Variant-resolved prediction of context-specific isoform variation with a graph-based attention model

Summary

In eukaryotes, most genes produce multiple transcript isoforms that diversify the transcriptome and proteome, serving as a key mechanism of functional regulation. Genetic variation can disrupt the RNA processing signals that shape isoform structure and abundance, yet modeling these effects at full-length isoform resolution remains challenging due to the complexity of transcript regulation. Here, we introduce Otari, an attention-based graph neural network framework trained on the human genomic sequence and long-read transcriptomes across 30 tissue types and brain regions. Otari predicts tissue-specific differential isoform abundance by integrating sequence-derived epigenetic and post-transcriptional signals, enabling isoform-resolved variant effect interpretation. Applied to large-scale variant datasets, including an autism cohort, Otari uncovers patterns of isoform dysregulation undetectable at the gene level, such as variant-driven perturbations in isoform abundance and microexon usage implicated in autism pathophysiology. We provide Otari as a resource for powering isoform-level analyses across tissues at scale.

Graphical abstract

Highlights

• •Otari is a graph attention model of full-length, tissue-specific isoform regulation
• •Integrates regulatory features such as splicing, chromatin, and RNA-binding proteins
• •Provides isoform-level mechanistic hypotheses for human regulatory disease variants
• •Reveals characteristic patterns of isoform dysregulation in brain-related conditions

Litman et al. introduce Otari, a graph-based deep learning model of full-length, tissue-specific isoform regulation, structure, and abundance. Systematic analyses of isoform-level variant effects with Otari reveal widespread, characteristic patterns of isoform dysregulation in various tissue and disease contexts, offering interpretable mechanistic hypotheses for human regulatory variants.

Introduction

Understanding transcript isoform diversity is essential for capturing the precise spatiotemporal expression of the genome in human health and disease.^1,2,3,4 This diversity arises from tightly regulated programs of isoform usage and switching across tissues and cell types, and their disruption has been implicated in a wide range of conditions.^{5,6,7,8,9,10,11,12,13,14,15,16} Evaluating genetic variant effects at the isoform level can therefore uncover clinically relevant mechanisms and inform isoform-targeted therapies.^17,18,19,20

Recent studies have shown that isoform-level analyses often reveal stronger effect sizes and greater disease specificity than gene-level analyses.^8,9,10,21,22 For example, transcriptome profiling of the cerebral cortex in autism, schizophrenia, and bipolar disorder identified differentially expressed transcripts not detected at the gene level, uncovering new candidate disease genes.¹⁰ An integrative framework is needed to model full-length, context-specific isoform variation that arises from the complex interplay of splicing factors, regulatory elements, epigenetic modifications, and the spliceosome.^2,3,12,23,24 To this end, we developed Otari: a comprehensive and interpretable graph-based framework of isoform regulation, powering the characterization of transcriptomic diversity and isoform-level variant effects at scale.

While sequence-based deep learning models have been successful in a variety of biological and disease contexts,^{25,26,27,28,29} extending these methods to full-length transcript isoforms remains a major challenge. Existing approaches in this space have made significant progress in predicting individual regulatory features, such as splice sites (e.g., SpliceAI, MMSplice, and Pangolin)^{30,31,32,33,34,35} or RNA-binding protein (RBP) profiles (e.g., Seqweaver and RBPNet).^36,37 However, they fall short in capturing the regulatory complexity of full-length isoforms and are unable to predict isoform abundance levels. These limitations can be attributed in part to the challenges of this modeling task but also to constraints of short-read RNA sequencing (RNA-seq), which lacks the resolution to span full transcripts or multiple splice junctions.

Our integrative graph-based framework, Otari, addresses the challenges of modeling full-length transcript isoforms at scale. Otari is trained entirely on long-read transcriptomic data, which provides direct measurements of full isoforms even in data-sparse settings³⁸ and leverages advances in attention-based graph deep learning to capture whole-transcript regulation. Transcript isoforms are naturally phrased as graphs and are therefore well-suited for a message-passing graph neural network (MPNN) approach. Unlike convolutional or transformer-based models, MPNNs learn directly from dynamically structured graphs³⁹ while leveraging attention to capture both local and global regulatory signals shaping transcriptomic variation.

Otari effectively learns tissue-specific regulatory patterns across 30 human tissues and brain regions by embedding epigenetic and post-transcriptional regulatory signals into DNA sequence graphs, enabling accurate and explainable predictions of transcriptome-wide isoform-level abundance. We show that Otari generalizes to novel isoforms, to complex multi-exonic transcripts, and across both protein-coding genes and long non-coding RNAs (lncRNAs). We further extend the framework to predict variant effects on isoform abundance, enabling isoform-resolved variant interpretation at scale. Applied to thousands of variants from HGMD,⁴⁰ ClinVar,⁴¹ and GTEx,⁴² Otari reveals patterns of transcript-level variant effects that are undetectable at the gene level, providing hypotheses for previously unexplored regulatory mechanisms underlying human health and disease. Finally, in a case study of a large autism cohort,⁴³ Otari identifies variant-driven isoform misregulation, including altered microexon usage, as a prevalent feature of autism pathophysiology.

Design

Predicting the regulation of full-length isoforms presents a distinct challenge: isoforms are defined not only by their splice junctions but by long-range interactions across exons, regulatory elements, and post-transcriptional processes. Existing deep learning approaches have demonstrated strong performance on focused tasks such as splice site or RBP binding predictions, yet they are not designed to capture the complex regulatory landscape that determines isoform abundance. This motivates the need for a modeling framework that can integrate isoform structures with a variety of regulatory signals.

To address this, we developed Otari, an attention-based graph deep learning framework that embeds full transcripts into structure-aware representations, enabling isoform-level predictions across tissues (Figure 1A). Each transcript is represented as a directed graph, where nodes encode DNA sequence features for splice site regions and edges define exon connectivity, effectively modeling isoform splice structure. This graph-based design enables a range of downstream tasks, including prediction of isoform-resolved variant effects and context-specific analysis of isoform regulation (Figure 1B).

Figure 1.

The Otari framework predicts context-specific isoform regulation, differential abundance, and variant effects

(A) Data construction and model training protocol. Transcript graphs were constructed from the human genomic sequence and long-read RNA-seq data. Nodes represent sequence features surrounding exon splice site regions, and directed edges encode exon connectivity based on GENCODE v.47 annotations. Each node is assigned an attribute vector comprising outputs from three sequence-based models (ConvSplice, Sei, and Seqweaver) that capture 5′ and 3′ splice site region regulatory features. A graph attention network (GAT) propagates information across nodes through attention-based message passing. The resulting node embeddings are pooled into a graph-level representation and regressed onto a 30-dimensional vector of tissue-specific isoform abundances derived from long-read RNA-seq data. Otari outputs relative isoform abundances across 30 tissues. See also Figure S1–S3 and S5. Created with BioRender.com.

(B) Otari enables prediction of isoform-level variant effects. For a given variant, transcript graph node sequences are mutated, and node attributes are recomputed and updated. Variant impact is estimated as the log fold change between predictions from reference and alternative graphs. These predictions are combined with interpretability metrics to identify transcript nodes affected by cis-regulatory variation and impacted underlying regulatory features. Otari facilitates downstream analyses of isoform regulation across tissues and contexts. Created with BioRender.com.

