DPLG-MFP: Enhanced prediction of multi-functional therapeutic peptides via dual-path learning of label sequences and label graphs.

Multi-functional therapeutic peptides play essential roles in diverse biological processes and hold great promise in drug discovery and biotechnology. Accurate prediction of peptide functions is crucial for understanding their biological activities and facilitating the design of peptide-based therapeutics. To this end, we propose DPLG-MFP, a deep learning framework for predicting multi-functional therapeutic peptides. The framework combines multi-scale sequence feature extraction and recurrent split-attention encoding to capture both local motif patterns and long-range contextual dependencies, while a dual-path label-sequence interaction mechanism is introduced to improve the modeling of correlations among functional labels and their associations with peptide sequences. In addition, a joint optimization strategy is employed to alleviate class imbalance and enhance the consistency between sequence representations and label semantics. Under a unified evaluation protocol on the MFTP benchmark dataset, DPLG-MFP obtained higher observed values than the evaluated comparison models across the five reported metrics. In addition to the original random-split benchmark, homology-controlled evaluations were performed using CD-HIT-based redundancy removal and cluster-aware data partitioning. Further ablation, visualization, and case-study analyses show that the proposed framework can more effectively capture discriminative sequence patterns and model label dependencies, providing more reliable predictions for complex multi-label peptides. In addition, residue attribution, masking, and physicochemical analyses provide se quence-level insights into the features underlying the model's predictions. Together, these results provide empirical support for DPLG-MFP under the evaluated settings, while the residue-level analyses offer sequence-level insights into its predictions. The source code and datasets are publicly available at: https://github.com/yaoli-code/DPLG-MFP.