Precision fragment addition: domain-specific DeepFrag2 models for smarter lead optimization

Abstract

This study introduces a series of machine-learning models based on DeepFrag, our previously published tool designed to guide small-molecule lead optimization through fragment addition. We demonstrate enhanced accuracy by training new DeepFrag models to predict optimizing fragments with specific sizes and chemical properties. Additionally, we show that DeepFrag accuracy improves when fine-tuned on specific receptor classes. These targeted models should prove valuable for medicinal chemists with predetermined insights into suitable molecular fragment characteristics (such as preferred size ranges, charge states, or aromaticity) or those conducting optimization campaigns against specific drug-target classes with many known ligands. To encourage adoption, we release DeepFrag2 under the open-source MIT license. Interested users can download DeepFrag2 free of charge without registration from https://durrantlab.com/deepfrag2/.