Otari integrates the regulatory landscape of each isoform, captured by constructing rich node embeddings that include splicing, chromatin, and RBP profiles. To model splicing signals, we developed ConvSplice, a convolutional neural network for splice site strength prediction (Figure S1A). ConvSplice achieved state-of-the-art performance, with area under the precision-recall curve (AUPRC) values of 0.98 for both donor and acceptor site prediction, surpassing SpliceAI’s³⁰ 0.97 (Figure S1B). In the top-k accuracy evaluation, ConvSplice achieved 0.94 for both site types, outperforming SpliceAI’s 0.93 and 0.92, with particularly improved accuracy at donor sites. The improved performance can be attributed to two factors: (1) the deeper architecture with 40 convolutional layers and (2) the larger 20 kb sequence context that allows the model to capture longer-range sequence dependencies affecting splicing regulation. To incorporate epigenetic signals in our node embeddings, we used the Sei model,²⁶ trained on chromatin immunoprecipitation sequencing (ChIP-seq) data, to predict tissue- and cell-type-specific cis-regulatory chromatin features. For post-transcriptional regulation-based node embeddings, we leveraged Seqweaver³⁶ to infer the binding affinities of over 100 RBPs that modulate splicing, processing, transport, stability, and translation.⁴⁴

These regulatory signals were embedded in a custom MPNN designed to capture both local and global transcript features. Our architecture combines attention, residual connections, and pooling layers (Figure S2), and a systematic architecture and hyperparameter search confirmed that this design outperforms alternative models (Figure S3).

Results

Otari accurately predicts tissue-specific isoform abundance

We trained Otari to regress pooled transcript graph embeddings onto relative isoform abundances (measured in transcripts per million [TPM]) across 30 human tissues and brain regions⁴⁵ and evaluated its performance in distinguishing high- versus low-expressed transcripts (n = 2,206 transcripts; see quantification and statistical analysis). Otari achieved strong predictive accuracy, with an average area under the receiver operating curve (AUROC) of 0.835 (SD = 0.021) and precision-recall curve (AUPRC) of 0.812 (SD = 0.031) on the chromosome 8 holdout test set from the Gao et al.⁴⁵ dataset (Figure 2A; mean Spearman’s r = 0.583; max p = 2.1 × 10⁻⁹⁴; Spearman’s r is shown in Figure S4), indicating that the model successfully captures the regulatory code underlying tissue-specific differential isoform abundance. Additionally, an ablation study was performed to quantify the relative contribution of each regulatory feature class to model performance (Figure S5).

Figure 2.

Otari accurately predicts isoform abundance on held-out and independent datasets

(A) Performance of Otari on the Gao et al. chromosome 8 holdout test set (n = 2,206 canonical transcripts). Measurements were taken from distinct samples. Relative isoform abundances were binarized to classify high- and low-expressed transcripts. Prediction performance is reported as AUROC values (x axis) across 30 tissues and brain regions (y axis). Mean AUROC = 0.835 (SD = 0.021). Each point represents a tissue-specific model. See also Table S1 and Figures S4 and S6–S9.

(B) Independent validation on an external dataset. Otari was evaluated on canonical transcripts from the Glinos et al. (n = 2,243 transcripts) and Leung et al. (n = 637 transcripts) validation test datasets (chromosome 8 holdout transcripts only). Glinos et al. tissues matched to Otari tissues are shown. Mean AUROC for Glinos et al. = 0.737 (SD = 0.029), and mean AUROC for Leung et al. = 0.657 (SD = 0.033).

(C) Otari performance stratified by transcript complexity. Chromosome 8 transcripts from the Gao et al. holdout test set (n = 2,206 transcripts) were grouped by exon count into 10 bins: nine bins for transcripts with 1–9 exons and one bin for transcripts with ≥10 exons. AUROC distributions (y axis) across tissues are shown for each bin. In all boxplots, the center lines indicate the median, the box limits denote the 25th and 75th percentiles, the whiskers extend to 1.5× the interquartile range, and individual points represent tissue-specific values. Mean AUROC by bin in ascending order: bin 1 = 0.749 (SD = 0.052), bin 2 = 0.894 (SD = 0.044), bin 3 = 0.896 (SD = 0.030), bin 4 = 0.904 (SD = 0.025), bin 5 = 0.867 (SD = 0.035), bin 6 = 0.850 (SD = 0.030), bin 7 = 0.813 (SD = 0.021), bin 8 = 0.855 (SD = 0.038), bin 9 = 0.765 (SD = 0.032), and bin 10 = 0.781 (SD = 0.021). See also Figure S10.

(D) Isoform-specific predictions outperform global gene-level comparisons (n = 2,206 transcripts). AUROC values for Otari’s isoform-level predictions (y axis) were compared to gene-level predictions computed using the predicted most abundant transcript per gene (x axis) in the Gao et al. holdout test set. Mean AUROC, isoform level = 0.835 (SD = 0.021). Mean AUROC, gene level = 0.766 (SD = 0.026). See also Figure S11.

The generalizability of our model was supported by rigorous evaluations on independent datasets,^46,47 including a total of 2,880 canonical transcripts from the held-out chromosome. Across matched tissues in the validation test Glinos et al. dataset,⁴⁶ Otari achieved a strong average AUROC of 0.737 (SD = 0.029; mean AUPRC = 0.730; Figure 2B) and performed similarly well in another validation test evaluation on the Leung et al.⁴⁷ dataset (mean AUROC = 0.657, SD = 0.033; mean AUPRC = 0.787; Table S1) despite differences in donors, sequencing platforms, and experimental conditions from the training data. We also evaluated Otari’s ability to predict expression of novel transcripts, which are often difficult to quantify due to low abundance or limited annotation. Despite these challenges, Otari achieved a median Spearman’s correlation of 0.375 across 196,283 novel isoforms (interquartile range [IQR] = 0.050; Figure S6), matching its performance on lower-abundance canonical transcripts (median Spearman’s r = 0.369, IQR = 0.044; p = 0.11; two-sided independent t test). Finally, on the Gao et al. chromosome 8 benchmark, Otari achieved significantly higher median correlation with measured transcript expression than AlphaGenome⁴⁸ (false discovery rate [FDR] = 3.85 × 10⁻⁵ compared to the RNA-seq track, FDR = 3.55 × 10⁻⁵ compared to the cap analysis of gene expression [CAGE] track; two-sided Mann-Whitney U tests with Benjamini-Hochberg correction; Otari median Spearman’s r = 0.587, IQR = 0.017; mean r = 0.541/0.540, IQR = 0.031/0.025 for AlphaGenome RNA-seq/CAGE; Figure S7).

lncRNAs undergo splicing and can generate multiple isoforms, contributing to transcriptomic complexity and functional regulation.^49,50 Despite their lower conservation compared to protein-coding genes, Otari achieved comparable performance on lncRNAs and protein-coding isoforms (lncRNA mean AUROC = 0.823; protein-coding mean AUROC = 0.808; Figure S8), supporting its strong generalizability across transcript types. Notably, Otari’s performance was also sustained when aggregating predictions per gene (gene-specific Spearman’s r mean = 0.496 and SD = 0.029; Figure S9) and was robust to transcript complexity, as demonstrated by accurate predictions even for transcripts with 10+ exons (Figure 2C; mean AUROC for 1–10 exons = 0.844; mean AUROC for 10+ exons = 0.781; transcript counts shown in Figure S10).

Importantly, Otari captures transcriptomic variation that isoform-agnostic models often miss. When benchmarked on the held-out set, AUROC values based on Otari’s isoform-specific predictions consistently outperformed gene-level baselines derived from the predicted most abundant transcript per gene (Figure 2D; mean AUROC: 0.835 versus 0.766; SD = 0.021/0.026). We also evaluated an alternative gene-level aggregation by summing predicted abundances across all isoforms for each given gene, which yielded even lower performance (mean AUROC: 0.719, SD = 0.029; Figure S11). These results underscore Otari’s ability to accurately model differential isoform abundance beyond gene-level expression, supporting its application in downstream isoform analyses.

