Accelerating Drug Discovery with Computational Chemistry

Graduate Student Guest Lecture Mikolai Fajer, PhD Senior Scientist

Transforming drug discovery and materials research Drug Discovery

Transforming drug discovery and materials research Designing Drugs is Hard

3 Designing Drugs is Hard

• Need to simultaneously optimize many properties: – Potency – Bioavailability – CYP inhibition – Toxicity – Selectivity – Clearance / half-life – hERG blockade – Solubility – Permeability – Synthesizability Property 10 Property Property 1 Property 2 Property 3 Property 4 Property 5 Property 6 Property 7 Property 8 Property 9 Property Mol 1 Mol 2 Mol 3 Mol 4

4 Designing Drugs is Hard

• Need to simultaneously optimize many properties: – Potency – Bioavailability – CYP inhibition – Toxicity – Selectivity – Clearance / half-life – hERG blockade – Solubility – Permeability – Synthesizability Statistical Thermodynamics Property 10 Property Property 1 Property 2 Property 3 Property 4 Property 5 Property 6 Property 7 Property 8 Property 9 Property Mol 1 Mol 2 Mol 3 Mol 4

5 Even Late in a Project, Activity Cliffs are a Problem

Representative Pre-FEP+ Project (2010) • ADMET tuning repeatedly leads to (12% of molecules tight binding) losses of potency even in late in project 11 • Primary series liabilities may force 10 the project team to effectively start over 9 – Core-related toxicity 8 – Competitor IP filings 7 • Earlier SAR may not transfer to backup series 6

Compound Potency (log units) 4 1 Time (2+ Years / 1,390 cmpds)

6 Most Drug Discovery Projects Fail to Deliver a Drug into the Clinic

34% Succeed Send molecule into the clinic 44% Fail Due to Ligands of Insufficient Quality 34% Ligand toxicity problems; efficacy not observed at 44% achievable dose; poor PK/PD

22% 22% Fail Due to Poor Initial Target Selection Target engagement not efficacious for treating disease; on-target toxicity; corporate portfolio rebalancing

7 Data extracted from: Nature Review Drug Discovery, 2016, 14, 475; Nature Review Drug Discovery, 2014, 13, 419; Nature Review Drug Discovery 2010, 9, 203. Current State of the Art

N O O H N N N O N O O H H N N N N N O O N O O O N N H N O O H H NH N N O N N N O O O N N N N O N N N O O H O N N O N H H N N N O N N N O N N N O O O O N N H O N N O N H O N N N O O O N N N O O N O N O N O O O O N Trial-and-errorO N O O

N O O N O O H H N N N O N N N O H H N N N O N N N O H H O N N N N N O O N N N N N O HO H H H O N N N O O N N N N O N N O O N H O N N H H 0.41% N N N O H HO O 10%O H N N N O N N N O HO O N O N N N O HO O N H O H O N H N N N O O N N N O O N N O N N N O N N N O H HO N O N probability a potent HO O O N N N N O HO O 34% probability the N O O N O O N O N Cl HO O compound meets O O N designed compound chance of Cl is Cpotentl enough other project criteria drug Cl 2,000 200 entering H potentH clinic H H N designN N ideasO N N N O N N N O HO N N N O H HO H N N N O N N N O synthesized synthesizedN N HO O O N O HNO O O compounds O O N O N

Cl Cl

8 Deduced from: Nature Review Drug Discovery, 2016, 14, 475; Nature Review Drug Discovery, 2014, 13, 419; Nature Review Drug Discovery 2010, 9, 203. Free Energy Perturbation

Transforming drug discovery and materials research How to Rigorously Compute Affinity – Compute All the Terms

Molecule in Molecule in bound Desolvated water conformation molecule

∆G(1) ∆G(3)

∆G(5) Molecule bound to protein

∆G(2) ∆G(4)

∆Gbind = ∆G(1) + ∆G(2) + ∆G(3) + ∆G(4) + ∆G(5) Protein in Protein in bound Desolvated water conformation protein

