Poster #84 - Aditya Birla

PREDICTION OF BINDING AFFINITIES FOR PROTEIN-PEPTIDE SYSTEMS USING SE(3) EQUIVARIANT NEURAL NETWORKS

Aditya Birla (PhD Candidate- Rutgers University), Guillaume Lamoureux (Associate Professor- Rutgers University)

Protein-peptide interactions are fundamental to biology but difficult to predict due to peptide flexibility and limited structural data. We present a deep learning framework based on SE(3)-equivariant convolutional neural networks that estimates binding affinity directly from 3D atomic density representations of receptor-peptide complexes. The model describes interaction energies using correlations of scalar and vector features, and preserves rotational and translational symmetries. Trained on PepBDB structures, it produces energy-based affinity proxies and discriminates between plausible and implausible binding poses. Ongoing work will extend the approach to PDBbind with experimental ΔG values, enabling more accurate and benchmarkable affinity predictions.