1996; Bird, Brough et al. 2006). 1.4. Computational analysis of potent and selective aggregate-binding ligands Although efforts to target the tau-bearing NFTs with benzothiazole derivatives have been reported (Wischik, Harrington et al. 2004; Kemp, LJ et al. 2010), to date the issue of binding selectivity for NFTs over Aß plaques has not been addressed. Most efforts have centered around structure-activity relationship (SAR) guided potency optimizations of already known scaffolds, such as derivatives of ThT or Congo Red, and novel leads discovered via high-throughput experimental screening methods (Cai, Innis et al. 2007). In silico approaches to de novo scaffold design and optimization, such as quantitative structure-activity relationships (QSAR) that have traditionally been used to elucidate the nature of ligand affinity and selectivity for globular proteins, have not been fully utilized for aggregate targets. QSAR attempts to find consistent correlations between the variations of molecular descriptor properties and the biological activity within a series of compounds for rational compound design and activity prediction (Stone and Jonathan 1993; Peltason and Bajorath 2009). It commonly takes the form of a linear equation:

Biological Activity = Const + (C1• P1) + (C2 • P2) + (C3 • P3) + ... (1.1) where the parameters P1 through Pn are the molecular properties/descriptors computed for each molecule in the series and the coefficients C1 through Cn are the weights of these descriptors calculated by fitting the numbers into the biological activity equation. For

example, compound charge, hydrophobicity, molecular volume, and dipole moment are simple 2D descriptors used in classical QSAR models; on the other hand, 3D-QSAR descriptors take advantage of molecular volume of ligands to assign regions of hydrophobicity, steric bulk, electrostatics, etc. to various compound substituents (Stone and Jonathan 1993; Peltason and Bajorath 2009). The former generally describe an overall “global” character of the molecule, whereas the latter can pinpoint “local” features, such as R-group substituents that enhance the affinity of a ligand for its target. The generic procedure for developing a QSAR includes the construction of a dataset of compound structures, the acquisition of activity values for the compounds against a target, e.g., AC50, the computation of a set of molecular properties/descriptors for compound structures, and the calculation of weights/coefficients for descriptors by fitting to an activity equation (Stone and Jonathan 1994). In the case of 3D-QSAR, alignment of the compound structures to a common substructure template, or docking to a binding site, and generation of comparative molecular similarity indices (CoMSIA) fields for the alignment, precede the regression analysis (Klebe and Abraham 1999). The strength of this type of analysis lies in its calibration, where the observed vs. the corresponding calculated datapoints are assessed by the correlation coefficient R2. Accordingly, the model’s predictive power is assessed on the basis of external validation, Q2, and the ligand optimization guidelines obtained from the results (Konovalov, Llewellyn et al. 2008). 1.5. QSAR approaches to aggregate binding targets Most QSAR studies for AD consist of the conventional ‘lock and key’ structural approach that aims to optimize various compounds or inhibitors for enzyme binding pockets, such as acetylcholinesterase (AChE) (Figure 1.4A), butyrylcholinesterase (BChE), and γ-secretrase, where the structure of the target and the ligand binding site are well-known (Kaur and Zhang 2000; Keerti, Kumar et al. 2005; Castilho, Guido et al. 2007). However, there exist few QSAR studies for the filamentous aggregates that resemble dye-cellulose adsorption models with variable binding site density (Figure 1.4B). These QSAR models for amyloids have focused on various benzothiazole (BTA) derivatives, trans-stilbene (TSB) compounds (Wang, Zhang et al. 2005; Chen 2006; Kim,

Choo et al. 2007), and other scaffolds that have strong affinity not only for Aß plaques but also tau fibrils. One of the first QSAR studies (Chen 2006) correlated classical molecular descriptors for 22 TSB compounds with affinities for Aß; highest occupied molecular orbital (HOMO) energy (degree of compound stability/reactivity), charge sum, molecular volume, and clogP (R2= 0.86) were identified as the major TSB molecular properties for amyloid affinity. This model was supported by a corresponding docking result using an NMR structure of the Aß fibril (Petkova, Ishii et al. 2002), where the most favorable binding mode was inside the fibril, where the ligand formed hydrophobic interactions with aromatic residues. The authors repeated this approach in a recent study of 21 flavonoid derivatives (Yang, Zhu et al. 2010), and extended it to include an analysis of ligand substituents with 3D-QSAR. Consistent with their previous study, hydrophobic substituents in all positions were favored, whereas H-bond donors and acceptors were largely disfavored, in agreement with their docking result. The docked mode of the flavonoids was consistent with TBS docking results (Chen 2006) and identified π- stacking interactions with aromatic residues in the hydrophobic core of the fibril as major drivers of binding. However, there is no experimental evidence that this binding site would be accessible to the ligands. A comprehensive 3D-QSAR analysis by Kim, et al. (Kim, Choo et al. 2007) on 73 BTA and TSB derivatives supported previous findings that hydrophobic and aromatic ligands potently bind amyloid structures. Moreover, the study identified planarity, length and H-bond donors of ligands as key features of these diverse compounds. The 3D-QSAR by Kim et al. is consistent with an SAR analysis of BTAs by Leuma Yona et al., (Yona, Mazeres et al. 2008) where para-positioned substituents were favored, whereas meta- and ortho-positioned bulky substituents were detrimental to amyloid affinity. H-bonding interactions were predicted, as several potent ligands possessed hydrogen donating moieties, however, the corresponding receptor hydrogen acceptors were not elucidated. In summary, QSAR studies identified steric, hydrophobic and H-bond donating features of potent amyloid ligands; however, no rationale of binding mechanism, compound binding site density or selectivity was addressed. The similarities between adsorption and amyloid-binding are not limited to the binding interactions; the ligands involved in these processes also share distinct structural

characteristics. Most of the compounds are planar aromatic dyes of intense color, where electron-rich and -deficient moieties are connected though a π-bridge, also known as donor-π-acceptor systems (see Chapter 2) or polymethines (Sczepan, Rettig et al. 2003). The push-pull of electron density between donor and acceptor moieties gives rise to strong compound polarizability, an important descriptor of polymethine character (Marder, Sohn et al. 1991). The interactions between polymethine dyes and cellulose surfaces have been routinely modeled using QSAR methodology in order to optimize adsorption efficiency. For example, in one study of 27 cellulose-binding anionic azo dyes, lipophilicity, molecular volume and polarizability along with aromatic substituents were identified as key compound features for efficient adsorption (Timofei, Kurunczi et al. 1996). This study was further refined using more comprehensive semi-empirical calculations in order to identify potency-generating substituents using molecular descriptors such as atomic charges, vdW volumes, and fragmental hydrophobicities and polarizabilites (Kurunczi, Funar-Timofei et al. 2007), yielding a predictive model as based on Q2 = 0.82 and R2 = 0.94. In addition to hydrophobic and steric features, highly polarizable moieties positioned at the ends of dyes emerged as key features of adsorption. Polarizability was also directly correlated (R2 = 0.79) for various acidic dyes as a function of ligand-cellulose distance in a study conducted by Bird et al (Bird, Brough et al. 2006). Taken together, the QSAR results of dye-cellulose interactions indicate that hydrophobicity and polarizability are important for adsorption phenomena, and may be essential for amyloid-binding compounds as well. 1.6. Computational modeling of ligand-aggregate binding interactions The fundamental limitation of classical QSAR modeling is that only compound structures are described in terms of molecular descriptors correlating to binding affinity – characteristics of target binding sites are not modeled. Therefore, QSAR models should be interpreted in the context of experimental or additional computational data that would elucidate the binding interactions of the receptor. Due to the lack of experimental amyloid-ligand complex structures, molecular dynamics (MD) simulations provide the context for the nature of the binding interaction, depicting atomic-level models of ligands in complex with cross-ß-sheet structure (Rodriguez-Rodriguez, Rimola et al. 2010; Wu,

Bowers et al. 2011). The results from two different computational studies revealed a high affinity binding site where compounds directly interact with the β-sheet backbone of the Aβ protomer in an extra wide channel formed by a Gly33 residue flanked by Ile31 and Met35 (Figure 1.5A). The deep insertion of the molecule in the binding channel favors polarizability-driven dispersion and induced-dipole effects, as the ligand directly interacts with the hydrogen-bonding network of the β-sheet core (Jensen, Cisek et al. 2011). Modeling studies have indicated that neutral molecules have the strongest affinity at this site as well as being more deeply buried than charged molecules within the glycine cleft (Wu, Bowers et al. 2011). This suggests a candidate binding site for radiotracer optimization, as neutral compounds are more suitable to meet the criteria of imaging agents (Kim, Jensen et al. 2010). Interestingly, glycine residues in the aggregated core regions of the protofibrils that expose the β-sheet backbone for this binding mode are present not only in the Aß sequence, but in tau as well. Moreover, the 2N4R tau sequence has a Gly-Gly-Gly motif in each of its four repeat domains, which would form an extra wide binding channel that would expose a bigger surface area of the fibril’s backbone for ligand binding, potentially providing the means to design a tau-selective ligand. To address this issue, the Aß structural model was adapted within the context of binding sites created by 2N4R tau sequence (Figure 1.5B). Indeed, an extra wide binding channel accessible to solvent would be able to accommodate ligands that might clash sterically with narrower binding channels, such as the single Gly cleft in Aß (Figure 1.5C and 1.5D). A highly polarizable and wide ligand can take advantage of a greater surface area contact with the receptor at this site and may be able to generate stronger binding affinity. Moreover, the presence of four Gly-Gly-Gly motifs in the 2N4R tau sequence has implications for ligand binding site density, such that a highly potent ligand optimized for this binding site may achieve high signal over the white matter uptake (non specific binding, background), which has been a significant challenge of amyloid imaging agents (Rowe and Villemagne 2011). Taken together, the characterization of the most important molecular properties of highly potent ligands as well as the context of the hypothetical binding modes can provide insight into rational design of diagnostic agents.

1.7. Summary Tau-directed imaging is a promising approach for premortem differential diagnosis of AD. To leverage this method using small molecule diagnostic agents, the probes must be capable of potently binding the molecular targets with high binding site density. Moreover, the lack of atomic-level experimental data on the nature of the binding interaction between cross-ß-sheet targets and small molecules complicates the design and optimization studies, as well as the discovery of novel ligands. Computational methods, such as QSAR, have the advantage of analyzing hundreds of molecular properties of aggregate-binding ligand libraries, such that statistically significant ligand features can be identified for further study. However, current QSAR studies have only identified generic trends of cross ß-sheet binding ligands, such as planarity and hydrophobicity and are therefore inadequate for rational ligand design and optimization. These retrospective QSAR models have failed to identify molecular features responsible for ligand selectivity among different aggregate targets. Furthermore, the models have not been integrated with in silico models of ligands interacting with cross ß-sheet surfaces and have not provided insights into the variable binding site densities of different small molecules. Therefore, there still exists a need for unambiguous predictive models that could be interpreted in mechanistic terms with maximum simplicity, transparency, and portability, to generate optimal small molecule diagnostic agents.

1.8. Figures

Figure 1.1. AD pathology in human brain. (A) Cartoon depiction of healthy neuronal cells (B) Cartoon depiction of AD brain: flame-shaped intra-neuronal NFT (green arrow), and extracellular amyloid deposits (red arrow), adapted from www.alz.org. (C) Confocal image of authentic AD brain tissue lesions stained with ThT, showing NFTs (green arrow) and amyloid deposits (red arrows). Confocal image courtesy of Kristen Funk.

Figure 1.2. Diversity of cross-ß-sheet binding ligands. Compounds identified as potent against tau fibrils are compounds are mainly dyes of very intense color, including triarylmethines (green, purple), benzothiazoles (yellow), phenothiazines (blue), rhodanines (red), and imidazothiazoles (orange). The color of each structure represents the corresponding frequencies of light that are emitted by the compound (the actual color of the dye).

Figure 1.3. Tau-bearing fibrils in AD pathology. (A) TEM image of an authentic PHF isolated from AD brain with darker regions showing the twist of the fibril. (B) In silico computational model of the cross--sheet structure of a filament showing channels on the surface of the fibril formed by the side chains of solvent-exposed residues, as well as the overall twist of the fibril, adapted from (Paparcone and Buehler 2009). Image was created using UCSF Chimera Alpha Version 1.5 (build 31329) software.

A B

Figure 1.4. Ligand – protein binding models. (A) Classical “lock and key” model of ThT, where carbon (grey), nitrogen (blue) and sulfur (yellow) are depicted in space- filling ball format, interacting with AChE enzyme, depicted by secondary structure (multicolored ribbon) in a hydrophobic pocket (pdb ID 2J3Q). (B) Proposed multiple binding site “adsorption” binding model, where many ThT molecules interact with the surface of a trimeric Aβ fibril (Petkova, Buntkowsky et al. 2004), adapted from (Paparcone and Buehler 2009). Images were created using UCSF Chimera Alpha Version 1.5 (build 31329) software.

A B

Met Gly Gly Gly Ile Gly

C D

Ile Met Gly Gly Gly

Gly

Figure 1.5. Ligand selectivity for cross-ß-sheet structures. Hypothetical fibril binding sites where protein atoms are colored grey (carbon), blue (nitrogen), and red (oxygen), and hypothetical binding modes of a “wide” benzothiazole compound, where ligand atoms are colored violet and magenta (carbon), blue (nitrogen), and yellow (sulfur), and both ligand and protein are overlaid with mesh and transparent molecular surfaces, respectively. (A) A narrow binding channel containing a Gly residue in amyloid structure (B) An extra wide binding channel containing Gly-Gly-Gly in a hypothetical tau structure. (C) A wide benzothiazole compound sterically clashes with the Ile residue (yellow lines) in the amyloid structure. (D) The wide benzothiazole compound is accommodated by the Gly-Gly-Gly binding site motif. Cross-ß-sheet structures adapted from (Luhrs, Ritter et al. 2005), where the original residues were mutated in silico to model a hypothetical tau fibril. Images were created using UCSF Chimera Alpha Version 1.5 (build 31329) software.

CHAPTER 2

QSAR STUDIES FOR PREDICTION OF CROSS- SHEET AGGREGATE BINDING AFFINITY AND SELECTIVITY

2.1. Introduction Whole-brain imaging is a powerful approach for premortem diagnosis of AD and potentially other neurodegenerative disorders associated with protein misfolding (Nordberg, Rinne et al. 2010). Although a radiotracer capable of binding AD lesions composed of A aggregates called Amyvid has been approved by the FDA (Hsiao, Huang et al. 2012) and many others are in advanced clinical trials, neither their mechanism of binding nor their target binding specificity is well defined. This is because the protein aggregates that appear in neurodegenerative disorders share cross-ß sheet conformation, characterized by parallel, in register  sheets oriented perpendicular to fibril axes (Sawaya, Sambashivan et al. 2007). Unlike three-dimensional pockets that appear on traditional globular protein targets, fibril surfaces present only shallow grooves and channels that are shaped primarily by amino acid residue side chains (Krebs, Bromley et al. 2005; Groenning 2010). The resulting common structural organization would appear to limit opportunities for selective molecular interactions. Nonetheless, recent studies have reported binding selectivity for small molecules among aggregates at the level of binding affinity, suggesting the feasibility of tuning binding interactions (Honson, Johnson et al. 2007; Fodero-Tavoletti, Okamura et al. 2011). If the problem of binding selectivity were overcome, the utility of whole-brain imaging for differential diagnosis and staging of individual neurodegenerative disorders could be greatly improved.

Few experimentally-derived structural models of small molecule binding sites on amyloid fibrils have been disclosed to date (Krebs, Bromley et al. 2005), and structure- activity relationship studies, which aim to rationalize compound affinity for biological targets, have not been interpreted in the context of target selectivity (Wang, Zhang et al. 2005; Chen 2006; Kim, Choo et al. 2007; Leuma Yona, Mazeres et al. 2008). For this project, investigation of binding selectivity has focused on protein aggregate targets associated with Alzheimer’s and Lewy body diseases such as A, tau, and -synuclein (Honson, Johnson et al. 2007). Using a competition assay with Thioflavin T (ThT), a fluorescent probe of cross- sheet structure, cationic polymethine dyes as especially potent displacers of ThT from AD lesions composed of tau protein were identified (Jensen, Cisek et al. 2011). These compounds share a planar, fairly rigid structure combined with highly delocalized aromatic π-electrons. Because of these properties, the most potent compounds investigated were highly polarizable, and therefore capable of supporting strong van der Waals interactions with flat surfaces exposed on fibrils. These data suggested compound polarizability as a descriptor for the tau-fibril binding affinity of dyes (Jensen, Cisek et al. 2011). Nonetheless, it was not clear from the correlations whether polarizability could be leveraged to generate selective binding among protein aggregates, or how this parameter could be maintained in neutral analogs capable of crossing the blood brain barrier. To address these questions and to identify additional descriptors of aggregate binding affinity, I used quantitative structure activity relationship (QSAR) analysis. The experimental input data for this computational study was compiled from the literature, and consisted of closely related neutral and cationic benzothiazole derivatives tested for their ability to displace ThT from synthetic aggregates prepared from A40 and insulin (Caprathe, Gilmore et al. 1999). The datasets, which included 50 compound structures and their associated AC50 values (Table 2.1), were chosen for QSAR analysis because they probed two aggregate targets of distinct composition while spanning three orders of magnitude in affinity for each. Second, the benzothiazole scaffold is under extensive investigation for whole-brain imaging purposes, with demonstrated affinity for aggregates composed of tau or A (Cai, Innis et al. 2007; Kemp, Storey et al. 2010).

Importantly, the potency of certain compounds in the series extend to radiotracer concentrations (low nanomolar). Finally, certain members of the series have been extensively characterized with respect to optical and electrical properties (Liu, Liu et al. 2000; Benkova, Cernusak et al. 2006; Hrobarik, Horvath et al. 2007; Sigmundova, Zahradnik et al. 2007; Hrobarik, Sigmundova et al. 2010). The results confirm polarizability as a major descriptor of relative binding affinity of the series for cross- sheet aggregates, and show how this parameter can be maximized in neutral analogs. 2.2. Experimental Procedures 2.2.1. Bioactivity data

AC50 data for displacement of ThT probe from aggregates composed of A (4.6 μM protomer) and insulin (1.4 μM protomer) were taken from the literature (Caprathe, Gilmore et al. 1999). Test compound concentrations ranged from 0.001 – 30 μM. The assays were performed at the apparent probe Kd for each protein substrate to facilitate direct comparison (0.5 and 20 μM for insulin and A, respectively). Within this in vitro format, it was assumed that displacement efficacy was complete for all compounds, and that total compound concentration approximated free concentration at equilibrium. 2.2.2. Chemical structures and calculation of molecular descriptors Compound structures (Table 2.1) were built (Chem3D Pro 12 software) and minimized (Allinger’s molecular mechanics MM2 force field (Allinger 1977)) using default convergence criteria of 0.100 for the minimum RMS gradient and 10,000 iterations. Molecular descriptors and properties were then generated using various semi- empirical and ab initio methods. First, semi-empirical descriptors were generated with E- DRAGON 1.0, an online implementation of the DRAGON 5.4 molecular descriptor generator (Mauri, Consonni et al. 2006) that computes >1,600 descriptors categorized into 20 logical blocks (Consonni and Todeschini 2000). Second, because of the variable accuracy of molecular lipophilicity predictions among various semi-empirical methods (owing to inadequate training sets and method parameterization (Mannhold, Poda et al. 2009)), the E-DRAGON descriptor set was augmented with clogP and topological polar surface area estimations calculated using the

highly parameterized and robustly trained fragment-based algorithm of the Molinspiration Property Calculation Service (www.molinspiration.com). Finally, compound dipole moment and polarizability were calculated at the quantum level using density functional theory methods implemented in Gaussian 09 (G09) (Frisch, Trucks et al. 2009) software package available on Ohio Supercomputing Center clusters. Each compound structure was evaluated using the three-step approach of Perpete, et al. (Jacquemin and Perpete 2006; Perpete and Jacquemin 2009) consisting of: i) a ground- state geometry optimization with 3x10-4 a.u. residual mean square convergence criteria (default OPT threshold); ii) confirmation of ground-state geometry with vibrational spectrum determination (structure minima verified by real vibrational frequencies); iii) calculation of α in a static (ω = 0) external electric field (default 1 a.u. in principal axes), at the optimized ground-state geometry. All calculations were performed using hybrid density functional B3LYP and the 6-311++G(d,p) basis set. Bulk solvent effects were implicitly modeled with the polarizable continuum model (Tomasi, Mennucci et al. 2005) (G09 keywords SCRF=(Solvent=Methanol)). Polarizabilities are reported as the mean, <α>, or the average of the three polarizability tensor quantities that correspond to x, y, and z components of parallel external field principal axes (Marder, Sohn et al. 1991):

<α> = ⅓ ( αxx + αyy + αzz) (2.1) given in units of polarizability volume (Å3). All calculated descriptors were merged prior to model generation. 2.2.3. Model generation Genetic algorithm-PLR analysis was performed using the Virtual Computational Chemistry Laboratory (VCCL), an online portal for computational chemistry tools available at www.vcclab.org (last accessed 1 November 2011)(Tetko, Gasteiger et al. 2005). Calibration analyses (i.e., training set models) for Aß and insulin molecular targets were performed using default parameters (minimum residual variance of factors = 0.0010; number of latent t-variables  12). Descriptors were identified as redundant (i.e., constant or flat) and deleted if ≥90% of the subject compounds shared identical values, yielding 84 descriptors for the final genetic algorithm fitting. PLR models were

2 optimized on the basis of leave-one-out cross validation (Q loo) as implemented in the VCCL. MLR was performed using the open source statistical software R version 2.13.0, available at www.r-project.org (last accessed 1 November 2011). Training set fits were generated using the linear model lm function, and cross validated using the crossval function in the bootstrap package. The descriptor correlation matrix was generated using the cor function in the bootstrap package. 2.2.4. Model validation Internal validation of MLR models was performed in R version 2.13.0. Correlations were cross validated using bootstrap resampling (Wehrens, Putter et al. 2000) as implemented in the crossval function of the R bootstrap package, whereas variance inflation factors were calculated using the vif function. Y-randomization (Rucker, Rucker et al. 2007) was performed in R in two steps. First, new MLR models were developed by randomly shuffling (250 shuffles) the 41 dependent variables (i.e., AC50 values for the training set) while keeping the independent variables (i.e., α, μ, clogP, and RBF descriptor values) constant. Dependent variables were shuffled using the sample function, whereas refitting to independent variables was performed using the lm function. In the second step, which assessed potential bias in selection of independent variable sets, randomly shuffled dependent variables were correlated with stepwise selected subsets of four out of the original 84 descriptors identified through PLR screening. This was performed using a custom-written R script that incorporated the sample function to shuffle dependent variables and the regsubsets function (exhaustive search algorithm) in the leaps package to fit each set of independent variables to the shuffled dependent variables. Both steps of y-randomization were run for 250 iterations, with the original dependent variables being independently shuffled with each iteration. The linear model 2 resulting from each shuffle was cross validated (Q boot) using the crossval function. 2 2 External validation was performed as described previously, where R0 , R'0 , k, and k' correspond to the correlation coefficients and slopes of linear regressions constrained through the origin (Golbraikh and Tropsha 2002). 2.2.5. Statistical methods

