BioGeoFormer: A deep learning approach to classify unknown genes associated with critical biogeochemical cycles

Remote functional annotation continues to impede progress in microbial ecology, as alignment-based approaches still leave over one-third of microbial sequences functionally unresolved. In contrast, pre-trained natural-language–processing approaches have shown strong potential for inferring functions from diverse biological sequences, and here we introduce a protein language modeling approach allowing us to classify sequences into 37 defined key pathway categories involved in 4 major biogeochemical cycles (methane, sulfur, nitrogen and phosphorus cycles). To do so, we fine-tuned ESM2-8m using databases curated for biogeochemical cycling pathways. Our resultant BioGeochemical cycling transFormer (BioGeoFormer or BGF) was high-performing on validation and test sets, producing embeddings that exhibit an ability to infer protein function at a metabolic pathway level. BGF was applied to a dataset of metagenome-assembled genomes (MAGs) constructed from methane-fueled, deep-sea, “cold seep” environments to demonstrate its utility in contrast to current informatics approaches. We employed multiple gene assignments to identify gene function within these MAGs. A total of 1.05M genes were assigned biogeochemical functions, with BGF alone suggesting putative ecosystem roles for 0.49M (46%) of these at a confidence of 85% or greater; these genes were classified as unknown by the other approaches. Across the pathways of interest, BGF identified 6 times as many genes, on average, as Hidden Markov models (HMMs) as well as alignment based approaches across the various pathways. BGF provides a novel tool that is capable of informing process-based hypotheses in diverse systems, highlighting cryptic proteins most notably linked to methane, nitrogen, and phosphorus cycling while uncovering the mysteries within microbial dark matter.