CLABP: a contrastive learning framework integrating protein language models and structural information for antibacterial peptide prediction.

Antibacterial peptides (ABPs) represent a promising alternative strategy for combating antimicrobial resistance. Graph neural networks and their variants have shown substantial potential in ABP identification. However, most existing graph-based methods focus on node features while neglecting edge features. Edge features play an important role in characterizing peptide conformation and spatial stability, thereby directly influencing interactions with bacterial membranes and antibacterial activity. Moreover, node and edge features extracted by heterogeneous encoders reside in disparate feature spaces, making their effective integration nontrivial. These challenges highlight the need for an explicit alignment mechanism to bridge heterogeneous representations. In this study, we propose CLABP, a contrastive learning-based framework for ABP identification that integrates edge features. Specifically, backbone dihedral angles, ProtT5 embeddings, and Define Secondary Structure of Proteins-derived secondary structures were employed as node features, while inter-residue distance, motion vectors, and rotation quaternions were used to construct edge features. Node and edge representations were independently extracted using dedicated encoders and subsequently aligned into a shared latent space via contrastive learning, which minimizes the distance between paired representations derived from the same peptide. The aligned features were then fused through a dual cross-attention mechanism for downstream prediction. CLABP outperforms other state-of-the-art methods, achieving an accuracy of 92.6% and a Matthews correlation coefficient of 0.853. Ablation studies further confirm that both edge features and the alignment mechanism are critical to model performance. The datasets and code for the CLABP model are available at https://github.com/IBelieve1234/CLABP.