Predictive Signatures from ToxCast Data for Chronic, Developmental and Reproductive Toxicity Endpoints Richard Judson, Matthew Martin, Thomas Knudsen, Nicole Kleinstreuer, David Dix, Keith Houck, Robert Kavlock U.S EPA, ORD, National Center for Computational Toxicology

Goals Cancer / Chronic Toxicity Developmental Toxicity Reproductive Toxicity

The goal of the ToxCast and Tox21 programs are: Rat Liver Tumors 1.Screening of hundreds to thousands of environmental chemicals using in Cleft Palate (CLP) Reproductive Performance vitro HTS methods Rodent liver endpoints such as proliferative CLP was selected as a prototype to build and evaluate lesions and tumors are among the most Chemicals tested in multigenerational reproductive toxicity studies (primarily carried out for 2 2.Development of predictive models or signatures by training against ROC Curve a predictive model for dev-tox. CLP-actives are listed commonly observed critical effects in generations) have been evaluated for their effects on reproductive performance. Reproductive chemicals with known toxicity Select model by categorical Lowest Effect Level (cLEL) else chronic / cancer studies. performance is an endpoint made up of many specific effects, including effects on mating, 3.Signatures should have relatively high sensitivity (few false negatives) that balances developmental LEL (dLEL) for rat and rabbit studies. sensitivity and fertility, gestation, implantation, and pregnancy. Each assay (where applicable) has been 4.Prioritization of chemicals for further targeted testing, based on positive This machine-generated signature includes specificity ToxCast Chemical CASRN mg/kg/day mapped to a gene and the average activity across all assays for a given gene have been used Univariate analysis in ToxMiner v16 showed 137 rat CLP rat cLEL rabbit dLEL to develop this model, predicting reproductive performance effects. response for signatures indicating potential for particular modes of toxicity the well known processes of PPAR significant associations with ToxCast™ in vitro assays. Cyproconazole 94361‐06‐524> 50 signaling and AR activity. Among the Propiconazole 60207‐90‐190> 400 Most (131) were rat-specific but 6 were rabbit-specific. Tri‐allate 2303‐17‐590not tested tested set of chemicals, sensitivity was Spiroxamine 118134‐30‐8100> 80 Reproductive Performance Gene Model The significant targets were used as inputs to build Triadimefon 43121‐43‐3100> 120 st nd nd high. There were many false positives multivariate Toxicity Signatures, grouped by signaling Triclopyr 55335‐06‐3200not tested 1 or 2 Generation 2 Generation Only (chemicals positive for the signature but Fluazinam 79622‐59‐6250> 12 Abstract Methods and Data system. No single association mapped to both species Flusilazole 85509‐19‐9250> 12 not causing rat liver tumors), but many of although close matches emerged for Forkhead Box Mancozeb 1/7/8018 512 > 80 these do cause either mouse liver tumors Dibutyl 84‐74‐2750not tested transcription factor networks [FOXM1, FOXA] and Dibutyl phthalate 84‐74‐2750not tested or pre-neoplastic liver lesions in rat or rabbit CLP rat dLEL rabbit cLEL The EPA ToxCast program is using in vitro assay data Univariate Association Analysis Biological / Toxicological signaling events mediated by Histone Deacetylase Cymoxanil 57966‐95‐7> 150 32 mouse. Dichlobenil 1194‐65‐6not tested 135 and chemical descriptors to build predictive models for Use AC50 for assays as input (~500 assays) Processes [HDAC-II, HDAC-I]. Class with in vivo toxicity endpoints. In vitro assays measure Log transform [-log10(AC50(M)/100000)] PPAR Signaling Literature Assay Component - Rabbit P-value activity of chemicals against molecular targets such as Inactive values set to zero Androgen Receptor Activity PS_PathwayCommons_NCI_NATURE_FOXM1_transcription_factor_network <0.001 Information PS_PathwayCommons_NCI_NATURE_Signaling_events_mediated_by_HDAC_Class_II <0.001 enzymes and receptors (measured in cell-free and cell- For dichotomous calculations, any activity=1 