Biological foundation models illuminate annotation blind spots in evolutionarily divergent genomes

Area of Science:

• Genomics
• Bioinformatics
• Evolutionary Biology

Background:

• Functional annotation of genomes from non-model organisms is challenging due to sequence divergence.
• Protein structural similarity often persists despite low amino acid sequence identity, offering a potential avenue for annotation.
• Existing annotation methods struggle with evolutionarily distant species where sequence homology is reduced.

Purpose of the Study:

• To develop and validate a hybrid annotation framework that integrates ab initio gene prediction with structure-based similarity mapping.
• To improve gene model discovery and functional annotation for evolutionarily divergent genomes, using the sea lamprey as a model.
• To enhance the utility of genomic data for downstream analyses like comparative genomics and single-cell RNA sequencing.

Main Methods:

• A hybrid framework combining Evo2 for ab initio exon-intron structure prediction and ESM-2 for protein-embedding-based structural similarity mapping.
• Application to the sea lamprey genome to identify cross-species similarity assignments and expand the structural gene catalog.
• Joint alignment-structure classification to identify structurally supported loci missed by traditional sequence alignment methods.
• Integration into a single-cell RNA sequencing workflow using a UTR-aware reference for improved gene recovery.

Main Results:

• The framework assigned high- or medium-confidence cross-species similarity to 73,485 Evo2-derived protein models in the sea lamprey.
• The deduplicated structural catalog was expanded to 31,286 loci, adding 20,871 loci absent from the Ensembl baseline.
• A joint alignment-structure classification identified 21,391 structurally supported loci, including many with no detectable human protein-sequence match.
• A UTR-aware Ensembl+Evo2 reference improved gene recovery and annotation resolution in sea lamprey immune cell states.

Conclusions:

• Structural protein signals can be detected beyond the limits of pairwise sequence alignment in evolutionarily divergent genomes.
• An embedding-based annotation layer effectively extends the detectable structural signal, improving comparative and single-cell analyses.
• The hybrid framework significantly enhances the functional annotation of non-model organisms, enabling more detailed biological insights.