Differences between regression MLR coefficients were assessed by z-test: x  x z  1 2 (2.2) 2 2 (S x1 )  (S x2 )

where x1 ± Sx1 and x2 ± Sx2 are the pair of estimates ± standard error being compared, and z is the 1-α point of the standard normal distribution. 2.3. Results 2.3.1. Experimental data, descriptor sets, and workflow To identify potential drivers of binding affinity and selectivity for A40 and insulin aggregates, a three-step workflow was adopted (Figure 2.1). First, the >1600 descriptors available in E-DRAGON were calculated for all Series A and B molecules (shown in Table 2.1.1). Because these compounds were flat and dye-like, the starting E-DRAGON descriptor set was pruned on the basis of reported dye adsorption QSAR studies (Kurunczi, Funar-Timofei et al. 2007; Metivier-Pignon, Faur et al. 2007) to yield a focused molecular property set comprising 278 descriptors representing five logical blocks: 48 constitutional descriptors, 33 connectivity indices, 154 functional group counts, 14 charge descriptors, and 29 molecular properties. The pruned descriptor set was then augmented with separately calculated clogP and topological polar surface area values to improve estimation accuracy (Mannhold, Poda et al. 2009), and polarizability () and dipole moment (μ) values to explicitly capture the contribution of these quantum parameters to biological activity. In the second step, a partial least squares regression (PLR) QSAR approach was taken to screen for candidate descriptors for displacement activity. Finally, the top descriptors identified by PLR (Table 2.2) were used to build and validate multiple linear regression (MLR) models for both A40 and insulin targets. This was done to generate unambiguous predictive models that could be interpreted in mechanistic terms with maximum simplicity, transparency, and portability. 2.3.2. Descriptors for relative ThT displacement affinity To identify the best combination of the descriptors described above for predicting

ThT displacement AC50, the datasets were subjected to QSAR analysis using a genetic algorithm-PLR method. The optimal A40 model consisted of 28 molecular descriptors

(x variables) collapsed into 11 linear combinations (latent t variables), whereas the insulin model consisted of 22 x variables collapsed into four latent t variables (Table 2.3). The resulting Y-correlations were adequately strong (as judged by the correlation coefficient, 2 2 R ) and stable (on the basis of bootstrap cross validation; Q boot) for both target models (Table 2.3)(Konovalov, Llewellyn et al. 2008). The four highest-weighted and therefore top-ranked descriptors for the A40 model represented four different logical blocks, and included average high-order valence connectivity 5χv, α, clogP, and rotatable bond fraction (RBF) (Table 2.4). Valence connectivities 4χv and 5χv describe the summed contributions of contiguous four- and five- bond fragments to activity, and are particularly sensitive to the presence of high-valence heteroatoms and double bonds (Kier and Hall 1986; Hall and Kier 2001). Although lacking a straightforward chemical interpretation in the current context, the negative coefficient for average 5χv indicates that increasing higher order average valence connectivity correlates with decreasing ThT displacement activity. The polarizability term α describes how easily electron density can shift about the molecule when exposed to an external electric field, such as an adjacent dipole or ion (Marder, Sohn et al. 1991). The positive coefficient indicates that increasing polarizability correlates with increasing potency. ClogP is the log of the calculated octanol/water partition coefficient (Consonni and Todeschini 2000). Although logP^2 descriptors were part of the molecular descriptor set screened with PLR, a linear dependence on hydrophobicity yielded the best correlation of the dataset. The negative sign of clogP indicates that hydrophobicity decreases displacement potency in the context of Series A and B molecules. RBF is the fraction of rotatable bonds (i.e., the ratio of rotatable to total number of bonds) (Consonni and Todeschini 2000). The positive sign of RBF indicates that increasing torsional freedom of atoms outside the rigid aromatic core allows these compounds to achieve maximal displacement potency. The model for insulin aggregate displacement activity identified molecular descriptors complementary to the Aβ40 model including hydrophilic factor, topological electronic factor, charge polarization factor, and average 4χv (Table 2.4). The hydrophilic factor is an empirically determined index based on the number of hydrophilic groups in

molecules (Consonni and Todeschini 2000). The positive coefficient for this parameter indicates that hydrophilic groups increased displacement activity against insulin aggregates. The topological electronic factor calculates the differences in partial atomic charges with respect to interatomic distance (Consonni and Todeschini 2000). It correlated weakly with polarizability in Series A and B molecules (R2 = 0.63). The positive sign indicates that displacement affinity increased in parallel with this parameter. Finally, charge polarization, which incorporates the effect of large heteroatoms in electrotopology (Consonni and Todeschini 2000), also correlated with polarizability in the datasets (R2 = 0.75). The positive coefficient indicates that displacement potency against insulin aggregates increased in parallel with this descriptor. Overall, these results are consistent with both molecular targets sharing a common cross-β sheet structure presenting similar yet distinct binding sites along their surfaces. 2.3.3. Construction and validation of predictive MLR models To generate predictive QSAR models, the top-ranked chemical, quantum, and constitutional descriptors identified by PLR screening (α, clogP and RBF) were analyzed using an MLR approach. Dipole moment magnitude (μ) was used in place of topological descriptors so that each of the four final descriptors had simple chemical interpretations. Although dipole moment magnitude was not identified as a candidate descriptor in the PLR screen, it has been proposed to contribute to the aggregate binding affinity of small molecules (Barrio, Satyamurthy et al. 2009), and so was included in the MLR analysis for this reason. The final four individual descriptors intercorrelated only weakly when compared pairwise in a correlation matrix (Table 2.5), and were appropriate in number for the size of training sets used in the analysis (i.e., the ratio of training compounds:descriptors was > 8:1) (Topliss and Costello 1972). In addition, variance inflation values were low for all parameters (Table 2.5), indicating that the variance associated with each descriptor was independent of the others. These data indicated that the final descriptor set was appropriate for modeling the affinity and selectivity characteristics of Aβ40 and insulin aggregates. To prepare MLR models, the Aβ40 and insulin datasets were split into training and test sets of 44 and six molecules, respectively, with initial MLR calibration performed on

the former. However, subsequent modeling identified compounds 3A, 14A and 48B as outliers (i.e., the calculated pAC50 deviated from the observed pAC50 by more than 1 log unit for both targets), and so these were excluded from further analysis. The final training sets consisted of 41 compounds, composed of 38 actives/3 inactives for Aβ40, and 40 actives/1 inactive for insulin. Final MLR calibration yielded two equations for the prediction of AC50: pAC50 Aβ40 = 0.045 (0.007)α – 0.062 (0.023)μ + 0.102 (0.023)%RBF – 0.362 (0.053)clogP – 6.10 (0.59) (2.3)

pAC50 insulin = 0.029 (0.008)α – 0.075 (0.027)μ + 0.085 (0.026)%RBF – 0.516 (0.063)clogP – 3.02 (0.70) (2.4)

These models were adequately correlated (as judged by R2 = 0.73 and 0.74 for Aβ40 and insulin training sets, respectively; Figure 2.2), consistent (on the basis of residual standard error; s = 0.38 and 0.44 for Aβ40 and insulin, respectively) and stable (on the 2 basis of bootstrap cross validation; Q boot = 0.63 and 0.60 for Aβ40 and insulin, respectively). To test statistical robustness, the models were subjected to Y- randomization tests. When the response data for each calibration compound was randomly shuffled 250 times and correlated with unchanged descriptor values, observed 2 Q boot values ranged from 0 - 0.453 for A40 and 0 - 0.507 for insulin, with only two 2 models generating Q boot > 0.4. The poor correlations indicated that the probability of eqs 1 and 2 occurring by chance was low. When the randomized response data were correlated with different subsets of four descriptors (out of the original set of 84 2 descriptors identified through PLR screening), observed Q boot values for A40 and insulin ranged from 0 - 0.470 and 0 - 0.391, respectively, with only one model generating 2 Q boot > 0.4. These poor correlations indicate that the chance of eqs 1 and 2 occurring through biased selection of descriptors (x variables) was low. The predictive capability of the MLR models was validated using an external test set (i.e., six compounds not used in the calibration) and the statistical criteria proposed by Golbraikh and Tropsha (Golbraikh and Tropsha 2002). The resulting correlations show

that the models met target slope and goodness of fit criteria for predictive utility (Table 2.6; Figure 2.2). Together, the internal and external validation experiments indicated that an acceptable characterization of compound bioactivity against both targets over the low nanomolar-low micromolar concentration range was achieved. 2.3.4. Descriptors for relative ThT displacement selectivity Although eqs 1 and 2 show that the descriptors of displacement activity against insulin and A40 aggregates paralleled each other, selectivity was observed for many library members (Table 2.1). To test whether selectivity correlated with any of the four descriptors quantified by MLR QSAR, the magnitude of their coefficients in the A40 and insulin MLR models was compared directly with each other. First, the t-statistic for each descriptor coefficient (defined as the ratio of the coefficient to its standard error) was calculated. All values were ≥ 2.7, corresponding to rejection of the null hypothesis at p  0.01, and confirming that the coefficients were determined with precision. Second, coefficient magnitudes from the A and insulin models were directly compared by z-test (Figure 2.3). Results showed that compound hydrophobicity and polarizability were the major drivers of displacement selectivity in these datasets. As the clogP coefficient increased, pAC50 increased for both molecular targets, but ~1.4-fold more strongly for insulin than for A40. The differences were significant at p < 0.05. These data indicate that insulin aggregates were more sensitive to compound hydrophobicity than were A40 aggregates. Similarly, as α increased, pAC50 decreased for both molecular targets, but ~1.6-fold more strongly for A40 than for insulin. The difference only trended toward statistical significance (p = 0.064), but suggested that binding to A40 aggregates was incrementally favored relative to insulin aggregates by highly polarizable compounds. In contrast, RBF and μ coefficient magnitudes differed by only ~20%, with poor statistical significance (p > 0.3; Figure 2.3), suggesting that these parameters did not contribute to differences in displacement potency between A40 and insulin aggregates. Together these data show that selective binding to insulin and A40 aggregates of up to ~2 log

AC50 units is feasible within the benzothiazole series investigated, and that compound hydrophobicity and polarizability are two candidate drivers of this selectivity.

2.4. Discussion This QSAR study indicates that Aβ and insulin aggregates yield populations of binding sites that are dominated by differing modes of interaction. It further suggests strategies for maximizing the affinity and selectivity of ligands for cross- sheet aggregates of defined composition. In the context of the benzothiazole series investigated here, affinity for A relative to insulin aggregates was modulated by compound hydrophobicity and polarizability. For example, starting compounds 1A – 3A displaced ThT from A only weakly. Like other members of Series A and B, these compounds shared a donor-π-acceptor architecture, where a dimethylamine electron donor was connected to the benzothiazole electron acceptor through a π-electron rich bridge containing a vinyl linker (Figure 2.4). The push-pull character of this architecture leads to the delocalization of π-electrons that drives compound polarizability (Albert, Marks et al. 1997). Displacement potency for Aß was increased by combining donor and acceptor groups so as to maximize polarizability while modulating the strength and orientation of the dipole moment (Liu, Liu et al. 2000). The inverse correlation between dipole moment magnitude and displacement potency suggests that affinity is driven by an ability to form induced dipoles in conjunction with binding surfaces rather than by the existence of a ground-state permanent dipole. However, the results also may reflect oversimplification of the dipole descriptor, which captured dipole moments as scalar magnitudes instead of vectors with magnitude and directionality. Induced dipoles, which are directly correlated to polarizability, were not represented at all. These considerations may rationalize why polarizability but not dipole magnitude correlated with displacement potency. The most efficacious affinity-driving modification was to quarternize N3 in the benzothiazole heterocycle so as to create a stronger electron acceptor. Although many different substituents on N3 served this function in Series A, a simple methyl substituent was adequate to drive affinity (5A – 11A). In fact, large hydrophobic N-substituents tended to slightly weaken potency (perhaps through steric effects), leading to the inverse correlation of clogP with affinity in this series. However, higher potency could be fostered without introducing a quaternary nitrogen by replacing the vinyl linker with an

azo linker (12A – 14A). On the basis of ab initio calculations, the azo linker has been reported to increase compound polarizability by acting as an auxiliary electron acceptor to the benzothiazole ring (Shuto 1996). Alternatively, displacement affinity could be raised by increasing the surface area of the π-bridge (48B and 49B). Both of these π- bridge modifications promote polarizability in the context of neutral molecules. Overall, to generate electronic properties that favor A relative to insulin aggregate ThT displacement potency, a neutral Series A or B compound should contain a strong electron donor flanking a π-bridge that maximizes push-pull electronic structure while retaining compound planarity. The nonspecific cross- sheet binding agent [1-(6-{[2- fluoroethyl](methyl)aminaphthalen-2-yl)ethylidene]propanedinitrile (FDDNP) shares a similar structural organization (Barrio, Satyamurthy et al. 2009), suggesting these affinity driving concepts can be extended to other scaffolds besides benzothiazoles. Because Series A and B potency was quantified in displacement format, our QSAR approach interrogated only those sites occupied by ThT probe. On the basis of molecular dynamics simulations, both ThT and its neutral analog Pittsburgh Compound B (PIB) bind up to six distinct sites on A protofilaments (Wu, Wang et al. 2008; Wu, Bowers et al. 2011). Total binding energy decomposition analysis identified van der Waals forces and non-electrostatic solvation energy (i.e., the hydrophobic effect) as the dominant descriptors of binding energy at these sites. In contrast, electrostatic interactions were found to antagonize binding. The two sites predicted to yield the highest binding energy for ThT and PIB are reproduced in Figure 2.5. In one calculated mode (Site A), the ligands preferentially bound shallow, hydrophobic clefts formed by the side chains of aromatic residues, resulting in favorable van der Waals interactions between their benzothiazole and benzaminic ring systems and the planar surface created by in-register aromatic sidechains of Phe20 (Figure 2.5A). Interestingly, a neighboring negatively charged residue (Glu22) did not interact with the ThT tertiary amine despite it being accessible in this binding mode. In the second hypothetical binding mode (Site B, Figure 2.5B), ThT and PIB ligands entered extra wide hydrophobic channels formed by Gly residues exposed on the aggregate surface. Deep insertion of ligand into these channels would allow direct interaction with the hydrogen-bonded β-sheet core through

dispersion and induced-dipole effects (Rodriguez-Rodriguez, Rimola et al. 2010; Jensen, Cisek et al. 2011; Jensen, Cisek et al. 2011). QSAR analysis predicts that highly polarizable Series A and B compounds should be especially well suited for competing with this mode of ThT interaction. Consistent with this hypothesis, PIB binding energy was predicted to be strongest at this site (Wu, Wang et al. 2008). In addition, this binding mode is consistent with the computational modeling of Rodriguez-Rodriguez et al., where a quantum-refined docked pose of ThT was predicted to preferentially occupy the wide channels formed by Gly residues of the Aβ42 protofibril (Rodriguez-Rodriguez, Rimola et al. 2010). These modeling studies highlight the importance of planar aromatic moieties of ligands for ThT-like binding interactions, as well as the need for adequate rotatability of neighboring groups so that surface contact at the binding sites is maximized. They also highlight the heterogeneity of binding sites that results from varying side chain composition despite commonality in main chain secondary structure. Overall, the proposed interaction of Series A and B compounds with cross- sheet aggregates is reminiscent of the interaction between cellulose-based textiles and acidic dyes, both of which present flat surfaces that support adsorption. Interaction is reportedly mediated by van der Waals interactions, with the enthalpic portion of binding energy reflecting the flatness of the dyes and how close their planes can contact the cellulosic surface (Bird, Brough et al. 2006). Both Series A and B ligands that displace ThT from A aggregates with submicromolar affinity resemble acid dyes in being highly polarizable and flat, and by having their binding affinity influenced by compound surface area. Although these structural features also support compound self association (Murakami 2002; Necula, Chirita et al. 2005), benzothiazoles were assayed at concentrations where interference with probe displacement was minimal (Honson, Jensen et al. 2009).

Radiotracer utility depends on binding site density (Bmax) as well as on binding affinity (Laruelle, Slifstein et al. 2003). Large differences in Bmax have been reported for A-aggregate binding ligands, although the structural basis for these observations is not clear (Lockhart, Ye et al. 2005). In the context of tau cross- sheet structure, polarizability as a potential descriptor for Bmax as well as for Kd has been identified

(Jensen, Cisek et al. 2011). It is conceivable that differences between active and inactive

Series A compounds reflect the contribution of polarizability to Bmax in addition to Kd. In summary, this QSAR study reveals that aggregates composed of Aβ and insulin present binding sites that interact differentially with small molecules, and that binding selectivity at the level of affinity can be tuned by leveraging the molecular properties indentified by the MLR model, including compound polarizability and hydrophobicity. These concepts are likely to be applicable to a range of scaffolds not limited to benzothiazoles.

2.5. Tables

AC50 a b b c # R1 R2 R3 X Aß Insulin Sl (nM) (nM) 5-Cl --- NMe C >100000 >100000 1 1A 2

2A H --- NMe2 C >100000 1200 83

3A 6-NMe2 --- NMe2 C >100000 6000 16

4A H Et NMe2 C 400 6 67

5A H Me NMe2 C 1000 3 333

6A 5-Cl Me NMe2 C 430 5.2 83

7A 5-F Me NMe2 C 1000 10 100

8A 5-Me Me NMe2 C 400 7.5 53

9A 6-Me Me NMe2 C 180 6 30

10A 6-OMe Me NMe2 C 300 8 38

11A 6-NO2 Me NMe2 C 1000 12 83

12A H --- NMe2 N 22000 1200 18

13A 6-OMe --- NMe2 N 3200 700 5

14A 6-Cl --- NMe2 N 1300 1300 1

15A 6-OMe Me NMe2 N 410 10 41

16A H Me NMe2 N 1300 60 22

17A H NMe2 C 110 12 9

18A H CH2CHCH2 NMe2 C 300 12 25

19A H (CH2)3Me NMe2 C 160 27 6

20A H (CH2)6Me NMe2 C 93 83 1 a R2 group abbreviations: none (---), methyl (Me), ethyl (Et). b AC50 is the concentration of compound needed to decrease ThT probe fluorescence by 50%. c Selectivity Index (SI) is the ratio of A40:insulin AC50 values.

Table 2.1. Compound structures and characteristics Continued

Table 2.1. Continued

AC50

# R1 R2 R3 X Aß Insulin Sl (nM) (nM)

21A 5-F NMe2 C 170 32 5

22A 6-Me NMe2 C 130 50 3

23A 6-OMe (CH2)6Me NMe2 C 140 40 4

24A H NMe2 C 120 42 3

25A H NMe2 C 240 34 7

26A H NMe2 C 100 80 1

27A H NMe2 C 230 130 2

28A NMe C 120 180 1 H 2

29A NMe C 200 210 1 H 2

30A H NMe2 C 100 80 1

31A H NMe2 C 460 210 2

32A H NMe2 C 140 52 3

33A H NMe2 C 170 93 2

34A H NMe2 C 170 80 2

35A H NMe2 C 170 80 2

36A H NMe2 C 120 52 2

37A H (CH2)6Me NEt2 C 140 120 1

38A H (CH2)6Me N((CH2)3Me)2 C 600 90 7

Continued 32

Table 2.1. Continued

AC50

# R1 R2 R3 X Aß Insulin Sl (nM) (nM)

39A H (CH2)6Me pyrrolyl C 160 42 4

40A (4,5) benzyl Me NMe2 C 210 53 4

41A (4,5) benzyl Me NMe2 C 210 41 5

42A (5,6) benzyl Me NMe2 C 120 120 1

43A (6,7) benzyl Me NMe2 C 210 41 5

44A H NMe2 C 120 13 9

45B H 3-(CH2)6Me NMe2 C 160 40 4

46B 6-OMe 3-Me NMe2 C 430 26 17

47B (4,5) benzyl 1-Me NMe2 C 250 42 6

48B H --- NH2 N 120 130 1

49B 6-OMe --- NMe2 N 2500 340 7

50B (4,5) benzyl --- NMe2 N 52000 16000 3

# αa μb clogPc RBFd # αa μb clogPc RBFd 1A 70.7 10.9 5.24 0.079 26A 106.2 5.74 4.76 0.094 2A 66.7 7.71 4.58 0.079 27A 109.2 9.12 4.86 0.119 3A 79.1 3.69 4.66 0.087 28A 106.4 9.54 4.75 0.108

4A 79.9 3.24 1.80 0.089 29A 95.5 6.78 3.74 0.119 5A 77.7 3.10 1.42 0.071 30A 99.1 6.67 3.77 0.105 6A 82.1 7.58 2.08 0.071 31A 111.9 10.3 5.15 0.123 7A 78.2 5.74 1.56 0.071 32A 98.8 5.60 3.27 0.143 8A 80.9 3.54 1.85 0.067 33A 98.9 7.41 3.64 0.143 9A 81.2 3.47 1.85 0.067 34A 100.5 10.9 4.52 0.138 10A 83.6 5.23 1.45 0.087 35A 100.5 10.9 4.15 0.138 11A 85.1 13.1 1.42 0.091 36A 93.3 4.74 4.16 0.097 12A 72.3 10.7 4.79 0.083 37A 97.5 3.80 5.13 0.167 13A 79.7 9.57 4.82 0.100 38A 108.0 4.59 7.25 0.192 14A 77.1 13.3 5.44 0.083 39A 89.0 9.54 4.53 0.148 15A 87.6 8.82 1.15 0.091 40A 92.1 4.70 2.58 0.061 16A 79.0 7.35 1.12 0.075 41A 94.8 5.00 2.96 0.077 17A 91.6 4.00 3.02 0.094 42A 96.5 7.80 2.58 0.061 18A 82.5 2.47 2.07 0.109 43A 93.9 6.15 2.96 0.077 19A 84.7 2.46 2.86 0.118 44A 103.1 6.96 4.20 0.095 20A 91.7 5.31 4.37 0.150 45B 103.4 3.29 5.31 0.134 21A 92.6 8.42 3.16 0.094 46B 96.0 4.41 2.76 0.089 22A 95.3 4.57 3.44 0.089 47B 114.0 4.27 3.37 0.068 23A 97.4 4.41 4.41 0.156 48B 78.0 5.27 5.28 0.054 24A 91.4 6.27 3.18 0.094 49B 94.6 9.42 5.93 0.085 25A 107.1 8.80 4.81 0.094 50B 104.5 9.69 7.06 0.060 a,b Calculated using quantum approach as described 2.2.2., measured in units Å³ and Debye, respectively. c,d Calculated using semi-empirical approaches as described 2.2.2., unitless properties.

Table 2.2. Compound MLR descriptors

PLR statistic Aß model insulin model

t variables 10 4

x variables 28 22

Y correlation 0.96 0.89

X correlation 0.99 0.96

RMSEloo 0.48 0.51

2 Q loo 0.77 0.74

Table 2.3. PLR models and statistics for Aβ and insulin

Logical Rank Aß40 descriptors Logical block Insulin descriptors block

1 (-) average 5χv Topological (+) hydrophilic factor (Hy) Chemical

2 (+) polarizability (α) Quantum (+) topological electronic factor (TE1) Topological

3 (-) clogP Chemical (+) average 4χv Topological

4 (+) RBF Constitutional (+) charge polarization factor (Qmean) Topological

Table 2.4. Top-ranked PLR descriptors for Aβ40 and insulin pAC50 (+, direct correlation; -, inverse correlation)

Descriptor α μ clogP RBF VIFb

α 1 1.48

μ 0.08 1 1.13

clogP 0.34 -0.02 1 1.36

RBF 0.56 0.27 0.48 1 1.88 a Correlation coefficient (Rij); where 0.5 < Rij < 0.8 signifies weak intercorrelation, and Rij < 0.5 signifies little or no intercorrelation. b 2 Variance inflation factor (VIF) = 1/(1-R j); VIF < 10 signifies weak multicollinearity

Table 2.5. Correlation matrixa

2 2 2 2 2 2 2 2 2 2 2 R R0 R'0 (R -R0 )/R (R -R'0 )/R |R0 - R'0 | k k'

A40: 0.79 0.79 0.74 0.000 0.06 0.05 1.02 0.96

Insulin: 0.74 0.73 0.61 0.004 0.17 0.12 0.88 1.07

Target:a >0.6 — — either <0.1 <0.3 0.85 < either < 1.15

a Target values are taken from Golbraikh and Tropsha (Golbraikh and Tropsha 2002).

Table 2.6. MLR external validation

2.6. Figures

Figure 2.1. QSAR modeling workflow, in which chemical structure and affinity data (circles) were integrated with calculated molecular descriptors (pentagons) and screened using PLR methods (diamond). Top descriptors were then subjected to MLR to create the final QSAR models (rectangle).