Modeling isoform-level variant effects

Genetic variants reshape transcriptomic architecture by altering splicing patterns and isoform abundance, thereby contributing to diverse disease outcomes.^{9,12,14,17,18,51,52,53,54} To systematically quantify these effects, we extended Otari to predict isoform-specific consequences of single-nucleotide variants (Figure 1B). For each variant, Otari mutates the transcript graph node sequences, recomputes node features, and evaluates the reference and alternative graphs to estimate log fold changes in tissue-specific isoform abundance. Furthermore, by analyzing the graph structure, Otari identifies affected nodes, which often correspond to alternatively spliced regions (see below). These predictions, coupled with the interpretability of the underlying disrupted regulatory features, provide mechanistic insight into how variants perturb isoform profiles.

To validate Otari’s variant effect predictions, we leveraged fine-mapped expression and splicing quantitative trait loci (eQTLs and sQTLs) from GTEx v.10.⁴² Notably, Otari correctly inferred the direction of expression change for thousands of eQTLs across 10 tissues (Figure 3A). Although eQTLs are defined at the gene level, aggregated isoform-level predictions achieved 90% accuracy for variants with strong predicted effects (normalized summed score > 36). Otari also successfully captured transcript-specific expression changes associated with sQTLs: transcripts overlapping or flanking variant-associated spliced regions showed significantly greater predicted impact than non-overlapping transcripts (Figure 3B; max FDR = 2.73 × 10⁻⁴; one-sided independent t tests with Benjamini-Hochberg correction; n = 7,412 variants).

Figure 3.

Otari captures interpretable functional effects of GTEx QTLs

(A) Directionality prediction of fine-mapped expression quantitative trait loci (eQTLs). Variant sets from GTEx v.10 across ten distinct tissues (n = 4,181 variants) were analyzed. For each variant, Otari-predicted isoform-level effects were aggregated per gene and compared to the direction of the eQTL allelic fold change (aFC). Prediction accuracy (y axis) was assessed across a range of summed-effect thresholds (x axis). Shaded regions represent ±1 standard deviation from 1,000 bootstrap samples.

(B) Otari predicts differential abundance effects associated with fine-mapped splicing QTLs (sQTLs). sQTLs from six GTEx v.10 tissues (n = 7,412 variants) were analyzed by grouping transcripts into overlapping or non-overlapping categories based on their exon overlap with the sQTL-associated spliced region. Absolute effect sizes (y axis) were compared between groups for each tissue. One-sided independent t tests were used for hypothesis testing, with Benjamini-Hochberg correction applied for multiple comparisons. Scores were normalized relative to a background distribution (n = 78,369 variants). Circles denote means; error bars represent the standard error of the mean (SEM). Sample sizes (overlapping/non-overlapping transcripts), means, and standard error for representative tissues are as follows: brain cortex (n = 2,947/353, mean = 1.01/0.28, SEM = 0.08/0.10), liver (n = 1,795/214, mean = 0.53/0.15, SEM = 0.06/0.03), whole blood (n = 4,848/580, mean = 0.61/0.33, SEM = 0.04/0.07), lung (n = 6,696/909, mean = 0.81/0.27, SEM = 0.04/0.03), heart (n = 4,382/512, mean = 0.55/0.17, SEM = 0.03/0.03), and colon (n = 4,868/599, mean = 0.87/0.26, SEM = 0.06/0.04). Test statistics with confidence intervals, degrees of freedom (dof), and FDR values are as follows: brain cortex (5.56 ± 0.26, dof = 3,183, FDR = 2.15 × 10⁻⁸), liver (5.82 ± 0.13, dof = 1,953, FDR = 5.44 × 10⁻⁹), whole blood (3.47 ± 0.16, dof = 5,213, FDR = 2.73 × 10⁻⁴), lung (10.62 ± 0.10, dof = 7,372, FDR = 1.29 × 10⁻²⁵), heart (8.03 ± 0.09, dof = 4,690, FDR = 1.71 × 10^-−15), and colon (8.97 ± 0.13, dof = 5,309, FDR = 7.03 × 10⁻¹⁹).

(C) Functional feature burden analysis of fine-mapped QTLs (n = 15,914 variants). Otari-estimated variant effect scores were used to select the top-impacted regulatory features for each eQTL and sQTL dataset (six tissues each, y axis). Features were annotated to either chromatin-associated (Sei) or RNA-binding protein (RBP)-associated (Seqweaver) regulatory categories. Category counts were min-max normalized within each category. The color scale indicates normalized burden per category. Mean Z scores for eQTLs = 0.35/0.65 (SD = 0.25). Mean Z scores for sQTLs = 0.92/0.08 (SD = 0.07).

Otari additionally outperformed existing splicing models (SpliceAI, Pangolin, and MTSplice),^30,34,35 which are limited to predicting individual splice sites, on variant effect prediction. Across six tissues, Otari achieved consistently higher performance in sQTL causality prediction, roughly doubling the average precision across all tissues (AUPRC = 0.201–0.299 compared to 0.077–0.141 for baseline models), with AUROC values ranging from 0.739 to 0.823 compared to 0.616–0.771 for baseline models (Figure S12). These results demonstrate the advantage of context-specific isoform-level modeling for functional variant interpretation.

Finally, Otari offers interpretable insights into the regulatory mechanisms driving variant effects. For each QTL (n = 15,914), we identified the most disrupted node features, weighted by predicted effect sizes, and annotated the top features in each variant set to transcriptional (Sei) or post-transcriptional (Seqweaver) regulatory categories. Otari accurately captured the distinct regulatory signatures of QTLs: sQTL effects were primarily driven by RBP-associated features (Figure 3C; mean Seqweaver score = 0.92; Sei = 0.08; SD = 0.07), whereas eQTL effects were mostly attributed to chromatin-based signals (Seqweaver = 0.35; Sei = 0.65; SD = 0.25). Together, these findings demonstrate that Otari enables accurate and interpretable predictions of variant-driven isoform changes.

Disease mutations differentially impact isoforms

We next applied Otari to 2,270 disease-associated functional polymorphisms (DFPs) curated in the Human Genome Mutation Database (HGMD)⁴⁰ to assess how known pathogenic variants affect isoform profiles. Compared to control variants without disease association, DFPs showed significantly greater predicted effects on isoform abundance, highlighting strong variant-level specificity (Figure 4A; FDR = 3.20 × 10⁻³²; one-sided independent t tests with Benjamini-Hochberg correction). Otari also uncovered isoform-specific differences in variant impact within genes. For each disease variant, we identified the isoform with the strongest predicted effect. We found that many of these isoforms were annotated as non-principal based on structural, functional, and evolutionary metrics in APPRIS.⁵⁵ When we compared the predicted effects of DFPs on these highly affected isoforms versus principal transcripts, we observed a significantly lower impact on the principal set (Figure 4A; FDR = 4.42 × 10⁻⁵; mean effect size = 1.37 for principal and 2.08 for top ranked; standard error of the mean [SEM] = 0.12/0.14), despite similar baseline expression levels (Figure S13; mean TPM: 5.25/3.89). These analyses demonstrate the specificity of Otari’s predictions at both the variant and isoform levels and suggest important functional roles for individual transcripts in disease.

Figure 4.

