BioGeoFormer is a protein language model that predicts and classifies microbial proteins driving the methane, sulfur, nitrogen, and phosphorus cycles. Built on ESM-2 and fine-tuned on four curated metabolic pathway databases, it extends sequence-to-function inference beyond traditional homology-based tools — recovering remote homologues that classical methods miss.

The problem

Biogeochemical cycles — the microbial transformations of methane, sulfur, nitrogen, and phosphorus — underpin planetary and ecosystem health. Identifying which genes carry out these transformations is central to understanding them, but classical annotation relies on homology to known sequences and struggles when a gene is only distantly related to anything in a reference database.

As metagenomic and genomic datasets grow, a large fraction of candidate genes remain hypothetical or unassigned. A method that can infer function directly from sequence — and generalize to remote homologues — is needed to complement existing mining tools.

The model

BioGeoFormer fine-tunes the ESM-2 transformer on four curated cycle databases — MCycDB, NCycDB, PCycDB, and SCycDB — together spanning 610 unique gene families across 37 metabolic pathways. A calibrated confidence function accompanies each prediction so users can gauge reliability.

"BioGeoFormer" is really a family of 8 fine-tuned models, each defined by a clustered identity split — training, validation, and test sets held dissimilar at 10% intervals from 20% to 90%. The 70% split model is the most effective at precisely identifying remote homologues and is recommended for most use cases.

Cycle databases MCycDB NCycDB PCycDB SCycDB Fine-tune ESM-2 8 identity splits BioGeoFormer 610 gene families 37 pathways
Training overview. ESM-2 is fine-tuned on four curated biogeochemical-cycle databases, yielding a family of models spanning 610 gene families across 37 metabolic pathways.
Why it matters

By inferring function from sequence rather than homology alone, BioGeoFormer surfaces hypothetical gene function related to biogeochemical cycling that classical tools overlook. It is best used as a complementary method alongside traditional approaches for metagenome and genome mining — expanding the set of genes we can confidently connect to the cycles that sustain life on Earth.