Figure 2.2. Correlation plots for MLR models of ThT displacement from Aβ40 and insulin aggregates. Each point represents observed vs. predicted log AC50 values for training and test sets of 41 and six compounds, respectively, whereas the lines represent linear regression of the data points (solid lines, training set; dashed lines, test set). The quality of regressions is indicated by R2 values.

Figure 2.3. Comparison of Aβ and insulin MLR equation coefficients. The bars represent the absolute value of coefficient magnitude  standard error for Aβ40 and insulin molecular targets, whereas the p value corresponds to z-test of null hypothesis between these two targets.

Figure. 2.4. Structural features influencing ThT displacement activity. The donor-π- acceptor organization generates a delocalized electron distribution and a permanent ground state dipole oriented parallel to the long molecular axis.

Figure 2.5. Reported binding modes for Pittsburgh Compound B (PIB, a neutral benzothiazole derivative) on synthetic Aβ40 protofibrils (Petkova, Yau et al. 2006) as deduced by molecular dynamics simulation (Wu, Bowers et al. 2011). A40 atoms are colored grey (carbon), blue (nitrogen), and red (oxygen), whereas PIB atoms are colored orange (carbon), blue (nitrogen), and yellow (sulfur), and overlaid with a transparent molecular surface area. Black arrows mark the points of contact between ligand and protein. (A) Site A consists of a shallow channel flanked by Phe20 and Val18 side chains whereas (B) Site B consists of a wide channel created by Gly33 flanked by Ile31 and Met35. Site B supports deeper insertion of ligand so that it can interact with the hydrogen-bond network of the cross-β sheet backbone through dispersion and induced- dipole effects. These models show how side chain composition can influence depth and width of channels dispersed along the surface of cross- sheet aggregates. Images were created using UCSF Chimera Alpha Version 1.5 (build 31329) software.

CHAPTER 3

LIGAND ELECTRONIC PROPERTIES MODULATE TAU FILAMENT BINDING SITE DENSITY

3.1. Introduction Tauopathic neurodegenerative diseases, including AD and certain forms of frontotemporal lobar degeneration, are defined in part by the accumulation of tau-bearing neurofibrillary lesions in characteristic brain regions (reviewed in (Sergeant, Bretteville et al. 2008)). Because the spatial and temporal distribution of tau aggregation correlates with disease progression, it is used to differentially diagnose and stage tauopathies in postmortem samples (reviewed in (Braak and Braak 1995; Feany and Dickson 1996)). In addition to being a useful biomarker, tau aggregates are potential mediators of neurodegeneration owing to toxicity associated with their accumulation (Mocanu, Nissen et al. 2008). Thus tau aggregates are promising targets for drug discovery efforts as well. Structurally, the tau fibril targets adopt cross--sheet structure characterized by parallel in register -sheets oriented perpendicular to the fibril long axis (Sawaya, Sambashivan et al. 2007). The folded core of the fiber is composed of ~90 amino acid residues located in the C-terminal microtubule binding repeat region of each tau protomer (Novak, Kabat et al. 1993). The solvent exposed side chains of these residues form shallow channels oriented parallel to the fiber axis that are capable of binding small molecules including thioflavin dyes (reviewed in (Jensen, Cisek et al. 2011)). Small molecules capable of binding these sites may support diagnostic and therapeutic applications. To date, the search for tau-binding ligands has focused primarily on affinity, and compounds with Kd in the mid-nanomolar range have been identified (Honson, Johnson et al. 2007; Fodero-Tavoletti, Okamura et al. 2011). However, compound utility depends

on binding site density (Bmax) as well as affinity. For example, the signal generated by whole brain imaging agents is proportional to Bavail/ Kd ratio, or “binding potential”, where Bavail depends on both target concentration and Bmax (Laruelle, Slifstein et al. 2003). Thus, occupancy of high density sites can be important for detection of targets present at low concentration in vivo. In fact, the binding potential needed for a putative tau imaging agents will likely be higher than for established amyloid radiotracers owing to the low molar concentration of the tau binding target in diseased brain relative to A aggregates (Schafer, Kim et al. 2012). Ligand Bmax also may be important for tau fibril- directed therapeutics. For example, one strategy for preventing toxicity associated with aggregate formation involves coating fibril surfaces with ligands (Inbar and Yang 2006;

Habib, Lee et al. 2010). High Bmax ligands would be preferable for maximizing the efficacy of this approach. Despite its importance, little is known about binding site density on protein aggregates, except that substantial variability exists for individual ligands. For example, the presence of at least three binding modes on A filaments that differ in Bmax has been reported (Lockhart, Ye et al. 2005). In the case of tau aggregates, Bmax has been reported to vary over the range of 0.001 – 0.05 mol/mol (i.e., 1:1000 – 1:20 mol ligand/mol tau protomer), suggesting that a wide range of binding site densities appear on tau filaments too (Rojo, Alzate-Morales et al. 2010; Fodero-Tavoletti, Okamura et al. 2011; Matsumura, Ono et al. 2012). Although SAR analysis has revealed structural features on certain scaffold classes that mediate binding affinity (Cai, Innis et al. 2007; Leuma Yona,

Mazeres et al. 2008), descriptors associated with various Bmax modes of interaction are not understood. To better characterize the interaction between small molecules and cross--sheet aggregates, I used benzothiazole derivatives as model ligands and QSAR analysis as a means to identify their binding descriptors. Using this approach, ligand polarizability as an important descriptor of thioflavin dye displacement potency from synthetic cross-- sheet aggregates composed of insulin and A were identified (Cisek and Kuret 2012). Moreover, a correlation between ligand polarizability and thioflavin dye displacement potency that extended to synthetic tau aggregates was found (Jensen, Cisek et al. 2011).

Interestingly, as thioflavin probe concentrations rose, the displacement efficacy of weakly polarizable ligands was greatly attenuated whereas the efficacy of highly polarizable ligands was unchanged (Jensen, Cisek et al. 2011). These data suggested that the modes of interaction between aggregates and fluorescent thioflavin dye probes vary with probe concentration, and that ligand polarizability is a potential descriptor of binding sites occupied at high probe concentrations. Overall, these data suggest that polarizability is a 18 potential descriptor of Bmax as well as of Kd. Consistent with this possibility, an F- labeled analog of one of most highly polarizable benzothiazole analogs in our series,

FPPDB, has achieved the highest Bmax yet reported for a synthetic tau filament binding agent (~1:20 tau protomer:ligand ratio)(Matsumura, Ono et al. 2012). Here the polarizability hypothesis is tested using as two closely related benzothiazole derivatives as ligands and both synthetic and authentic disease derived tau filaments as binding target. The results indicate that incorporation of polarizability into ligand design may yield agents with superior diagnostic and therapeutic performance. 3.2. Experimental Procedures 3.2.1. Materials Octadecyl sulfate (ODS) was from Research Plus (Bayonne, NJ), ThT, ThS and N- laurylsarcosine were from Sigma-Aldrich (St. Louis, MO), Congo red (ultrapure grade) was from Anaspec (Fremont, CA), and compounds 1 and 2 were custom synthesized as described (Honson, Jensen et al. 2009). Both 125I-radiolabeled (in 40% ethanol) and unlabeled IMSB (5) were the generous gifts of Avid Pharmaceuticals (Philadelphia, PA). 3.2.2. Protein preparation

Recombinant human His6-2N4R tau protein was prepared as described previously (Carmel, Leichus et al. 1994). Synthetic tau filaments were prepared from recombinant protein (5 μM) incubated (24 h at 37°C) in assembly buffer (10 mM HEPES, pH 7.4, 100 mM NaCl, 5 mM dithiothreitol) in the presence of 50 μM ODS inducer (Necula, Chirita et al. 2003; Honson, Johnson et al. 2007), and stored in aliquots at -80°C until used. Authentic PHF-tau was prepared from AD tissue as described previously (Jensen, Cisek et al. 2011). The tau content of PHF preparations was determined by dot blot using recombinant 2N4R tau as standard. Aliquots were diluted in 0.1% SDS and 0.1% 2-

mercaptoethanol, then spotted in triplicate onto 0.2 μM nitrocellulose membranes. After blocking in 4% nonfat dry milk, blots were developed using monclonal antibody Tau5 as described previously (Chang and Kuret 2008). PHF-tau weights were converted to moles assuming a mean molecular mass of 40,238 g/mol, which reflects the estimated ratios of all six human isoforms (Goedert and Jakes 1990; Hong, Zhukareva et al. 1998) 3.2.3. Filament length distributions Filament populations were viewed and digitally captured on a Tecnai G2 Spirit BioTWIN transmission electron microscope (FEI, Hillsboro, OR) operated at 80kV and 68,000x magnification. Filament lengths were measured using ImageJ software (National Institutes of Health) and segregated into bins as described previously (Necula and Kuret 2004). Length distributions were fit to the log normal function:

2  ln(x / x )    0   0.5    b   y  ae  (3.1) where y is the number of filaments per bin of length interval midpoint x, a is the number of filaments per bin at distribution mode x0, and b is an estimate of distribution skew. 3.2.4. Fluorescence displacement assays ThS fluorescence assays were performed as described previously (Honson, Johnson et al. 2007). Synthetic 2N4R-tau and authentic PHF-tau filaments (1 uM total tau concentration) were incubated (2 h at 37°C) in assembly buffer with varying concentrations of ThS probe (1.7 x10-11 – 10-5 M) and test ligand (0 – 10-5 M). Fluorescence was measured using a FlexStation microplate reader (Molecular Devices,

Sunnyvale, CA) at ʎex = 440 nm; ʎem = 490 nm; filter = 475 nm. Thioflavin dye displacement (net fluorescence) was determined by subtracting no protein controls from protein and compound fluorescence readings. Net fluorescence readings were normalized to percent displacement using no compound controls. 3.2.5. Radioligand binding assays [125I]IMSB binding and displacement from tau filaments was assessed using a filter trap assay (Klunk, Debnath et al. 1994; Lockhart, Ye et al. 2005). Tau filaments (230 nM

synthetic 2N4R-tau; 300 nM authentic PHF-tau) were incubated (4h at 37°C with hourly agitation) with varying concentrations of [125I]IMSB and ligand in PBS buffer (10 mM

Na2HPO4/KH2PO4, 137 mM NaCl, 2.7 mM KCl, pH 7.4). For determination of nonspecific binding, ligand was replaced with 8 μM of Congo Red. Aliquots were then vacuum filtered (5 – 10 s) over glass fiber filters (0.3 μm pore diameter, Sterlitech) in a Millipore 1225 Sampling Manifold using a Pharmacia LKB VacuGene Pump operated at 50 mbar. Each sample was then washed four times with 3-ml volumes of PBS. Filters were collected in polypropylene tubes and counted using a Packard A5003 gamma counter (Packard Instrument Company, Meriden, CT) with 84% efficiency. 125I-IMSB specific binding was determined by subtracting non-specific binding from total binding. Under these assay conditions, the specific binding signal accounted for ~80% of the total radioactivity. 3.2.6. Spectrophotometry Absorbance spectra were collected in methanol solvent using a CARY50Bio UV-VIS spectrophotometer and recorded with the Cary WinUV Scan Application version 3.00(182). Spectra were fit to a double Gaussian function as described previously (Necula, Chirita et al. 2005). 3.2.7. Analytical methods Binding and displacement potency of ligands in radioactivity and fluorescence assay formats was determined from the function:

Fmax  Fmin F  Fmin  (3.2) 110(log AC50x)n

For direct binding assays, F represents the net signal at concentration x of probe, Fmin is the signal at zero probe concentration, Fmax is the net signal at infinite probe concentration, n is the Hill coefficient, and AC50 corresponds to probe Kd. For displacement assays conducted at constant probe concentration, F represents the probe signal in the presence of varying concentrations of test ligand, Fmin is the signal at infinite ligand concentration, Fmax is the signal in the absence of test ligand, AC50 is the concentration of test compound that reduces probe signal by 50%. 3.2.8. Computational chemistry

Quantum property calculations were performed using density functional theory methods implemented in Gaussian 09 (G09) (Frisch 2009) and Turbomole V6.3.1 (Ahlrichs, Bar et al. 1989) software packages available on Ohio Supercomputing Center clusters. All calculations were performed using hybrid density functional B3LYP and the 6-311++G(d,p) basis set. Dipole moments were calculated in G09 after optimizing ground state geometries using G09 as described previously (see Chapter 2) (Cisek and Kuret 2012). Electrostatic potential surface calculations were performed on compound ground-state geometries using population analysis with atomic charge assignments produced according to the ChelpG scheme (Breneman and Wiberg 1990) (G09 keyword pop=chelpg) as implemented in G09. Surface plots were generated in GaussView 4.1.2 by mapping electrostatic potentials onto SCF (self consistent field) total density surfaces with isodensity contour values of  0.0004 a.u. To facilitate direct comparison between compounds, electrostatic potential energy scales were adjusted to a standard color scale ( 55.8 kcal/mol) such that all potentials resided within the standard extremes. For all G09 calculations, bulk solvent effects were implicitly modeled with the polarizable continuum model (Tomasi, Mennucci et al. 2005) (G09 keywords SCRF =(Solvent = Methanol)). Electronic difference density calculations were performed in Turbomole for first excited (S1) and ground states (S0) on compound ground-state geometries under C1 molecule symmetry (control keywords $scfinstab and $soes) as implemented in the Turbomole egrad module. Because this module does not implement implicit solvent effects, these calculations were performed in gas phase. Electronic difference maps were generated using UCSF Chimera Alpha Version 1.5 (build 31329) (Pettersen, Goddard et al. 2004) software with isodensity contour values  0.004 a.u. Contour volumes were measured using Chimera’s Volume tool. 3.3. Results 3.3.1. Tau filament populations for binding studies Synthetic filaments composed of recombinant 2N4R tau were prepared to provide binding sites for this study. The 2N4R isoform was chosen because it efficiently

aggregates in vitro and has been used in previous investigations of ligand binding (Honson, Johnson et al. 2007; Jensen, Cisek et al. 2011; Matsumura, Ono et al. 2012). Tau filaments prepared from AD brain also were prepared so that interaction with authentic aggregates could be compared. These filaments differed from synthetic filaments by being composed of all six CNS isoforms and by containing extensive posttranslational modifications (reviewed in (Sergeant, Bretteville et al. 2008)). Both preparations displayed clear filamentous character when viewed in transmission electron micrographs (Figure 3.1A and 3.1B). However, they differed in morphology, with synthetic filaments displaying straight morphology whereas the authentic filaments retained the classic morphology of PHF (Figure 3.1A and 3.1B). Moreover, the two filament populations differed in length distribution. The PHF population averaged 193 nm in length and displayed a clear mode at 123  2 nm when fit to a log-normal distribution (Figure 3.1C). These dimensions, which were consistent with literature values for filaments isolated by differential centrifugation (Chirita and Kuret 2004), reflect the effects of breakage and shearing during isolation. In contrast, synthetic 2N4R- tau filaments averaged 776 nm in length and were predicted on the basis of a log-normal distribution to have a mode within the standard error of estimate of zero (Figure 3.1C). These data indicate that synthetic filaments adopted a nearly exponential length distribution (Chirita, Necula et al. 2003), and that this distribution was skewed toward longer lengths relative to PHFs. This would suggest that synthetic filaments may provide more available binding sites for small molecules, both as a function of longer lengths of these fibrils relative to PHFs, as well as more accessible due to lack of posttranslational modifications which are prevalent on PHFs. 3.3.2. Displacement probes To characterize the binding surfaces presented by tau filaments, the interaction of each filament preparation with fluorescent thioflavin dyes (ThT and ThS)(Kelenyi 1967) and radioprobe [125I]IMSB (Zhuang, Kung et al. 2001) was investigated. These compounds were chosen as binding probes because they had been reported to bind tau- bearing neurofibrillary lesions in brain sections (King, Ahuja et al. 1999; Kung, Skovronsky et al. 2003) yet differed substantially in structure. For example, the major

component of ThS adopts a donor-π-acceptor (D-π-A) configuration, composed of strong benzothiazolium electron acceptors linked to a strong dimethylamine electron donor through a planar conjugated π network (Table 3.1). In contrast, IMSB has two weak methoxy electron donors feeding into a conjugated aromatic π network yielding a donor- π-donor (D-π-D) configuration (Table 3.1). As a result, ThS was a colored compound with a strong absorbance maximum at 377 nm and a weak shoulder at 420 nm, whereas IMSB appeared colorless with an absorbance maximum centered at 350 nm (Figure 3.2). Overall, ThS and IMSB probes were charged molecules of similar size and shared planar, highly conjugated organization (Table 3.1). However, they differed in hydrophobicity and electronic structure, especially in the magnitude of delocalization of their π electron systems. When assayed in filter trap format (300 nm pore diameter) at constant tau protomer concentration, both 2N4R aggregates and PHF-tau bound [125I]IMSB with affinities estimated in the low nanomolar range (Figure 3.3A). However, Bmax was only 1:112  25 ([125I]IMSB:tau) for 2N4R tau and 1:2990  160 ([125I]IMSB:tau) for PHF. Because the percentage of total aggregation distributed into filaments ≥300 nm in length was higher for synthetic filaments (90%, Figure 3.1C) than for PHFs (31%, Figure 3.1C), a portion of the observed Bmax difference may have resulted from differences in filament length distribution. However, the majority of the 23  2 fold difference in Bmax reflected a lower density of high affinity IMSB sites on PHFs relative to synthetic filaments. When synthetic filaments were assayed in fluorescence format with near saturating concentrations of ThS and ThT, both probes underwent excitation and emission Stokes shifts and fluoresced brightly (Figure 3.3B). These results were consistent with previous reports (Friedhoff, Schneider et al. 1998). However, only ThS fluoresced strongly in the presence of PHF, with ThT fluorescence intensity being 15  1 fold lower than in the presence of equimolar concentrations of 2N4R filament protomer (Figure 3.3B). To maintain tau protomer concentrations at similar levels for interaction assays involving synthetic and authentic filaments, all subsequent studies involving thioflavin dye probes used ThS exclusively. To clarify the relationship between IMSB and ThS sites, the ability of probes to

displace each other was investigated. Despite its high binding affinity, IMSB proved incapable of displacing ThS from either synthetic tau filaments (Figure 3.4A) or authentic PHF-tau (Figure 3.4B) by 50% when present at saturating concentrations up to 10 μM. In contrast, ThS completely displaced [125I]IMSB from synthetic filaments with

AC50 of 249  60 nM (Figure 3.5). Together these data indicated that IMSB and ThS differentially interacted with tau filaments, with IMSB selectively probing low density sites capable of supporting high affinity binding. In contrast, ThS interacted with IMSB sites but also additional sites that were present at similar densities on both PHF and synthetic filaments. 3.3.3. Benzothiazole ligands differentially interact with tau aggregates To further characterize IMSB and ThS modes of interaction, the ability of neutral benzothiazole derivatives 1 and 2 (Table 3.1) to displace each probe was investigated. The two ligands differed only by the structure of the bridge that connected the benzothiazole and aryl moieties: 1 contained an azo bridge whereas 2 contained an alkene bridge. As a result, the two compounds shared similar size, hydrophobicity and neutral net charge (Table 3.1). Nonetheless, bridge structure greatly affected compound absorbance, with compound 1 appearing red and having an intense lowest energy charge- transfer absorption band centered at 508 nm whereas 2 appeared yellow and absorbed at 393 nm (Figure 3.2B). Consistent with previous results (Honson, Jensen et al. 2009), 1 efficiently displaced ThS probe from synthetic 2N4R filaments with an AC50 of 31  6 nM (Figure 3.4A). In contrast, 2 did not displace ThS by 50% at concentrations up to 10 μM. A similar pattern was observed for ThS displacement from PHF, where 1 displaced

ThS with AC50 of 48  4 nM and 2 displaced with AC50 > 10 μM (Figure 3.4B). These data indicate that 1 had far greater ability than 2 to interact with binding sites occupied by ThS, and that this pattern of interaction held for authentic PHFs as well as for synthentic filaments. On the basis of ThS probe displacement described above, compound 2 appeared to bind poorly to tau filaments. Yet when tested in [125I]IMSB competition format, 1 and 2 both displaced probe with similar affinity (AC50s of 29  6 and 45  11 nM, respectively), and with nearly equal efficacy (Figure 3.5). These data reveal that both 1 and 2 were

capable of binding tau filaments with high affinity at the low density sites bound by IMSB, but that they differed greatly in their ability to bind the additional sites occupied by near saturating concentrations of ThS probe. 3.3.4. Electronic properties of tau aggregate binding ligands To characterize their electronic properties, 1 and 2 were analyzed using a quantum chemical modeling approach. First, the ground state geometries of the E and Z geometric isomers of 1 and 2 were optimized in polar solvent at the B3LYP/6-311++G(d,p) level of theory. Isomerization was found to only modestly influence calculated properties (e.g., dipole moment). Nonetheless, all subsequent analysis focused on E geometric isomers because they had been reported to weakly predominate in polar solvents (Sigmundova, Zahradnik et al. 2007). Second, to assess overall compound polarity in the ground state, calculated molecular electrostatic surface potentials were mapped onto electron density isosurfaces and plotted along with dipole moment. Both compounds presented electronegative surfaces along the faces of their conjugated π systems (Figure 3.6A) However, for compound 1, negative electrostatic potential localized most intensely on the nitrogen atoms of the benzothiazole heterocycle and azo bridge, with positive electrostatic potential localized primarily on the dimethylamine electron donor group. This pattern was consistent with extensive delocalization of electrons through the conjugated π system from donor groups toward thiazole and azo group nitrogen atoms. The resulting separation of negative and positive charges created a permanent ground state dipole moment of 7.8 D oriented approximately parallel to the compound long axis (Figure 3.6A). In contrast, compound 2 supported lower charge polarity, with localized negative charge being limited to the nitrogen atom of the benzothiazole moiety (Figure 3.6A). As a result, ground state dipole moment decreased 30% relative to 1. These data indicate that the azo bridge fostered more effective coupling between electron donor and acceptor groups than the alkene bridge, resulting in more charge transfer character in the ground state of 1 relative to 2 in polar solvent. Proposed resonance forms that contribute to charge transfer character are shown in Figure 3.6B. Finally, to model how bridge composition affected electron movement to higher

energy states, vertical excitations were calculated for 1 and 2 at the B3LYP/6- 311++G(d,p) level of theory and used to compute differences in electron density between the ground (S0) and lowest identified singlet excited states (S1). Analysis was limited to gas phase owing to computational constraints. The resultant S1 – S0 difference maps identified the locations of electron density that dispersed from the ground state and accumulated upon excitation to the higher energy level (Figure 3.7). The S1 – S0 difference density map for 2 exhibited a change in electron density that alternated in sign along the conjugated alkene bridge, which is characteristic of a (π, π*) excited state. Electron density on the methoxy and dimethylamine donors shifted further toward the thiazole ring in the excited state. In contrast, the S1 – S0 difference density map for 1 was dominated by the depletion of electron density from the lone pairs of the bridge nitrogen atoms, implicating a (n, π*) excited state. Electron density redistributed toward the thiazole and phenyl ring systems and within the bridge. Overall, the difference maps showed that the volume of perturbed density was 1.7-fold greater for 1 relative to 2 (on the basis of identical isocontouring), consistent with increased charge transfer efficiency in this compound. Together these calculations show that simple replacement of the alkene bridge in 2 with an azo moiety to create 1 altered compound electronic properties by enhancing electron delocalization in the ground state and by supporting greater intramolecular charge transfer in response to photoexcitation. These characteristics, which were fostered while holding compound hydrophobicity, size, and net charge nearly constant, correlated with increased displacement efficacy against sites occupied by ThS probe relative to those occupied by the IMSB probe. 3.4. Discussion These data show that the efficacy of probe displacement from tau filaments is highly sensitive to the electronic properties of ligand. Importantly, they show that displacement of near saturating concentrations of ThS probe, which occupies sites over and above those occupied by IMSB, can be selectively enhanced by fostering a highly delocalized electronic structure. Thus, they suggest a route for rationally maximizing the binding potential of imaging radiotracers and the efficacy of interaction-inhibiting molecular coatings (Inbar and Yang 2006; Habib, Lee et al. 2010). Because PHF-tau shares this

behavior with synthetic 2N4R-tau filaments, the route can yield ligands with utility for human disease. Because electronic properties can be modulated in small molecules while maintaining zero net charge, the proposed strategy is consistent with the need to maintain blood brain barrier penetrability of ligands for diagnostic and therapeutic applications. A role for electronic structure in mediating filament binding is consistent with recent models of small molecule/cross--sheet aggregate interactions formulated on the basis of docking (Rodriguez-Rodriguez, Rimola et al. 2010; Bieschke, Herbst et al. 2012) and molecular dynamics simulation (Wu, Wang et al. 2008; Wu, Biancalana et al. 2009; Wu, Bowers et al. 2011). These computational models predict that binding occurs in shallow channels oriented parallel to the long filament axis formed by solvent exposed amino acid side chains. Thus, potential binding sites are predicted to vary with the dimensions of the channels and their surface chemical properties (e.g., amino acid side chain composition). The most energetically favorable interactions identified involved Phe residues, which presented a surface composed of aromatic moieties, and Gly residues, which exposed the delocalized π system of the cross--sheet main chain core (Rodriguez-Rodriguez, Rimola et al. 2010; Wu, Bowers et al. 2011). In the case of compounds investigated herein, two modes of binding with tau aggregates could be discerned. The first mode, which we call

