Graph Learning on Ensembles of Cyclic Peptides: An Investigation of Molecular Ensemble Modeling

Researchers at the University of Texas at Austin and Novo Nordisk introduce EnsembleEGNN, a molecular ensemble foundation model designed for cyclic peptides whose solution behavior depends on multiple Boltzmann-weighted conformers rather than one static structure. The model applies shared E(n)-Equivariant Graph Neural Network layers to each conformer, then combines representations using Boltzmann-informed attention and a Set Attention Block; experiments retain the top 5 conformers per peptide. Self-supervised pretraining on CREMP jointly performs masked atom-token recovery, noisy-coordinate denoising, and pairwise-distance reconstruction. On the 2979-example CREMP-CycPeptMPDB membrane-permeability benchmark, training EnsembleEGNN from scratch nearly fails, producing R2 = 0.005, while pretrained EnsembleEGNN reaches R2 = 0.477 and Pearson r = 0.699, exceeding the sequence-only BERT baseline at R2 = 0.439. Co-training the geometric encoder with the PeptideCLM-2 sequence encoder yields the strongest Hybrid model, with R2 = 0.538 and Pearson r = 0.737. The 11.3M-parameter Hybrid outperforms the 114M-parameter BERT-style architecture, suggesting that explicit ensemble geometry can improve data-efficient cyclic-peptide property prediction, although the study remains limited by a low-data benchmark and the absence of direct cross-conformer message passing.