Species-specific antimicrobial activity prediction with biological large language model-based methods

The rise of antimicrobial resistance, driven by the overuse and misuse of antimicrobial agents, has necessitated the search for alternative therapeutic candidates. Antimicrobial peptides (AMPs) have emerged as a promising alternative, produced by every organism across the tree of life and capable of targeting multiple organisms—including multidrug-resistant pathogens—through diverse modes of action. However, identifying AMPs and their specific targets requires labour-intensive, time-consuming, and costly wet-lab experiments. Consequently, in silico methods capable of narrowing down candidate peptides and pathogens are of significant value. Current computational methods are often limited by a lack of pathogen specificity or the exclusion of crucial proteomic and genomic contexts. In this study, we examine the potential of Biological Large Language Models (BLLMs) for species-specific peptide bioactivity prediction. Through overrepresentation analysis of BLLM embeddings, we confirmed that these models capture species-level antimicrobial activity in in-house validated peptides, as well as virulence profiles within an Escherichia coli cohort. Building on this, we developed a heterogeneous graph attention network-based link predictor. This method constructs a knowledge graph representing peptides and pathogens as distinct nodes, with similarity and bioactivity as edges. Information is extracted via semantic traversals of node types, prioritizing the importance of neighbouring nodes and their biological context. We benchmarked this approach against Support Vector Machines, Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost), including ensemble combinations, using both publicly mined and in-house validated databases. The ensemble of XGBoost and MLP emerged as the most robust method, achieving statistically superior or equivalent performance in 10 out of 16 comparisons, including statistical equivalence to the best method in Macro F1-Score on both datasets. Furthermore, this ensemble significantly outperformed or matched state-of-the-art methods across most metrics, with the exception of Precision. Overall, this work highlights the capability of BLLMs to capture species-specific bioactivity patterns. Code for this study is publicly available at https://github.com/BirolLab/triAMPh.