Class I, is characterized by high affinity but low Bmax binding. Class I activity is shared by diverse structures composed of planar and highly conjugated ring systems that present π-electron rich surfaces, including all four of the probes and ligands investigated herein. Their π-electron systems can but need not be highly delocalized, and as a result these compounds are frequently colorless (e.g., IMSB). The location of their binding sites on tau aggregates is not established, but π-rich heterocycles have been reported to support cation-π, π-π, and CH-π interactions (reviewed in (Meyer, Castellano et al. 2003)). A tau-aggregate directed radiotracer with Class I electronic character currently undergoing clinical development, THK523 (Fodero-Tavoletti, Okamura et al. 2011), shares a D-π-D electronic organization with IMSB. Although it binds synthetic tau aggregates with a low nanomolar Kd, its Bmax is only 0.002 mol/mol (i.e., one ligand molecule for every 460 tau protomers). This binding site density is within ~4-fold of IMSB, but >20-fold lower than FPPDP, an analog of 1. On the basis of mathematical modeling, we have predicted that

this combination of binding parameters (i.e., binding potential) will have difficulty overcoming nonspecific background associated with whole-brain imaging (Schafer, Kim et al. 2012). Imaging performance will be weaker still if, as shown here for [125I]IMSB,

Bmax for authentic PHF is lower than for synthetic tau filaments. In contrast, the second binding mode, which we term Class II, is shared by compounds comprising planar and highly conjugated π systems flanked by electron donor and acceptor groups (Jensen, Cisek et al. 2011). As a result, these compounds, which can be intensely colored (e.g, compound 1), have more highly delocalized electrons in the ground state and greater electron mobility in response to photoexcitation than Class I compounds. These electronic characteristics are ideal for supporting attractive van der Waals interactions with flat binding surfaces on tau and other protein aggregates (reviewed in (Jensen, Cisek et al. 2011)). Conversely, Class I compounds lack this feature and displace Class II ligands poorly. We propose that this mode of interaction is responsible for the high binding site densities associated with a radiolabeled derivative of 1 (Matsumura, Ono et al. 2012), which was reported to bind synthetic tau aggregates with 125 Bmax up to one order of magnitude greater than Class I compounds [ I]IMSB (reported herein) or [18F]THK523 (Fodero-Tavoletti, Okamura et al. 2011). On the basis of mathematical modeling, we predict that this Bmax (when accompanied by low nanomolar affinity) is adequate for whole-brain imaging applications (Schafer, Kim et al. 2012). Moreover, we have reported that compound 1, but not 2, is capable of blocking toxicity associated with tau aggregation in a biological model of tauopathy (Honson and Kuret 2008). Together these data suggest that the Class II mode of interaction offers translational opportunities for tauopathic neurodegenerative diseases. Our reported QSAR analysis of benzothiazole derivatives (Cisek and Kuret 2012), along with quantum chemistry analysis of compounds 1 and 2 reported herein, provide insight into the structural features associated with Class II activity. The most important feature is the introduction of electron acceptor and donor moieties that foster electron delocalization. In the case of 1, this was achieved by placing an azo bridge adjacent to the weak benzothiazole acceptor in a conjugated π system flanked by electron donating groups. The resulting D-π-A-π-D architecture strongly enhanced intramolecular charge

transfer, leading to extensive π electron delocalization in the ground state and lower energy requirements for transition to the first excited state. The resulting increase in dipole moment and polarizability in 1 relative to 2 reflect these features. The dispersion effects supported by this architecture may be especially important for high density binding to tau aggregates, which contain multiple Gly and aromatic residues in their folded core (Novak, Kabat et al. 1993). Nonetheless, compound 1 does not discriminate between A and tau on the basis of ThS displacement activity (Honson, Johnson et al. 2007), and this has been confirmed by direct radiobinding assays (Matsumura, Ono et al. 2012). It remains to be seen whether Class II interactions can be harnessed to drive selectivity among protein aggregates composed of different protein protomers. In summary, here we found that highly delocalized electronic structure fostered potent displacement of cross--sheet binding probes from tau aggregates, but only from a subset of binding sites marked by near-saturating concentrations of ThS probe. The high density nature of the sites, combined with the compatibility of delocalized electronic structure with overall neutral compound charge, suggests a route for optimization of binding properties for diagnostic and therapeutic applications.

3.5. Tables

a Compound Structure MW max clogP Volume

(nm) (Å3)

Probes

ThS 493 377 -4.3 427

IMSB 554 350 3.7 418

Ligands

1 310 508 4.8 277

2 312 393 4.6 285

aThS is a mixture of substances (Kelenyi 1967). The chemical structures of its major components are available in the PubChem Substance and Compound database through substance identifier number SID: 8137049 and/or unique chemical structure identifier CID: 415676 (ThS). Table 3.1. Compound structures and characteristics

3.6. Figures

Figure 3.1. Tau filament morphology and length distribution. Transmission electron microscope images of negatively stained A, synthetic filaments composed of 2N4R-tau and B, authentic tau filaments purified from AD brain were captured at 68,000-fold magnification. Synthetic filaments adopted straight morphology, whereas authentic filaments adopted classic paired helical morphology. C, the dimensions of all filaments ≥50 nm in length measured from electron micrographs were plotted as the fraction of total filament number that segregated into consecutive length intervals (total filament number n = 247 and 111 for 2N4R and PHF-tau, respectively), whereas each line represents the best fit of data points to a log normal function (Eq 1). The length distribution of synthetic 2N4R filaments skewed toward longer lengths relative to the PHF population.

Figure 3.2. Optical properties of probes and ligands. Experimental absorption spectra for A, probes ThS and IMSB and B, ligands 1 and 2 were determined in methanol solvent. Each point represents absorbance normalized for cuvette pathlength and compound concentration (i.e., extinction coefficient) plotted as a function of wavelength (). See text for details.

Figure 3.3. Probe interactions with filamentous 2N4R- and PHF-tau. A, binding of radioprobe [125I]IMSB was determined after incubation with synthetic and paired helical tau filaments. Each point represents specific binding as a function of total [125I]IMSB concentration  SD (triplicate determination), whereas the lines represent the best fit of the data points to a binding isotherm (Eq 3.2). [125I]IMSB bound to both synthetic and paired helical filaments with low nanomolar affinity. ***, p < 0.001; n.s., p > 0.05 for each comparison. B, probes ThS (7.5 μM) and ThT (1 μM) were incubated with tau filaments composed of 2N4R- and PHF-tau (each at 1 μM), then subjected to fluorescence measurements at ex = 440 nm, em = 490 nm. Each bar represents net fluorescence yield associated with binding  SD (triplicate determination). ThS fluoresced strongly in the presence of both 2N4R filaments and PHFs, whereas ThT fluoresced most strongly in the presence of 2N4R filaments.

Figure 3.4. ThS displacement assays for tau filament binding. A, synthetic 2N4R-tau filaments and B, PHFs (both at 1 μM) were incubated with fluorescent ThS probe (7.5 μM) in the presence of varying concentrations of 1, 2, and unlabeled IMSB, then subjected to fluorescence measurements at ex = 440 nm, em = 490 nm. Each point represents mean net ThS fluorescence (F) (expressed as % fluorescence in the presence of vehicle alone) whereas the solid lines reflect best fit of data points to Eq 2. Only compound 1 decreased ThS fluorescence from both filament populations at submicromolar concentrations under these conditions.

Figure 3.5. 125I-IMSB displacement assays for tau filament binding. Binding of radioprobe 125I-IMSB to synthetic 2N4R-tau filaments (1 μM) was determined after incubation in presence of varying concentrations of 1, 2, and ThS. Each point represents specific binding (expressed as % specific binding in the presence of vehicle alone) as a function of total ligand concentration  range (duplicate determination), whereas the lines represent the best fit of the data points to Eq 2. All three compounds displaced 125I-IMSB specific binding under these conditions.

Figure 3.6. Ligand electrostatic potential surfaces. A, plots for 1 and 2 were generated at the B3LYP/6-311++G(d,p) level of theory and visualized in GaussView 4.1.2. The semi-transparent electrostatic potential surface, color-coded red (-55.8 kcal/mol) to blue (55.8 kcal/mol), overlays compound ball and stick atoms (carbon, grey; hydrogen, white; nitrogen, blue; oxygen, red; and sulfur, yellow). Calculated dipole moments are depicted as arrows. B, neutral and charge transfer resonance forms, where X = C or N. See text for details.

Figure 3.7. Ligand electronic difference density maps. Plots for 1 and 2 were generated at the B3LYP/6-311++G(d,p) level of theory and visualized in UCSF Chimera Alpha Version 1.5 (build 31329) with isocountour values normalized to  0.004 a.u. Green contours represent the accumulation of electron density in the S1 excited state whereas red contours depict the loss of electron density from the S0 ground state. Atom colors represent carbon (gray), hydrogen (white), nitrogen (blue), oxygen (red), and sulfur (yellow). See text for details.

CHAPTER 4

QSAR MODELING OF BENZOTHIAZOLE-ARYLS FOR TAU-DIRECTED IMAGING AGENT DEVELOPMENT

4.1. Introduction Many neurodegenerative diseases present similar clinical symptoms complicating the diagnoses of these disorders. However, the underlying pathology of these conditions are the characteristic cross-ß-sheet lesions that appear in different regions of the brain, such as the tau-bearing tangles and Aß deposits that accumulate in AD (reviewed in (Sergeant, Bretteville et al. 2008)). Because tau tangles precede the formation of Aß deposits in certain regions of the brain, such as the entorhinal cortex (Duyckaerts and Hauw 1997; Morsch, Simon et al. 1999), closely correlate with disease progression (Braak and Braak 1991; Ghoshal, Garcia-Sierra et al. 2002; Royall, Palmer et al. 2002) and guide the postmortem assessment of disease severity (Braak and Braak 1991), tracking the spatiotemporal pattern of tau deposition has an advantage over other methods of premortem diagnosis. This approach is feasible provided that a tau-selective diagnostic agent were available for whole brain imaging purposes. The greatest challenge in developing a selective probe for protein aggregates is the lack of experimental evidence and atomic-level models of the binding interactions between small molecules and these targets that could be used to guide the optimization process. Additionally, these molecular targets are not the typical lock-and-key binding models with stoichiometric one to one binding site density; instead, they present solvent- exposed surfaces that contain shallow grooves and clefts formed by the side chains of peptide residues (Petkova, Yau et al. 2006) that may support different modes of interactions and binding constants. Moreover, fibrils may adapt different morphologies

(Fandrich, Schmidt et al. 2011) such that binding site density (Bmax) may vary with fibril type (Groenning 2009). Therefore, in order to develop a tau-selective imaging agent without any direct experimental structural data about the binding target, QSAR models, which correlate binding affinity to compound molecular properties, will be crucial to guide the rational compound design and optimization process (Stone and Jonathan 1993; Peltason and Bajorath 2009). Recently, QSAR analysis was used to identify molecular properties driving binding affinity and selectivity of benzothiazole derivatives for cross ß-sheet aggregates (Cisek and Kuret 2012). It was found that compound polarizability, hydrophobicity, dipole moment and torsional freedom contributed to potent binding affinity, as well as that polarizability and hydrophobicity could be fine-tuned for aggregate selectivity. To address the questions of binding affinity and selectivity of the benzothiazole scaffold for tau aggregates, we have devised a two method QSAR approach. First, 2D “global” descriptors were calculated as part of a classical QSAR approach, followed by a 3D- QSAR to identify critical R-groups of highly potent ligands (Vanopdenbosch, Cramer et al. 1985; Stone and Jonathan 1993; Peltason and Bajorath 2009). The experimental input for this analysis included the structures and binding affinities of a small library of 42 benzothiazole-aryl compounds tested in a ThT displacement assay against in vitro 2N4R tau and Aß40 fibrils. Because more than half of the dataset tested inactive against Aß40 fibrils, the dataset was augmented with 13 benzothiazole-aryls from literature tested, which were against Aß40 fibrils in the same assay format (Leuma Yona, Mazeres et al. 2008). Thus, the compound library under investigation was structurally concise yet comprised of affinities for tau and Aß fibrils that varied over three orders of magnitude for each molecular target. Importantly, the selectivity of certain benzothiazoles for tau over Aß was over 50-fold and the affinities extended into the radiotracer concentrations (low nanomolar). The resulting QSAR analyses confirmed polarizability, molecular topology, hydrophobicity and certain steric features as affinity- and selectivity-driving features of the benzothiazole series under investigation and provided guidelines for compound design and tau-selectivity optimization.

4.2. Methods 4.2.1. Materials Thioflavin T (ThT), sodium phosphate monobasic, sodium phosphate dibasic, hexafluoroisopropanol (HFIP), 4-Morpholineethanesulfonic acid (MES), ethylene glycol- bis(2-aminoethylether)-N,N,N′,N′-tetraacetic acid (EGTA), and N-laurylsarcosine (Sarcosyl) were purchased from Sigma-Aldrich (St. Louis, MO). 1-11 and 13-42 were from the NIH Chemical Genomics Center (Rockville, MD), 12 was from Anaspec (Fremont, CA). Microfluor I 96-well black microplates, 4-(2-Hydroxyethyl)-1- Piperazineethanesulfonic Acid (HEPES), dimethyl sulfoxide (DMSO), isopropanol, magnesium chloride (MgCl2), and sodium chloride (NaCl) were from Fisher Scientific (Fair Lawn, NJ). Phenylmethylsulfonyl fluoride (PMSF) was from Roche Diagnostics (Indianapolis, IN). Sucrose was from Mallinckrodt (Paris, KY). Dithiothreitol (DTT) was from Gold Biotechnology (St. Louis, MO). Ethylenediaminetetraacetic acid (EDTA) was purchased from Amresco (Solon, OH). Octadecyl sulfate (ODS) were from Research Plus (Bayonne, NJ). Siliconized pipette tips were from BioPlas (San Rafael, CA) and siliconized 1.7 mL microcentrifuge tubes were from Midwest Scientific (St.

Louis, MO). Recombinant, His6-2N4R, was prepared and stored as previously described

(Carmel, Mager et al. 1996; Necula, Chirita et al. 2003). Aß1-42 was purchased as lyophilized solid from Biopeptide (San Diego, CA). 4.2.2. Fibrillization

Recombinant 2N4R tau was fibrillized as described in 3.2.2. Synthetic Aß1-40 was fibrillized using a modification of methods described previously (Zhuang, Kung et al. 2001; Stine, Dahlgren et al. 2003). Briefly, the peptides were dissolved in HFIP (1 mg/mL), and then HFIP was evaporated under argon leaving a dry film. The film was first dissolved in DMSO (1 mM), and then diluted (100 μM) with buffer (10 mM phosphate pH 7.4, 1 mM EDTA). The solution was aliquated into siliconized 1.7 mL microcentrifuge tubes, incubated with rotisserie agitation for 48 hours at 37°C and then stored at -80°C until needed for assays.

4.2.3. Displacement assays Displacement assays were performed as described in 3.2.4. and analyzed as described in 3.2.7. 4.2.4. Computational procedures and QSAR model generation Classical QSAR descriptors were calculated as described previously in section 2.2.2. Additional descriptors, compound surfaces and volumes were analytically computed using the ASV software (Petitjean 1994) available at http://petitjeanmichel.free.fr/ itoweb. petitjean.freeware.html. Molecular surface and volume parameters were calculated using Bondi radii (Rowland and Taylor 1996) and maximum precision (EPSTAB=0 and NPERM=0), and the values given in units of Å² and Å³ respectively. Multiple compound alignment was performed using the pharmACOphore program (Korb, Monecke et al. 2010) available at http://www.tcd.uni-konstanz.de/research/ pharmacophore.php on OSC clusters. The alignment was performed flexibly using default parameters, except the iteration scaling factor which was increased due to the size of the dataset (aco_sigma 40). 3D-QSAR was performed using the comparative similarity index analysis (CoMSIA) module implemented in the Sybyl Tripos version 8.1 package (Klebe and Abraham 1999) using a Silicon Graphics II workstation running IRIX 6.5 operating system. Aligned compound structures were assigned Mulliken partial charges and imported into Sybyl along with binding affinity data for each target in log units. A grid box with a resolution of 2.0 Å was generated for the aligned compounds, extending an additional 4.0 Å beyond the alignment. Each grid point in the lattice was assigned a numeric value representing a 3D property descriptor using an sp3 carbon atom probe with a +1 net charge. CoMSIA 3D descriptor fields for the five physicochemical properties (steric, electrostatic, hydrophobic, HBD, and HBA) were generated using PLR methods with default parameters as implemented in Sybyl. The cross-validated PLR model was evaluated for a 1-10 range of components using the bootstrap method, standard scaling and 2.0 kcal/mol column filtering. The non-validated final PLR model was generated using the optimal 5 components, standard scaling and 2.0 kcal/mol column filtering. Test set compound biological activity was predicted using the final PLR model.

4.3. Results 4.3.1. Workflow of QSAR models for affinity and selectivity In order to elucidate the dominant interactions of binding and target selectivity in the context of global and local structural features of the indolinones (Table 4.1), a two- method workflow was adapted (Figure 4.1). It consists of a classical QSAR which correlates diverse 2D molecular descriptors to binding affinity (ThT displacement AC50) and well as a 3D-QSAR that generates volume-based contours representing favorable and unfavorable physiochemical properties of a compound. The classical QSAR molecular descriptors were calculated using four different types of software packages including semi-empirical, ab-initio and analytical implementations such that accurate yet computationally tractable values describing the physiochemical features of the dataset could be obtained (Cisek and Kuret 2012). E-DRAGON was used to calculate 278 descriptors: 48 constitutional descriptors, 33 connectivity indices, 154 functional group counts, 14 charge descriptors, and 29 molecular properties (Mauri, Consonni et al. 2006). This semi-empirical approach was augmented with more accurate calculations of clogP and tPSA from Molinspiration. Quantum effects governing polarizability (α) and dipole moment (μ) were evaluated using ab initio methods implemented in G09, and compound surface area and molecular volume were analytically computed using ASV software (Petitjean 1994; Frisch, Trucks et al. 2009). Two types of regression were used to analyze the descriptor set such that a small number of descriptors contributing to target affinity could be determined. First, the descriptor set was processed using PLR in order to reduce the number of descriptors. Second, the top descriptors identified through PLR screening were fit to an MLR equation. This approach enabled a direct relationship between the descriptors (Table 4.2) and binding affinity that could be compared for selectivity among tau and Aß binding targets. The classical QSAR model was augmented with 3D-QSAR methods implemented in Sybyl, such that the global molecular trends from 2D descriptors could be interpreted in the context of specific structural regions for compound optimization (Klebe and Abraham 1999). First, a flexible compound alignment was performed on the chemical structures using the rigid indolinone core as a template providing a frame of reference (Figure 4.2).

Because no structural data for compound binding modes for the molecular targets was available, an exhaustive multiple compound alignment was performed using pharmACOphore, a program capable of flexibly aligning large datasets (Korb, Monecke et al. 2010). Next, a spatial grid encompassing the aligned molecules was generated for the CoMSIA steric, electrostatic, hydrophobic, hydrogen-bond donor (HBD) and hydrogen-bond acceptor (HBA) molecular fields. Each point in the grid was assigned a numerical value and correlated to target binding affinity using PLR. Subsequently, contour maps were created for each molecular field indicating favorable and unfavorable structural features for compound affinity. Finally, each contour map was overlapped with its counterpart for the other molecular target such that tau and Aß selectivity-driving 3D regions for the five physiochemical properties could be determined. 4.3.2. Classical PLR model for identification of most significant binding affinity descriptors Validated PLR calibration models were used to identify the most important 2D molecular descriptors for binding affinity for tau and Aß. For tau, the training set consisted of 35 compounds, whereas the Aß training set consisted of 40 compounds, both datasets spanning three orders of magnitude in affinity. The resulting PLR calibration models both consisted of 13 linear combinations (t variables); the tau model contained 27 molecular descriptors (x variables), whereas the Aß model contained 22 x variables (Table 4.3). The models were stable and predictive on the basis of correlation coefficients (Y correlation values of 0.98 and 0.97 for tau and Aß respectively, and X 2 correlation of 0.99 for both models) and leave-one-out cross-validation (Q loo 0.78 and 0.83 for tau and Aß, respectively; Table 4.3) (Konovalov, Llewellyn et al. 2008). The molecular descriptors (x variables) in the calibration models were assigned an importance factor by the PLR algorithm as they contributed to the independent variable

