Otari

Most human genes produce several transcript isoforms, and which isoform a tissue makes is often more informative than how much of the gene it expresses. Genetic variants perturb this layer by disrupting the RNA processing signals that determine transcript structure and abundance, but sequence-based deep learning has largely addressed the problem piecemeal — predicting individual splice sites or RNA-binding protein footprints rather than the abundance of full-length isoforms.

Otari, developed at Princeton University and the Flatiron Institute in Olga Troyanskaya's group, attacks the problem by treating a transcript as what it structurally is: a graph. Each isoform becomes a directed graph whose nodes carry the sequence context around each exon's 3' and 5' splice sites and whose edges encode exon connectivity from GENCODE v47. A graph attention network propagates information across those nodes, pools them into a transcript-level embedding, and regresses onto a 30-dimensional vector of tissue-specific isoform abundances measured by long-read sequencing. Because the model consumes sequence, any variant can be applied by mutating the affected node and re-running the graph.

That design places Otari alongside models such as Isoformer, which fuses DNA, RNA, and protein encoders to predict transcript expression, and AbSplice2, which scores splicing-disrupting variants across tissues. Otari's distinguishing move is predicting differential abundance for full-length isoforms and propagating variant effects through the transcript's own splice structure. It was posted as a preprint in September 2025 and published in Cell Genomics in 2026.

#Key Features

• Transcripts as directed graphs: Nodes represent splice-site sequence context and edges encode exon order, so the architecture reflects isoform structure directly rather than flattening it into a single sequence window.
• Composite regulatory node embeddings: Each node carries a 23,600-dimensional vector concatenating splice-site scores from a purpose-built CNN (ConvSplice), predicted binding affinities for over 100 RNA-binding proteins from Seqweaver, and 10,062 chromatin peak features from Sei.
• Isoform-resolved variant effects: Variant impact is the log fold change between predictions on the reference and alternative graphs, yielding a separate effect estimate for every annotated isoform in every one of the 30 profiled tissues.
• Mechanistic interpretability: Comparing node attributes before and after mutation identifies which exon is most disrupted and whether the cause is chromatin- or RBP-associated.
• Coverage of novel and non-coding transcripts: Performance holds on unannotated novel isoforms and on long non-coding RNAs, not just well-characterized protein-coding transcripts.

#Technical Details

The final architecture is three blocks of graph attention layers with residual connections in the second and third blocks, global max pooling, and a two-layer feed-forward head, selected by a systematic search over GAT versus GCN layers, pooling strategies, and block depth. Training used a hidden size of 512, two attention heads, dropout of 0.5, and a loss combining mean squared error with a triplet margin term, and completed in roughly eight hours on a single GPU. The training set comprised 135,638 canonical transcripts quantified by full-length Nanopore sequencing across 30 tissues and brain regions, with chromosome 8 held out. On that holdout, Otari separates high- from low-expressed transcripts with mean AUROC 0.835 and AUPRC 0.812 (mean Pearson r = 0.563), against 0.766 for a gene-level baseline built from each gene's most abundant predicted transcript. Independent long-read datasets from different donors and sequencing platforms give AUROCs of 0.737 and 0.657, and 196,283 novel isoforms reach a mean Pearson r of 0.368. lncRNA and protein-coding transcripts score comparably (0.823 versus 0.808). ConvSplice, the splicing component, reaches precision-recall AUC 0.98 for both donor and acceptor sites using a 20 kb context window.

#Applications

Otari is distributed as a variant-scoring tool: given a VCF or TSV of variants, it returns per-isoform, per-tissue effect sizes plus interpretability tables. On fine-mapped GTEx v10 eQTLs it recovers the direction of expression change with 87.5% accuracy for strongly predicted variants, and it assigns larger effects to transcripts overlapping sQTL-associated spliced regions. Applied to HGMD disease-associated polymorphisms, it separates them from neutral controls and often flags a non-principal isoform as the most affected — including a clinical PTEN variant whose predicted disruption localizes to the exon reported to be skipped. In the SPARK autism cohort, de novo variants in probands drove greater predicted isoform dysregulation than sibling variants despite equal variant burden, concentrated in brain tissues and in microexon-containing transcripts.

#Impact

Otari extends sequence-to-function modeling from genes and splice sites to full-length isoform abundance, a level at which many disease effects are visible and gene-level analyses are blind. The framework code is on GitHub and the trained weights and preprocessed resources are archived on Zenodo under CC BY 4.0, though the repository itself carries no license file. Scope is bounded: the model is human-only, resolves regulation at tissue rather than cell-type level, depends on GENCODE-annotated splice structures, and predicts transcript abundance without linking those changes to protein abundance, localization, or stability. The disease findings are computational and await experimental validation.