10 FEP Provides a Rigorous Method to Compute Relative Binding Affinity

Free Energy Perturbation (FEP) technology

• Rigorous calculation of the binding affinity difference between two ligands – Series of molecular dynamics simulations are run where Ligand A is alchemically transformed into Ligand B – Appropriate Stat. Mech. is used to rigorously compute the binding free energy difference between Ligands A and B

11 FEP and the Abstract λ-coordinate

Real starting state Alchemical intermediates Real final state

V0 V (λ) V1

λ = 0 λ = 0.5 λ = 1

0 G 0 0 0 ΔG G1

0 0 0 ΔG = G −G = −kT ln exp −ΔE (BAR or MBAR used in practice) 1 0 ( kT)

12 Dissecting Expectation Values

• Thermodynamic Integration • What do the various bits mean?

∆�=−��ln⟨exp[−∆�⁄�� ]⟩ �(�) – Forcefield • Free Energy Perturbation exp[−∆�⁄�� ] ��/ �(�) �� ∆�=∫0↑1▒⟨��/�� ⟩↓� �� – Forcefield, Perturbation Size

∫↑▒⋯�� ∫↑▒�↑−�(�) �� • Expectation Value

⟨�⟩=∫↑▒�(�)�↑−�(�) �� /∫↑▒ – Sampling �↑−�(�) ��

13 Schrödinger’s Approach: FEP+

Requirements for FEP calculations to enable drug discovery Schrodinger FEP+ Advances

High Accuracy Across Chemical Space Built an accurate force field (molecular energy function and interatomic potentials) that covers effectively all of 1. Predictions must be accurate relevant chemical space. 2. Calculation must take much less time than compound synthesis High Throughput: Speed & Efficiency Deployed GPU-enabled molecular dynamics engine - 3. Calculation setup must be ~50-100s performance advantage over CPUs. Invested in technology for running securely on the cloud. straightforward Extensive use of enhanced sampling methods. Automated Setup Tools Built automation tools to make jobs significantly easier to run and troubleshoot

~ 200 person-years of effort to make advances in the science

High Throughput: Fast and efficient calculations Secure Cloud Platform Invested heavily in internal GPU cluster and technology for running on the Cloud Performance and Validation

Transforming drug discovery and materials research Head-to-Head Study: FEP+ vs. Traditional Methods for Compound Prioritization

Study Overview Methods

• Cathepsin L is a lysosomal endopeptidase which plays an • 3,325 R-group idea library provided by the important role in protein degradation and apoptosis collaborator Good structural data available along with some SAR • • Design ideas were docked and scored using • The collaboration was used as a test of FEP+ scoring vs. Glide and Prime MM-GB/SA for idea triage traditional approaches to optimize a 200 nM lead compound • 92 molecules selected for FEP+ scoring • Head-to-head comparison:

Article

pubs.acs.org/jmc – 10 molecules prioritized by FEP+ scoring – 10 molecules prioritized by an Prospective Evaluation of Free Energy Calculations for the experienced med chemist (Med Chem) Prioritization of Cathepsin L Inhibitors # # § # § Bernd Kuhn,†, Michal Tichy,́ ‡, Lingle Wang, , Shaughnessy Robinson, Rainer E. Martin,† – 10 molecules prioritized an experienced § Andreas Kuglstatter,† Jörg Benz,† Maude Giroud,‡ Tanja Schirmeister,∥ Robert Abel,*, *,‡ *,† modeler using any technique other than Francoiş Diederich, and Jeró ̂me Hert free energy calculations (SBDD) †Roche Pharmaceutical Research and Early Development (pRED), Roche Innovation Center Basel, F. Hoﬀmann-La Roche Ltd., Grenzacherstrasse 124, 4070 Basel, Switzerland ‡Laboratorium für Organische Chemie, ETH Zurich, Vladimir-Prelog-Weg 3, 8093 Zurich, Switzerland – 10 molecules prioritized by docking and § Schrödinger, Inc., 120 West 45th Street, New York, New York 10036, United States ∥Institut für Pharmazie und Biochemie, Johannes Gutenberg-Universitaẗ Mainz, Staudinger Weg 5, 55128 Mainz, Germany filtering (Docking) *S Supporting Information