Otari predicts isoform-resolved impact of regulatory disease variants

(A) Predicted variant effects from Otari for regulatory disease-associated functional polymorphisms (DFPs) from the Human Gene Mutation Database (HGMD) (n = 2,270 variants). From right to left: distribution of DFP effects on the most impacted isoform per gene, DFP effects on the principal transcript of each gene, and a control set of neutral variants (n = 78,369 variants), showing absolute effects on the most impacted isoform per gene. Effects were retained for the top-impacted tissue per variant. Circles represent group means; error bars indicate the standard error of the mean (SEM). One-sided independent t tests were used for hypothesis testing with Benjamini-Hochberg correction for multiple comparisons. Asterisks (∗∗∗) indicate false discovery rate (FDR) < 0.01. Normalized to neutral isoform effects. Mean scores from left to right (before normalization): 0.41 (SEM = 0.01), 1.37 (SEM = 0.12), and 2.08 (SEM = 0.14). Test statistics with confidence intervals: 12.04 ± 0.27/8.11 ± 0.23/3.92 ± 0.36, degrees of freedom: 80,637/80,080/3,981, FDR values: 3.20 × 10⁻³²/6.96 × 10⁻¹⁶/4.42 × 10⁻⁵.

(B) Isoform-level predictions for clinical PTEN variant rs786203847. Top: effect size fold changes (y axis) across 30 tissue types (x axis) for four PTEN isoforms; the dashed line indicates no effect. Middle: splice structure of the affected isoform showing exon 3 skipping (lighter shade); the x axis corresponds to the genomic coordinates on chromosome 10. Bottom: L2-norm distances between graph node attributes in reference and alternative sequences across the nine exons; the largest impact was observed in exon 3. See also Figures S13–S15.

(C) Application of Otari to cancer-associated variants (rs217727 and rs2839698) and a cancer-protective variant (rs2107425) in the H19 long non-coding RNA. Log fold change distributions across n = 30 tissues (y axis) are shown separately for each variant and two annotated isoforms. The center lines denote medians, the box limits indicate the 25th and 75th percentiles, the whiskers span 1.5× the interquartile range, and individual tissue-specific values are plotted as points. All effect sizes were Z score normalized relative to a background distribution (n = 78,369 variants). Median scores: rs217727 (1.68/0.93, IQR = 0.30/0.18), rs2839698 (1.75/0.63, IQR = 0.32/0.14), and rs2107425 (−0.04/1.23, IQR = 0.02/0.28).

To further illustrate this, we examined a pathogenic clinical⁴¹ variant in PTEN, a gene that has been implicated in various neurodevelopmental disorders and cancer.^56,57,58 Otari predicted that variant rs786203847 selectively downregulates several annotated PTEN isoforms (Figure 4B, top). For example, two transcripts, ENST00000371953 and ENST00000693560, showed sharp, tissue-specific predicted reduction in expression (mean effects = −23.1 and −15.2). However, Otari predicted that transcript ENST00000688308 is only moderately downregulated (mean effect = −5.99), while isoform ENST00000700021 remained unaffected (mean = 0.01; all transcripts are shown in Figure S14). Otari further identified mechanisms for these changes. Analysis of the graph nodes before and after introduction of the variant revealed that the most disrupted node corresponds to exon 3 of PTEN, consistent with previously reported exon 3 skipping for this variant⁵⁹ (Figure 4B, bottom). Otari also profiled the underlying disrupted regulatory signals, including loss of the donor splice site signal, increased signal of the repressive histone mark H3K27me3, and altered RBP activity involved in splicing and transcript stability (Figure S15).

We also demonstrate that Otari captures isoform-specific dysregulation in lncRNAs. For example, previous studies have linked three H19 lncRNA polymorphisms to cancer susceptibility, as reported in a meta-analysis of over 48,000 individuals.⁶⁰ We applied Otari to these H19-associated variants: rs217727 and rs2839698 (risk alleles) and rs2107425 (protective allele) (Figure 4C). Among the two annotated H19 isoforms, Otari predicted that the two risk variants selectively increase expression of isoform ENST00000691195 (H19-213), consistent with reported overexpression of H19 in human cancers.^60,61 In contrast, the protective variant increased expression of a distinct isoform, ENST00000710492 (H19-227), without affecting H19-213. These findings support the hypothesis that cancer risk and susceptibility may be mediated in part by differential regulation of H19 isoforms and propose a specific underlying mechanism for this association.

Together, these case studies highlight Otari’s ability to resolve variant-driven isoform dysregulation, including transcript-specific expression changes, splicing disruptions, and associated regulatory mechanisms that are largely undetectable in isoform-agnostic analyses.

Otari identifies isoform misregulation in autism

Transcript isoform diversity and regulation are particularly relevant to neurodevelopmental disorders, as the brain exhibits the highest levels of alternative splicing among all human tissues.^8,13,62 Otari enables direct mapping of genetic variants to isoform-specific consequences, allowing mechanistic interrogation of cis-regulatory variation implicated in autism and related conditions.^27,63

We applied Otari to a large autism cohort, SPARK.⁴³ Whole-genome de novo variant (DNV) calling was performed for 3,507 autistic probands and 2,206 unaffected siblings. While overall DNV burdens were comparable between groups (Figure 5A; mean counts: 70.92 and 70.44; SEM: 0.25 and 0.31), Otari predicted significantly greater isoform dysregulation associated with proband DNVs in Satterstrom genes⁶⁴ compared to unaffected siblings (Figure 5B; p = 7.51 × 10⁻⁸; one-sided independent t test). These effects were strongly tissue specific, with brain regions showing significantly higher predicted disruption in isoform abundance compared to non-brain tissues (Figure 5C; p = 3.11 × 10⁻⁹; one-sided independent t test), particularly in the fetal brain, corpus callosum, and cerebellum, all of which have been associated with autism.^57,63

Figure 5.

Characterizing patterns of isoform dysregulation in autism and brain-related conditions

(A) De novo variants were called from whole-genome sequencing of complete trios in the SPARK cohort (n = 3,507 probands, n = 2,206 siblings). Families are either simplex (one affected child) or multiplex (multiple affected children). Measurements were taken from distinct samples. The genome-wide de novo variant burden (y axis) was computed for each proband and sibling. Bar heights indicate group means, with error bars representing the standard error of the mean (SEM). Means = 70.92/70.44; SEM = 0.25/0.31.

(B) Otari-predicted effect size fold changes (fc; y axis) for de novo variants in SPARK probands (right; n = 1,200 variants) and siblings (left; n = 698 variants), limited to genes in the Satterstrom ASD-associated gene set. Effects were retained for the top-impacted brain tissue per isoform. In all analyses, circles represent means; error bars show SEM. A one-sided independent t test was used for hypothesis testing (p = 7.51 × 10⁻⁸). Means = 0.163/0.094; SEM = 0.012/0.006. Test statistic with confidence interval = 5.26 ± 0.03. Degrees of freedom = 15,551. Effect sizes were Z score normalized relative to a background distribution (n = 78,369 variants).

(C) For each tissue, the log2 fc of the mean absolute Otari score for probands compared to siblings is shown (x axis). The inset compares the distribution of log2 fc between brain/CNS tissues (n = 14) and non-brain tissues (n = 14). A one-sided independent t test was used for hypothesis testing (p = 3.11 × 10⁻⁹). Means = 0.508/0.360; SEM = 0.029/0.049. Test statistic with confidence interval = 9.37 ± 0.03. Degrees of freedom = 26.

