AI-Driven Genome Annotation Systems

Abstract

Genome annotation -- the identification of genes, regulatory elements, and functional features within raw DNA sequences-- is the foundational step linking genomic data to biological knowledge. With over 30,000 species genomes now sequenced, annotation has become the primary bottleneck in comparative and functional genomics. Current annotation pipelines (MAKER, Prokka, NCBI PGAP) combine ab initio gene prediction, homology search, and transcript evidence, achieving 85-95% gene sensitivity for well-studied organisms but dropping to 60-75% for non-model species lacking reference transcriptomes. Deep learning models trained on evolutionary sequence patterns promise to improve annotation accuracy, particularly for non-coding functional elements (enhancers, promoters, lncRNAs) that homology-based methods cannot detect. This study developed and benchmarked six genome annotation approaches - ab initio gene prediction (Augustus), homology-based pipeline (MAKER2), hybrid evidence-based (NCBI PGAP), genomic language model (DNABERT-2 fine-tuned), sequence-to-annotation transformer (Enformer-derived), and a proposed Foundation Model for Genome Annotation (FMGA) combining a 1.2-billion-parameter DNA language model pre-trained on 3,200 genomes with multi-task fine-tuning for simultaneous gene structure prediction, regulatory element detection, and functional annotation -- across 12 benchmark genomes spanning prokaryotes, fungi, plants, invertebrates, and vertebrates. FMGA achieved the highest mean gene-level F1 (0.924) across all genomes, significantly outperforming MAKER2 (0.862; p < 0.001), PGAP (0.886; p < 0.001), and DNABERT-2 (0.892; p = 0.002). For non-model organisms (species without reference transcriptomes), FMGA's advantage was largest: F1 0.894 versus MAKER2 0.742 (p < 0.001), demonstrating that evolutionary patterns learned from diverse genomes transfer effectively to unannotated species. FMGA additionally detected regulatory elements (enhancers, promoters) with F1 0.782 -- a capability absent from traditional annotation pipelines. These results establish foundation models as the next generation of genome annotation technology.