16 Source: Kuhn et al. J. Med. Chem. 2017, 60(6):2485-2497.

ABSTRACT: Improving the binding affinity of a chemical series by systematically probing one of its exit vectors is a medicinal chemistry activity that can benefit from molecular modeling input. Herein, we compare the effectiveness of four approaches in prioritizing building blocks with better potency: selection by a medicinal chemist, manual modeling, docking followed by manual filtering, and free energy calculations (FEP). Our study focused on identifying novel substituents for the apolar S2 pocket of cathepsin L and was conducted entirely in a prospective manner with synthesis and activity determination of 36 novel compounds. We found that FEP selected compounds with improved affinity for 8 out of 10 picks compared to 1 out of 10 for the other approaches. From this result and other additional analyses, we conclude that FEP can be a useful approach to guide this type of medicinal chemistry optimization once it has been validated for the system under consideration.

■ INTRODUCTION is to engage in parallel synthesis: A common setting involves a ff fi Free energy calculation approaches, such as free energy sca old with one or several de ned exit vectors and a set of perturbation (FEP), have been around for a long time1−5 but chemical reactions with the goal of optimizing side chains. The had only limited impact in the drug discovery process so far. number of suitable reactants accessible (internally or Likely reasons for their historically restrained use include purchasable) can be very large (hundreds, thousands, or more). The task of molecular modeling consists then in lengthy simulation times not practical in fast-paced project ffi environments combined with overstated accuracy levels based prioritizing building blocks with respect to binding a nity in on small test set retrospective analyses which did not translate order to limit the amount of synthesis and experimental testing when employed prospectively in real-world systems. FEP has required. now taken advantage of improved sampling algorithms6,7 and A prerequisite for this exercise is the availability of an initial force-field quality8 and is profiting from the increased ligand together with structural information, an experimental availability of low-cost parallel computing. Speed and accuracy cocrystal structure describing the binding mode to the protein. appear to have progressed significantly.7,9,10 This has in turn led We picked human cathepsin L (hCatL), a cysteine protease, to recent accounts of successful industrial applications of FEP which can be inhibited by ligands with an activated nitrile group in active drug discovery projects.11−13 forming a covalent thioimidate adduct with the catalytic Cys25. Here, we investigate the application of FEP in a typical drug Previous SAR and structural studies with aryl nitriles (Figure 1) discovery use case where the goal is to prioritize compounds for synthesis. One way for therapeutic project teams to further Received: December 28, 2016 explore the structure−activity relationship (SAR) of a hit series

FEP+ Med Chem SBDD Docking

Cmpd. ID Exp Ki (nM) Cmpd. ID Exp Ki (nM) Cmpd. ID Exp Ki (nM) Cmpd. ID Exp Ki (nM) 3 12 3 12 3 12 22 77 37 25 11 279 14 217 29 304 31 27 9 515 16 505 28 358 33 30 7 952 13 1010 26 411 35 77 6 1800 20 1020 23 671 34 91 4 3020 17 2790 13 1010 30 123 8 (cis) 3500 15 3860 20 1020 38 167 5 >5100 18 >5100 24 5100 36 1430 8 (trans) >5100 19 >5100 25 >5100 32 1750 10 >5100 21 >5100 27 >5100

8/10 tighter binding 1/10 tight binding 1/10 tight binding 1/10 tight binding than the reference than the reference than the reference than the reference

17 Note: Molecule 3 was very close to the known SAR, tight binding was unsurprising FEP+ Selected Molecules were Diverse and Manifested Bond Topologies not Previously Known to Bind Cathepsin L

31 37 3 32 Expt. Ki = 25 nM Expt. Ki = 27 nM Expt. Ki = 12 nM Expt. Ki = 1750 nM Pred. Ki = 24 nM Core Pred. Ki = 15 nM Pred. Ki = 18 nM Pred. Ki = 15 nM

33 34 35 30 36 Expt. Ki = 30 nM Expt. Ki = 91 nM Expt. Ki = 77 nM Expt. Ki = 123 nM Expt. Ki = 1430 nM Pred. Ki = 16 nM Pred. Ki = 103 nM Pred. Ki = 19 nM Pred. Ki = 11 nM Pred. Ki = 18 nM 18 FEP+ Domain of Applicability

