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

Traditional molecular property prediction models often rely on a single static conformation to represent a molecule. However, flexible molecules like cyclic peptides exist in solution as a Boltzmann-weighted ensemble of multiple conformations. This paper investigates whether explicitly encoding these conformational ensembles into a single, thermodynamically informed embedding can improve the accuracy of property prediction tasks, such as membrane permeability.

The authors introduce EnsembleEGNN, a foundation model designed to process multiple conformers of the same molecule. The architecture consists of two main components:

1. Conformer-wise Encoding: Each conformer is processed through shared Equivariant Graph Neural Network (EGNN) layers, which maintain rotational and translational equivariance without requiring expensive coordinate alignment.
2. Ensemble Fusion: The resulting representations are aggregated using a Set Attention Block (SAB). This block uses a Boltzmann-informed attention mechanism, allowing the model to prioritize conformers based on both their thermodynamic weight and their relevance to the specific downstream task.

The model was pretrained on the CREMP dataset using a multi-task objective that includes masked token recovery, noisy-coordinate reconstruction, and pairwise distance reconstruction.

The study demonstrates that pretraining is essential for performance; the model fails when trained from scratch but achieves strong results after self-supervised pretraining. On the CREMP-CycPeptMPDB benchmark, the pretrained EnsembleEGNN outperforms a sequence-only BERT baseline. Furthermore, a hybrid model that co-trains the EnsembleEGNN with a BERT sequence encoder achieves the highest predictive performance (R² = 0.538, Pearson r = 0.737), suggesting that geometric ensemble information and sequence data provide complementary signals for cyclic peptide property prediction.