PocketFlow: an autoregressive flow model incorporated with chemical knowledge for generating drug-like molecules inside protein pockets

Abstract

Abstract Identifying an active seed compound against a specific target protein is the first but challenging step for initiating a new drug development project. Newly emerging deep generative models (DGMs) providea rapid strategy to directly generate potential seed compounds inside protein pockets. However, the poor quality of generated molecules remains a major challenge, and whether these DGMs can generate bioactive molecules has not yet been wet-lab verified. We herein propose a new structure-based DGM, PocketFlow, which is an autoregressive flow model with chemical knowledge incorporated in molecular generation. PocketFlow can generate high-quality drug-like molecules with 100% chemical validity. In various evaluations, PocketFlow outperforms the current state-of-the-art DGMs. We applied PocketFlow to two new target proteins, HAT1 and YTHDC1, and successfully obtained wet-lab validated bioactive lead compounds. Overall, PocketFlow is the first wet-lab verified structure-based molecular DGM, highlighting the great potential of structure-based DGMs in drug discovery.