1. At least one high quality crystal structure with cocrystalized series ligand

2. Reasonable expectation of a conserved binding mode across the series

3. Minimal tautomeric, ionization-state, and stereochemistry uncertainties across the series

4. High reliability experimental binding data from the same assay for all compounds

5. Assay data and crystal structures are for the same protein construct

19 FEP+ Domain of Applicability

• Failing to meet these criteria doesn’t necessarily guarantee failure, but it does make it more likely

• In general, the cleanliness of the experimental data is much more important than the identity of the target or the type of ligand modification

20 FEP+ as a ’Computational Assay’

• Accuracy: beginning to approach experimental error for amenable targets

• Speed: ~50X faster per molecule – 6 hours/compound vs. ~3 weeks/compound

• Throughput: ~250X higher – e.g., 5000 molecules in 1 week vs. 5 years

• Project synthesis resources can be focused where they will be most productive

21 How Can a Rigorous Computational Binding Assay Impact Preclinical Drug Discovery Projects?

• Faster potency optimization with fewer synthesized 1 compounds à improve efficiency of MedChem cycles

• Ability to de-risk challenging chemical synthesis and 2 affordably explore a greater diversity of chemical space

• Better maintain potency while simultaneously tuning 3 ADMET properties during lead optimization: – Binding selectivity – Mutational resistance – Solubility – Membrane permeability

22 FEP+ Provides Unprecedented Control of Compound Affinity

Representative Pre-FEP+ Project (2010) Post-FEP+ Project (2016) (12% of molecules tight binding) (66% of molecules tight binding)

11 11

10 10

9 9

8 8

7 7

6 6

5 5 Compound Potency (log units) 4 Compound Potency (log units) 4 1 0 Time (2+ Years / 1,390 cmpds) Time (11 Months, 509 cmpds)

23 FEP+ was used to simultaneously optimize ligand binding potency, binding selectivity, and aqueous solubility

FEP+ Selec > 1000x FEP+ pKi=11

FEP+ Sol < 0.5 uM This approach has allowed us rapidly identify multiple highly potent, selective, and soluble ligands to advance the project

FEP+ Selec < 0.1x

pKi > 9 > 100x selectivity > 20 uM solubility

FEP+ pKi=8 FEP+ Sol > 100 uM

24 FEP+ scoring accurately captures highly counter intuitive potency, selectivity, and solubility trends

• FEP+ scoring correctly predicted molecule D would be much more selective and soluble than molecule F

Mol. D: Mol. E: pKi > 9 pKi > 9 Selec. < 100x Selec. > 100x Solub. < 10 µM Solub. > 20 µM

Core Core

• Traditional Med. Chem. design strategies have great difficult anticipating such trends

25 Drug Discovery Guided by a Rigorous Computational Method

N O O H N N N O N O O H H N N N N N O O N O O O N N H N O O H H NH N N O N N N O O O N N N N O N N N O O H O N N O N H H N N N O N N N O N N N O O O O N N H O N N O N H O N N N O O O N N N O O N O N O N O O O O N FEP+ scoringO N O O

N O O N O O H H N N N O N N N O H H N N N O N N N O H H O N N N N N O O N N N N N O HO H H H O N N N O O N N N N O N N O O N H O N N H H 0.41% N N N O H HO O >>10%O H N N N O N N N O HO O N O N N N O HO O N H O H O N H N N N O O N N N O O N N O N N N O N N N O H HO N O N probability a potent HO O O N N N N O HO O >>34% probability the N O O N O O N O N Cl HO O compound meets O O N designed compound chance of Cl is Cpotentl enough other project criteria drug Cl 2,000 >>200 entering H potentH clinic H H N PoolN ofN designO ideas, N N N O N N N O HO N N N O H HO H N N N O N N N O synthesized rankedN by FEP+ scoring N HO O O N O HNO O O compounds O O N each med chem cycle;O N best 2,000 synthesized Cl over the project Cl

