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

Area of Science:

• Biochemistry
• Computational Chemistry
• Drug Discovery

Background:

• Proteolysis-targeting chimeras (PROTACs) offer a novel therapeutic strategy for targeted protein degradation.
• Predicting the formation of the ternary complex (PROTAC, target protein, E3 ligase) is crucial but challenging.
• Current deep learning models struggle with accurate 3D structural modeling and leveraging sequence information for PROTAC efficacy prediction.

Purpose of the Study:

• To develop an advanced deep learning model, SE(3)-PROTACs, for accurate prediction of PROTAC-mediated protein degradation.
• To address the limitations of existing methods in modeling ternary complex 3D arrangements and utilizing sequence-derived context.

Main Methods:

• Utilized a geometric deep learning architecture with an SE(3)-equivariant transformer to encode PROTAC components as molecular graphs.
• Incorporated pretrained Evolutionary Scale Modeling (ESM) embeddings to provide sequence-based functional and structural context for proteins.
• Developed a Pairwise Interaction Mechanism to compute residue-level compatibility scores, conditioned on the PROTAC scaffold.

Main Results:

• Achieved 80.81% accuracy on a held-out random-split test set using a benchmark dataset of 1979 PROTAC samples.
• Demonstrated robust performance across cluster-split (65.62%) and temporal-split (64.08%) evaluations, indicating strong generalization.
• Outperformed baseline models in various evaluation scenarios, highlighting its reliability for new targets and compounds.

Conclusions:

• SE(3)-PROTACs provides a reliable computational tool for predicting PROTAC degradation efficiency.
• The model serves as an effective pre-filter for prioritizing potential PROTAC drug candidates before experimental validation.
• This approach enhances the efficiency and success rate of PROTAC-based drug discovery efforts.