(D) Microexon misregulation is associated with proband de novo variants in SPARK. Otari-predicted effects were computed for de novo variants mapping to brain-expressed genes (n = 57,929 variants in probands, n = 35,642 variants in siblings). Effect sizes (mean ± SEM) are shown for variants located near splice sites for microexons (3–27 nt; n = 650/355 transcripts) and long exons (>27 nt; n = 63,786/38,794 transcripts). Effects were retained for the top-impacted brain tissue per isoform. Significance is marked as follows: ∗FDR < 0.1, ∗∗FDR < 0.05, and ∗∗∗FDR < 0.01. One-sided independent t tests were used for hypothesis testing, with Benjamini-Hochberg correction applied for multiple comparisons. Means = 1.427/0.911/0.515/0.498; SEM = 0.256/0.160/0.010/0.012. FDR values: 0.044/6.005 × 10⁻⁴/7.800 × 10⁻³. Test statistics with confidence intervals: 1.707 ± 0.592/3.558 ± 0.502/2.576 ± 0.314. Degrees of freedom = 1,003/64,434/39,147. Effect sizes were Z score normalized relative to a background distribution (n = 78,369 variants).

(E) Otari-predicted effect size fc (y axis) for Alzheimer’s disease variants from HGMD (right; n = 550) and control variants (left; n = 698). Effects were retained for the top-impacted brain tissue per isoform. A one-sided independent t test was used for hypothesis testing (p = 2.678 × 10⁻⁵). Means = 0.170/0.094; SEM = 0.018/0.006. Test statistic with confidence interval = 4.047 ± 0.037. Degrees of freedom = 7,858. Effect sizes were Z score normalized relative to a background distribution (n = 78,369 variants).

(F) Otari-predicted effect size fc (y axis) for schizophrenia-associated variants from HGMD (right; n = 1,409) and control variants (left; n = 698). A one-sided independent t test was used for hypothesis testing (p = 8.724 × 10⁻⁶). Means = 0.174/0.094; SEM = 0.018/0.006. Test statistic with confidence interval = 4.300 ± 0.037. Degrees of freedom = 9,813. Effect sizes were Z score normalized relative to a background distribution (n = 78,369 variants).

To investigate some of the mechanisms underlying these effects, we examined microexons, which are highly conserved, neuron-specific short exons (3–27 nt) previously implicated in autism and other neurodevelopmental disorders.^65,66,67,68 We classified all brain-expressed exons as either microexons or long exons (>27 nt) and assessed the impact of nearby DNVs (n = 93,571). Notably, transcripts containing microexons near proband variants showed significantly greater dysregulation than those with nearby long exons (Figure 5D; FDR = 6.0 × 10⁻⁴; one-sided independent t tests with Benjamini-Hochberg correction). Proband variants near microexons also drove significantly stronger expression changes than sibling variants near microexons (FDR = 0.04), suggesting altered microexon inclusion in autistic individuals. Together with prior small-scale experimental studies,^65,66,67 our findings from an analysis of a large cohort support variant-driven microexon misregulation as a prevalent feature of autism pathophysiology.

Finally, application of Otari to Alzheimer’s disease variants (n = 550; Figure 5E) and schizophrenia-associated variants (n = 1,409; Figure 5F) from the HGMD revealed significantly elevated predicted isoform effects compared to control variants (p = 2.68 × 10⁻⁵/8.72 × 10⁻⁶; one-sided independent t tests), suggesting that Otari captures isoform-level dysregulation across brain-related conditions.

Discussion

Otari is a sequence-based model capable of resolving isoform-level regulatory and expression changes at scale. This attention-based graph deep learning framework offers a powerful approach for transcriptome-wide, isoform-level variant effect analysis, driving insights into the regulation of isoform usage in various biological contexts. Unlike prior models that focus primarily on predicting individual splice sites or RBP binding, Otari learns rich, structure-aware representations of full transcripts, capturing the complex regulatory code underlying differential isoform abundance. Trained on long-read transcriptomic data spanning diverse tissue types, Otari provides high-resolution predictions across canonical and novel isoforms, generalizes to complex multi-exonic structures, and enables isoform-resolved variant interpretation.

Otari also provides interpretable insights into how genetic variation alters isoform usage. Across large-scale mutation datasets, Otari enabled high-specificity isoform-level variant effect predictions. For example, Otari predicted that a regulatory clinical variant in PTEN led to minimal to modest downregulation of some transcripts but stronger, more tissue-specific effects on alternative transcripts, along with exon skipping and regulatory disruptions involving H3K27me3 and RBPs. In H19, Otari revealed that cancer risk and protective variants modulate the expression of distinct lncRNA isoforms, demonstrating its capacity to resolve functional mechanisms even in low-conservation regions. These findings open an avenue for systematic studies of lncRNA isoforms in future research.

Applying Otari to a large autism cohort further demonstrated its potential for mechanistic discovery. Despite similar genome-wide variant burdens, we observed significant brain-specific isoform dysregulation in autistic probands compared to unaffected siblings, particularly in brain regions previously implicated in autism. Otari also captured altered microexon usage in probands compared to unaffected siblings based on an analysis of a large cohort, further supporting the role of this mechanism in autism pathophysiology previously reported in small-scale experimental studies. These findings point to a broader role for isoform misregulation in neurodevelopmental disorders. Notably, Otari’s demonstrated ability to interpret regulatory variants at the isoform level, which are challenging to study experimentally and at scale, offers a promising path for isoform-resolved variant interpretation in a wide variety of context-specific studies.

In conclusion, Otari establishes a foundation for transcriptome-wide, isoform-level studies, enabling interpretability and precision in large-scale analyses of isoform regulation, diversity, and variant impact. To facilitate community use, we release Otari as an open-source package. Future work with a focus on experimental validation will be critical for translating these insights into isoform-targeted therapeutic strategies.

Limitations

Notably, variant effect evaluations currently focus on SNPs, and extension to insertions or deletions (indels) and structural variants, which can significantly impact transcript regulation, is an important future direction. Furthermore, while Otari captures transcript-level effects, it depends on transcript annotations for either annotated or novel isoforms of interest for graph construction. Novel isoform prediction in Otari is performed relative to an extended transcript reference constructed from long-read RNA-seq data across 30 tissues. While this captures isoforms absent from standard annotations such as Ensembl, it remains dependent on transcript structures observed in available data or novel structures provided by the user.

In future work, integrating single-cell long-read sequencing will be critical for capturing isoform regulation at cell-type resolution. While Otari models tissue-level regulation, resolving cell-type-specific regulation is essential for investigating heterogeneous tissues. Incorporating these data could also improve the detection and analysis of rare, novel, or cell-type-specific isoforms that may play important functional roles. Isoform regulation also varies across developmental stages and conditions. As more data become available, modeling these context-dependent regulatory dynamics will help reveal how isoform usage is altered across diverse biological and disease contexts. Another challenge lies in translating isoform-level effects to functional outcomes. While Otari provides high-specificity predictions of changes in transcript isoform abundance, linking these effects to downstream consequences, such as protein abundance, function, localization, or stability, remains non-trivial, especially for novel isoforms. This underscores the need for integration of multi-modal data types to drive prediction across multiple layers of regulation.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Olga G. Troyanskaya (ogt@princeton.edu).

Materials availability

This study did not generate new, unique reagents.

Data and code availability

• •This paper analyzes existing publicly available data. Training, validation, and testing data were obtained from the GENCODE v.47 human reference catalog,^69,70 as well as the Gao et al., Glinos et al., and Leung et al. datasets.^45,46,47 eQTL and sQTL datasets were obtained from the GTEx v.10 release.⁴² Pathogenic and benign human regulatory variants were obtained from ClinVar.⁴¹ Links to datasets are listed in the key resources table. Autism cohort data were obtained from SPARK.⁴³ Approved researchers can obtain the SPARK population genetic dataset described in this study by applying at https://base.sfari.org. Human regulatory disease mutations were obtained from HGMD (2024.1 release).⁴⁰
• •The Otari framework code generated during this study is available at https://github.com/FunctionLab/otari, and the model and associated data files can be downloaded from Zenodo or by following the instructions in the GitHub repository. The code for the manuscript analyses is available at https://github.com/FunctionLab/otari-manuscript.
• •Resources and trained models have been uploaded to Zenodo (DOIs: https://doi.org/10.5281/zenodo.16432270 and https://doi.org/10.5281/zenodo.16433545).

