Abstract
BACKGROUND: Accurate annotation of protein-coding genes is critical for genome analysis in non-model organisms. However, limited RNA-Seq data and incomplete protein resources can lead to errors, including chimeric gene mis-annotations, where two or more adjacent genes are incorrectly fused into a single model. These errors often persist due to annotation inertia, where mistakes are propagated and amplified through data sharing and reanalysis, and leads to cases where the mis-annotated model becomes favoured over the correct model. This complicates almost all downstream genomic analyses such as gene expression studies and comparative genomics. RESULTS: We investigated chimeric mis-annotations across 30 recently annotated genomes spanning invertebrates, vertebrates, and plants, identifying 605 confirmed cases. The majority of these errors occurred in invertebrates and plants. Using structural prediction and splicing assessment, we demonstrated that utilising machine-learning annotation tools (such as Helixer) provides an approach which can identify mis-annotations. CONCLUSIONS: This study highlights the prevalence of chimeric mis-annotations in genomic datasets and showcases the potential of machine-learning tools such as Helixer to refine gene models for highly variable gene families with mis-annotations present in databases. By addressing these annotation errors, we improve genomic data reliability and facilitate a deeper understanding of non-model organisms.