LLM-Assessed Relatedness of Microbiome Study Descriptions Aligns more Strongly with Functional than with Taxonomic Profile Similarity.

Microbiome studies reveal the taxonomic and functional composition of microbial communities inhabiting many diverse environments. Comprehensive microbiome repositories, such as MGnify, organize data into studies, each consisting of multiple sequencing runs or assemblies and accompanying metadata. This structure enables integrative, large-scale, cross-study analyses, leading to broader insights across ecosystems, hosts, and experimental contexts. Despite extensive microbiome research, methods for defining similarity between studies and validating those similarity metrics, remain insufficiently established, especially for large-scale analyses. To address this, we evaluate whether taxonomic and functional similarities from MGnify can serve as reliable indicators of study relatedness between study pairs, testing multiple metrics against conceptual relatedness (e.g., shared environments, goals, or methods). To scale validation, we introduce a framework that applies a Large Language Model (LLM) to study descriptions, categorizing study pairs by relatedness. Our results show that functional similarity correlates more strongly with LLM-inferred study relatedness than taxonomic similarity, highlighting both the promise and limitations of current metrics. Via the above, we demonstrate the value of combining microbial profiles with LLM-driven semantic reasoning to navigate the expanding landscape of metagenomic research.