SE(3)-PROTACs: Geometric deep learning for PROTAC degradation prediction

Abstract Proteolysis-targeting chimeras (PROTACs) are a valuable therapeutic method for degrading target proteins of interest. Their success depends on forming a stable ternary complex composed of PROTAC, the target protein, and the E3 ligase—a relationship that is inherently difficult to predict. Existing deep learning methods often fail to accurately model the 3D arrangement of this ternary structure or overlook the critical background information provided by sequences, which is essential for accurate degradation predictions. Herein, we introduce SE(3)-PROTACs, a PROTAC degradation prediction model designed specifically using a geometric deep learning architecture. An SE(3)-equivariant transformer encodes PROTAC substructures—the warhead, linker, and E3 ligand—as molecular graphs invariant to translation and rotation while retaining geometrical features relevant to chemistry. Pretrained evolutionary scale modeling (ESM) embeddings provide functional and structural context to target proteins and E3 ligases directly from sequences, thus bypassing the limitations inherent in incomplete structural data. A Pairwise Interaction Mechanism computes all-pairs compatibility scores between target and E3 ligase residue positions conditioned on the PROTAC scaffold, using mean-pooled sigmoid weighting to reweight each protein representation before classification. A benchmark dataset of 1979 curated PROTAC samples sourced from PROTAC-DB was used for model development and evaluation. SE(3)-PROTACs achieved a test accuracy of 80.81% on a held-out random-split test set, with consistent performance across the cluster-split (65.62%) and temporal-split (64.08%) evaluations. SE(3)-PROTACs outperform baseline models across random, cluster-based, and temporal evaluations, demonstrating strong generalization to new targets and compounds, and serving as a reliable computational pre-filter for prioritizing candidate degraders before experimental validation.