STSC-GNN: A sequence-derived topology-aware graph neural network for solvent-conditioned prediction of peptide structural composition.

Peptide structural composition is strongly influenced by the surrounding solvent environment, making environment-dependent structural prediction important for understanding peptide conformational behavior and structural plasticity. However, many existing computational approaches primarily rely on sequence representations or static structural information and do not explicitly integrate sequence-derived topology with physicochemical properties of the solvent. To address this gap, we developed STSC-GNN, a sequence-derived topology-aware graph neural network for solvent-conditioned prediction of peptide structural composition. The framework represents each peptide as a sequence-derived residue graph incorporating sequential, local 2-hop, and long-range biochemical-similarity relationships, while combining amino-acid and biochemical features with contextual sequence representations. Explicit graph-level topology descriptors and continuous solvent physicochemical features are integrated through bidirectional peptide-solvent cross-attention and solvent-conditioned feature gating, followed by joint prediction of α-helical, β-like, and unstructured structural fractions. Three graph neural network backbones, GCN, GAT, and GIN, were evaluated on the Test set using MAE and Accuracy at a tolerance of 0.10, with performance assessed across five independent runs with different random seeds. GIN achieved the lowest Test MAE and highest Test Accuracy among the evaluated STSC-GNN variants on the independent Test set, with an MAE of 0.0977 ± 0.0005 and an Accuracy of 0.6567 ± 0.0022. Component-wise analysis showed that ESM-2, biochemical-similarity relationships, explicit topology descriptors, peptide-solvent cross-attention, and solvent-conditioned gating provide complementary contributions to prediction performance. STSC-GNN provides a unified framework for integrating peptide sequence, sequence-derived topology, and solvent information, offering a computational approach for environment-dependent peptide structural prediction and the investigation of peptide structural plasticity.