PI-024 - PROMISING CAPACITY OF MACHINE LEARNING MODELS TO PREDICT OATP1B1- AND OATP1B3-MEDIATED INHIBITION POTENTIAL IN EARLY DRUG DISCOVERY.

A. Sadighi¹, M. Parvez¹, S. Keller², J. Enoru²; ¹AbbVie, South San Francisco, CA, USA, ²AbbVie, South San Francisco, United States.

Background: Organic anion transporting polypeptide (OATP)-1B1 and -1B3 mediated drug-drug interactions (DDIs) are clinically relevant. Therefore, screening for OATP1B1- and 1B3-mediated DDI potential at early stages is beneficial. The in vitro cell-based OATP1B1 and 1B3 IC50 assay used for estimation of OATP1B1- and 1B3-mediated DDI is laborious, time consuming, and expensive. In contrast, machine learning (ML) based predictions are fast, reliable, and inexpensive and could be an alternative to in vitro assays. In this study, we explored the performance of conventional ML models to predict IC50 values for OATP1B1 and 1B3 transporters.

Methods: The initial ML models were constructed using RDKit physicochemical features for a list of objects and precipitants relevant to OATP1B1 and 1B3 IC50, which were retrieved from the Certara Drug Interaction Database (DIDB®). The regression ML models including decision trees (DTs), k-nearest neighbors (kNN), random forest (RF), support vector regression (SVR), and eXtreme gradient boosting (XGBoosting) were built using Scikit-learn Python library. Data curation was accomplished by removing zero variance features and normalizing the features of all observations before model development followed by hyperparameters optimization.

Results: RF and SVR performed the best among all the investigated models in predicting IC50 values for OATP1B1 and 1B3, respectively. Overall models predicted IC50 values (Figure 1) for 70% and 64% of the compounds falls within ±3-fold vs observed for OATP1B1 and 1B3. For weak inhibitors (IC50 = 10-50 µM), the RF and SVR models showed more sensitivity (±2-fold) for 63% and 60% of the compounds, respectively. Furthermore, 78% of OATP1B1 moderate inhibitors (IC50 = 1-10 µM) and 80% of OATP1B3 weak inhibitors (IC50 = 10-50 µM) were predicted within ±3-fold.

Conclusion: Models suggest that the lipophilicity of the precipitants ranks on the top five important RDKit physicochemical properties relevant to OATP1B1 and 1B3 inhibition. The current ML-based model predicted in vitro experimental IC50 and could be an alternative to in vitro assessments in early DDI screening.