Accurate, comprehensive gene annotation and ortholog identification across thousands of vertebrate genomes with TOGA2

Inferring orthologs and annotating coding genes remain central challenges in genomics, evident by the growing gap between assembled and annotated genomes. TOGA (Tool to infer Orthologs from Genome Alignments) addresses this challenge by integrating gene annotation and orthology inference. Here, we present TOGA2, the next generation of TOGA, which substantially improves annotation completeness, accuracy, scalability, and orthology inference. TOGA2 leverages exon-level orthology and introduces an exon-wise annotation procedure that reduces memory usage 513-fold and runtime 6.1-fold. We show that human-trained deep learning models for splice site prediction generalize across vertebrates. Integrating these predictions enables robust handling of evolutionary changes in exon-intron structure, including splice site shifts, intron deletions, and exonization of introns. A new gene tree reconciliation step refines orthology inference, and UTR annotation improves gene model completeness. Across mammals, birds, turtles, and percomorph fishes, TOGA2 annotations generally achieve higher gene completeness than transcriptome-informed RefSeq annotations. TOGA2 identifies previously unannotated exons in mouse, assigns informative gene symbols, and annotates V(D)J segments of antigen receptors. TOGA2 scales to thousands of genomes, which we demonstrate by generating comprehensive comparative genomics resources for 2,162 vertebrate assemblies, including gene annotations, ortholog sets, gene losses and duplications, retrogene candidates, and outputs supporting downstream analyses. Together, TOGA2 provides a scalable and versatile framework for comparative genomics that bridges the genome annotation gap.