Acknowledgments

We are grateful to all the families in SPARK, the SPARK clinical sites, and the SPARK staff. We appreciate obtaining access to the SPARK genetic datasets on SFARI Base. Approved researchers can obtain the SPARK population dataset described in this study by applying at https://base.sfari.org. This work was conducted utilizing the computing resources, supported by the Scientific Computing Core, at the Flatiron Institute. We acknowledge dbGaP project #30901. This work was supported by funding from the NIH NIGMS no. R01GM071966 (O.G.T.), Simons Foundation grant 395506 (O.G.T.), and the NIH NHGRI training grant T32HG003284 (A.L.).

Author contributions

Conceptualization, A.L. (equal), Z.P. (equal), K.S. (equal), N.S. (supporting), C.Y.P. (supporting), C.L.T. (supporting), and O.G.T. (equal); data curation, A.L. (equal), Z.P. (equal), and T.M. (supporting); formal analysis, A.L. (lead), Z.P. (supporting), and J.F. (supporting); investigation, A.L. (lead); methodology, A.L. (equal), Z.P. (equal), and J.F. (supporting); project administration, A.L. (equal), C.L.T. (supporting), and O.G.T. (equal); software, A.L. (equal), Z.P. (equal), K.S. (supporting), and J.F. (supporting); visualization, A.L. (lead); writing – original draft, A.L. (lead), Z.P. (supporting), J.F. (supporting), C.Y.P. (supporting), C.L.T. (supporting), and O.G.T. (supporting); writing – review & editing, A.L. (lead), Z.P. (supporting), N.S. (supporting), C.Y.P. (supporting), C.L.T. (supporting), and O.G.T. (supporting); supervision, C.L.T. (supporting) and O.G.T. (lead).

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT in order to improve language and readability only. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Otari model	This manuscript	Zenodo: 10.5281/zenodo.16432270
GENCODE v47	Kaur et al.69	https://www.gencodegenes.org/human/release_47.html
APPRIS	Rodriguez et al.55	https://apprisws.bioinfo.cnio.es/appris_mane_matches/gencode47/
ESPRESSO long-read RNA-seq data	Gao et al.45	GEO: GSE192955
GTEx long-read RNA-seq data	Glinos et al.46	https://gtexportal.org/home/downloads/adult-gtex/long_read_data
Cortex long-read RNA-seq data	Leung et al.47	Table S2; Leung et al.47
HGMD variants	Stenson et al.40	N/A
GTEx QTLs	GTEx Consortium42	https://gtexportal.org/home/downloads/adult-gtex/qtl
ClinVar	Landrum et al.41	https://www.ncbi.nlm.nih.gov/clinvar/
Satterstrom gene set	Satterstrom et al.64	Table S2; Satterstrom et al.64
SPARK cohort VCFs	SPARK Consortium43	N/A
Software and algorithms
Otari	This manuscript	GitHub: https://github.com/FunctionLab/otari and https://github.com/FunctionLab/otari-manuscript
Sei	Chen et al.26	GitHub: https://github.com/FunctionLab/sei-framework
Seqweaver	Park et al.36	https://hb.flatironinstitute.org/seqweaver
ConvSplice	This manuscript	GitHub: https://github.com/FunctionLab/otari
MTSplice	Cheng et al.34	GitHub: https://github.com/gagneurlab/MMSplice_MTSplice
SpliceAI	Jaganathan et al.30	GitHub: https://github.com/Illumina/SpliceAI
Pangolin	Zeng et al.35	GitHub: https://github.com/tkzeng/Pangolin
AlphaGenome	Avsec et al.48	GitHub: https://github.com/google-deepmind/alphagenome
HAT	Ng et al.71	GitHub: https://github.com/TNTurnerLab/HAT
GATK	McKenna et al.72	GitHub: https://github.com/broadinstitute/gatk

Experimental model and study participant details

We received approval to access and analyze de-identified genetic data from the SPARK autism cohort from SFARI Base and the Princeton University IRB Committee in the Office of Research Integrity. We analyzed the genetic data of 3,507 probands and 2,206 non-autistic siblings from SPARK with available trios (mother, father, child) and whole genome sequencing (WGS). Any child with complete trio WGS passing QC was included in these analyses. Children with professional diagnoses of autism were assigned to the “probands” group, and children without professional diagnoses of autism were assigned to the “siblings” group. Additional details and cohort characteristics can be found in the SPARK Consortium manuscript.⁴³

Method details

Data

Training and testing

A total of 135,638 unique canonical transcripts meeting filtering criteria were retrieved from the Gao et al.⁴⁵ dataset. Gao et al. provide robust quantification of transcript isoforms using full-length long-read Nanopore sequencing across 30 distinct human tissues and brain regions. Measurements were taken from distinct samples. They report a total of 623 million ONT 1D cDNA reads (reads per tissue range from 13.2 to 30.6 million). Each sample included tissues pooled from between 2 and 64 individuals (see Table S5 in Gao et al.). Only transcripts annotated as full splice matches in the GENCODE v47 human reference catalog (GRCh38.p14, basic annotation)^69,70 were used for training, validation, and evaluation. Transcripts located on the X, Y, and mitochondrial (M) chromosomes were excluded from all training, validation, testing, and downstream analyses. The test set consisted of all transcripts from chromosome 8 annotated as protein-coding or lncRNAs (n = 2,206 transcripts), and all other transcripts from other autosomes were used for training and validation. Isoform abundance values (in TPM) were log₂-transformed with a pseudocount of 0.01 across all datasets.

Validation

Model validation was conducted using three independent transcript sets: novel isoforms from Gao et al., and canonical isoforms from the Glinos et al.⁴⁶ and Leung et al.⁴⁷ datasets. The Gao et al. set comprised 196,283 unannotated novel transcripts with annotated splice sites, profiled across 30 tissues. The Glinos et al. validation test set included 2,243 transcripts (protein-coding and lncRNA only) from chromosome 8, profiled via Nanopore sequencing across multiple tissues. The Leung et al. validation test set consisted of 637 chromosome 8 transcripts (protein-coding and lncRNA only) profiled in fetal and adult cortex using both Nanopore and PacBio platforms. For each dataset, transcript expression was reported in TPM (transcripts per million), and values were averaged across samples and/or replicates within tissues. Only transcripts with full splice match annotations were included. Transcripts were annotated as protein-coding or lncRNA based on the GENCODE v47 reference to assess gene biotype distribution. All measurements were taken from distinct samples.

Graph construction

Graph structure

We constructed a directed graph for each transcript across all datasets, where nodes represent the sequence context surrounding the 3′ and 5′ splice sites of each exon, and edges represent the sequential ordering of exons based on the GENCODE reference annotation. In the absence of node features, the resulting gene structure resembles a standard splice graph. Node sequences were generated in 5′–3′ order. For transcripts on the positive strand, nodes were ordered by increasing genomic position; for those on the negative strand, nodes were ordered by decreasing position.