(AC50), such that the descriptors could be ranked based on significance. The top-ranked tau PLR descriptors represented constitutional and quantum logical blocks and included reciprocal distance connectivity index (RDCHI), polarizability (α), sum of conventional bond orders without hydrogen (SCBO), and fourth-order solvation connectivity terms (4χsol). RDCHI term increases with molecular size but decreases with molecular branching, indicating the level of connectivity and complexity of a compound (Consonni 71

and Todeschini 2000). The positive coefficient of this descriptor indicates that larger compounds with decreased molecular branching have an advantage in binding affinity over smaller aliphatic chain- containing ligands. Polarizability (α) emerged as the second most important descriptor for the tau model, confirming that quantum electronic effects are important for tau-binding compounds. The positive coefficient of this property suggests that an increase in a compound’s electron density capable of undergoing a perturbation when acted on by external forces (Marder, Sohn et al. 1991) increases tau potency. SCBO is the degree of unsaturation of the structure, indicating the patterns of bond orders in the molecule (Consonni and Todeschini 2000). The negative sign of this parameter indicates that higher order bonds, such as triple bonds, are detrimental to compound potency. Lastly, the fourth-ranked molecular descriptor for the tau PLR model was 4χsol, addressing the solvation, bonding and branching nature of the scaffold’s molecular topology (Kier and Hall 1986; Hall and Kier 2001). Although based on graph theory and intended for a black-box QSAR approach, 4χsol had a negative coefficient, indicating that hydrophilicity is detrimental to binding. The top-ranked Aß PLR descriptors represented constitutional and analytical logic blocks and included sum of Kier-Hall electrotopological states (Ss), SCBO, analytical volume measurement (Vol), and 0th-order connectivity valence (0χv). Ss indicates the ratio of π and lone pairs to σ electrons (Consonni and Todeschini 2000) and correlates with polarizability (R2 = 0.84). The positive coefficient of this term indicates that much like polarizability for tau, an increase in a π-network of a ligand increases Aß affinity. SCBO was a common descriptor for tau and Aß, and also had a negative sign in the Aß PLR model indicating that high order bonds are detrimental to binding affinity for both molecular targets. Another top-ranked descriptor was the molecular volume of the ligand; a positive coefficient for this parameter indicated that an increase in molecular volume, and therefore, molecular size was beneficial for Aß-binding ligands. Much like the tau model, the fourth-ranked descriptor was a valence connectivity term with a negative coefficient; however, unlike the tau descriptor, the 0th-order connectivity term suggested a simple topology and branching pattern (Kier and Hall 1986; Hall and Kier 2001) was detrimental for Aß affinity. Because 0χv was highly correlated to tau’s top-

ranking descriptors RDCHI and α (R2 = 0.96 and 0.90, respectively), this suggests that for both molecular targets, a common pattern for potent ligands includes a larger, complex topology with delocalizable π-electrons. Although tau and Aß PLR models had just one descriptor in common out of four highest-weighted descriptors, other top-ranked descriptors for each molecular target were highly correlated and had similar physical interpretations (Table 4.4). This finding is indicative of the similarities between the tau and Aß molecular targets, where the available binding sites for compounds are very similar between the two, with the differences reflecting distinct binding sites created by the unique primary structures of these aggregates. 4.3.3. Classical MLR model for identification of binding affinity and selectivity descriptors The top-ranked 2D molecular descriptors from PLR analysis (Table 4.4) were used in a subsequent MLR analysis to create a model with maximum transparency. The training datasets used for PLR analysis for tau and Aß were also used for MLR analysis. Although PLR calibration models for tau and Aß were similar in terms of their top- ranked descriptors, the same descriptors could not be used to generate a statistically 2 2 significant model for both molecular targets (as judged by R and Q boot). Therefore, the optimal MLR model for tau was generated using α, 4χsol and SCBO, whereas the model for Aß was generated using surface area (SA), 0χv, and SCBO. Although SA was not among the top-ranked PLR descriptors for Aß, it was the only descriptor generating a statistically significant MLR model for Aß. SA is the analytical calculation of the compounds’ surface area and is incidentally highly correlated to α and 4χsol, the descriptors in tau’s MLR model. In fact, all of the MLR descriptors share significant intercorrelation as well as fairly high variance inflation factors (Table 4.5), indicating that the benzothiazole-aryl dataset is a tightly structurally-focused library. However, the training set to descriptor ratio was >8:1, which is optimal for this type of analysis (Barrio, Satyamurthy et al. 2009). Taken together, these results indicate that these MLR descriptors were appropriate for generating focused and transparent QSAR models for “global” affinity characteristics of tau and Aß aggregates. However, due to different

descriptors being used to build the MLR models, direct comparisons between tau and Aß parameters for selectivity were not possible. Final MLR calibration yielded two equations for the prediction of AC50:

4 sol pAC50 2N4R tau = 0.3(±0.03)α - 1.72(±0.33) χ - 0.11(±0.06)SCBO - 2.39(±0.72) (4.1) 0 v pAC50 Aß40 = 0.09(±0.01)SA - 1.3(±0.26) χ + 0.41(±0.05)SCBO - 5.35(±0.53) (4.2)

The resulting statistics from the models (Table 4.3) indicate that the MLR equations are adequately calibrated (R2 = 0.71 for both training sets), however, the tau model is less 2 stable and consistent than the Aß (R RMSEloo = 0.58 and 0.47 for tau and Aß, respectively). However, the cross-validated correlation indicates that both models may have some predictive utility for assessing compound affinity for the respective molecular 2 targets (Q boot = 0.65 and 0.60 for tau and Aß, respectively). Additionally, statistical robustness was tested using two types of Y-randomization. To test the probability that the MLR equations occurred by chance due to the distribution of the AC50 data, these data were independently shuffled 250 times and correlated back to the original three 2 MLR descriptors. The corresponding Q boot values were poor, ranging from 0 to 0.51 for tau and 0 to 0.44 for Aß, with only two models each generating values > 0.50 and > 0.40, respectively, validating that original MLR equations were generated using an unbiased protocol. Next, to test the probability that the three MLR descriptors occurred by chance, MLR equations were rebuilt 250 times using all 60 descriptors from PLR analysis with 2 independently shuffled response data. The new MLR models were very poor, with Q boot values ranging from 0 to 0.41 for tau and 0 to 0.42 for Aß, with only two models each 2 generating Q boot > 0.40, validating an unbiased selection of the descriptors (x variables) for the optimal MLR models. 4.3.4. 3D-QSAR model for affinity and selectivity in the context of 2D descriptors 3D-QSAR models can predict which R-group moieties are beneficial and detrimental to binding affinity and selectivity for a given structural dataset. They have utility in interpreting the “global” molecular trends from classical QSAR models in terms of specific regions on the molecules, by assigning volumetric physiochemical attributes to those regions. To identify the 3D steric, electrostatic, hydrophobic, hydrogen-bond donor

(HBD) and hydrogen-bond acceptor (HBA) molecular features (CoMSIA fields) of the benzothiazole scaffold, a CoMSIA PLR analysis was performed for tau and Aß. The training set compounds were different from the dataset used in PLR and MLR calibration, to balance the number of active and inactive compounds in the alignment, such that the volume-based PLR analysis was not biased toward inactive compounds. The best CoMSIA PLR models consisted of 5 components (linear combinations of descriptors) for both molecular targets, which were sufficiently trained (R2 = 0.91 for both tau and Aß 2 training sets), stable and consistent (R RMSEloo = 0.37 and 0.27 for tau and Aß, 2 respectively), as well as predictive (Q boot = 0.97 and 0.95 for tau and Aß, respectively; Table 4.3). The CoMSIA contour maps (overlays of the tau and Aß graphical results) for the steric, electrostatic, hydrophobic, HBD and HBA were superimposed onto a benzothiazole-aryl template for visualization of the 3D regions (Figure 4.3). Each contour map was color-coded according to molecular target and type of property, with green and red contours indicating favorable and unfavorable regions for tau-binding ligands and magenta and blue indicating the favorable and unfavorable regions for Aß- binding ligands. The steric contour maps (Figure 4.3A) indicate that both tau and Aß prefer long planar molecules, with constituents in the para position of the aryl ring. Moreover, both molecular targets have unfavorable steric regions; while the tau model disfavors substituents in the ortho position of the aryl ring, the Aß model disfavors substituents in the meta position. This suggests that tau selectivity may be created by allowing substituents in the meta, but not the ortho positions of the aryl ring. The electrostatic contour maps (Figure 4.3B) indicate that tau is permissive to polar moieties in the meta position, whereas Aß prefers ligands with polar substituents in the para position. Therefore, tau selectivity may be achieved by adding polar substituents in the meta position of the aryl ring. The hydrophobic contour maps (Figure 4.3C) indicate that both molecular targets, although Aß to a much greater extent, favor hydrophobic substituents in the aryl ring, except for substituents directly in the para position. Therefore, tau selectivity may be achieved by lowering the overall hydrophobicity of the compound. Based on the HBD and HBA contour maps (Figure 4.3D and 4.3E) the aryl

ring favors hydrogen bonding interactions of the para substituents on the aryl ring for both tau and Aß, with no discernible selectivity differences. In summary, the contour maps suggest structural basis for affinity, where most affinity modifying features for the benzothiazole-aryl library in this study are the R-group moieties off the aryl ring of the scaffold, where longer, planar ligands are predicted to generate more binding affinity. Additionally, selectivity for tau may be enhanced with polar substituents of the aryl moiety in the meta position and an overall less hydrophobic character of the ligand. 4.3.5. QSAR predictivity assessment using external validation To assess the predictive capability of the classical PLR, MLR and the 3D-QSAR models, external validation sets were used to evaluate the models on the basis of statistical criteria proposed by Golbraikh and Tropsha (Golbraikh and Tropsha 2002). The validation datasets consisted of six compounds each that were not part of the training datasets used to calibrate the models. Instead, the external validation datasets were independently chosen prior to model calibration such that they spanned three orders of magnitude in affinity for each molecular target. The classical PLR and MLR models were validated using the same external test sets for tau and Aß, respectively, whereas the 3D- QSAR was evaluated using an external test set balanced for the number of active and inactive compounds. The results of the validation meet the statistical criteria and confirm the models are predictive on the basis of external dataset correlation coefficients (all R2 ≥ 0.78) as well as additional criteria for goodness of fit and slope (Table 4.6). Therefore, the models can be used for predictive assessment of compound potency over a low nanomolar to low micromolar concentration range. 4.4. Discussion The QSAR workflow developed in this study includes two different types of analysis, classical (PLR and MLR) as well as 3D-QSAR, and all analyses converged on similar potency-generating trends defined by “global” and “local” benzothiazole-aryl features. In the context of the benzothiazole scaffold, tau affinity and selectivity are driven by compound polarizability and molecular electronegativity (4χsol solvation connectivity term), as well as an increase in the size of the ligand while preserving the topological

complexity of the molecule. Planar, heterocyclic as well as aromatic ligands favor substituents that increase the size of the molecule along its length, as well as meta substituents in the aryl ring that increase the ligand’s width. Moreover, the significance of the molecular electronegativity parameter is supported by the hydrophobicity contour map, where tau selectivity is predicted to favor ligands with less overall hydrophobic character. On the other hand, Aß-selective ligands tend to be large and hydrophobic, with the surface area extending along the length, rather than the width of the compound. On the basis of the QSAR analyses, both molecular targets, favor complex topology of aromatic heterocycles, disfavor higher bond orders (e.g., triple bonds). Overall, the combined classical and 3D-QSAR workflow suggests clear design guidelines for the optimization of the benzothiazole-aryl library for tau affinity and selectivity. Optimization of the benzothiazole-aryl scaffold for tau-selectivity will involve balancing between strong polarizability and optimal hydrophobicity (ideally 3 < clogP < 4). The benzothiazole-aryl scaffold has a classic donor-π-acceptor architecture, where the benzothiazole ring acts as an electron acceptor that is connected to a strong electron donor via a π-bridge (see Chapter 2) (Liu, Liu et al. 2000). Compound planarity, balanced aromaticity of the π-network, as well as the strength of the donor and acceptor moieties contributes to polarizability (Gompper and Wagner 1988). For example, the most potent tau compounds in this study, 12, 29 and 34, possess very strong mono- and dimethylamine donor moieties, and have been previously characterized for their Aß- binding characteristics (Mathis, Wang et al. 2003; Wang, Mathis et al. 2003; Kitts and Bout 2009; Kitts and Vanden Bout 2010). However, it may be possible to increase tau potency and selectivity of these compounds by adding polar substituents in the meta position of the aryl ring. Another potent and very selective compound in this library, 23, is unique in that it’s a long and wide molecule; however, its affinity for tau may be maximized by coupling a stronger donor, such as a methylamine, to the aryl moiety. These compound design and optimization guidelines can be further integrated with existing data on the nature of the binding interaction between cross-ß-sheet structures and small molecules.

Benzothiazole-aryls remain the most researched of scaffolds for whole brain imaging, and due to lack of experimental structures of these small molecules binding cross-ß-sheet aggregates, several computational MD studies have tried to rationalize this multi-modal interaction (see Chapter 2). Two key studies predict van der Waals and dispersion driven binding rather than electrostatic interactions for neutral as well as charged benzothiazoles (Wu, Wang et al. 2008; Rodriguez-Rodriguez, Rimola et al. 2010). The in silico models predict that the compounds will preferentially bind either the exposed cross-ß-sheet backbone of Aß created by Glycine residues or hydrophobic surfaces created by aromatic residues, such as Phenylalanine (Cisek and Kuret 2012). Much like Aß, tau filament protomers adopt a cross-ß-sheet structure; however, on the basis of the primary structure, tau fibrils are predicted to present unique double wide Glycine channel binding sites (see Chapter 1) that rationalize selectivity in the context of a wide, highly polarizable ligand. These models support the QSAR results in that the molecular features identified in the tau models suggest that tau-binding ligands potently bind solvent-exposed shallow binding sites or clefts (as opposed to binding “pockets”), where the ligands share dispersion interactions (α descriptor) with the molecular target, as well as the solvent (4χsol descriptor). Although the QSAR analysis in this study is able to address the questions of ligand affinity and selectivity of the compound dataset under investigation, the limitation of the approach is lack of information about binding site density (Bmax) in the context of the ligand’s binding mode. Strong polarizability supports various types of dispersion interactions, and moreover, the degree or amount of a compound’s polarizability may modulate its preference for a specific binding mode, giving rise to various Bmax. For example, the strongly polarizable compound ((E)-2-[[4-(dimethylamino)phenyl]azo]-6- methoxybenzothiazole) (compound 1 Chapter 3) may potently bind two binding modes, the exposed Glycine channel site, as well as aromatic residues, contributing to a high

Bmax binding constant (see Chapter 3) (Matsumura, Ono et al. 2012). On the other hand, ((E)-2-[2-[4-(dimethylamino)phenyl] ethenyl]-6-methoxybenzothiazole) (compound 2 Chapter 3), as well as THK523, although both share a planar π-rich aromatic network and strong amine donors, they lack strong acceptor moieties for maximum polarizability

(Fodero-Tavoletti, Okamura et al. 2011). Therefore, these ligands may be more suitable for lower Bmax binding sites via cation-π or π-stacking interactions with aromatic side chains. Taken together, this suggests that polarizability may modulate not only binding affinity but Bmax as well. In summary, this QSAR study identifies molecular properties and features of potent and selective cross-ß-sheet binding benzothiazole-aryl compounds for two types of aggregates. These results can be leveraged in the process of rational compound design of potent compounds against tau aggregates.

4.5 Tables

AC50 (nM) a a b b c # R1 R2 Tau Aß Sl

1 H 51 ± 7 17 ± 4 0

2 H 3600 ± 570 >10000 3

3 H 130 ± 34 450 ± 120 3

4 6-OMe 310 ± 28 870 ± 330 3

5 5-OMe >10000 >10000 1

6 6-OMe 52 ± 5 1200 ± 440 23

7 6-OMe 30 ± 3 52 ± 10 2

8 5-OMe 350 ± 36 >10000 29

9 6-OMe 250 ± 21 670 ± 290 3

a R2 group abbreviations: methyl (Me) b AC50 is the concentration of compound needed to decrease ThT probe fluorescence by 50%. Inactive compounds, defined as AC50  100,000 nM, were arbitrarily assigned an AC50 of 100,000 nM for QSAR training and testing. c Selectivity Index (SI) is the ratio of tau:A40 AC50 values.

Table 4.1. Compound structures and characteristics

Continued 80

Table 4.1. Continued

AC50 (nM)

# R1 R2 Tau Aß Sl

10 H >10000 >10000 1

11 5-OMe 190 ± 16 >10000 53

12 H 8.3 ± 1.2 1100 ± 300 133

13 H 690 ± 200 1500 ± 570 2

14 H 3000 ± 430 >10000 3

15 5-OMe >10000 >10000 1

16 5-OMe 2000 ± 840 >10000 5

17 H 1900 ± 470 >10000 5

18 6-OMe 140 ± 19 780 ±190 6

19 6-OMe 120 ± 12 110 ± 11 1

20 5-OMe 260 ± 26 >10000 38

21 H 4700 ± 2100 >10000 2

22 6-OMe >10000 >10000 1

Continued

Table 4.1. Continued

AC50 (nM)

# R1 R2 Tau Aß Sl

23 H 16 ± 2 1869 ± 340 117

24 H 320 ± 35 317 ± 79 1

25 6-OMe 300 ± 60 340 ± 160 1

26 H 340 ± 26 >10000 29

27 H >10000 >10000 1

28 5-OMe 75 ± 8 240 ± 150 3

29 6-OMe 2.7 ± 0.2 200 ± 64 74

30 H 770 ± 150 >10000 13

31 5-OMe >10000 >10000 1

32 H >10000 >10000 1

33 H 97 ± 7 290 ± 210 3

34 H 2.6 ± 0.3 260 ± 61 100

35 H >10000 >10000 1

36 H >10000 >10000 1

Continued

Table 4.1. Continued

AC50 (nM)

# R1 R2 Tau Aß Sl

37 5-OMe >10000 >10000 1

38 5-OMe 1200 ± 400 >10000 8

39 6-OMe 27 ± 2 370 ± 120 14

40 5-OMe >10000 >10000 1

41 5-OMe 69 ± 5 740 ± 130 11

42 H >10000 >10000 1

43 H ND 1529 ± 340 ND

44 H ND 31 ± 6 ND

45 H ND 272 ± 57 ND

46 H ND 1303 ± 340 ND

47 H ND 39 ± 6 ND

48 H ND 374 ± 68 ND

49 H ND 374 ± 91 ND

50 H ND 431 ± 91 ND

Continued 83

Table 4.1. Continued

AC50 (nM)

# R1 R2 Tau Aß Sl

51 H ND 238 ± 57 ND

52 H ND 238 ± 45 ND

53 H ND 36 ± 11 ND

54 H ND 79 ± 11 ND

55 H ND 19 ± 6 ND

4 sol b 4 sol b # αa χ SCBOc # αa χ SCBOc 1 70.4 9.47 40.0 29 55.9 6.87 30.0 2 49.7 6.62 30.0 30 44.9 6.13 29.0 3 57.4 7.76 32.0 31 54.6 8.17 36.5 4 47.5 6.51 28.0 32 44.1 6.43 33.0 5 48.1 6.90 35.0 33 47.2 6.35 29.0 6 46.7 6.48 28.0 34 51.9 6.35 28.0 7 74.3 9.99 42.0 35 57.4 8.20 37.0 8 49.1 6.82 31.0 36 50.7 7.70 34.5 9 43.6 6.19 28.0 37 43.1 6.95 27.5 10 43.7 6.42 27.0 38 51.6 7.16 32.0 11 59.7 8.32 38.0 39 61.4 8.28 34.0 12 48.6 6.13 27.0 40 43.2 6.64 25.5 13 58.0 7.76 32.0 41 52.9 6.59 30.0 14 45.1 6.35 29.0 42 39.2 6.49 25.5 15 61.5 8.66 39.0 43 39.4 5.67 25.0 16 60.7 8.49 39.0 44 41.3 5.82 26.0 17 55.9 7.85 36.0 45 39.2 5.67 26.0 18 61.9 8.28 34.0 46 46.6 6.35 30.0 19 48.6 6.65 29.0 47 45.1 5.82 26.0 20 49.1 6.59 31.0 48 45.1 5.82 26.0 21 56.8 8.02 37.0 49 38.2 5.67 25.0 22 42.8 6.19 27.0 50 39.3 5.67 26.0 23 58.6 7.96 37.0 51 38.0 5.67 25.0 24 44.3 6.13 27.0 52 35.7 5.49 23.5 25 43.9 6.19 28.0 53 47.0 7.41 31.0 26 42.9 5.99 26.0 54 44.8 6.42 27.0 27 38.1 5.67 25.0 55 45.0 6.18 27.0 28 51.3 6.81 31.0 a Calculated using quantum approach as described 4.2.2., measured in units Å³. b, c Calculated using semi-empirical approaches as described 4.2.2., unitless properties.

Table 4.2. Compound MLR descriptors

PLR Tau model Aß model

N 35 40 t variables 13 13 x variables 27 22 Y correlation 0.98 0.97 X correlation 0.99 0.99

RMSEloo 0.47 0.41 2 0.78 0.83 Q loo

MLR N 35 40 x variables 3 3 2 Q boot 0.65 0.60 2 Q boot RMSE 0.64 0.89 R2 0.71 0.71 2 R RMSEloo 0.58 0.47 F-statistic 25.10 33.90 p-value 1.97E-08 1.40E-10

3D-QSAR N 29 32 components 5 5 intercept 0 0.024 2 Q boot 0.97 0.95 2 Q boot RMSE 0.21 0.18 R2 0.91 0.91 2 R RMSEloo 0.37 0.27 F-statistic 45.5 52

Table 4.3. Training set statistics for Tau and Aß

2 2 2 2 2 2 2 2 2 2 2 PLR R R0 R'0 (R -R0 )/R (R -R'0 )/R |R0 -R'0 | k k' n 2N4R tau 0.927 0.904 0.865 0.02 0.07 0.04 0.94 1.03 6 Aβ40 0.980 0.930 0.949 0.05 0.03 0.02 0.97 1.03 6

MLR 6 2N4R tau 0.988 0.988 0.988 0.000 0.000 0.00 0.997 1.000 6 Aβ40 0.837 0.811 0.837 0.031 0.000 0.026 0.974 1.000

3D-QSAR 2N4R tau 0.888 0.873 0.888 0.017 0.000 0.02 0.87 1.13 6 Aβ40 0.783 0.783 0.722 0.000 0.078 0.06 0.90 1.11 6

0.85 < k or Target >0.6 Either <0.1 <0.3 k' < 1.15

Table 4.4. Test set statistics for Tau and Aß

Rank Tau descriptors Logical Block Aß descriptors Logical Block 1 (+) RDCHI Constitutional (+) Ss Topological 2 (+) Polarizability (α) Quantum (-) SCBO Constitutional 3 (-) SCBO Constitutional (+) Vol Analytical 4 (-) 4χsol Connectivity (-) 0χsol Connectivity

Table 4.5. Top-ranked PLR descriptors for tau and Aß40 pAC50 (+, direct correlation; -, inverse correlation)

Correlation Matrix a VIF b

α 4χsol SCBO Area 0χsol α 1 9.39 4χsol 0.945 1 17.53 SCBO 0.883 0.939 1 7.75 Area 0.958 0.962 0.933 1 53.35 0χsol 0.959 0.961 0.922 0.989 1 46.02 a Correlation coefficient (Rij); where 0.5 < Rij < 0.8 signifies weak intercorrelation, and Rij < 0.5 signifies little or no intercorrelation. b 2 Variance inflation factor (VIF) = 1/(1-R j); VIF < 10 signifies weak multicollinearity

Table 4.6. Correlation matrixa

4.6. Figures

Figure 4.1. QSAR modeling workflow, in which chemical structure and affinity data (circles) were integrated with calculated molecular descriptors (pentagons) and screened using PLR methods (diamonds). Top descriptors were then subjected to MLR to create the final QSAR models (flags). For 3D-QSAR (right side), structure alignment (rhombus) and contour generation (triangle) were part of the workflow.

Figure 4.2. Pharmacophore alignment of the benzothiazoles for the 3D-QSAR analysis. Each compound is depicted in stick format and different colors and by heteroatom (oxygen, red; nitrogen, blue; sulfur, yellow; fluorine, neon). The 55 BTA were aligned using the rigid benzothiazole core as a template, while giving full rotational freedom to the flexible side groups. Images were created using UCSF Chimera Alpha Version 1.5 (build 31329) software.

4.3. 3D-QSAR contour maps for BTA compounds generated using Sybyl software superimposed on top of BTA template, where the ligand atoms are carbon (blue), nitrogen (navy), sulfur (yellow). The tau favorable and unfavorable contours are green and red, respectively, whereas Aß are colored magenta and blue, respectively for (A) steric, (B) electrostatic, (C) hydrophobic, (D) HBD and (E) HBA characteristics of the ligand.

