FAMUS: A Few-Shot Learning Framework for Large-Scale Protein Annotation

Predicting gene function is a pivotal and challenging step in genomic and metagenomic data analysis. Current automatic annotation tools typically rely on the single most similar sequence from the query database and struggle to robustly set hit thresholds for annotation. The sparsity of proteins per annotation makes it challenging to confidently assign gene function for underrepresented families. Here, we present a contrastive learning framework for functional annotation. FAMUS (Functional Annotation Method Using Supervised contrastive learning) compares query sequences to a full array of profile Hidden Markov Models and transforms the similarity scores into a condensed vector space that minimizes the distance of proteins from the same family. The similarity scores of a query to all profiles are used for its representation instead of considering only the top-ranking hit. Unannotated sequences are incorporated as negative examples during training, enabling robust detection of proteins that fall outside the scope of the reference database without requiring a user-defined threshold. Using this approach, FAMUS outperformed KEGGs native KofamScan for KEGG Orthology annotation and InterPros InterProScan for PANTHER family annotation. We thus created four protein annotation models using protein families from the KEGG Orthology, InterPro family, OrthoDB, and EggNOG databases. All four models are available as a conda package and via our user-friendly web server, allowing users to annotate large-scale datasets. FAMUS is the first comprehensive and modular annotation framework based on contrastive learning. It supports both pre-defined and user-specific databases for tailored annotation, and can be easily integrated into any genomic and metagenomic analysis pipeline to facilitate accurate, large-scale functional annotation.