Node attributes

To create a rich feature representation for each node, we constructed a concatenated vector of biological signals using three deep learning models applied to the DNA sequence context around each splice site. For splicing signals, we used a custom model named ConvSplice (described below) to score the likelihood that each site represents a splice donor or acceptor. For post-transcriptional regulation, we generated Seqweaver³⁶ predictions as part of the node embedding vector. Seqweaver provides predicted binding affinity scores for >100 RNA-binding proteins (RBPs), including regulators of splicing, stability, localization, and translation, based on 217 Cross-Linking and Immunoprecipitation sequencing (CLIP-seq) datasets. For each splice site, we generated predictions across an extended 1 kb window using eight 50-nt bins spanning from 200 bp upstream to 200 bp downstream. Finally, we used the Sei model²⁶ to predict chromatin-associated regulatory features as part of the node embeddings, including histone marks, transcription factors, and DNAse sensitivity. We retained the 10,062 cis-regulatory peak scores for histone marks (e.g., H3K27ac, H3K4me3), which have been previously linked to splicing and isoform control.²³ Predictions were obtained using a 4 kb sequence window centered on each splice site. These outputs were concatenated to form a final node feature vector of dimension 23,600, structured as follows: 3′ SpliceConv(2), 3′ Seqweaver(1,736), 3′ Sei(10,062), 5′ SpliceConv(2), 5′ Seqweaver(1,736), and 5′ Sei(10,062). The numbers in the parentheses indicate the feature dimension for each component. All genomic sequences were extracted from the GRCh38 human reference genome.

ConvSplice model

ConvSplice is a deep learning model developed to predict splice donor and acceptor sites directly from DNA sequence. It was trained using principal transcripts from GENCODE v40⁷⁰ annotated by APPRIS.⁵⁵ Positive examples included known donor (5′) and acceptor (3′) splice sites; all other positions in the transcripts were treated as negatives. To evaluate the model’s performance, we used chromosomes 1, 3, 5, 7, and 9 as the test set, with the remaining chromosomes used for training and validation.

The ConvSplice architecture builds upon the framework introduced by SpliceAI³⁰ but incorporates several key improvements. The model takes as input a 20kb sequence window centered on the position of interest (10kb on each side). Each nucleotide position in the input sequence is one-hot encoded as a 4-dimensional vector representing A, C, G, and T. The model outputs three scores for each position, representing the probability of that position being a splice acceptor, splice donor, or neither. The architecture consists of 20 dilated convolution layers with residual connections applied every four layers. Each layer includes batch normalization and ReLU activation functions followed by convolution operations. This deeper architecture with a larger sequence context enables the model to capture more complex splicing patterns compared to previous approaches.

For model training, we randomly split the training data into 80% training and 20% validation sets. The model was trained using the Adam optimizer with an initial learning rate of 0.0001. We utilized the ReduceLROnPlateau (PyTorch v2.2.2)⁷³ learning rate scheduler that monitored validation loss to adaptively adjust the learning rate during training. To ensure robust predictions, we trained 5 independent models using different random seeds and used the average of their predictions as the final output. Performance compared to SpliceAI can be found in Figure S1B.

MPNN training

Overview

The model was trained to predict relative isoform abundances across 30 distinct tissue types and brain regions by mapping a learned graph embedding to the target abundance values. Training was performed on mini-batches of transcripts, each encoded as a directed graph where nodes represent the sequence context surrounding the 3′ and 5′ splice sites, and edges represent exon connectivity as annotated in GENCODE v47. The model output is continuous, predicting log2-transformed relative isoform abundances. Chromosome 8 was held out as a test set, while all other autosomes were used for training and validation (X, Y, and M were excluded). The implementation was built with PyTorch Geometric v2.5.2⁷⁴ and PyTorch Lightning v2.3.0, with training tracked using Weights and Biases v0.17.0.

Architecture search

To optimize model performance, an architecture search was conducted to evaluate different configurations. Candidate models were trained with either 2, 3, or 4 stacked attention-based message-passing blocks, and comparisons were made between models using graph attention networks (GAT) and graph convolutional networks (GCN) to assess the added benefit of attention mechanisms. Global pooling strategies, including add, mean, and max, were benchmarked, along with the effects of concatenating versus averaging attention head outputs. Residual connections were also tested to evaluate whether propagating information from previous layers improved performance. All models were trained for 50 epochs during architecture benchmarking.

The final selected architecture consisted of three sequential components: (1) three blocks of graph attention layers with residual connections; (2) a global pooling operation using maximum aggregation; and (3) a two-layer fully connected output head. Each attention block included two GAT layers followed by batch normalization, ReLU activation, and dropout. Residual connections were applied in the second and third attention blocks to enhance information flow across layers. After attention-based message passing, node embeddings were pooled to produce a graph-level embedding, which was then processed by a feedforward head consisting of linear layers, dropout, and non-linear activations to generate a prediction for each transcript across the 30 tissues.

Hyperparameter sweeps

Hyperparameter sweeps were conducted using Bayesian optimization via the Weights and Biases sweep utility. The search explored a range of learning rates (from 0.001 to 0.000001), hidden channel dimensions (180, 220, 320, 440, 512, 840, 932), dropout rates (0.1–0.5), batch sizes (32 and 64), and attention head counts (2, 4, and 8). For each sweep configuration, 80% of transcripts (excluding those on chromosome 8) were used for training and 20% for validation, with a maximum of 15 training epochs.

The final model was trained using the selected hyperparameters: learning rate of 0.0002646, hidden channel size of 512, dropout rate of 0.5, batch size of 64, and 2 attention heads. Training was performed on an NVIDIA V100 GPU and took approximately 8 h. To enhance the model’s learning on challenging examples, hard example mining was incorporated during training. Specifically, a weighted hard loss was computed using the TripletMarginLoss function in PyTorch v2.2.2 for each batch, and this was combined with the mean squared error loss to form the total training objective:

$$L_{total}=MSE+\lambda ·L_{triplet}$$

Where:

$$MSE=\frac{1}{n}\sum _{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2$$

$$L_{triplet}=\mathit{max}\left\{d\left(a_i,p_i\right)-d\left(a_i,n_i\right)+margin,0\right\}$$

$$d\left(x_i,y_i\right)=\Vert x_i-y_i|{|}_2$$

In this equation, MSE represents the mean squared error between the predicted value $\left(\widehat{y_i}\right)$ and the experimental measured value $\left(y_i\right)\text{.}$ The triplet loss term $L_{triplet}$ forces the model to learn the patterns where similar transcripts are closer together and dissimilar ones are further apart. We set $\lambda$ = 0.25 and $margin$ = 1.0. For each training batch, anchor $\left(a_i\right)$, positive $\left(p_i\right)$, and negative $\left(n_i\right)$ examples were selected based on transcript expression patterns, with L2 norm as the distance metric $d$.

Ablation study

Ablation models were trained for each regulatory feature category: RBP features (Seqweaver), chromatin features (Sei), and splicing features (ConvSplice). Notably, for each ablation model, the relevant feature subset was masked to zero, ensuring that the total embedding dimensionality remained unchanged.

Variant effects on isoforms

Variant effect prediction

To quantify how variants impact differential isoform abundance, we extended the model to incorporate cis-regulatory variant effects. This was done by introducing sequence mutations and updating all node features using ConvSplice, Seqweaver, and Sei, allowing the model to capture functional disruptions affecting splicing and abundance regulation. Abundance predictions were made for both reference and alternative graphs, and isoform-level variant effects were defined as the log-fold change between the alternative and reference predictions per tissue: $s={\mathit{log}}_2\left(\frac{2^{{pred}^{alt}}+1}{2^{{pred}^{ref}}+1}\right)$.

