BioIT Drug induced liver injury talk 2011

A Predictive Ligand-Based Bayesian Model for Human Drug Induced Liver Injury Sean Ekins Collaborations in Chemistry, Jenkintown, PA. Collaborative Drug Discovery, Burlingame, CA. School of Pharmacy, Department of Pharmaceutical Sciences, University of Maryland.

Pharma reached a productivity tipping point Cost of drug development high Failure in clinic due to toxicity How to predict failure earlier

Why We Need Models Define structure activity relationship (SAR) in vitro models - limited throughput in silico – in vitro approach has value in targeting testing of compounds. Computers can beat humans at chess and Jeopardy why not help with predicting toxicity?

The future: crowdsourced drug discovery Williams et al., Drug Discovery World, Winter 2009

What is DILI? Drug metabolism in the liver can convert some drugs into highly reactive intermediates, In turn can adversely affect the structure and functions of the liver. Drug-induced liver injury (DILI), is the number one reason drugs are not approved and also the reason some of them were withdrawn from the market after approval Estimated global annual incidence rate of DILI is 13.9-24.0 per 100,000 inhabitants, and DILI accounts for an estimated 3-9% of all adverse drug reactions reported to health authorities Herbal components can cause DILI too https://dilin.dcri.duke.edu/for-researchers/info/

Drug Examples for DILI + and - Troglitazone DILI + Pioglitazone DILI - Rosiglitazone DILI - Sulindac DILI + Aspirin DILI - Diclofenac DILI + Xu et al., Toxicol Sci 105: 97-105 (2008).

Limitations of DILI? Compound has to physically have been made and be available for testing. The screening system is still relatively low throughput compared with any primary screens Whole compound or vendor libraries cannot be cost effectively screened for prioritization. Screening system should be representative of the human organ including drug metabolism capability. Prediction of human therapeutic C max is often imprecise before clinical testing in actual patients. Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010

Statistical Methodologies Non Linear regression Genetic algorithms Neural networks Support vector machines Recursive partitioning (trees) Sammon maps Bayesian methods Kohonen maps A rich collection of descriptors. Public and proprietary data. Problems to date – small datasets Understanding applicability chemical space Tools for big datasets P-gp +ve P-gp -ve Balakin et al.,Curr Drug Disc Technol 2:99-113, 2005.

DILI Computational Models 74 compounds - classification models (linear discriminant analysis, artificial neural networks, and machine learning algorithms (OneR)) Internal cross-validation (accuracy 84%, sensitivity 78%, and specificity 90%). Testing on 6 and 13 compounds, respectively > 80% accuracy. (Cruz-Monteagudo et al., J Comput Chem 29: 533-549, 2008). A second study used binary QSAR (248 active and 283 inactive) Support vector machine models – external 5-fold cross-validation procedures and 78% accuracy for a set of 18 compounds (Fourches et al., Chem Res Toxicol 23: 171-183, 2010). A third study created a knowledge base with structural alerts from 1266 chemicals. Alerts created were used to predict results for 626 Pfizer compounds (sensitivity of 46%, specificity of 73%, and concordance of 56% for the latest version) (Greene et al., Chem Res Toxicol 23: 1215-1222, 2010).

DILI data Tested a panel of orally administered drugs at multiples of the maximum therapeutic concentration ( C max ), taking into account the first-pass effect of the liver and other idiosyncratic toxicokinetic/toxicodynamic factors. The 100-fold C max scaling factor represented a reasonable threshold to differentiate safe versus toxic drugs for an orally dosed drug and with regard to hepatotoxicity. Concordance of the in vitro human hepatocyte imaging assay technology (HIAT) for 300 drugs and chemicals, ~ 75% with regard to clinical hepatotoxicity, with very few false-positive results Xu et al., Toxicol Sci 105: 97-105 (2008).

CRIMALDDI Meeting 2010 www.collaborativedrug.com Linking databases www.chemspider.com ~25 million compounds. Linking to >300 data sources Underpinning the semantic web.patents and publications, chemical suppliers etc. host for crowdsourced data Focus on data curation quality Used multiple databases to validate structures [email_address] Data curation

Bayesian machine learning Laplacian-corrected Bayesian classifier models were generated using Discovery Studio (version 2.5.5; Accelrys). Training set = 295, test set = 237 compounds Uses two-dimensional descriptors to distinguish between compounds that are DILI-positive and those that are DILI-negative ALogP ECFC_6 Apol logD molecular weight number of aromatic rings number of hydrogen bond acceptors number of hydrogen bond donors number of rings number of rotatable bonds molecular polar surface area molecular surface area Wiener and Zagreb indices Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010 Extended connectivity fingerprints

Bayesian machine learning Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010 Bayesian classification is a simple probabilistic classification model. It is based on Bayes’ theorem h is the hypothesis or model d is the observed data p ( h ) is the prior belief (probability of hypothesis h before observing any data) p ( d ) is the data evidence (marginal probability of the data) p ( d|h ) is the likelihood (probability of data d if hypothesis h is true) p ( h|d ) is the posterior probability (probability of hypothesis h being true given the observed data d ) A weight is calculated for each feature using a Laplacian-adjusted probability estimate to account for the different sampling frequencies of different features. The weights are summed to provide a probability estimate

