DeepDegradome: A structure-aware deep learning framework for PROTAC and ligand generation against protein targets

Abstract

Targeted protein degradation is a promising strategy for drug discovery, but designing effective PROTACs remains challenging, especially for proteins without well-defined binding sites. Current methods rely on modifying linkers between fixed ligands, which limits the diversity and innovation of the overall molecular architecture of PROTAC. Here, we introduce DeepDegradome, an AI-powered method that automates the structure-aware design of both small-molecule ligands and PROTACs. It employs a large fragment library constructed from public databases and applies an in-house docking method (iFitDock) to obtain initial binding fragments. DeepDegradome builds ligands by assembling these fragments based on the shape and physicochemical features of the target protein pocket. It can further construct PROTACs from these generated ligands, eliminating the dependency on predefined warheads or E3 ligands. Compared to other AI models, DeepDegradome produces more valid, drug-like molecules with higher predicted binding affinity. We demonstrate DeepDegradome’s effectiveness by designing and validating multiple potency inhibitors and PROTACs for two protein targets: WDR5 and CDK9. One synthesized compound showed excellent agreement between predicted and actual binding conformation confirmed by X-ray crystallography. By combining ligand and PROTAC design in one system, DeepDegradome offers a scalable and reliable tool for discovering new drugs against protein targets.