Interpretable Antimicrobial Peptide Identification using Retrieval-Augmented Generation

Sammanfattning

Introduction: Antimicrobial resistance (AMR) is making conventional antibiotics less effective, driving interest in antimicrobial peptides (AMPs) as alternatives. While protein language models (PLMs) predict AMPs accurately, they rarely explain why a sequence is classified as antimicrobial, limiting their value for research and experimental validation. Research Question: The research question of this study is “How can a retrieval-augmented generation (RAG) framework be used to generate biologically grounded, post-hoc explanations for peptide sequences predicted as antimicrobial or non-antimicrobial by a PLM?” Two sub-questions guide the evaluation are (1) “How well does the fine-tuned PLM classify peptide sequences as AMP or non-AMP?” and (2) “How faithfully do the RAG-generated explanations reflect the retrieved biological evidence?” Method: We fine-tuned a ProtBERT-BFD classifier and integrated a RAG module that retrieves similar sequences from curated sources, passing them to Mistral-7B-Instruct large language model (LLM) to generate biological explanations. We evaluated classifier performance using F1-score and assessed explanation faithfulness using two independent LLM judges. Results: The classifier achieved an F1-score of 0.903 on the held-out test set. The RAG module generated biological explanations achieved mean faithfulness scores of 0.888 (Qwen2.5:7b) and 0.823 (Claude Haiku). Discussion: Overall, pairing PLM-based prediction with knowledge retrieval makes AMP classifications easier to interpret. Future work should explore multiclass AMP prediction and more diverse retrieval datasets to make the approach more robust and generalisable.