Discovery of antimicrobial peptides targeting Acinetobacter baumannii via a pre-trained and fine-tuned few-shot learning-based pipeline

Abstract

Abstract Acinetobacter baumannii, a robust Gram-negative bacterium known for inducing nosocomial infections and displaying multidrug resistance, remains a formidable challenge to combat. The limited arsenal of antimicrobial peptides targeting this pathogen underscores the need for innovative strategies. Here, we report a pioneering few-shot learning-based pipeline designed to identify potent antimicrobial peptides targeting A. baumannii. This pipeline effectively scans through the entire libraries of hexapeptides, heptapeptides and octapeptides, encompassing tens of billions of candidates, despite the extreme scarcity of available training data (148 sequences). Comprising classification, ranking, and regression modules as an integration, each module is trained using a few-shot learning strategy involving pre-training and multiple fine-tuning steps while incorporating both similar and true data fine-tuning. This methodology mitigates the potential overfitting concerns, due to the small size of the training samples, then enhances the predictive capability of the pipeline. The leading peptides predicted showcase robust activity against multiple A. baumannii strains, while demonstrating low off-target toxicity and negligible susceptibility to drug resistance. Additionally, the EME7(7) exhibits efficacy in controlling A. baumannii infections within a mouse pneumonia model, notably without inducing kidney injury—a contrast to the observed effects of polymyxin B. This work provides a paradigm for addressing the challenges posed by limited data availability.