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

Abstract Understanding microbial phenotypes from genomic data is crucial in areas of research including co-evolution, ecology and pathology. This study proposes a new approach to integrate literature-derived information with genomic data to study microbial traits, combining natural language processing (NLP) with functional genome analysis. We applied this methodology to publicly available data to overcome current limitations and provide novel insights into microbial phenotype prediction. We fine-tuned specialized transformer-based large language models to analyze 3.3 million open-access scientific articles, extracting a network of phenotypic information linked to bacterial strains. The network maps relationships between bacterial strains and traits such as pathogenicity, metabolic capacity, and host and biome preference. By functionally annotating reference genome assemblies for strains in the phenotypic network, we were able to predict key genes influencing phenotypes. Our findings align with known phenotypes and reveal novel correlations, leading to the identification of microbial genes relevant in particular disease and host-association phenotypes. The interconnectivity of strains within the network provided further understanding of microbial community interactions, leading to the identification of hub species by inferring trophic connections—insights challenging to extract by means of experimental work. This study demonstrates the potential of machine learning methods to uncover cross-species patterns in microbial gene-phenotype correlations. As the number of sequenced strains and literature descriptions grows exponentially, such methods become crucial for extracting meaningful information and advancing microbiology research. Competing Interest Statement The authors have declared no competing interest. Footnotes Data Availability: The full pipeline to reproduce all analyses and train all models in this manuscript is found in GitHub. Competing Interests The authors declare no competing interests. Acronyms - NLP - Natural Language Processing - LLMs - Large Language Models - RE - Relation Extraction - NER - Named Entity Recognition - PMC - PubMed Central - BERT - Bidirectional Encoder Representations from Transformers