26 Drug Discovery Guided by a Rigorous Computational Method

N O O N O O H H N N N O N N N O H H N N N O N N N O H H O N N N N N O O N N N N N O HO H H H O N N N O O N N N N O N N O O N H O N N H H 0.41% N N N O H HO O >>10%O H N N N O N N N O HO O N O N N N O HO O N H O H O N H N N N O O N N N O O N N O N N N O N N N O H HO N O N probability a potent HO O O N N N N O HO O >>34% probability the N O O N O O N O N Cl HO O compound meets O O N designed compound chance of Cl is Cpotentl enough other project criteria drug Cl >>2,000 >>200 entering H potentH clinic H H N PoolN ofN designO ideas, N N N O N N N O HO N N N O H HO H N N N O N N N O synthesized rankedN by FEP+ scoring N HO O O N O HNO O O compounds O O N each med chem cycle;O N best 2,000 synthesized Cl over the project Cl

27 In the Trenches

Transforming drug discovery and materials research The Problem

Article BACE1 cleaves amyloid precursor pubs.acs.org/JCTC • protein yielding the β-amyloid peptide Acylguanidine Beta Secretase 1 Inhibitors: A Combined Experimental that aggregates into plaques in and Free Energy Perturbation Study Alzheimer’s disease § Henrik Keranen,̈ † Laura Perez-Benito,́ ‡, Myriam Ciordia,⊥ Francisca Delgado,⊥ Thomas B. Steinbrecher,∥ # § Daniel Oehlrich, Herman W. T. van Vlijmen,† Andreś A. Trabanco,⊥ and Gary Tresadern*, †Computational Chemistry, Janssen Research & Development, Janssen Pharmaceutica N. V., Turnhoutseweg 30, B-2340 Beerse, Belgium ‡Laboratori de Medicina Computacional Unitat de Bioestadistica, Facultat de Medicina, Universitat Autonoma de Barcelona, 08193, • We have: Bellaterra, Spain §Computational Chemistry, Janssen Research and Development, Janssen-Cilag, c/ Jarama 75A, 45007, Toledo, Spain – A crystal structure with a reference ⊥Neuroscience Medicinal Chemistry, Janssen Research and Development, Janssen-Cilag, c/ Jarama 75A, 45007, Toledo, Spain ∥Schrödinger GmbH, Dynamostrasse 13, 68165 Mannheim, Baden-Württemberg, Germany, compound #Neuroscience Medicinal Chemistry, Janssen Research & Development, Janssen Pharmaceutica N. V., Turnhoutseweg 30, B-2340 Beerse, Belgium – A suggestion that changing the size of *S Supporting Information the ring in the P2’ pocket could improve ABSTRACT: A series of acylguanidine beta secretase 1 (BACE1) inhibitors with modified scaffold and P3 pocket substituent was synthesized and studied with free results energy perturbation (FEP) calculations. The resulting molecules showed potencies in enzymatic BACE1 inhibition assays up to 1 nM. The correlation between the predicted activity from the FEP calculations and the experimental activity was good for the P3 – Enumeration of the other end of the pocket substituents. The average mean unsigned error (MUE) between prediction and experiment was 0.68 ± 0.17 kcal/mol for the default 5 ns lambda window simulation molecule for further affinity tuning time improving to 0.35 ± 0.13 kcal/mol for 40 ns. FEP calculations for the P2′ pocket substituents on the same acylguanidine scaffold also showed good agreement with experiment and the results remained stable with repeated simulations and increased simulation time. It proved more difficult to – Knowledge that the protein flap and use FEP calculations to study the scaffold modification from increasing 5 to 6 and 7 membered-rings. Although prediction and experiment were in agreement for short 2 ns simulations, as the simulation time increased the results diverged. This was 10s loop are flexible improved by the use of a newly developed “Core Hopping FEP+” approach, which also showed improved stability in repeat calculations. The origins of these differences along with the value of repeat and longer simulation times are discussed. This work provides a further example of the use of FEP as a computational tool for molecular design.