Angiogenesis (MCP1) PS_Gene_ESR1 <0.001 based systems in binding, agonist and antagonist Oxidative Stress PS_Ingenuity_Estrogen_Receptor_Signaling <0.001 Perform two univariate tests for each assay-endpoint pair ATG_PXRE_CIS 0.002 modes), in addition to cellular phenotypes and T-test, using continuous assay data, dichotomous Inflammatory response (IL8) CLZD_CYP2B6_48 0.031 pharmacokinetic-related parameters. endpoint Xenobiotic removal (SLCO1B1) Assay Component - Rat P-value PS_Gene_TGFB1 0.035 Chi-squared using dichotomous assay and endpoint PS_KEGG_TGF_beta_signaling_pathway 0.014 PS_PathwayCommons_CELL_MAP_TGFBR 0.017 Conclusions Over 600 separate assays were run in concentration data GENE ASSAY COMPONENT PS_PathwayCommons_REACTOME_Signaling_by_TGF_beta 0.035 Assay Score ATG_AR_TRANS 1. Improved model statistics when using response format to derive AC50 or LEC values for each PS_KEGG_Renal_cell_carcinoma 0.016 Positive = -log(AC50)+6 NCGC_AR_Agonist nd chemical-assay combination. This collection of AC50 Chemical by Chemical Results PS_Ingenuity_LPSIL_1_Mediated_Inhibition_of_RXR_Function 0.025 Negative = 0 AR NCGC_AR_Antagonist 2 Generation Only Endpoint Multivariate Signature Development PS_Ingenuity_PPARaRXRa_Activation 0.006 NVS_NR_hAR 2. Simple linear model does not over-fit values were used to build statistically significant Simple linear model (linmod) PS_Ingenuity_RAR_Activation 0.022 NVS_NR_rAR Rat Liver Tumorigen PS_PathwayCommons_NCI_NATURE_Retinoic_acid_receptors_mediated_signaling 0.043 Gene Scores ATG_ERa_TRANS 3. Comparable sensitivity and specificity predictors of in vivo toxicity endpoints derived from Calculate t-test p-value for all assays PS_PathwayCommons_NCI_NATURE_FOXA_transcription_factor_networks 0.041 {AR,ESR1,PPARA} = Mean(Assay Score) ATG_ERE_CIS PS_PathwayCommons_NCI_NATURE_Signaling_events_mediated_by_HDAC_Class_I 0.040 NCGC_ERalpha_Agonist chronic/carcinogenicity studies in rats and mice, ESR1 even though dataset is unbalanced Sort assays by increasing p-value Rat Proliferative lesions or mouse liver PS_Ingenuity_Aryl_Hydrocarbon_Receptor_Signaling 0.033 NCGC_ERalpha_Antagonist AR = Androgren Receptor 4. Biologically relevant targets identified prenatal developmental toxicity studies in rats and Calculate ROC curve as function of: tumorigen PS_Ingenuity_Wntbeta_catenin_Signaling 0.035 NVS_NR_bER PS_Ingenuity_Tight_Junction_Signaling 0.024 ESR1 = ERa = Receptor Alpha NVS_NR_hER 5. Predicts, accurately, many chemicals rabbits, and two-generation reproduction studies in rats. PS_KEGG_Neuroactive_ligand_receptor_interaction_Rattus_norvegicus 0.044 PPARA = Peroxisome proliferator-activated ATG_PPARa_TRANS number of assays included and cutoff value (hit No Rat or mouse major liver effects PPARA NCGC_PPARa_Agonist PS_PathwayCommons_NCI_NATURE_PDGFR_alpha_signaling_pathway 0.031 receptor alpha not used in training or testing of model vs. no-hit) NVS_ADME_hCYP2C19 0.043 NVS_NR_hPPARa A collection of 309 unique chemicals were tested and No animal data NVS_ADME_rCYP2D2 0.034 (e.g., HPTE, ) Only 2 free parameters PS_Gene_CYP2C19 0.050 used for predictive signature development. Machine Select model with best sensitivity and specificity PS_Gene_Cyp2d2_Rattus_norvegicus 0.034 learning techniques were used to build multivariate PS_Ingenuity_Xenobiotic_Metabolism_Signaling 0.042 Use k-fold cross validation 5 false positive are low-dose PS_PathwayCommons_REACTOME_Host_Interactions_with_Influenza_Factors 0.035 Reproductive Performance Gene Model (2nd Generation Only) predictive signatures. K-fold cross validation was PS_PathwayCommons_REACTOME_Influenza_Infection 0.035 True Positives (Total: 24) False Positives (Total: 59) Model=intercept + sum(-log10(p-value) for included cholinesterase inhibitors PS_PathwayCommons_REACTOME_Influenza_Virus_Induced_Apoptosis 0.035 PPV performed to quantify prediction accuracy. assays CLP pathways common to rat and rabbit ; Lactofen; ; Fenthion; Imazalil; Vinclozolin; Diclofop‐methyl; Fenhexamid; Bensulide; Flumetralin;

