Target-Aware Drug Activity Model: A Deep Learning Approach to Virtual HTS

Virtual screening is a common approach in computer aided drug discovery (CADD) to identify potentially active compounds in large molecular libraries. These libraries have grown over the years and currently count several billions commercially available compounds. This raises the need for high-throughput virtual screening approaches that can handle these sizes in a reasonable amount of time. In this paper we introduce our Target-Aware Drug Activity Model (TADAM), able to predict the activity of a compound towards any target protein pocket on a very large scale. TADAM focuses on the pocket residues involved in ligand interaction, remaining lightweight compared to other deep learning alternatives. The model performs better than standard computational chemistry techniques and contemporary AI models, including those that reached state-of-the-art results in several benchmarks. The validity of its predictions were further confirmed with prospective in vitro assays.