Artificial Intelligence-Based Scoring Functions for Polypharmacology Prediction – POLYPHAI

PubChem and ChEMBL are well-known public repositories which provide valuable experimental data for predicting a molecule’s potential in polypharmacology, a concept reflecting its ability to simultaneously interact with two or more targets. These databases have been leveraged by artificial intelligence (AI), applying state-of-the-art machine learning (ML) techniques, to build target-specific scoring functions (SFs) that predict possible binders of protein targets. However, the development of such effective SFs is still limited, due to the time and computing resource required for data retrieval/curation and for training predictive models tailored to large data sets. In this project, we aim, first, to exhaustively retrieve and process experimental data of all PubChem/ChEMBL proteins present in these public repositories. We will then assess diverse modeling approaches incorporating different structural encoding schemes, ML algorithms (including deep learning) and model optimization strategies for developing AI/ML models, both binary classifiers and regression models, specific for each available protein target. This project is expected to help us accurately predict not only molecules which may exhibit a certain bioactivity toward a target, but also those that may have polypharmacological profiles, in order to anticipate their therapeutic effects and/or possible adverse reactions. We will also offer a user-friendly web service where our SFs are integrated for use by other researchers, and validate, in prospective settings, the predictions made by our models on two kinase proteins through in vitro assays.