Flusilazole; Triadimefon; Fipronil; Triflumizole; Fluazinam; Hexaconazole; Fludioxonil; Pyridaben; Malathion; Phosalone; =

A total of 687 in vitro assay endpoints were used from ToxCast. Diagnostic Pathway TP BA

18/21 rat liver tumorigens identified 29% rat CLP Clomazone; Diazinon; Emamectin benzoate; Triticonazole; Carbaryl; Linuron; Tetraconazole; ; Oxyfluorfen; Three sets of mechanistic signatures are presented. This includes multiple time points for several assays. All assays PS_PathwayCommons_NCI_NATURE_Signaling_events_mediated_by_HDAC_Class_I 11 0.72 Fenbuconazole; Maneb; Folpet; Boscalid; Resmethrin; Acetochlor; Triadimenol; Dithiopyr; Propiconazole; Fenpropathrin; were run in concentration-response format and we extracted a 3 false negatives are active in no PS_PathwayCommons_NCI_NATURE_FOXA_transcription_factor_networks 10 0.65 1.Chronic rat liver tumors based on assays including rabbit CLP Sulfentrazone; Propetamphos; Methidathion; Benfluralin Dibutyl phthalate; Etoxazole; 38 more... characteristic concentration for each chemical-assay pair. These model assays Consensus model PS_PathwayCommons_NCI_NATURE_Signaling_events_mediated_by_HDAC_Class_II 2 0.82 targets in the peroxisome proliferator-activating receptor PS_PathwayCommons_NCI_NATURE_FOXM1_transcription_factor_network 2 0.83 False Negatives (Total: 13) True Negatives (Total: 146)

were either AC50 or LEC (lowest effective concentration, or the NPV (PPAR) pathway. Trichlorfon; Dichlobenil; Mevinphos; Dichlorprop; Butafenacil; Ethalfluralin; Cycloate; Myclobutanil; Ethofumesate; Cyclanilide; lowest concentration at which the response was significantly A simple linear model predicted CLP for all CLP-actives except Tri-allate, Triclopyr, Cymoxanil

Carboxin; Daminozide; Dichlorvos; Norflurazon; Fluazifop‐butyl; Cyromazine; MGK; Azinphos‐methyl; Bendiocarb; Oxasulfuron; =