29 ■ INTRODUCTION applications are emerging,12−18 including work from our laboratories investigating FEP applied in lead optimization.19,20 The accurate prediction of protein−ligand binding affinities is Here, we explore FEP predictions of the binding energies of of major interest.1 Rigorous approaches can calculate the β-secretase 1 (BACE1) inhibitors in the different subpockets of binding free energy difference between two structurally similar fi the active site. Cleavage of amyloid precursor protein by this ligands by making use of alchemical structural modi cations. aspartyl protease leads to increases of β-amyloid (Aβ) peptides Free-energy perturbation (FEP) or thermodynamic integration that aggregate forming senile plaques, one of the neuro- (TI), using molecular dynamics (MD) or Monte Carlo ’ 21 2 pathological features in Alzheimer s disease (AD). Contem- simulations, are among the widely used approaches. These porary BACE1 inhibitors contain an amidine/guanidine moiety methods are ideal for drug discovery lead optimization where within a heterocycle of varying size.22 This protonated group close structural analogues are synthesized and their properties forms an optimal hydrogen-bonding network to the catalytic compared to previous best leads. Computation of accurate aspartate dyad, Figure 1. Also, a quaternary alpha sp3 carbon relative binding affinities can make a big impact in this costly provides an ideal vector to fill the P1−P3 and P2′ pockets of phase. Also, it avoids the computationally expensive prediction the binding site (Figures 1 and 2).23,24 These molecules have of absolute binding free energies. Calculating relative protein− improved drug like properties and multiple examples are in 25 ligand binding affinities in this way dates back at least thirty clinical trials. Given the huge pharmaceutical interest in years.3−8 Lately, new sampling algorithms, improved force finding new treatments for AD, BACE1 is a very well-studied fields and low-cost parallel computing (often graphics processing units GPU), have improved accuracy and turn- Received: November 22, 2016 around time.9−11 Reports of large scale and industrial Published: January 19, 2017

• Change the size of the ring • Change the R-groups

Perturbation Size

30 Testing the Ring Size Next

ΔΔG @ 2ns ΔΔG @ 5ns ΔΔG @ 10ns ΔΔG @ 40ns R N Compound (kcal/mol) (kcal/mol) (kcal/mol) (kcal/mol) -3.2±0.3 0→1 8a→17a -5.2±2.6 -2.0±2.3 -1.5±0.7 -2.3±3.3 -1.3±0.3 0→2 8a→27a -1.3±1.8 1.1±1.5 -0.2±1.9 1.9±2.6 -3.5±0.1 0→1 8c→17c -3.0±1.3 -5.0±2.5 -2.3±1.7 -3.7±1.6 -1.8±0.1 0→2 8c→27c -2.1±0.5 -0.9±2.5 -1.2±1.6 0.1±3.2 -1.9±0.1 0→1 8g→17g -4.6±3.0 -4.4±1.3 -1.8±0.9 -3.3±1.3 -0.9±0.1 0→2 8g→27g -1.8±0.9 0.7±1.2 2.2±1.8 0.9±1.9

31 What causes large deviations between runs? Back

ΔΔG @ 2ns ΔΔG @ 40ns Compd. (kcal/mol) (kcal/mol)

8c→27c -2.1±0.5 0.1±3.2

32 Testing the R-group enumeration Back

R N Compound 2ns 5ns 10ns 40ns

-11.4±0.1 1 17a -10.8±0.1 -11.6±0.3 -11.4±0.3 -11.7±0.1

-12.1±0.4 1 17c -11.7±0.1 -11.8±0.2 -12.3±0.1 -12.1±0.1

-12.6±0.2 1 17g -14.2±0.4 -13.9±0.3 -13.5±0.3 -13.7±0.1

33 Minimize the Perturbation

Full Ring Perturbation “Core Hopping”

“Core Hopping” Perturbation

34 Evaluation

• Quantitative accuracy? • What have we learned for the next round? • Predictive accuracy?

ΔΔG @ 40ns ΔΔG exp. Compound (kcal/mol) (kcal/mol)

8a→17a -2.3±3.3 -3.2±0.3

8a→27a 1.9±2.6 -1.3±0.3