Integrating natural language processing and genome analysis enables accurate bacterial phenotype prediction.

Understanding microbial phenotypes from genomic data is crucial for studying co-evolution, ecology, and pathology. This study presents a scalable approach that integrates literature-extracted information with genomic data, combining natural language processing and functional genome analysis. We applied this method to publicly available data, providing novel insights into predicting microbial phenotypes. We fine-tuned transformer-based language models to analyze 3.83 million open-access scientific articles, extracting a phenotypic network of bacterial strains. This network maps relationships between strains and traits such as pathogenicity, metabolism, and biome preference. By annotating their reference genomes, we predicted key genes influencing these traits. Our findings align with known phenotypes, reveal novel correlations, and uncover genes involved in disease and host associations. The network's interconnectivity provides deeper understanding of microbial communities and allowed identification of hub species through inferred trophic connections that are difficult to infer experimentally. This work demonstrates the potential of machine learning for uncovering cross-species gene-phenotype patterns. As microbial genomic data and literature expand, such methods will be essential for extracting meaningful insights and advancing microbiology research. In summary, this integrative approach can accelerate discovery and understanding in microbial genomics. Ultimately, such techniques will facilitate the study of microbial ecology, co-evolutionary processes, and disease pathogenesis to an unprecedented depth.