While the predictive performance of supervised learning models has improved over the years, the interpretability of modern machine learning algorithms remains poorly understood, especially in computational chemistry. Access to this interpretation can be crucial in characterizing roles of various molecular features or descriptors that determine structure-activity relationships (SAR)—a critical aspect of the drug discovery process. In this work, we obtain interpretability of model predictions by primarily using a local interpretation method called SHAP (SHapley Additive exPlanations). This method is model agnostic, is rooted in solid mathematical foundations, and the explanation of a prediction is given by a set of values that represent the fair contribution of each feature to the prediction. We first demonstrate the utility of calculating SHAP values of a single molecule and a dataset for a machine learning model using both Morgan Fingerprints and Molecular Descriptors. Next, we discuss examples that illustrate the advantages of our implementation of the SHAP method with the current state-of-the-art chemoinformatics interpretability method, Similarity Maps, implemented in RDKit. Finally, we discuss how SHAP values can help chemists (i) acquire knowledge on the target or chemical properties from the model with no prior knowledge; (ii) check the model quality when the expert knowledge of the target exists; (iii) debug the model to ensure it is capturing known SAR, and if it is biased due to any spurious correlations in the data; and (iv) offer suggestions to improve the featurization and/or the model. This method captures valuable information on compound properties and features, and can be implemented on compound libraries. It allows for more efficient decision making in chemical design, and can help accelerate and optimize the drug discovery process. Future work involves integration of this method with Makya, our tool for AI generative modeling for de novo drug design.

Interpreting quantitative structure-activity relationship models to guide drug discovery

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.