GENATATOR

Genome annotation — locating genes within a chromosome and reconstructing their internal exon-intron organization — is a prerequisite for nearly every downstream genomic analysis, yet classical ab initio tools such as AUGUSTUS and Helixer rely on hand-engineered gene models derived largely from protein-coding sequence. These rule-based systems struggle with non-coding transcripts and often require species-specific retraining or external RNA-seq and protein evidence. GENATATOR, developed at AIRI (Artificial Intelligence Research Institute) in Moscow and posted as a bioRxiv preprint in June 2026, reframes annotation as a supervised sequence-labeling problem learned directly from DNA by fine-tuning DNA language models.

Rather than encoding explicit segmentation rules, GENATATOR learns to infer transcript boundaries and gene structure from labeled reference annotations. It builds on AIRI's DNA language model lineage — the GENA-LM family and its successor ModernGENA — and also fine-tunes the Caduceus backbone, applying these pretrained sequence models to the annotation task. A distinguishing capability is the simultaneous annotation of both protein-coding and long non-coding RNA (lncRNA) genes, including segmentation of non-coding transcripts and the untranslated regions of protein-coding genes that rule-based annotators frequently miss.

The model was trained on human plus 38 additional mammalian assemblies and reported to generalize zero-shot to genomes separated from the training set by hundreds of millions of years of evolution — including fruit fly (Drosophila), Arabidopsis, and budding yeast (Saccharomyces cerevisiae) — without any species-specific fine-tuning. The work was developed by Aleksei Shmelev, Artem Shadskiy, Yuri Kuratov, Mikhail Burtsev, Olga Kardymon, and Veniamin Fishman.

#Key Features

• Ab initio annotation from raw DNA: Predicts gene structure directly from a genomic FASTA sequence and emits standard GFF3 annotation files, without requiring RNA-seq alignments or homology evidence.
• Multi-stage annotation pipeline: Combines an edge model that detects transcription start sites and polyadenylation signals, a region model that filters candidate intervals, a transcript-type classifier that labels intervals as mRNA or lncRNA, and a segmentation model that predicts exon, intron, and coding-sequence structure.
• Joint coding and non-coding annotation: Annotates lncRNA and protein-coding genes in a single pass, recovering non-coding transcripts and untranslated regions that protein-centric annotators typically overlook.
• Zero-shot cross-species transfer: A checkpoint trained on mammalian genomes annotates evolutionarily distant genomes — insect, plant, and fungal — without retraining.
• DNA language model backbones: Fine-tunes pretrained DNA language models including ModernGENA and Caduceus, transferring self-supervised sequence representations to gene-structure prediction.
• Biologically grounded evaluation: Introduces metrics anchored on transcription start and termination sites, alongside public benchmarking datasets and leaderboards for ab initio annotation and segmentation.

#Technical Details

GENATATOR is a pipeline of fine-tuned DNA language models rather than a single monolithic network. Training uses a gene-segmentation dataset providing nucleotide-level labels for UTRs, exons, introns, and coding sequences across 39 mammalian assemblies, with human transcripts and a 1.45-million-example multispecies split; human chromosomes 8, 20, and 21 are held out for validation and final evaluation. The distributed pipeline pairs boundary-detection and region models with transcript-type and segmentation heads. Backbone variants span transformer-based DNA language models (ModernGENA, GENA-RMT) and the Caduceus architecture; the released Caduceus-PS segmentation variant is a 14M-parameter model. The reference pipeline runs on CUDA GPUs with float32 inference. AIRI publicly released the fine-tuned models, the training and gene-finding datasets, and leaderboards on HuggingFace.

#Applications

GENATATOR targets the annotation of newly sequenced eukaryotic genomes, particularly non-model organisms where curated training data and RNA-seq evidence are scarce. Because a fixed checkpoint transfers across species, genomics groups can annotate divergent genomes without assembling species-specific pipelines. Its joint treatment of coding and non-coding transcripts makes it well suited to studies of the non-coding transcriptome, including lncRNA discovery and delineation of transcript boundaries and untranslated regions. The published datasets and leaderboards also give the community a standardized benchmark for comparing ab initio annotation methods.

#Impact

GENATATOR extends DNA language models from element-level prediction tasks — promoter, splice-site, and chromatin classification — to full gene-structure annotation, a workflow long dominated by hidden Markov model and rule-based tools. By demonstrating that fine-tuned sequence models outperform established annotators and generalize across large evolutionary distances, it positions learned representations as a practical alternative for de novo genome annotation. The openly released models, datasets, and leaderboards lower the barrier to reproduction and comparison. The work is a preprint awaiting peer review, the reference pipeline currently requires GPU execution, and annotation quality is expected to decline on genomes most divergent from the mammalian training distribution.