Examples of using Bayesian Models Integrated in Silico-in Vitro Strategy for Addressing Cytochrome P450 3A4 Time-Dependent Inhibition Zientek et al., Chem Res Toxicol 23: 664-676 (2010) Challenges predicting ligand-receptor interactions of promiscuous proteins: the nuclear receptor PXR Ekins S, Kortagere S, Iyer M, Reschly EJ, Lill MA, Redinbo MR and Krasowski MD, PLoS Comput Biol 5(12): e1000594, (2009) . Computational models for drug inhibition of the human apical sodium-dependent bile acid transporter Zheng X, et al., Mol Pharm, 6: 1591-1603, (2009) Quantitative structure activity relationship for inhibition of human organic cation/carnitine transporter Diao et al., Mol Pharm, 7: 2120-2131, (2010)

Features in DILI + Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010 Avoid Long aliphatic chains Phenols Ketones Diols  -methyl styrene Conjugated structures Cyclohexenones Amides ?

Features in DILI - Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010

Results Fingerprints with high Bayesian scores that are present in many DILI compounds appeared to be reactive in nature, Could cause time-dependent inhibition of cytochromes P450 or be precursors for metabolites that are reactive and may covalently bind to proteins. Why are long aliphatic chains important for DILI generally hydrophobic and perhaps enabling increased accumulation? may be hydroxylated and then form other metabolites that are in turn reactive? Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010

Test set analysis compounds of most interest well known hepatotoxic drugs (U.S. Food and Drug Administration Guidance for Industry “Drug-Induced Liver Injury: Premarketing Clinical Evaluation,” 2009), plus their less hepatotoxic comparators, if clinically available. Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010

Training vs test set PCA Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010 Yellow = test Blue = training retinyl palmitate

Compare to newer drugs Extracted small molecule drugs from 2006 to 2010 from the Prous Integrity database Structure validation resulted in a set of 77 molecules (mean molecular weight 427.05 ± 280.31, range 94.11–1994.09) These molecules were distributed throughout the combined training and test sets (N = 532), representative of overlap These combined analyses suggest that the test and training sets used for the DILI model are representative of current medicinal chemistry efforts. Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010

Fingolimod (Gilenya) for MS (EMEA and FDA) Paliperidone for schizophrenia Pirfenidone for Idiopathic pulmonary fibrosis Roflumilast for pulmonary disease Predictions for newly approved EMEA compounds Can we get DILI experimental data for these?

SMARTS FIlters Smartsfilter kindly provided by Dr. Jeremy Yang (University of New Mexico, Albuquerque, NM, http://pasilla.health.unm.edu/tomcat/biocomp/smartsfilter). Substructure Alerts used to filter libraries – remove reactive groups etc. Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010

SMARTS Filters vs Rule of 5 Ekins and Freundlich, Pharm Res, In press 2011 Correlation between the number of SMARTS filter failures and the number of Lipinski violations for different types of rules sets with FDA drug set from CDD (N = 2804) Suggests # of Lipinski violations may also be an indicator of undesirable chemical features that result in reactivity

Open source tools for modeling Open source descriptors CDK and C5.0 algorithm ~60,000 molecules with P-gp efflux data from Pfizer MDR <2.5 (low risk) (N = 14,175) MDR > 2.5 (high risk) (N = 10,820) Test set MDR <2.5 (N = 10,441) > 2.5 (N = 7972) Could facilitate model sharing? DILI dataset - another example Gupta RR, et al., Drug Metab Dispos, 38: 2083-2090, 2010 $ $$$$$$

Open source software for molecular descriptors and algorithms Spend only a fraction of the money on QSAR Selectively share your models with collaborators and control access Have someone else host the models / predictions The next opportunities for crowdsourcing… Models Inside company Collaborators Current investments >$1M/yr >$10-100’s M/yr Databases, servers Commercial Descriptors Algorithms In house data generation Data

Pfizer Merck GSK Novartis Lilly BMS Could combining models give greater coverage of ADME/ Tox chemistry space and improve predictions? Lundbeck Allergan Bayer AZ Roche BI Merk KGaA Expanding computational model coverage of chemical space

More collaborations, integrating models into scientific social networks Drug Disc Today, 14: 261-270, 2009

Conclusions First large-scale testing of DILI machine learning models Concordance lower than with in vitro model Statistics similar to Structural alerts from Pfizer paper SMARTS can be used to filter libraries Machine learning models better than SMARTS Could use models to filter compounds for further testing in vitro Use published knowledge to predict DILI Combinations of models Combine datasets – create models with Open descriptors and algorithms – equivalent to commercial?? Make models widely available Collaborations to cover wider chemical space Ekins, Williams and Xu, Drug Metab Dispos 38: 2302-2308, 2010

Acknowledgments Merck Jim Xu Royal Society of Chemistry Antony J. Williams Pfizer Rishi Gupta Eric Gifford Ted Liston Chris Waller Accelrys CDD [email_address] http://www.slideshare.net/ekinssean http://www.collabchem.com/ http://www.collaborations.com/CHEMISTRY.HTM

BioIT Drug induced liver injury talk 2011

More Related Content

What's hot (20)

Viewers also liked (18)

Similar to BioIT Drug induced liver injury talk 2011 (20)

More from Sean Ekins (20)

Recently uploaded (20)

BioIT Drug induced liver injury talk 2011

Editor's Notes