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

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Antimicrobial resistance (AMR) necessitates novel therapeutic strategies.
• Antibacterial peptides (ABPs) show promise against resistant pathogens.
• Existing graph neural network methods for ABP identification often overlook critical edge features.

Purpose of the Study:

• To develop an advanced framework for ABP identification that incorporates edge features.
• To address the challenge of integrating heterogeneous node and edge representations in graph-based models.
• To improve the accuracy and effectiveness of computational ABP discovery.

Main Methods:

• Proposed CLABP, a contrastive learning-based framework for ABP identification.
• Utilized backbone dihedral angles, ProtT5 embeddings, and secondary structures as node features.
• Employed inter-residue distance, motion vectors, and rotation quaternions as edge features.
• Aligned node and edge representations into a shared latent space using contrastive learning.
• Fused aligned features via a dual cross-attention mechanism for prediction.

Main Results:

• CLABP achieved high performance in ABP identification, with 92.6% accuracy and a Matthews correlation coefficient of 0.853.
• Ablation studies validated the critical importance of both edge features and the alignment mechanism.
• The proposed method outperforms existing state-of-the-art approaches.

Conclusions:

• Integrating edge features and employing a contrastive learning-based alignment mechanism significantly enhances ABP identification.
• CLABP offers a powerful new tool for discovering novel antibacterial peptides.
• The framework provides a valuable contribution to combating antimicrobial resistance.