CHAPTER 5

QSAR MODELING OF 3-SUBSTITUTED INDOLINONES FOR TAU- DIRECTED IMAGING AGENT DEVELOPMENT

5.1. Introduction Whole brain imaging methods, such as PET, leverage brain-penetrant radiolabeled probes to detect the spatial distribution of biological targets in situ (Baert 2008). In AD, these targets are extracellular deposits composed of Aß peptides and intracellular NFTs composed of the microtubule-associated protein tau (Buee, Bussiere et al. 2000). The filamentous protein deposits form as tau monomers, much like Aß peptides, aggregate into cross ß-sheet structure composed of parallel, in register ß-sheets perpendicularly oriented to the fibril axis (Petkova, Yau et al. 2006). The solvent-exposed surfaces of these cross ß-sheet fibrils contain shallow grooves and clefts formed by the side chains of peptide residues (Petkova, Yau et al. 2006), where many compounds can bind to various binding sites in different ways (binding modes) (Groenning 2009). To utilize the properties of this molecular target in developing an imaging agent for AD diagnosis will require a ligand capable of specifically binding the lesions with adequate binding potential, which is gauged by sufficient compound affinity and binding site density (Bmax) (Nordberg, Rinne et al. 2010). Because no crystal structures or atomic level experimental models exist that could elucidate the nature of the binding interaction between compounds and fibrils, QSAR models are crucial for rational compound design and optimization (Stone and Jonathan 1993; Peltason and Bajorath 2009). Recently, QSAR analysis was used to identify molecular properties driving binding affinity and selectivity of benzothiazole derivatives for cross ß-sheet aggregates (see Chapter 2 and Chapter 4) (Cisek and Kuret 2012). It

was found that compound polarizability, hydrophobicity, dipole moment and torsional freedom contributed to potent binding affinity, and that polarizability and hydrophobicity could be fine-tuned for aggregate selectivity. However, in the Thioflavin displacement assay, the benzothiazole-aryl scaffold displaced only low Bmax binding modes (Jensen, Cisek et al. 2011). Previously, an indolinone compound known as SU-4312 was identified (Sun, Tran et al. 1998) in a Thioflavin dye displacement assay as a potent binder of cross ß-sheet fibrils (Honson, Johnson et al. 2007). The indolinone compound was capable of complete Thioflavin displacement, even at high concentrations of the probe (Jensen, Cisek et al. 2011), suggesting that it could support high binding affinity and binding site density (high Bmax). In addition to favorable binding potential, a representative member of this scaffold has been radiolabeled and found to exhibit favorable pharmacokinetic properties desirable for PET imaging agents (Kim, Jensen et al. 2010), such as brain penetration and washout kinetics (Kniess, Bergmann et al. 2009). To address the questions of binding affinity and selectivity for the indolinone scaffold, we have devised a two method QSAR approach. A classical QSAR approach was used to identify 2D descriptors for overall “global” molecular trends, and a 3D- QSAR was used to identify critical R-groups of highly potent ligands (Vanopdenbosch, Cramer et al. 1985; Stone and Jonathan 1993; Peltason and Bajorath 2009). The experimental input for this analysis included the structures and binding affinities of a small library of 44 indolinone compounds tested in a ThT displacement assay against in vitro 2N4R tau and Aß42 fibrils. The library was ideal for a QSAR study because the indolinone structures varied within a concise series yet their affinities for tau and Aß fibrils varied over three orders of magnitude for each molecular target. Importantly, the affinity of certain indolinones extended into the radiotracer concentrations (low nanomolar), especially for the tau molecular target. The QSAR analyses confirm polarizability, molecular electronegativity and hydrophobicity as affinity-driving features of the indolinone series and provide guidelines for indolinone optimization. New indolinone analogues designed on the basis of QSAR results and tested in the fluorescence assay verify the predictive utility of the models.

5.2. Methods 5.2.1. Materials All reagents are as described in section 4.2.1. Compounds 1-14 were obtained from Avid Pharmaceuticals (Philadelphia, PA), 15-44 came from AAChembio (San Diego, CA), 26 and 27 were purchased from ChemDiv (San Diego, CA). Compound 28 was custom synthesized by Michael Darby (Columbus, OH), and compounds 45-50 were custom synthesized by Dr. Tom Li (Columbus, OH). 5.2.2. Fibrillization Fibrillization was performed as in section 3.2.2. 5.2.3. Displacement assays Displacement assays were performed as in section 3.2.3.

5.2.4. Calculation of AC50 values

AC50 values were calculated as in section 2.2.4. 5.2.5. Computational methods and model construction All computational calculations and model generation were performed as in 4.2.4. 5.3. Results 5.3.1. Workflow of QSAR models for affinity and selectivity To elucidate the dominant interactions of binding and target selectivity in the context of global and local structural features of the indolinones (Table 5.1), a two-method workflow was adapted (Figure 5.1). It consists of a classical QSAR which correlates diverse 2D molecular descriptors to binding affinity (ThT displacement AC50) and well as a 3D-QSAR that generates volume-based contours representing favorable and unfavorable physiochemical properties of a compound. The classical QSAR molecular descriptors were calculated using four different types of software packages including semi-empirical, ab-initio and analytical implementations such that accurate yet computationally tractable values describing the physiochemical features of the dataset could be obtained (Cisek and Kuret 2012). E-DRAGON was used to calculate 278 descriptors: 48 constitutional descriptors, 33 connectivity indices, 154 functional group counts, 14 charge descriptors, and 29 molecular properties (Mauri, Consonni et al. 2006). This semi-empirical approach was augmented with more accurate

calculations of clogP and tPSA from Molinspiration. Quantum effects governing polarizability (α) and dipole moment (μ) were evaluated using ab initio methods implemented in G09, and compound surface area and molecular volume were analytically computed using ASV software (Petitjean 1994; Frisch, Trucks et al. 2009). Two types of regression were used to analyze the descriptor set such that a small number of descriptors contributing to target affinity could be determined. First, the descriptor set was processed using PLR in order to reduce the number of descriptors. Second, the top descriptors identified through PLR screening were fit to an MLR equation. This approach enabled a direct relationship between the descriptors (Table 5.2) and binding affinity that could be compared for selectivity among tau and Aß binding targets. The classical QSAR model was augmented with 3D-QSAR methods implemented in Sybyl, such that the global molecular trends from 2D descriptors could be interpreted in the context of specific structural regions for compound optimization (Klebe and Abraham 1999). First, a flexible compound alignment was performed on the chemical structures using the rigid indolinone core as a template providing a frame of reference (Figure 5.2). Because structural data for compound binding modes for the molecular targets was unavailable, an exhaustive multiple compound alignment was performed using Pharmacophore, a program capable of flexibly aligning large datasets (Korb, Monecke et al. 2010). Next, a spatial grid encompassing the aligned molecules was generated for the CoMSIA steric, electrostatic, hydrophobic, hydrogen-bond donor (HBD) and hydrogen- bond acceptor (HBA) molecular fields. Each point in the grid was assigned a numerical value and correlated to target binding affinity using PLR. Subsequently, contour maps were created for each molecular field indicating favorable and unfavorable structural features for compound affinity. Finally, each contour map was overlapped with its counterpart for the other molecular target such that tau and Aß selectivity-driving 3D regions for the five physiochemical properties could be determined. 5.3.2. Classical PLR model for identification of most significant binding affinity descriptors The most important 2D molecular descriptors for binding affinity were chosen on the basis of validated PLR calibration models. For each molecular target, the training dataset

contained 38 compounds spanning three orders of magnitude. The resulting tau PLR calibration model consisted of 10 linear combinations (t variables) containing 26 molecular descriptors (x variables), whereas the Aß model consisted of 11 t variables containing 24 x variables (Table 5.3). The models were stable and predictive on the basis of correlation coefficients (Y, X correlation values of 0.95 and 0.99, respectively, for 2 both models) and leave-one-out cross-validation (Q loo 0.85 and 0.79 for tau and Aß, respectively; Table 5.3)(Konovalov, Llewellyn et al. 2008). Each x variable in the PLR set of equations was assigned an importance factor by the model as it contributed to the independent variable (AC50) such that the molecular descriptors used in the models could be ranked based on significance. Tau and Aß PLR models had four identical descriptors in common out of six highest-weighted descriptors, albeit with different rank order for the two targets (Table 5.4). Polarizability (α) emerged as the most important descriptor for the tau model, however it ranked in the 6th position for the Aß model. The polarizability term α describes the movement of electron density within a molecule when influenced by an external electric field, dipole or ion (Marder, Sohn et al. 1991). The positive coefficient suggests that increasing polarizability increases binding affinity for tau as well as Aß. Both descriptors ALOGP and MLOGP2 are a measure of compound hydrophobicity (Consonni and Todeschini 2000); ALOGP is an implementation of the Ghose-Crippen classical algorithm (a summation of atomic contributions to molecular hydrophobicity) whereas MLOGP2 is the square term of a parameterized linear hydrophobicity model with correction factors (a summation of weighted atomic contributions to molecular hydrophobicity) (Mannhold, Poda et al. 2009). In the tau PLR model, ALOGP and MLOGP2 are ranked 2nd and 3rd, respectively, whereas in the Aß model they ranked 5th and 2nd, respectively. In both models ALOGP had a positive coefficient, whereas MLOGP2 had a negative coefficient, suggesting that hydrophobicity has a biphasic relationship with binding affinity for both aggregate targets. High-order connectivity terms, average 2χv, 4χsol, 2χv and average 4χv address the solvation, bonding and branching nature of the scaffold’s molecular topology (Kier and Hall 1986; Hall and Kier 2001). These descriptors ranked 4th and 5th in the tau model, however, they were more significant in the Aß model ranked 1st, 3rd, and 4th.

Connectivity indices are based on graph theory and lack straightforward physical interpretation; they are intended for a black-box QSAR approach wherein the parameters of ligand optimization are not known, only the optimal structures. Nevertheless, the results, where the 4χsol terms had negative coefficients, and the other connectivity terms had positive signs, suggest that high hydrophilicity is detrimental to binding, whereas the presence of complex topology, such as heterocycles, is beneficial for affinity for both molecular targets. Only the tau PLR model contained the tPSA descriptor, ranked in the 6th position, with a negative sign of tPSA, indicating that increasing polar surface area is disfavored for indolinone affinity. In summary, the most important molecular descriptors correlated to indolinone- mediated ThT displacement affinity (AC50) are very similar for tau and Aß, mainly differing in the rank order for the two molecular targets. These results emphasize the structural similarities and therefore binding site similarities between the molecular targets, that arise as the ß-sheet core common to these aggregates forms from each respective peptide. 5.3.3. Classical MLR model for identification of binding selectivity descriptors The top-ranked 2D molecular descriptors from PLR analysis (Table 5.4) were used in a subsequent MLR analysis to identify key descriptors indicative of target selectivity. The training datasets used for PLR analysis for tau and Aß were also used for MLR analysis. The optimal MLR models for tau and Aß were generated using α, 0χsol, ALOGP and MLOGP2 molecular descriptors in the equations. Although the 0th order solvation was not among the top-ranked PLR descriptors, it is the most simplistic connectivity term generating the best MLR models for both targets. The 0χsol term describes the total measure of molecular electronegativity and its effect on the solvent as a function of topological structure and is intercorrelated with α; ALOGP and MLOGP2 are also intercorrelated (Table 5.5). However, the variance inflation factors were low for all MLR descriptors, indicating that the variance associated with each descriptor was independent of the others. Moreover, the training set to descriptor ratio was >8:1, which is optimal for this type of analysis (Barrio, Satyamurthy et al. 2009). Taken together, these results indicate that the MLR descriptors were appropriate for generating predictive

and transparent QSAR models for affinity and selectivity characteristics of tau and Aß aggregates. Final MLR calibration yielded two equations for the prediction of AC50:

0 sol pAC50 2N4R tau = 0.13(±0.02)α - 0.47(±0.07) χ + 1.12(±0.21)ALOGP - 0.10(±0.03)MLOGP2 - 4.54(±0.63) (5.1) 0 sol pAC50 Aß42 = 0.10(±0.01)α - 0.35(±0.06) χ + 1.10(±0.20)ALOGP - 0.10(±0.03)MLOGP2 - 5.35(±0.69) (5.2)

The resulting statistics from the models (Table 5.3) indicate that the MLR equations are adequately calibrated (R2 = 0.74 and 0.71 for tau and Aß training sets, respectively), 2 stable and consistent (R RMSEloo = 0.45 and 0.41 for tau and Aß, respectively), as well 2 as predictive (Q boot = 0.63 and 0.60 for tau and Aß, respectively). Additionally, statistical robustness was tested using two types of Y-randomization. To test the probability that the MLR equations occurred by chance due to the distribution of the

AC50 data, these data were independently shuffled 250 times and correlated back to the 2 unchanged four MLR descriptors. The corresponding Q boot values were poor, ranging from 0 to 0.46 for tau and 0 to 0.49 for Aß, validating that original MLR equations were optimal. Next, to test the probability that the four MLR descriptors occurred by chance, MLR equations were rebuilt 250 times using all 76 descriptors from PLR analysis with 2 independently shuffled response data. The new MLR models were very poor, with Q boot values ranging from 0 to 0.42 for both molecular targets, with only one model each 2 generating Q boot > 0.40, validating an unbiased selection of the descriptors (x variables) for the optimal MLR models. The direct relationship of molecular descriptors to binding affinity allowed the two models to be compared to each other and the descriptors evaluated for selectivity between tau and Aß. To test whether the MLR models were predictive for selectivity in addition to affinity, the magnitudes of the coefficients were directly compared using t-test (defined as the ratio of the coefficient to its standard error) and z-test (defined as the difference of coefficients divided by the square root of the sum of squared errors). All t-test values were ≥ 3.2, rejecting the null hypothesis at p ≤ 0.01 and confirming that the coefficients were precise. The coefficient magnitudes were also directly compared using z-test,

however, all values were ≥ 0.07, failing to reject the null hypothesis at p ≤ 0.1 and revealing no statistical significance between the two MLR models with respect to selectivity. These results reflected the narrow range of selectivity among tau and Aß for the indolinone series, suggesting that the MLR descriptors were not sensitive enough to detect trends in selectivity. 5.3.4. 3D-QSAR model for affinity and selectivity in the context of 2D descriptors 3D-QSAR models have the advantage of resolving the physiochemical attributes of R-groups contributing to binding affinity and selectivity in the context of the “global” molecular trends determined from classical PLR and MLR modeling. To identify the 3D steric, electrostatic, hydrophobic, hydrogen-bond donor (HBD) and hydrogen-bond acceptor (HBA) molecular features (CoMSIA fields) of the indolinone scaffold, a CoMSIA PLR analysis was performed for tau and Aß. Although the training set compounds were different from the dataset used in PLR and MLR calibration, it was necessary to balance the number of E- and Z-isomers in the training sets while retaining three orders of magnitude in affinity. The best CoMSIA PLR models consisted of 5 components (linear combinations of descriptors) for both molecular targets, which were sufficiently trained (R2 = 0.91 and 0.86 for tau and Aß training sets, respectively), stable 2 and consistent (R RMSEloo = 0.27 and 0.29 for tau and Aß, respectively), as well as 2 predictive (Q boot = 0.93 and 0.90 for tau and Aß, respectively; Table 5.3). The CoMSIA contour maps (overlays of the tau and Aß graphical results) for the steric, electrostatic, hydrophobic, HBD and HBA were superimposed onto a 3-substituted indolinone template for visualization of the 3D regions (Figure 5.3). Each contour map was color-coded according to molecular target and type of property, with green and red contours indicating favorable and unfavorable regions for tau-binding ligands and magenta and blue indicating the favorable and unfavorable regions for Aß-binding ligands, respectively. The steric contour maps (Figure 5.3A) indicate that both tau and Aß prefer Z- over E-isomers, as the steric region in the E-isomer position is disfavored. Moreover, both molecular targets favor para substituents in the aryl ring, especially Aß, where ortho positioned R-groups are disfavored. The electrostatic contour maps (Figure 5.3B) indicate that tau is permissive to polar moieties in the meta position, but not the

100

para position, whereas the case is reversed for Aß. The hydrophobic contour maps (Figure 5.3C) indicate that both molecular targets, although tau to a greater extent, favor hydrophobic substituents in the aryl position. Both molecular targets, with Aß to a greater extent, disfavor hydrophobic substituents of the E-isomer, as well as disfavor hydrophobic substituents off the indolinone ring. Based on the HBD and HBA contour maps (Figure 5.3D and 5.3E) the indolinone ring does not favor any hydrogen bonding interactions, however, Z-isomers can be enhanced by the presence of hydrogen bond donor and acceptor groups on the meta position of the aryl ring for both tau and Aß. In summary, the contour maps suggest structural basis for affinity, where the Z- isomer is preferred over the E-isomer, as well as selectivity, where binding to tau may be enhanced with polar substituents of the aryl moiety in the meta position. 5.3.5. QSAR predictivity assessment using external validation and novel analogue design To assess the predictive capability of the classical PLR, MLR and the 3Q-QSAR models, external validation sets were used to evaluate the models on the basis of statistical criteria proposed by Golbraikh and Tropsha (Golbraikh and Tropsha 2002). The validation datasets consisted of six compounds each that were not part of the 38 compound datasets used to calibrate the models. Instead, the external validation datasets were independently chosen prior to model calibration such that they spanned three orders of magnitude for each molecular target. The classical PLR and MLR models were validated using the same external test sets for tau and Aß, respectively, whereas the 3D- QSAR was evaluated using an external test set balanced for the number of E- and Z- isomers. The results of the validation meet the statistical criteria and confirm the models are predictive on the basis of external dataset correlation coefficients (all R2 ≥ 0.83) as well as additional criteria for goodness of fit and slope (Table 5.6). Therefore, the models can be used for predictive assessment of compound potency over a low nanomolar to low micromolar concentration range. The last step in QSAR analysis is independent assessment of the model by using the results to guide the design and optimization of new potent analogues. On the basis of the classical QSAR, the novel structures were optimized for higher polarizability and optimal

101

hydrophobicity. On the basis of the 3D-QSAR, the analogues were optimized for the Z- isomer and planarity. Polarizability was fostered by extending the π-network between the dimethylamine donor and indolinone heterocycle acceptor. The novel analogues were synthesized and tested in in vitro fluorescence displacement assays for tau and Aß. In agreement with the QSAR predictions, the novel compounds (Table 5.1) were highly potent against tau, with the best compound having AC50 of 2.6 nM against tau but not as potent against Aß with AC50 of 36 nM. The high potency of the novel compounds against the tau molecular target independently confirmed the predictive capability of the QSAR models. 5.4. Discussion The QSAR approach outlined in this study provides strategies for ligand optimization taking into account the classical “global” features of compounds as well as 3D “local” regions of potent ligands. Both the classical and 3D-QSAR results are in agreement for the two molecular targets under investigation and pinpoint the global and local ligand features that can be further optimized. In the context of the indolinone scaffold, compound polarizability, molecular electronegativity (0χsol solvation connectivity term), and an optimum hydrophobicity emerged as the main drivers of aggregate binding. Polarizability and solvation connectivity are complementary descriptors and are therefore intercorrelated. Whereas polarizability describes the electronic malleability of the compound, 0χsol models the behavior of the solvent (entropy and dispersion) caused by the ligand (Zhao, Boriani et al. 2008). Therefore, the most potent indolinone compounds maintained a balance between strong polarizability and low hydrophilicity. Strong polarizability was achieved by planar compounds with neutral polar moieties contributing to the delocalization of π-electrons. The presence of halogens on the indolinone heterocycle modestly improved affinity by a factor of ~2, trending by electronegativity (F > Cl > Br > I). This was also reflected in the 3D-QSAR result, where the totally planar Z-isomer, as well as certain polar R-group regions emerged as affinity driving features of the indolinone ligands. The balance between polarizability and hydrophilicity was also reflected in the biphasic behavior of hydrophobicity, capturing the optimum hydrophobicity of high affinity ligands (3 < clogP < 5 log units, based on MLR

102

equations) required a quadratic relationship of potency to hydrophobicity. All of these molecular features suggest that the ligands potently bind solvent-exposed shallow binding sites or clefts (as opposed to binding “pockets”), where the ligands share dispersion interactions with the molecular target, as well as the solvent. Moreover, larger and more hydrophobic ligands may favor intramolecular interactions within this binding mode, where the ligand may preferably adapt a folded conformer involving π-stacking of heterocyclic R-groups or packing of aliphatic chains (Lewis, Ito et al. 2010). In the context of the indolinone series, maximum potency was achieved by compounds with a donor-π-acceptor architecture, wherein a strong electron donating moiety, such as a dimethylamine, was attached to the indolinone electron accepting heterocycle through a π-bridge. The most efficacious π-bridge consisted of an aryl ring with low electronegativity (phenyl, pyrrole > furan). The π-network was balanced with respect to aromaticity, forming an equilibrium of resonance; the conversion of aromatic aryl moiety to nonaromatic quinoid structure was balanced by formation of aromatic pyrrole moiety in the indolinone heterocycle. The oscillation of π-electron density within this balanced framework supported strong polarizability (Albert, Marks et al. 1997), as well as ligand planarity, driving Z-isomerization owing to steric clashes. The Z-isomer also could be selectively stabilized by the formation of an intramolecular hydrogen bond. Conversely, compounds that lacked a strong electron donating moiety, or deviated from planarity in E-isomer form, displaced ThT from protein aggregates only weakly. Interestingly, in addition to binding potency, polarizability-driven interactions support binding modes with high binding site density (Bmax), much like ((E)-2-[[4- (dimethylamino)phenyl]azo]-6-methoxybenzothiazole) (compound 1 from Chapter 3). In contrast, the cross-ß-sheet binding ligand THK523, which lacks a strong acceptor needed to drive high polarizability, displays low Bmax binding to both tau and Aβ aggregates. Taken together, these data suggest that the electronic structure of the indolinone scaffold may be ideal for generating high binding affinity at high Bmax sites. In summary, this QSAR study identifies molecular properties and features of substituents that can be leveraged in the process of rational compound design for potent binding against tau aggregates.

