Multimodal molecular features enable interpretable umami peptide screening and structure-activity analysis.

Umami peptides are important taste-active components in protein-rich foods, but their computational discovery is still constrained by predictors that rely mainly on sequence information and provide limited mechanistic interpretability. In this study, BioPP-GFD was developed as an interpretable multimodal deep learning framework for umami peptide screening and structure-activity analysis. The model integrates molecular graphs, molecular fingerprints, and physicochemical descriptors through adaptive channel attention. In the umami-bitter peptide classification task, BioPP-GFD achieved an accuracy of 0.932 ± 0.028, sensitivity of 0.919 ± 0.023, specificity of 0.944 ± 0.041, and Matthews correlation coefficient of 0.864 ± 0.056, outperforming Umami-Transformer and UniDL4BioPep in most overall metrics. Supplementary evaluation across additional bioactivity-specific peptide datasets further indicated the cross-task applicability of the multimodal architecture to other peptide classification tasks. Interpretability analysis showed that the fingerprint and descriptor branches contributed strongly to the discrimination of umami peptides from bitter peptides. Directional occlusion and fingerprint evidence mapping distinguished umami-supporting substructures from bitter-associated cues. Atom-pair attention analysis of strongly active umami peptides, with activity thresholds ranging from 0.003 to 0.1998 mM, localized prediction-relevant regions to carboxylate-bearing termini, amide-linked segments, and other heteroatom-rich environments. Molecular docking with the T1R1 and T1R3 receptor model further supported the receptor-level relevance of these hotspots, with docking energies ranging from -6.846 to -8.325 kcal/mol and recurrent hydrogen-bonding, electrostatic, and hydrophobic contacts. These findings suggest that BioPP-GFD provides an interpretable strategy for umami peptide screening and receptor-level structure-activity analysis, supporting the rational discovery of taste-active functional food peptides.