By 2028, many genome teams may use EviAnn as the "annotate this new creature before lunch" button, which is slightly rude to lunch but great for biology.

Sequencing a genome is no longer the sci-fi part. You can read an organism's DNA and end up with a gigantic string of A, C, G, and T. The problem comes next: finding the genes, where they start and stop, which exons survive the editing room, which introns get cut like a deleted scene from Dune, and which transcripts never become proteins but still do useful work.

That process is genome annotation. Without it, a genome is basically a 3-billion-letter PDF with no search bar.

The Old Way: Educated Guessing In A Lab Coat

For years, eukaryotic annotation leaned hard on ab initio gene finders. These are machine-learning-flavored models that scan raw DNA for statistical signals: splice sites, codon patterns, open reading frames, the usual genomic breadcrumbs. They became popular because RNA evidence used to cost real money and arrive slowly, like a Netflix DVD but wetter.

The catch: DNA alone can be ambiguous. Eukaryotic genes are chopped up, alternatively spliced, duplicated, pseudogenized, and occasionally arranged like a Christopher Nolan timeline. A predictor can make a plausible guess and still miss the exact exon-intron structure. That matters because a wrong gene model can scramble downstream work in expression analysis, protein prediction, comparative genomics, conservation, agriculture, and disease studies. One bad annotation can start a whole group chat of confused biologists.

EviAnn's Move: Ask The Witnesses

EviAnn, from Aleksey Zimin, Daniela Puiu, Mihaela Pertea, James Yorke, and Steven Salzberg, flips the emphasis. Instead of treating transcript and protein evidence mainly as training fuel for an ab initio predictor, it builds gene structures directly from the evidence: RNA-seq/transcript alignments and protein homology from related species. If ab initio prediction is a detective staring at footprints, EviAnn is the detective saying, "What if we also interview the person holding the shoe?"

The paper reports that, using the same input data, EviAnn outperformed leading packages including BRAKER3, MAKER2, and FINDER, while using far less compute. The headline number lands with sitcom timing: a mammalian genome can be annotated in under an hour on one multicore server. Not "after a moon ritual and a campus cluster reservation." Under an hour.

EviAnn also outputs GFF3, handles protein-coding and long noncoding RNA annotations, supports short and long transcript reads, can include UTRs, labels processed pseudogenes, and ships as open-source software through GitHub and Bioconda. That last bit matters. Tools do not help much if installing them feels like solving a Dark Souls side quest with Perl dependencies.

Why This Lands Now

The timing is not random. Projects like the Earth BioGenome Project want to scale from beautiful one-off reference genomes to a flood of species-level data. EMBL-EBI reported that the project aims to reach 3,000 new genomes per month on the path to sequencing known eukaryotic life by 2035. Assembly has sped up. Annotation has become the bouncer at the club.

Recent tools show the field attacking that bottleneck from multiple angles. BRAKER3 and GeneMark-ETP integrate RNA-seq and protein evidence into strong automated predictors. LiftOn transfers annotations by combining DNA and protein alignments. Long-read RNA-seq benchmarks show that longer, more accurate transcript reads can improve transcript discovery. Deep-learning tools like Helixer and ANNEVO are pushing sequence-only gene prediction forward too, with main-character energy and a suspiciously large GPU appetite.

EviAnn's argument is not "AI is canceled." It is more practical: when good evidence exists, use it directly. Your phone's autocomplete should not guess what your friend said if you are literally holding the message.

The Fine Print, Because Biology Enjoys Chaos

EviAnn depends on evidence quality. If your RNA-seq misses a tissue, developmental stage, stress condition, or rare transcript, EviAnn may not see that gene unless protein homology rescues it. If related protein evidence comes from organisms too far away, alignments get weaker. Weird genomes will continue to be weird. Biology has franchise rights to "the sequel is messier."

But this is exactly why EviAnn feels useful rather than flashy. It targets a boring-sounding pain point that quietly controls everything downstream: turning raw sequence into a trustworthy parts list. If its results hold up across more organisms and messy field datasets, it could make non-model species research faster, cheaper, and less dependent on bespoke pipeline wizardry.

That means better crop genomes, sharper biodiversity studies, improved conservation genetics, and fewer "gene not found" moments when the gene was there the whole time wearing a fake mustache.

References

• Zimin, A. V. et al. "Efficient evidence-based genome annotation with EviAnn." Nature Methods (2026). DOI: 10.1038/s41592-026-03156-0, PMID: 42399474.
• Gabriel, L. et al. "BRAKER3: fully automated genome annotation using RNA-seq and protein evidence." Genome Research 34, 769-777 (2024). DOI: 10.1101/gr.278090.123.
• Bruna, T. et al. "GeneMark-ETP significantly improves the accuracy of automatic annotation of large eukaryotic genomes." Genome Research 34, 757-768 (2024). DOI: 10.1101/gr.278373.123, PMCID: PMC11216313.
• Chao, K.-H. et al. "Combining DNA and protein alignments to improve genome annotation with LiftOn." Genome Research 35, 311-325 (2025). DOI: 10.1101/gr.279620.124.
• Pardo-Palacios, F. J. et al. "Systematic assessment of long-read RNA-seq methods for transcript identification and quantification." Nature Methods 21, 1349-1363 (2024). DOI: 10.1038/s41592-024-02298-3.
• Djossou, A. "An overview of computational methods for gene prediction in eukaryotes." Bioinformatics Advances 5, vbaf222 (2025). DOI: 10.1093/bioadv/vbaf222.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.