Skip to content

Integrating multi-modal language models with graph structural learning for bioactive peptide prediction.

Nov 2026 · Food Chemistry · Vol 533 Pt 2, pp. 151971 · 0 citations · 58 references
Medicine

Abstract

Bioactive peptides are important targets in functional food development, biomedicine and molecular design because of their potential antimicrobial, anticancer, antihypertensive, antioxidant and immunomodulatory activities. Accurate prediction of peptide function is essential for rapid candidate prioritization and for understanding sequence-structure-function relationships, yet many protein language model-based methods treat pretrained embeddings as fixed inputs and do not explicitly couple complementary residue states with adaptive residue-relation learning. Here, we present MFGCN-Pep, a multimodal graph learning framework that integrates pretrained protein language models with adaptive graph learning. From a unified amino acid sequence input, the model integrates ProtT5, ESM-2 and AAindex representations into a task-adaptive weighted sequence graph, and combines graph convolution, convolutional refinement and multi-head attention to capture non-local residue dependencies, local sequence patterns and sequence-level evidence. Across seven bioactive peptide tasks, MFGCN-Pep achieves competitive and generally balanced performance. Ablation analyses support the complementary contributions of multimodal representation and adaptive relational learning, while residue-level attribution and structural mapping provide interpretable evidence for model predictions.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.