High-quality gene structure annotation: The last mile of reference genomes

Over the past two decades, rapid advances in omics technologies have generated an unprecedented volume of genomic data. Coupled with substantial improvements in genome assembly algorithms, these advancements have enabled the production of telomere-to-telomere genome assemblies and phased super-pangenomes for an increasing number of species.1 However, achieving high sequence completeness alone does not ensure that a genome assembly represents a robust and reliable reference. Gene structure annotation, which involves the identification of gene-encoding genomic regions and the accurate delineation of their exon-intron organization, remains a major challenge.2 For most species, existing gene structure annotations continue to contain substantial inaccuracies, including missing gene loci, erroneous exon-intron structures, and inaccurately defined exon boundaries. Such inaccuracies can compromise a broad range of downstream analyses, including analyses of gene gain and loss, high-resolution genetic mapping, and gene expression quantification. More importantly, inaccurate gene annotations can mislead gene functional studies, leading to erroneous biological interpretations and unnecessary experimental effort before the errors are recognized. Therefore, establishing a true reference genome requires the integration of a high-quality genome sequence assembly with accurate, comprehensive, and well-supported gene structure annotation. In this context, as genome assembly quality continues to improve, the generation of accurate gene structure annotations has become a major bottleneck.

Why automated gene structure annotation still falls short

A variety of state-of-the-art algorithms and annotation pipelines have been developed for gene structure annotation. Recently developed tools, including BRAKER3, EviAnn, Helixer, and ANNEVO, have achieved substantial advances in gene annotation performance, especially in terms of annotation completeness, as evidenced by improved Benchmarking Universal Single-Copy Orthologs (BUSCO) scores. However, achieving accurate gene structure annotation remains a major unresolved challenge, primarily attributable to three factors.

First, the inherent complexity of eukaryotic transcription and our limited understanding of its regulatory and structural diversity pose major challenges to accurate gene structure annotation. In a simplified view, a eukaryotic gene is typically depicted as a linear arrangement of exons and introns, with exons containing both untranslated regions and coding sequences. In reality, however, genome organization and transcription are substantially more complex.3 In gene-dense genomes, adjacent genes may be separated by only short intergenic regions or overlap at the transcriptional level. Certain protein-coding genes may be completely nested within the introns of other genes. Furthermore, gene structures exhibit substantial variation in scale, with introns in certain genes spanning hundreds of kilobases, whereas exons in others may consist of only a few nucleotides. Such extensive structural diversity is difficult to capture with simplified annotation models, posing a major challenge to accurate gene structure reconstruction and limiting the base-level accuracy of automated gene structure annotation across diverse genomes.

Second, the limitations in supporting evidence pose significant challenges to accurate gene structure annotation. Homology-based protein alignment and RNA-seq read alignment are the two most widely used evidence-based approaches for gene structure annotation. However, the performance of homology-based annotation is affected by evolutionary distance, parameter settings, and the presence of lineage- or species-specific genes in reference datasets. RNA-seq data are also inherently incomplete owing to tissue-, developmental-stage-, and growing-condition-specific gene expression patterns. Consequently, many genes may lack sufficient transcriptomic support within the available RNA-seq datasets. Ab initio gene prediction is therefore indispensable, particularly for the identification of highly divergent or novel orphan genes. Although deep-learning approaches, can improve gene prediction, their performance remains dependent on the availability of high-quality datasets for model training and fine-tuning. At present, large-scale, high-quality gene structure annotation datasets remain scarce.

Third, the optimization objectives of genome-scale data exploration differ fundamentally from those of experimental biology. For many genome-scale and omics analyses, a broadly complete set of gene models is often sufficient for computational analyses. However, functional studies, particularly those involving wet-lab experimentation, require substantially higher annotation accuracy. Gene cloning and functional characterization require highly accurate gene models, including precisely defined coding sequence boundaries. Consequently, annotation algorithms optimized for genome-wide completeness may not achieve the base-level accuracy required for gene-level experimental studies.