103

5.5 Tables

AC50 (nM)

a a b b c # R1/R2 R3 Tau Aß Sl

1 H/H 8.1 ± 0.1 95 ± 4 12

2 H/H 17 ± 1 730 ± 74 43

3 Br/H 32 ± 3 310 ± 14 10

4 H/H 160 ± 26 1100 ± 140 7

5 H/H 550 ± 110 4000 ± 610 7

6 Cl/H 57 ± 8 1100 ± 98 19

7 Br/H 44 ± 3 192 ± 12 4

8 Cl/H 24 ± 3 200 ± 8 8

Table 5.1. Compound structures and characteristics

Continued

104

Table 5.1. Continued

AC50 (nM)

a a b b c # R1/R2 R3 Tau Aß Sl

9 H/H 6.0 ± 0.4 38 ± 3 6

10 H/H 16 ± 1 56 ± 6 4

11 H/H 420 ± 91 370 ± 30 1

12 H/H 5.0 ± 0.4 27 ± 1 5

13 H/H 317 ± 47 2846 ± 840 9

14 H/H 4300 ± 1300 4000 ± 1100 1

15 H/H 323 ± 69 7254 ± 3156 22

16 H/H 37 ± 2 310 ± 22 8

17 H/H 12 ± 2 210 ± 9 18

18 H/H 1433 ± 229 11597 ± 6262 8

Continued

105

Table 5.1. Continued

AC50 (nM)

# R1/R2 R3 Tau Aß Sl

19 H/F 5 ± 1 30 ± 2 6

20 H/F 16 ± 1 71 ± 11 4

21 H/F 4 ± 1 20 ± 1 5

22 H/F 3.0 ± 0.2 37 ± 2 12

23 H/F 151 ± 23 1493 ± 206 10

24 H/F 11 ± 0.4 81 ± 7 7

302 ± 38 530 ± 121 2 25 H/F

26 Cl/H 6 ± 1 61 ± 4 10

27 Cl/H 13 ± 2 117 ± 16 9

28 Cl/H 155 ± 21 266 ± 34 2

29 Cl/H 12 ± 2 51 ± 3 4

Continued

106

Table 5.1. Continued

AC50 (nM)

a a b b c # R1/R2 R3 Tau Aß Sl

30 Cl/H 331 ± 35 386 ± 56 1

31 I/H 4.4 ± 0.4 90 ± 10 20

32 Br/H 170 ± 21 520 ± 49 3

33 NO2/H 380 ± 59 2200 ± 430 6

34 I/H 370 ± 72 480 ± 42 1

35 Br/H 62 ± 11 670 ± 78 11

36 NO2/H 1500 ± 590 >10000 7

37 I/H 1000 ± 120 4900 ± 2400 5

38 I/H 44 ± 2 220 ± 12 5

39 NO2/H 54 ± 10 250 ± 36 5

40 Br/H 7.8 ± 0.6 72 ± 3 9

41 Br/H 13 ± 2 260 ± 28 20

42 H/H 120 ± 23 420 ± 17 4

43 I/H 410 ± 40 570 ± 44 1

Continued

107

Table 5.1. Continued

AC50 (nM)

a a b b c # R1/R2 R3 Tau Aß Sl

44 Br/H 120 ± 15 300 ± 21 3

45 N/H 41 ± 3 410 ± 24 10

46 NO2/H 7.5 ± 0.6 150 ± 41 20

47 NH2SO2/H 630 ± 52 1800 ± 100 3

48 H/H 4.2 ± 0.3 12 ± 0.9 3

49 H/H 2.6 ± 0.1 36 ± 4 14

50 H/H 31 ± 9 440 ± 100 14

108

4 sol b 4 sol b # αa χ ALOGPc MLOGP2d # αa χ ALOGPc MLOGP2d 1 60.1 14.1 2.89 10.3 26 68.7 17.2 4.29 18.1 2 56.2 13.2 2.54 8.81 27 70.3 20.0 4.95 18.3 3 48.7 13.2 3.43 9.69 28 51.7 16.3 3.83 10.2 4 43.3 11.0 2.12 2.61 29 53.2 14.9 3.69 11.4 5 43.2 11.8 2.32 3.53 30 54.4 16.2 3.36 10.1 6 46.0 13.2 2.99 5.78 31 67.1 16.5 3.47 15.6 7 44.0 12.8 2.87 5.19 32 51.0 14.4 3.21 11.3 8 46.3 12.3 2.78 4.61 33 53.5 15.0 2.36 7.38 9 65.5 15.8 3.62 14.1 34 53.7 14.9 3.04 12.1 10 68.7 18.6 4.28 14.5 35 60.0 17.2 2.99 10.6 11 52.9 14.9 3.16 8.72 36 62.6 17.8 2.13 7.23 12 65.6 14.9 3.35 10.8 37 51.4 14.9 2.43 8.11 13 51.0 14.1 2.45 5.90 38 72.4 17.9 4.17 19.4 14 43.5 13.0 3.60 17.6 39 72.5 18.1 2.48 6.39 15 49.3 15.3 3.79 14.0 40 68.9 17.5 3.33 9.54 16 53.9 13.5 3.03 8.22 41 59.8 16.5 3.93 17.4 17 52.7 13.5 3.03 8.22 42 52.4 14.0 3.31 16.4 18 51.6 14.8 2.70 7.14 43 62.9 18.8 3.15 12.9 19 68.7 15.8 3.83 17.1 44 45.2 12.8 2.88 5.19 20 68.1 18.6 4.49 17.5 45 59.4 14.1 2.28 6.66 21 65.7 14.9 3.56 13.5 46 67.7 16.6 2.79 10.3 22 60.2 14.1 3.10 13.0 47 68.8 17.7 1.72 3.15 23 49.1 15.3 4.00 15.0 48 70.2 16.0 3.48 7.83 24 50.3 13.5 3.23 10.6 49 77.0 16.7 3.80 15.9 25 51.4 14.8 2.90 9.36 50 56.7 14.2 3.55 11.7 a Calculated using quantum approach as described 2.2.2., measured in units Å³. b, c,d Calculated using semi-empirical approaches as described 2.2.2., unitless properties.

Table 5.2. Compound MLR descriptors

109

PLR Tau model Aß model N 38 38 t variables 10 11 x variables 26 24 Y correlation 0.95 0.95 X correlation 0.99 0.99

RMSEloo 0.39 0.45 2 0.85 0.79 Q loo

MLR N 38 38 x variables 4 4 intercept 4.54 5.35 2 0.63 0.60 Q boot 2 0.54 0.48 Q boot RMSE R2 0.74 0.71 2 0.45 0.41 R RMSEloo F-statistic 24.04 20.3 p-value 2.21E-09 1.615E-08

3D-QSAR N 38 38 field components 5 5 intercept -0.03 -0.05 2 0.93 0.90 Q boot 2 0.22 0.25 Q boot RMSE R2 0.91 0.86 2 0.27 0.29 R RMSEloo F-statistic 64.88 38.65

Table 5.3. Training set statistics for Tau and Aß

110

2 2 2 2 2 2 2 2 2 2 2 PLR R R0 R'0 (R -R0 )/R (R -R'0 )/R |R0 -R'0 | k k' n

2N4R tau 0.90 0.89 0.86 0.008 0.040 0.029 1.00 0.86 6

A42 0.88 0.88 0.86 0.000 0.025 0.022 0.98 0.86 6

MLR

2N4R tau 0.97 0.97 0.97 0.000 0.000 0.000 0.97 1.00 6

A42 0.96 0.99 0.96 -0.031 0.000 0.030 1.00 1.00 6

3D-QSAR

2N4R tau 0.88 0.83 0.87 0.049 0.006 0.038 0.98 1.00 6

A42 0.94 0.83 0.87 0.108 0.068 0.038 0.98 1.00 6

0.85 < k or Target >0.6 Either <0.1 <0.3 k' < 1.15

Table 5.4. Test set statistics for Tau and Aß

111

Logical Logical Rank Tau descriptors block Aß descriptors block 1 (+) polarizability (α) Quantum (+) 2χv Connectivity 2 (+) ALOGP Molecular (-) MLOGP2 Molecular 3 (-) MLOGP2 Molecular (-) 4χsol Connectivity 4 (+) average 2χv Connectivity (+) average 4χv Connectivity 5 (-) 4χsol Connectivity (+) ALOGP Molecular 6 (-) tPSA Molecular (+) polarizability (α) Quantum

Table 5.5. Top-ranked PLR descriptors for tau and Aß40 pAC50 (+, direct correlation; -, inverse correlation)

112

Correlation Matrix a VIF b

Descriptor α 0χsol ALOGP MLOGP2 α 1 3.49 0χsol 0.83 1 3.74 ALOGP 0.43 0.54 1 3.71 MLOGP2 0.52 0.54 0.84 1 3.81 a Correlation coefficient (Rij); where 0.5 < Rij < 0.8 signifies weak intercorrelation, and Rij < 0.5 signifies little or no intercorrelation. b 2 Variance inflation factor (VIF) = 1/(1-R j); VIF < 10 signifies weak multicollinearity

Table 5.6. Correlation matrixa

113

5.6. Figures

Figure 5.1. QSAR modeling workflow, in which chemical structure and affinity data (circles) were integrated with calculated molecular descriptors (pentagons) and screened using PLR methods (diamond). Top descriptors were then subjected to MLR to create the final QSAR models (flags). For 3D-QSAR (right side), structure alignment (rhombus) and contour generation (triangle) were part of the workflow.

114

Figure 5.2. Pharmacophore alignment of the indolinones for the 3D-QSAR analysis. Each compound is depicted in stick format and different colors and by heteroatom (oxygen, red; nitrogen, blue; sulfur, yellow; fluorine, neon; bromine, russet; iodine, meline). The 44 indolinones were aligned using the rigid indolinone heterocycle as a template, while giving full rotational freedom to the flexible side groups. Images were created using UCSF Chimera Alpha Version 1.5 (build 31329) software.

115

5.3. 3D-QSAR contour maps from Sybyl software superimposed on top a wireframe model of the indolinone scaffold (Z geometric isomer), where the ligand atoms are carbon (blue), nitrogen (navy), sulfur (yellow). The tau favorable and unfavorable contours are green and red, respectively, whereas Aß are colored magenta and blue, respectively for (A) steric, (B) electrostatic, (C) hydrophobic, (D) HBD and (E) HBA characteristics of the ligand.

116

CHAPTER 6

CONCLUSIONS AND FUTURE DIRECTIONS

At the beginning of this year the first amyloid-targeting imaging agent, 18F- Amyvid was approved by the Food and Drug Administration for use in the clinic. 18F- Amyvid and other radiolabeled PET probes (Klunk, Engler et al. 2004; Choi, Golding et al. 2009; Johnson, Jeppsson et al. 2009) that are undergoing clinical trials for premortem detection of AD all target amyloid plaques, which have been shown to poorly correlate with neurodegeneration (Braak and Braak 1991; Terry, Masliah et al. 1991; Masliah, Mallory et al. 1993). These ligands were derived from known cross-ß-sheet-binding dye compounds that were serendipitously discovered to stain tissue, and so their development did not require rational discovery or optimization pipelines based on computational methods. The few QSAR studies performed on various aggregate-binding compound libraries were mostly retrospective, identifying the readily apparent planarity, hydrophobicity and some steric features that described potent ligands; however, there was no interpretation of selectivity or binding site density. In contrast, by developing a hypothesis-driven, two-step classical and 3D-QSAR, I built a comprehensive model with robust predictive power for affinity and NFT-selectivity to serve as guide in the design and optimization process of lead scaffolds. My QSAR models also integrated information from hypothetical binding models and recently-identified trends of highly potent ß-sheet binding ligands (Wu, Biancalana et al. 2009; Kim, Jensen et al. 2010; Rodriguez-Rodriguez, Rimola et al. 2010). In the benzothiazole and indolinone libraries,

117

potent compounds had characteristic electronic configurations supporting electron delocalization and planarity, making the compounds very polarizable (Kim, Jensen et al. 2010). Consequently, these ligands favored dispersion-driven binding interactions, consistent with the hypothetical models of ThT binding (Wu, Biancalana et al. 2009; Rodriguez-Rodriguez, Rimola et al. 2010). In addition, the hypothetical binding models were integrated into the QSAR result to provide insight into the sterical constraints of the ligand as well as isomer preference for binding interactions. They rationalized how tau selectivity could be achieved by “wider” compounds, taking advantage of the Glycine binding channel predicted by Rodriguez-Rodriguez, et al. computational modeling (see Chapter 1). In the case of the indolinone compounds, the QSAR model successfully predicted potency-enhancing molecular features of this scaffold. The purpose of this project was to design and optimize structures for novel analogs based on QSAR modeling. In silico analogs were optimized based on predictions from QSAR models to increase affinity and selectivity for tau filaments. For the indolinone scaffold, affinity for tau filaments was generated by introducing favorable local (e.g. extended π-network) and global (e.g. higher polarizability) compound characteristics predicted from tau QSAR models. However, the initial compound library had low affinity selectivity between tau and Aß that was insufficient to predict true selectivity for tau filaments on the basis of unfavorable local and global compound characteristics from amyloid QSAR. However, QSAR modeling is an iterative process; an initial model is built using a compound dataset and biological activity against a given target, then novel analogues are designed using the parameters identified with this calibration. After synthesis and testing of the novel analogues, this data can be integrated with the initial dataset and the model recalibrated. On the basis of mathematical modeling, it has been found that a tau-selective ligand should have a binding potential that corresponds to a 25- fold binding selectivity over amyloid at the level of Kd or 50-fold binding selectivity at the level of Bmax (Schafer, Kim et al. 2012). Therefore, as each iteration of QSAR calibration, as necessary, produces guidelines for ever improving ligand affinity, the cutoff values for the most promising radiotracer ligands are already available via independent a priori analysis. Once novel analog structures are predicted and ranked

118

using the QSAR, the next step in the scaffold design and optimization pipeline should narrow down the candidates based on imaging agent criteria. The in silico analogs should be further refined to exclude compounds that do not meet the central nervous system (CNS) drugability criteria (e.g. blood-brain barrier (BBB) penetration) (Wager, Chandrasekaran et al. 2010) and then re-ranked for optimal IND drugability (e.g. ADME properties) (Carpenter, Pontecorvo et al. 2009) and synthetic chemistry criteria (e.g. feasibility of radio-labeling) (Ertl and Schuffenhauer 2009). In silico candidates that are predicted to bind with high affinity to tau fibrils only and have the best IND drugability criteria should be chosen for experimental validation. Following fluorescence competition assays performed on synthetic tau and amyloid fibrils, and the experimental affinities should be compared to the QSAR-predicted affinity values in order to validate the QSAR models. The most potent and selective compounds against synthetic tau filaments should be verified using authentic PHFs and those with adequate affinity and selectivity for the authentic target should be chosen for radio-labeling and subsequent ex vivo and in vivo testing. This type of iterative optimization and testing pipeline is ideal for a streamlined and simple imaging agent discovery effort. Although QSAR modeling is a very powerful method that can streamline the design and optimization process for novel scaffolds, there are some notable limitations. First, the compound dataset under investigation must contain enough datapoints such that the tested biological affinity covers at least three orders of magnitude, preferably extending into a highly potent range (1 nM is a typical affinity limit for radiotracers (Fodero- Tavoletti, Okamura et al. 2011)). Similarly, the molecular descriptor set must cover the structural diversity of the compound library, such that potentially significant features are not overlooked in the analysis. For example, while classical QSAR models can identify statistically significant structural features from among hundreds of molecular descriptors, such as charge or hydrophobicity, they frequently cannot capture the individual contributions of substituents. Therefore, integrating a 3D-QSAR ensures that the compound library has the best possible description of the scaffold’s molecular features. However, the molecular descriptors must be simple and transparent; for example, certain molecular connectivity indexes have been difficult to interpret, let alone use as a

119

guideline for ligand optimization, creating a “black box” QSAR model that can predict, but not elucidate, potency-generating molecular features. Also, descriptors should be normalized to facilitate direct comparison between scaffolds, or a frame of reference such that ligands can be ranked according to a specific property, but not simplified to the extent that they lose their physical meaning (see Chapter 2). Current approaches to QSAR modeling have the capacity to encode the structural and biological information of a ligand in order to define the binding affinity and selectivity of ligands for a particular target binding site. Classically, these targets have been globular proteins with stoichiometric ligand-to-target ratios. However, due to the complexity of our binding target, these methods do not capture the variable binding site density created by the arrangement of solvent-exposed cross-ß sheet residues. Lastly, the QSAR model should be seamlessly integrated with the synthesis and experimental validation modules of the optimization pipeline such that it becomes an efficient iterative protocol. In addition to imaging agent discovery efforts, many studies are also focusing on discovering novel inhibitors of aggregation as a possible treatment for AD. Interestingly, many ligands under investigation for this purpose are the same ligands that bind cross-ß- sheet aggregates. Moreover, these ligands also share the highly delocalized planar structure that gives rise to strong polarizability. I have found that polarizability correlates with inhibitory potency for four different inhibitor scaffolds (data not shown), which suggests that high polarizability may have implications for ligand-based fibril disaggregation and fibril formation. As such, ligand polarizability may be a consideration not only for novel imaging agents, but inhibitors as well. This suggests that in addition to a diagnostic agent design, the compound features indentified herein may be utilized for rational optimization of aggregate-targeted drug treatments for AD.

120

References

(1998). "Consensus report of the Working Group on: "Molecular and Biochemical Markers of Alzheimer's Disease". The Ronald and Nancy Reagan Research Institute of the Alzheimer's Association and the National Institute on Aging Working Group." Neurobiol Aging 19(2): 109-116. Ahlrichs, R., M. Bar, et al. (1989). "Electronic-Structure Calculations on Workstation Computers - the Program System Turbomole." Chemical Physics Letters 162(3): 165-169. Albert, I. D. L., T. J. Marks, et al. (1997). "Large molecular hyperpolarizabilities. Quantitative analysis of aromaticity and auxiliary donor-acceptor effects." J Am Chem Soc 119(28): 6575-6582. Albert, M. S., S. T. DeKosky, et al. (2011). "The diagnosis of mild cognitive impairment due to Alzheimer's disease: Recommendations from the National Institute on Aging-Alzheimer's Association workgroups on diagnostic guidelines for Alzheimer's disease." Alzheimers & Dementia 7(3): 270-279. Allinger, N. L. (1977). "Conformational-Analysis .130. Mm2 - Hydrocarbon Force-Field Utilizing V1 and V2 Torsional Terms." J Am Chem Soc 99(25): 8127-8134. Assoc, A. (2012). "2012 Alzheimer's disease facts and figures." Alzheimers & Dementia 8(2): 131-168. Baert, A. L. (2008). Encyclopedia of diagnostic imaging. Berlin ; London, Springer. Barrio, J. R., N. Satyamurthy, et al. (2009). "Dissecting molecular mechanisms in the living brain of dementia patients." Acc Chem Res 42(7): 842-850. Benkova, Z., I. Cernusak, et al. (2006). "Theoretical study of static electric properties of benzothiazole containing push-pull systems as potential candidates for NLO materials." Struct Chem 17(3): 287-300. Berchtold, N. C. and C. W. Cotman (1998). "Evolution in the conceptualization of dementia and Alzheimer's disease: Greco-Roman period to the 1960s." Neurobiol Aging 19(3): 173- 189. Bieschke, J., M. Herbst, et al. (2012). "Small-molecule conversion of toxic oligomers to nontoxic ß-sheet-rich amyloid fibrils." Nat Chem Biol 8(1): 93-101. Bird, J., N. Brough, et al. (2006). "Understanding adsorption phenomena: investigation of the dye-cellulose interaction." J Phys Chem B 110(39): 19557-19561. Boxer, A. L., M. D. Geschwind, et al. (2006). "Patterns of brain atrophy that differentiate corticobasal degeneration syndrome from progressive supranuclear palsy." Arch Neurol 63(1): 81-86. Braak, H. and E. Braak (1991). "Neuropathological Staging of Alzheimer-Related Changes." Acta Neuropathologica 82(4): 239-259. Braak, H. and E. Braak (1995). "Staging of Alzheimer's disease-related neurofibrillary changes." Neurobiol Aging 16(3): 271-278; discussion 278-284. Bramblett, G. T., J. Q. Trojanowski, et al. (1992). "Regions with abundant neurofibrillary pathology in human brain exhibit a selective reduction in levels of binding-competent tau and accumulation of abnormal tau-isoforms (A68 proteins)." Lab Invest 66(2): 212-222. Breneman, C. M. and K. B. Wiberg (1990). "Determining Atom-Centered Monopoles from Molecular Electrostatic Potentials - the Need for High Sampling Density in Formamide Conformational-Analysis." Journal of Computational Chemistry 11(3): 361-373. Buee, L., T. Bussiere, et al. (2000). "Tau protein isoforms, phosphorylation and role in neurodegenerative disorders." Brain Research Reviews 33(1): 95-130. Cai, L., R. B. Innis, et al. (2007). "Radioligand development for PET imaging of b-amyloid (Ab)- -current status." Curr Med Chem 14(1): 19-52.

121

Cairns, N. J., E. H. Bigio, et al. (2007). "Neuropathologic diagnostic and nosologic criteria for frontotemporal lobar degeneration: consensus of the Consortium for Frontotemporal Lobar Degeneration." Acta Neuropathol 114(1): 5-22. Caprathe, B. W., J. L. Gilmore, et al. (1999). Method of imaging amyloid deposits. United States Patent. United States. Carmel, G., B. Leichus, et al. (1994). "Expression, purification, crystallization, and preliminary x- ray analysis of casein kinase-1 from Schizosaccharomyces pombe." J. Biol. Chem. 269(10): 7304-7309. Carmel, G., E. M. Mager, et al. (1996). "The structural basis of monoclonal antibody Alz50's selectivity for Alzheimer's disease pathology." J Biol Chem 271(51): 32789-32795. Carpenter, A. P., M. J. Pontecorvo, et al. (2009). "The use of the exploratory IND in the evaluation and development of 18F-PET radiopharmaceuticals for amyloid imaging in the brain: a review of one company's experience." Quarterly Journal of Nuclear Medicine and Molecular Imaging 53(4): 387-393. Castilho, M. S., R. V. C. Guido, et al. (2007). "Classical and hologram QSAR studies on a series of tacrine derivatives as butyryleholinesterase inhibitors." Letters in Drug Design & Discovery 4(2): 106-113. Chang, E. and J. Kuret (2008). "Detection and quantification of tau aggregation using a membrane filter assay." Anal Biochem 373(2): 330-336. Chen, X. J. (2006). "QSAR and primary docking studies of trans-stilbene (TSB) series of imaging agents for ß-amyloid plaques." J Mol Struc-Theochem 763(1-3): 83-89. Chirita, C. N. and J. Kuret (2004). "Evidence for an intermediate in tau filament formation." Biochemistry 43: 1704-1714. Chirita, C. N., M. Necula, et al. (2003). "Anionic Micelles and Vesicles Induce tau Fibrillization in vitro." J. Biol. Chem. 278: 25644-25650. Choi, S. R., G. Golding, et al. (2009). "Preclinical Properties of F-18-AV-45: A PET Agent for Aß Plaques in the Brain." Journal of Nuclear Medicine 50(11): 1887-1894. Cisek, K. and J. Kuret (2012). "QSAR studies for prediction of cross-ß sheet aggregate binding affinity and selectivity." Bioorg Med Chem 20(4): 1434-1441. Cleveland, D. W., S. Y. Hwo, et al. (1975). "Mechanism of Action of Tau Protein in Microtubule Assembly." Journal of Cell Biology 67(2): A72-A72. Consonni, V. and R. Todeschini (2000). Handbook of molecular descriptors. New York, Wiley- VCH. de Leon, M. J., S. DeSanti, et al. (2004). "MRI and CSF studies in the early diagnosis of Alzheimer's disease." Journal of Internal Medicine 256(3): 205-223. Dickson, D. W. (1999). "Neuropathologic differentiation of progressive supranuclear palsy and corticobasal degeneration." J Neurol 246 Suppl 2: II6-15. Dickson, D. W. (2001). "Neuropathology of Pick's disease." Neurology 56(11 Suppl 4): S16-20. Dickson, D. W., C. Bergeron, et al. (2002). "Office of Rare Diseases neuropathologic criteria for corticobasal degeneration." J Neuropathol Exp Neurol 61(11): 935-946. Dubois, B., H. H. Feldman, et al. (2007). "Research criteria for the diagnosis of Alzheimer's disease: revising the NINCDS-ADRDA criteria." Lancet Neurol 6(8): 734-746. Duyckaerts, C. and J. J. Hauw (1997). "Prevalence, incidence and duration of Braak's stages in the general population: Can we know?" Neurobiology of Aging 18(4): 362-369. Engler, H., A. Forsberg, et al. (2006). "Two-year follow-up of amyloid deposition in patients with Alzheimer's disease." Brain 129(Pt 11): 2856-2866. Ertl, P. and A. Schuffenhauer (2009). "Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions." J Cheminform 1(1): 8.

