Multi-view framework for multi-label bioactive peptide classification based on multi-modal representation learning

Abstract

The diversity and specific biological functions of bioactive peptides make them key regulators in various physiological processes and crucial contributors to the development of new anti-infective drugs. Although existing graph-based deep learning methods effectively model multi-label peptide representation, they often fail to incorporate multi-modal feature representation and extract multi-scale features from various views. To address these limitations, we present a multi-view framework for multi-label bioactive peptide classification based on multi-modal representation Learning by combining amino acid sequences and fusion molecular fingerprints. The peptide molecular graph is constructed by extracting the topological information and node features, respectively. Multi-view branches are designed by developing sequence-based and graph-based models to leverage their distinct and complementary strengths. Specifically, the protein language model ESM-2 is utilized to extract residue features from amino acid sequences deeply. Meanwhile, local features from molecular fingerprints are learned through a Filter Response Normalization layer and a Thresholded Linear Unit. At the same time, the Mamba module is innovatively employed to extract long-range dependencies and reduce time complexity. Our model demonstrates significantly enhanced and robust performance in multi-label bioactive peptide prediction tasks compared with state-of-the-art models, achieving 82.5% coverage, 80.9% precision and 80.3% accuracy on the MFBP dataset. Furthermore, visual analyses demonstrate that the model can effectively capture features from multiple views and highlight the interpretability of the model through the decision process.