2.Prenatal cleft palate in rat fetuses based on assays different than negative control). and Dichlobenil. On the other hand, the FOX and HDAC pathway models were of the highest 92% targeting Wnt signaling, retinoids, AhR, TGF-beta, Propoxycarbazone‐sodium; Prosulfuron; Tri‐allate Cymoxanil; Bifenazate; Dichloran; Pyriproxyfen; Diphenylamine; Chemical Predicted Positive for Rat Liver Tumors diagnostic value in terms of sensitivity-specificity, including: signaling, Forchlorfenuron; Acibenzolar‐S‐Methyl; Pymetrozine; Propazine; glucocorticoids and chemokines, identifying pathway Pathway score were calculated by taking the average log(AC50) for Based on Signature signaling events mediated by HDAC Class II, and the FOXM1 transcription factor network (BA Simazine; Tebufenpyrad; Icaridin; TCMTB; 2,4‐D; 122 more... level disruption of the Forkhead-family of transcription all assays that mapped to genes included in a pathway. Pathway = 0.83). Mutation or deregulation of FOX genes are associated with human congenital Sensitivity = 65% Specificity = 71% factors and chromatin remodeling. sources included KEGG, Ingenuity and PathwayCommons. Only Predicted Positive ‐ No ToxRefDB Study (Total: 26) Predicted Negative‐ No ToxRefDB Study (Total: 41) human pathways were included. ToxCast assays mapped to a total disorders, and CLP is one of the complications. 3.Reproductive fitness toxicity seen in rat reproduction HPTE*; Methoxychlor; Bromoxynil; Clorophene; Fenitrothion; Tefluthrin; Tebupirimfos; Anilazine; Spiroxamine; Cloprop; 6‐ of 303 human genes. studies and based on in vitro assays, pathways and Perfluorooctanoic acid; Perfluorooctane sulfonic acid; Parathion; Deisopropylatrazine; Oxytetracycline dihydrate; Clopyralid‐olamine; Napropamide; Prodiamine; MEHP*; Fluazifop‐P‐butyl; Tribufos; Abamectin; Fluroxypyr‐meptyl; Acifluorfen; Clofentezine; chemical descriptors. For these calculations, we did not use information on magnitude of Cinmethylin; Isazofos; Chlorpyrifos oxon*; Milbemectin; Bensulfuron‐methyl; Boric acid; 25 more... Any Lesion Pre-Neoplastic Neoplastic effect (Max fold change or Emax). Assay AC50/LEC values were Summary of Univariate Associations Conclusion Dimethomorph; Diniconazole; ; Alachlor; Esfenvalerate; Data Availability log transformed prior to performing statistical calculations. In 132 60 21 Fenoxycarb; Butralin; Hexythiazox; Disulfoton addition to the in vitro assay data, we also used chemical class *Designates metabolites within chemical set Look across disease progression information, limited physico-chemical values calculated from This toxicity signature posits a generalized role for http://www.epa.gov/ncct/toxcast chemical structure, and structure fragment composition fingerprints. Includes gene-disease linkage • Judson, et al., “A Comparison of Machine Learning Algorithms for Chemical Toxicity http://actor.epa.gov the Forkhead-family of transcription factors and Classification Using a Simulated Multi-Scale Data Mode” BMC Bioinformatics Vol. 9, 241 (2008) chromatin remodeling as targets for environmental • Martin, et al. “Profiling Chemicals Based on Chronic Toxicity Profiles from the U.S. EPA ToxRef Endpoints from rat, mouse and rabbit chronic-cancer, http://toxrefdb.epa.gov (coming soon) chemicals in the mechanism underlying cleft palate. Database”, Environmental Health Perspectives, Vol. 117, 392-399 (2009) multigenerational and developmental studies were extracted from • Knudsen, et al. “Profiling the Activity of Environmental Chemicals in Prenatal Developmental ToxRefDB and dichotomized (cause / do not cause endpoint) References Toxicity Studies using the U.S. EPA’s ToxRefDB”, Reproductive Toxicology, Vol. 31, 241 (2009) • Martin, et al. “Profiling the Reproductive Toxicity of Chemicals from Multigeneration Studies in the Toxicity Reference Database (ToxRefDB)”, Toxicological Sciences, Vol. 110, 181 (2009). • Judson, et al. “Predictive In Vitro Screening of Environmental Chemicals – The ToxCast Project”, Environmental Health perspectives (EPub ahead of Print Dec 2009)

Building a scientific foundation for sound environmental decisions This work may not necessarily reflect official Agency policy.