122

Fandrich, M., M. Schmidt, et al. (2011). "Recent progress in understanding Alzheimer's ß- amyloid structures." Trends in Biochemical Sciences 36(6): 338-345. Feany, M. B. and D. W. Dickson (1996). "Neurodegenerative disorders with extensive tau pathology: a comparative study and review." Ann. Neurol. 40(2): 139-148. Fodero-Tavoletti, M. T., N. Okamura, et al. (2011). "18F-THK523: a novel in vivo tau imaging ligand for Alzheimer's disease." Brain 134(Pt 4): 1089-1100. Forman, M. S., J. Farmer, et al. (2006). "Frontotemporal dementia: clinicopathological correlations." Ann Neurol 59(6): 952-962. Friedhoff, P., A. Schneider, et al. (1998). "Rapid assembly of Alzheimer-like paired helical filaments from microtubule-associated protein tau monitored by fluorescence in solution." Biochemistry 37(28): 10223-10230. Frisch, M. J., G. W. Trucks, et al. (2009). "Gaussian 09, Rivision A.1." Gaussian, Inc.(Wallingford, CT). Geser, F., M. Martinez-Lage, et al. (2009). "Amyotrophic lateral sclerosis, frontotemporal dementia and beyond: the TDP-43 diseases." J Neurol 256(8): 1205-1214. Ghoshal, N., F. Garcia-Sierra, et al. (2002). "Tau conformational changes correspond to impairments of episodic memory in mild cognitive impairment and Alzheimer's disease." Experimental Neurology 177(2): 475-493. Goedert, M. (1995). "Molecular Dissection of the Paired Helical Filament." Journal of Neurochemistry 65: S136-S136. Goedert, M. (2004). "Tau protein and neurodegeneration." Semin Cell Dev Biol 15(1): 45-49. Goedert, M., R. A. Crowther, et al. (1991). "Molecular Characterization of Microtubule- Associated Proteins-Tau and Map2." Trends in Neurosciences 14(5): 193-199. Goedert, M. and R. Jakes (1990). "Expression of separate isoforms of human tau protein: correlation with the tau pattern in brain and effects on tubulin polymerization." EMBO J 9(13): 4225-4230. Goedert, M., M. G. Spillantini, et al. (1991). "Tau proteins and neurofibrillary degeneration." Brain Pathol 1(4): 279-286. Golbraikh, A. and A. Tropsha (2002). "Beware of q2!" J Mol Graph Model 20(4): 269-276. Gompper, R. and H. U. Wagner (1988). "Donor Acceptor-Substituted Cyclic Pi-Electron Systems - Probes for Theories and Building-Blocks for New Materials." Angewandte Chemie- International Edition in English 27(11): 1437-1455. Greenberg, S. G. and P. Davies (1990). "A preparation of Alzheimer paired helical filaments that displays distinct tau proteins by polyacrylamide gel electrophoresis." Proc Natl Acad Sci U S A 87(15): 5827-5831. Groenning, M. (2009). "Binding mode of Thioflavin T and other molecular probes in the context of amyloid fibrils-current status." J Chem Biol. Groenning, M. (2010). "Binding mode of Thioflavin T and other molecular probes in the context of amyloid fibrils-current status." J Chem Biol 3: 1-18. Habib, L. K., M. T. Lee, et al. (2010). "Inhibitors of catalase-amyloid interactions protect cells from ß-amyloid-induced oxidative stress and toxicity." J Biol Chem 285(50): 38933- 38943. Hall, L. H. and L. B. Kier (2001). "Issues in representation of molecular structure the development of molecular connectivity." J Mol Graph Model 20(1): 4-18. Hampel, H., A. Mitchell, et al. (2004). "Core biological marker candidates of Alzheimer's disease - perspectives for diagnosis, prediction of outcome and reflection of biological activity." Journal of Neural Transmission 111(3): 247-272. Hong, M., V. Zhukareva, et al. (1998). "Mutation-specific functional impairments in distinct tau isoforms of hereditary FTDP-17." Science 282(5395): 1914-1917.

123

Honson, N. S., J. R. Jensen, et al. (2009). "Small-molecule mediated neuroprotection in an in situ model of tauopathy." Neurotox. Res. 15: 274-283. Honson, N. S., R. L. Johnson, et al. (2007). "Differentiating Alzheimer disease-associated aggregates with small molecules." Neurobiol Dis 28(3): 251-260. Honson, N. S. and J. Kuret (2008). "Tau aggregation and toxicity in tauopathic neurodegenerative diseases." J. Alzheimers Dis. 14(4): 417-422. Hrobarik, P., B. Horvath, et al. (2007). "The impact of p-electron conjugation on 15N, 13C and 1H NMR chemical shifts in push-pull benzothiazolium salts. Experimental and theoretical study." Magn Reson Chem 45(11): 942-953. Hrobarik, P., I. Sigmundova, et al. (2010). "Molecular Engineering of Benzothiazolium Salts with Large Quadratic Hyperpolarizabilities: Can Auxiliary Electron-Withdrawing Groups Enhance Nonlinear Optical Responses?" J Phys Chem C 114(50): 22289-22302. Hsiao, I. T., C. C. Huang, et al. (2012). "Correlation of early-phase (18)F-florbetapir (AV- 45/Amyvid) PET images to FDG images: preliminary studies." Eur J Nucl Med Mol Imaging 39(4): 613-620. Inbar, P. and J. Yang (2006). "Inhibiting protein-amyloid interactions with small molecules: a surface chemistry approach." Bioorg Med Chem Lett 16(4): 1076-1079. Iqbal, K., A. C. Alonso, et al. (1994). "Mechanism of neurofibrillary degeneration in Alzheimer's disease." Mol Neurobiol 9(1-3): 119-123. Iqbal, K. and I. Grundkeiqbal (1991). "Ubiquitination and Abnormal Phosphorylation of Paired Helical Filaments in Alzheimers-Disease." Molecular Neurobiology 5(2-4): 399-410. Jack, C. R., Jr., V. J. Lowe, et al. (2009). "Serial PIB and MRI in normal, mild cognitive impairment and Alzheimer's disease: implications for sequence of pathological events in Alzheimer's disease." Brain 132(Pt 5): 1355-1365. Jacquemin, D. and E. A. Perpete (2006). "Ab initio calculations of the colour of closed-ring diarylethenes: TD-DFT estimates for molecular switches." Chem Phys Lett 429(1-3): 147-152. Jensen, J. R., K. Cisek, et al. (2011). "Research towards tau imaging." J Alzheimers Dis 26(S3): 147-157. Jensen, J. R., K. Cisek, et al. (2011). "Ligand polarizability contributes to tau fibril binding affinity." Bioorg Med Chem 19(17): 5147-5154. Johnson, A. E., F. Jeppsson, et al. (2009). "AZD2184: a radioligand for sensitive detection of ß- amyloid deposits." Journal of Neurochemistry 108(5): 1177-1186. Josephs, K. A., J. L. Whitwell, et al. (2008). "Voxel-based morphometry in autopsy proven PSP and CBD." Neurobiol Aging 29(2): 280-289. Kaur, J. and M. Q. Zhang (2000). "Molecular modelling and QSAR of reversible acetylcholinesterase inhibitors." Current Medicinal Chemistry 7(3): 273-294. Keerti, A. R., B. A. Kumar, et al. (2005). "QSAR studies - potent benzodiazepine gamma- secretase inhibitors." Bioorganic & Medicinal Chemistry 13(5): 1873-1878. Kelenyi, G. (1967). "On the histochemistry of azo group-free thiazole dyes." J Histochem Cytochem 15(3): 172-180. Kelenyi, G. (1967). "Thioflavin S Fluorescent and Congo Red Anisotropic Stainings in Histologic Demonstration of Amyloid." Acta Neuropathologica 7(4): 336-&. Kemp, S., S. LJ, et al. (2010). Ligands for Aggregated Tau Molecules. U. S. P. Application. 034982 A1. Kemppainen, N. M., S. Aalto, et al. (2007). "PET amyloid ligand [11C]PIB uptake is increased in mild cognitive impairment." Neurology 68(19): 1603-1606. Kier, L. B. and L. H. Hall (1986). Molecular connectivity in structure-activity analysis. Letchworth, Hertfordshire, England, Research Studies Press, Wiley.

124

Kim, M. K., I. H. Choo, et al. (2007). "3D-QSAR of PET agents for imaging b-amyloid in Alzheimer's disease." Bull Kor Chem Soc 28(7): 1231-1234. Kim, S., J. R. Jensen, et al. (2010). "Imaging as a strategy for premortem diagnosis and staging of tauopathies." Curr Alzheimer Res 7(3): 230-234. King, M. E., V. Ahuja, et al. (1999). "Ligand-dependent tau filament formation: implications for Alzheimer's disease progression." Biochemistry 38(45): 14851-14859. Kitts, C. C. and D. A. V. Bout (2009). "Near-Field Scanning Optical Microscopy Measurements of Fluorescent Molecular Probes Binding to Insulin Amyloid Fibrils." Journal of Physical Chemistry B 113(35): 12090-12095. Kitts, C. C. and D. A. Vanden Bout (2010). "A Spectroscopic Study of 2-[4'- (Dimethylamino)phenyl]-benzothiazole Binding to Insulin Amyloid Fibrils." J Fluoresc. Klebe, G. and U. Abraham (1999). "Comparative molecular similarity index analysis (CoMSIA) to study hydrogen-bonding properties and to score combinatorial libraries." J Comput Aided Mol Des 13(1): 1-10. Klunk, W. E., M. L. Debnath, et al. (1994). "Development of small molecule probes for the ß- amyloid protein of Alzheimer's disease." Neurobiol Aging 15(6): 691-698. Klunk, W. E., H. Engler, et al. (2004). "Imaging brain amyloid in Alzheimer's disease with Pittsburgh Compound-B." Annals of Neurology 55(3): 306-319. Knibb, J. A., J. H. Xuereb, et al. (2006). "Clinical and pathological characterization of progressive aphasia." Ann Neurol 59(1): 156-165. Kniess, T., R. Bergmann, et al. (2009). "Synthesis and radiopharmacological investigation of 3- [4'-[(18)F]fluorobenzylidene]indolin-2-one as possible tyrosine kinase inhibitor." Bioorg Med Chem 17(22): 7732-7742. Knopman, D. S., S. T. DeKosky, et al. (2001). "Practice parameter: Diagnosis of dementia (an evidence-based review) - Report of the Quality Standards Subcommittee of the American Academy of Neurology." Neurology 56(9): 1143-1153. Konovalov, D. A., L. E. Llewellyn, et al. (2008). "Robust cross-validation of linear regression QSAR models." J Chem Inf Model 48(10): 2081-2094. Korb, O., P. Monecke, et al. (2010). "pharmACOphore: multiple flexible ligand alignment based on ant colony optimization." J Chem Inf Model 50(9): 1669-1681. Krebs, M. R., E. H. Bromley, et al. (2005). "The binding of thioflavin-T to amyloid fibrils: localisation and implications." J. Struct. Biol. 149(1): 30-37. Ksiezak-Reding, H. and J. S. Wall (2005). "Characterization of paired helical filaments by scanning transmission electron microscopy." Microscopy Research and Technique 67(3- 4): 126-140. Kudo, Y., M. Susuki, et al. (2006). Quinoline derivative as diagnostic probe for disease with tau protein accumulation. U. P. Office. Japan, BF Research Institute, Inc. Kung, M. P., D. M. Skovronsky, et al. (2003). "Detection of amyloid plaques by radioligands for Ab40 and Ab42: potential imaging agents in Alzheimer's patients." J. Mol. Neurosci. 20(1): 15-24. Kurunczi, L., S. Funar-Timofei, et al. (2007). "Application of the MTD-PLS method to Heterocyclic dye-cellulose interactions." Int J Quantum Chem 107(11): 2057-2065. Laruelle, M., M. Slifstein, et al. (2003). "Relationships between radiotracer properties and image quality in molecular imaging of the brain with positron emission tomography." Mol Imaging Biol 5(6): 363-375. Leuma Yona, R., S. Mazeres, et al. (2008). "Thioflavin derivatives as markers for amyloid-b fibrils: Insights into structural features important for high-affinity binding." ChemMedChem 3(1): 63-66.

125

Lewis, D. F., Y. Ito, et al. (2010). "Quantitative structure-activity relationships (QSARs) for inhibitors and substrates of CYP2B enzymes: importance of compound lipophilicity in explanation of potency differences." J Enzyme Inhib Med Chem 25(5): 679-684. Liu, Y., Y. Liu, et al. (2000). "Theoretical investigation of second-order nonlinear optical properties of substituted benzothiazole derivatives." Teor Eksp Khim 36(6): 333-341. Lockhart, A., L. Ye, et al. (2005). "Evidence for the presence of three distinct binding sites for the thioflavin T class of Alzheimer's disease PET imaging agents on ß-amyloid peptide fibrils." J Biol Chem 280(9): 7677-7684. Luhrs, T., C. Ritter, et al. (2005). "3D structure of Alzheimer's amyloid-ß(1-42) fibrils." Proceedings of the National Academy of Sciences of the United States of America 102(48): 17342-17347. Maccioni-Baraona, R., L. Rojo, et al. (2010). Benzimidazole-derived compounds used as markers in the case of neurodegenerative diseases. W. I. P. Organization. Chile, Neuroinnovation Scientific Services, Ltd. Mannhold, M., G. I. Poda, et al. (2009). "Calculation of molecular lipophilicity: state of the art and comparison of methods on more than 96000 compounds." Chem Cent J 3. Marder, S. R., J. E. Sohn, et al. (1991). Materials for nonlinear optics: chemical perspectives. Washington, DC, American Chemical Society. Masliah, E., M. Mallory, et al. (1993). "Quantitative synaptic alterations in the human neocortex during normal aging." Neurology 43(1): 192-197. Mathis, C. A., Y. Wang, et al. (2003). "Synthesis and evaluation of 11C-labeled 6-substituted 2- arylbenzothiazoles as amyloid imaging agents." J Med Chem 46(13): 2740-2754. Matsumura, K., M. Ono, et al. (2012). "18F-Labeled Phenyldiazenyl Benzothiazole for in Vivo Imaging of Neurofibrillary Tangles in Alzheimer's Disease Brains." ACS Med Chem Lett 3(1): 58-62. Mauri, A., V. Consonni, et al. (2006). "Dragon software: An easy approach to molecular descriptor calculations." MATCH-Commun Math Co 56(2): 237-248. Melnikova, I. (2007). "Therapies for Alzheimer's disease." Nat Rev Drug Discov 6(5): 341-342. Metivier-Pignon, H., C. Faur, et al. (2007). "Adsorption of dyes onto activated carbon cloth: Using QSPRs as tools to approach adsorption mechanisms." Chemosphere 66(5): 887- 893. Meyer, E. A., R. K. Castellano, et al. (2003). "Interactions with aromatic rings in chemical and biological recognition." Angew Chem Int Ed Engl 42(11): 1210-1250. Mocanu, M. M., A. Nissen, et al. (2008). "The potential for b-structure in the repeat domain of tau protein determines aggregation, synaptic decay, neuronal loss, and coassembly with endogenous Tau in inducible mouse models of tauopathy." J. Neurosci. 28(3): 737-748. Morsch, R., W. Simon, et al. (1999). "Neurons may live for decades with neurofibrillary tangles." Journal of Neuropathology and Experimental Neurology 58(2): 188-197. Murakami, K. (2002). "Thermodynamic and kinetic aspects of self-association of dyes in aqueous solution." Dyes and Pigments 53(1): 31-43. Necula, M., C. N. Chirita, et al. (2003). "Rapid anionic micelle-mediated a-synuclein fibrillization in vitro." J. Biol. Chem. 278: 46674-46680. Necula, M., C. N. Chirita, et al. (2005). "Cyanine dye n744 inhibits tau fibrillization by blocking filament extension: implications for the treatment of tauopathic neurodegenerative diseases." Biochemistry 44(30): 10227-10237. Necula, M. and J. Kuret (2004). "Electron microscopy as a quantitative method for investigating tau fibrillization." Anal. Biochem. 329(2): 238-246. Nordberg, A., J. O. Rinne, et al. (2010). "The use of PET in Alzheimer disease." Nat Rev Neurol 6(2): 78-87.

126

Novak, M., J. Kabat, et al. (1993). "Molecular characterization of the minimal protease resistant tau unit of the Alzheimer's disease paired helical filament." EMBO J. 12(1): 365-370. Paparcone, R. and M. J. Buehler (2009). "Microscale structural model of Alzheimer Aß(1-40) amyloid fibril." Applied Physics Letters 94(24). Peltason, L. and J. Bajorath (2009). "Systematic computational analysis of structure activity relationships: concepts, challenges and recent advances." Future Medicinal Chemistry 1(3): 451-466. Perpete, E. A. and D. Jacquemin (2009). "TD-DFT benchmark for indigoid dyes." J Mol Struc- Theochem 914(1-3): 100-105. Petitjean, M. (1994). "On the Analytical Calculation of Van-Der-Waals Surfaces and Volumes - Some Numerical Aspects." Journal of Computational Chemistry 15(5): 507-523. Petkova, A. T., G. Buntkowsky, et al. (2004). "Solid state NMR reveals a pH-dependent antiparallel ß-sheet registry in fibrils formed by a ß-amyloid peptide." Journal of Molecular Biology 335(1): 247-260. Petkova, A. T., Y. Ishii, et al. (2002). "A structural model for Alzheimer's ß-amyloid fibrils based on experimental constraints from solid state NMR." Proceedings of the National Academy of Sciences of the United States of America 99(26): 16742-16747. Petkova, A. T., W. M. Yau, et al. (2006). "Experimental Constraints on Quaternary Structure in Alzheimer's b-Amyloid Fibrils." Biochemistry 45(2): 498-512. Pettersen, E. F., T. D. Goddard, et al. (2004). "UCSF chimera - A visualization system for exploratory research and analysis." Journal of Computational Chemistry 25(13): 1605- 1612. Rabinovici, G. D. and B. L. Miller (2010). "Frontotemporal lobar degeneration: epidemiology, pathophysiology, diagnosis and management." CNS Drugs 24(5): 375-398. Rodriguez-Rodriguez, C., A. Rimola, et al. (2010). "Crystal structure of thioflavin-T and its binding to amyloid fibrils: insights at the molecular level." Chem. Commun. 46(7): 1156- 1158. Rojo, L. E., J. Alzate-Morales, et al. (2010). "Selective interaction of lansoprazole and astemizole with tau polymers: potential new clinical use in diagnosis of Alzheimer's disease." J Alzheimers Dis 19(2): 573-589. Rowe, C. C., S. Ng, et al. (2007). "Imaging b-amyloid burden in aging and dementia." Neurology 68(20): 1718-1725. Rowe, C. C. and V. L. Villemagne (2011). "Brain Amyloid Imaging." Journal of Nuclear Medicine 52(11): 1733-1740. Rowland, R. S. and R. Taylor (1996). "Intermolecular nonbonded contact distances in organic crystal structures: Comparison with distances expected from van der Waals radii." Journal of Physical Chemistry 100(18): 7384-7391. Royall, D. R., R. Palmer, et al. (2002). "Pathological determinants of the transition to clinical dementia in Alzheimer's disease." Experimental Aging Research 28(2): 143-162. Rucker, C., G. Rucker, et al. (2007). "y-Randomization and its variants in QSPR/QSAR." J Chem Inf Model 47(6): 2345-2357. Sawaya, M. R., S. Sambashivan, et al. (2007). "Atomic structures of amyloid cross-ß spines reveal varied steric zippers." Nature 447(7143): 453-457. Schafer, K. N., S. Kim, et al. (2012). "Selectivity requirements for diagnostic imaging of neurofibrillary lesions in Alzheimer's disease: A simulation study." Neuroimage 60(3): 1724-1733. Sczepan, M., W. Rettig, et al. (2003). "Spectral and photophysical characteristics of unsymmetric polymethine dyes as model compounds for the colour shift of visual pigments." Photochemical & Photobiological Sciences 2(12): 1264-1271.

127

Sergeant, N., A. Bretteville, et al. (2008). "Biochemistry of Tau in Alzheimer's disease and related neurological disorders." Expert Rev Proteomics 5(2): 207-224. Shaw, L. M., H. Vanderstichele, et al. (2009). "Cerebrospinal fluid biomarker signature in Alzheimer's disease neuroimaging initiative subjects." Ann Neurol 65(4): 403-413. Shuto, Y. (1996). "Quadratic hyperpolarizabilities of nitro-substituted pseudo-linear dye molecules with ethylenic and azo bridges." Int J Quantum Chem 58(4): 407-418. Sigmundova, I., P. Zahradnik, et al. (2007). "Synthesis and study of novel benzothiazole derivatives with potential nonlinear optical properties." Collect. Czech. Chem. C 72(8): 1069-1093. Stine, W. B., Jr., K. N. Dahlgren, et al. (2003). "In vitro characterization of conditions for amyloid-ß peptide oligomerization and fibrillogenesis." J Biol Chem 278(13): 11612- 11622. Stone, M. and P. Jonathan (1993). "Statistical Thinking and Technique for Qsar and Related Studies .1. General-Theory." Journal of Chemometrics 7(6): 455-475. Stone, M. and P. Jonathan (1994). "Statistical Thinking and Technique for Qsar and Related Studies .2. Specific Methods." Journal of Chemometrics 8(1): 1-20. Sun, L., N. Tran, et al. (1998). "Synthesis and biological evaluations of 3-substituted indolin-2- ones: a novel class of tyrosine kinase inhibitors that exhibit selectivity toward particular receptor tyrosine kinases." J Med Chem 41(14): 2588-2603. Ter-Pogossian, M. M., M. E. Phelps, et al. (1975). "A positron-emission transaxial tomograph for nuclear imaging (PETT)." Radiology 114(1): 89-98. Terry, R. D., E. Masliah, et al. (1991). "Physical Basis of Cognitive Alterations in Alzheimers- Disease - Synapse Loss Is the Major Correlate of Cognitive Impairment." Annals of Neurology 30(4): 572-580. Tetko, I. V., J. Gasteiger, et al. (2005). "Virtual computational chemistry laboratory - design and description." J Comput Aid Mol Des 19(6): 453-463. ThS National Center for Biotechnology Information. PubChem Compound Database; CID=415676. Timofei, S., L. Kurunczi, et al. (1996). "Lipophilicity in dye-cellulose fibre binding." Dyes and Pigments 32(1): 25-42. Tomasi, J., B. Mennucci, et al. (2005). "Quantum mechanical continuum solvation models." Chem Rev 105(8): 2999-3093. Topliss, J. G. and R. J. Costello (1972). "Chance correlations in structure-activity studies using multiple regression analysis." J Med Chem 15(10): 1066-1068. Vanopdenbosch, N., R. Cramer, et al. (1985). "Sybyl, the Integrated Molecular Modeling System." Journal of Molecular Graphics 3(3): 110-111. von Bergen, M., S. Barghorn, et al. (2006). "The core of tau-paired helical filaments studied by scanning transmission electron microscopy and limited proteolysis." Biochemistry 45(20): 6446-6457. Wager, T. T., R. Y. Chandrasekaran, et al. (2010). "Defining Desirable Central Nervous System Drug Space through the Alignment of Molecular Properties, in Vitro ADME, and Safety Attributes." Acs Chemical Neuroscience 1(6): 420-434. Wang, W. S., J. M. Zhang, et al. (2005). "QSAR study of 125I-labeled 2-(4- aminophenyl)benzothiazole derivatives as imaging agents for b-amyloid in the brain with Alzheimer's disease." J Radioanal Nucl Chem 266(1): 107-111. Wang, Y., C. A. Mathis, et al. (2003). "Effects of lipophilicity on the affinity and nonspecific binding of iodinated benzothiazole derivatives." J Mol Neurosci 20(3): 255-260. Wehrens, R., H. Putter, et al. (2000). "The bootstrap: a tutorial." Chemom Intell Lab Syst 54(1): 35-52.

128

Weingarten, M. D., A. H. Lockwood, et al. (1975). "Protein Factor Essential for Microtubule Assembly." Proceedings of the National Academy of Sciences of the United States of America 72(5): 1858-1862. Wischik, C., C. Harrington, et al. (2004). Neurofibrillary labels. U. S. P. Application. 0213736 A1. Wu, C., M. Biancalana, et al. (2009). "Binding Modes of Thioflavin-T to the Single-Layer ß- Sheet of the Peptide Self-Assembly Mimics." J Mol Biol 394(4): 627-633. Wu, C., M. T. Bowers, et al. (2011). "On the origin of the stronger binding of PIB over thioflavin T to protofibrils of the Alzheimer amyloid-ß peptide: a molecular dynamics study." Biophys J 100(5): 1316-1324. Wu, C., Z. Wang, et al. (2008). "The Binding of Thioflavin T and Its Neutral Analog BTA-1 to Protofibrils of the Alzheimer's Disease Ab16-22 Peptide Probed by Molecular Dynamics Simulations." J. Mol. Biol. 384(3): 718-729. Yang, Y., L. Zhu, et al. (2010). "Binding research on flavones as ligands of ß-amyloid aggregates by fluorescence and their 3D-QSAR, docking studies." Journal of Molecular Graphics & Modelling 29(4): 538-545. Yona, R. L., S. Mazeres, et al. (2008). "Thioflavin derivatives as markers for amyloid-ß fibrils: Insights into structural features important for high-affinity binding." Chemmedchem 3(1): 63-66. Zhao, C., E. Boriani, et al. (2008). "A new hybrid system of QSAR models for predicting bioconcentration factors (BCF)." Chemosphere 73(11): 1701-1707. Zhuang, Z. P., M. P. Kung, et al. (2001). "Radioiodinated styrylbenzenes and thioflavins as probes for amyloid aggregates." J. Med. Chem. 44(12): 1905-1914.

129