Variant case studies

We also conducted case studies on select clinical variants.^41,60 Isoform variant effect scores were scaled to the background distribution described above, and L2 norm distances were calculated between reference and alternative node attribute vectors to identify the most impacted graph nodes. Features with the highest predicted change in these nodes were reported to highlight the underlying regulatory mechanisms.

eQTL directionality

The accuracy of alignment between predicted and observed aFC directions was computed at each threshold, with variability estimated using 1,000 bootstrap replicates (sampling variants with replacement to generate bootstrap samples of the same size as the original dataset).

Quantification and statistical analysis

Model performance and evaluation

Abundance binarization

To assess the model’s ability to distinguish between low and high transcript expression across tissues, we binarized transcript abundances using tissue-specific percentile thresholds. Within each tissue, transcripts above the 70th percentile of relative expression were labeled as high abundance (positive examples), and those below the 30th percentile were labeled as low abundance (negative examples), based on ground-truth data. Percentile cutoffs were computed separately for each tissue according to its expression distribution. Each isoform’s relative abundance (Rᵢ) was compared to these thresholds and labeled accordingly, yielding one categorical label per tissue that reflects the isoform’s expression level relative to other isoforms within that tissue.

Predicted transcript abundances were ranked and scaled to a (0,1] interval by normalizing against the total number of transcripts. AUROC was then computed for each tissue. Model performance on this binarization task was assessed using the chromosome 8 holdout set from the Gao et al. dataset (n = 2,206; mean AUROC = 0.835, s.d. = 0.021), along with validation on chromosome 8 transcripts from the Glinos et al. (n = 2,243; mean AUROC = 0.737, s.d. = 0.029) and Leung et al. (n = 637; AUROC = 0.657, s.d. = 0.033) validation datasets. Tissue-specific thresholds were computed separately for each dataset, and tissues with equivalent Otari tissues were kept for evaluation. This included eight tissues in Glinos et al. and two tissues in Leung et al.

Evaluation on novel isoforms

The model was also evaluated on its ability to predict abundance levels of novel unannotated isoforms. A total of n = 196,283 novel isoforms from the Gao et al. dataset were included in this analysis, with transcript graphs generated using available exon annotations from Gao et al. Predictions for novel isoforms were compared to measured abundances from the study, and Spearman’s correlations were computed for each tissue (median Spearman’s correlation = 0.376, IQR = 0.050). In comparison, we assessed performance on lower-abundance canonical transcripts, defined as those with expression levels below the 80th percentile in each tissue (median Spearman’s correlation = 0.369, IQR = 0.044). More statistical details can be found in the Figure S6 legend.

Global- versus isoform-level predictions

To compare gene- and isoform-level performance, we computed AUROC scores at both levels for n = 2,206 transcripts. We compared isoform-specific performance to a baseline that assigns each gene the maximum predicted expression among its isoforms (max-predicted isoform). These gene-level maxima were evaluated against ground-truth binarized isoform expression across all transcripts (mean AUROC = 0.766, s.d. = 0.026). We additionally computed a gene-level baseline by summing predicted isoform abundances for each given gene (mean AUROC = 0.719, s.d. = 0.029). Isoform-specific AUROC scores were computed as described previously, by comparing predicted and true binarized abundances at the individual transcript level (mean AUROC = 0.835, s.d. = 0.021).

Transcript complexity

We stratified transcripts on chromosome 8 from the Gao et al. holdout test set (n = 2,206 transcripts) into ten bins based on exon count, including bins for transcripts with 1 through 9 exons and a final category for those with 10 or more exons. For each category, we reported the range of tissue-specific AUROC scores to understand how prediction performance is affected by exon count. Mean AUROC by bin in ascending order: bin 1 = 0.749 (s.d. = 0.052), bin 2 = 0.894 (s.d. = 0.044), bin 3 = 0.896 (s.d. = 0.030), bin 4 = 0.904 (s.d. = 0.025), bin 5 = 0.867 (s.d. = 0.035), bin 6 = 0.850 (s.d. = 0.030), bin 7 = 0.813 (s.d. = 0.021), bin 8 = 0.855 (s.d. = 0.038), bin 9 = 0.765 (s.d. = 0.032), bin 10 = 0.781 (s.d. = 0.021).

Performance within genes

We computed the Spearman’s correlation between predicted and observed isoform abundances for all genes with at least two isoforms in the Gao et al. holdout test set (Spearman’s r mean = 0.496, std = 0.029). This intra-gene analysis was performed across n = 2,649 transcripts with available data.

AlphaGenome benchmark

We benchmarked Otari against AlphaGenome on the Gao et al. chromosome-8 holdout set (n = 2,206 transcripts) by computing transcript-specific per-tissue Spearman’s correlations with ground truth data across 19 shared tissues. Analyses were restricted to protein-coding and long non-coding RNA transcripts. AlphaGenome was run for CAGE by centering a 1,048,576-bp window (auto-selected max supported length) on each transcript TSS, predicting CAGE tracks restricted to tissue biosamples in a curated list, and taking the promoter signal as the max within ±500 bp of the TSS. For RNA-seq, the 1,048,576-bp window was centered on the transcript midpoint, and predictions were aggregated as the sum over exonic bases and normalized for transcript length. If multiple CAGE/RNA-seq tracks exist for a tissue biosample, the final predictions were averaged over the tracks. Medians and IQRs: AlphaGenome CAGE (0.540/0.025), AlphaGenome RNA-seq (0.541/0.031), Otari (0.587/0.017). We compared the distributions of tissue-specific Spearman’s correlations between Otari and AlphaGenome and computed statistical significance using two-sided Mann-Whitney U tests followed by Benjamini-Hochberg multiple-hypothesis correction. More statistical details can be found in the Figure S7 legend.

Variant effects on isoforms

Evaluation on HGMD variants

We evaluated the model on 2,270 clinical variants from HGMD⁴⁰ (release 2024.1), focusing on disease-associated functional polymorphisms (DFPs). DFPs are primarily noncoding variants found to have a statistically significant association with a clinical phenotype such as gene expression or splicing. We retained only autosomal noncoding SNPs within gene bodies, 2kb of the transcription start site (TSS), or 2kb of the transcription end site (TES) (n = 2,270). Variant effect scores were computed for each isoform by comparing predictions on reference and mutated graphs and summarized per gene by taking the maximum absolute score across tissues and isoforms. Principal transcripts were defined using APPRIS “PRINCIPAL:1” annotations, and only the most affected principal transcript per gene was retained. For benchmarking, we compared pathogenic variant effect sizes to a control distribution, comprising the combined distribution of benign ClinVar variants and all SPARK sibling de novo variants passing filtering criteria (n = 78,369 variants, n = 304,851 transcript-variant combinations). All effect sizes were also Z score normalized to this background distribution. Statistical comparisons between groups were conducted using one-sided independent t-tests with Benjamini-Hochberg correction. Visualization was performed to determine if data met assumptions of the statistical method and showed no major deviations. More statistical details can be found in the Figure 4 legend.

Fine-mapped QTLs

Datasets

We analyzed fine-mapped eQTLs and sQTLs identified using SuSiE across 49 GTEx v10 tissues.⁴² Unless otherwise specified, analyses focused on six representative tissues with available models: brain cortex, liver, whole blood, lung, heart, and colon. All statistical tests were one-tailed independent t-tests, corrected for multiple comparisons using the Benjamini–Hochberg method. Visualization was performed to determine if data met assumptions of the statistical method and showed no major deviations. More statistical details can be found in the Figure 3 legend.

eQTL directionality

To assess directionality, we compared Otari’s aggregated transcript-level predicted effects (normalized to the background distribution described above) to the known allelic fold-change (aFC) direction of n = 4,181 tissue-specific eQTLs. Only single-nucleotide variants within 2 kb of the TSS were retained for this analysis. For normalized summed-score thresholds ranging from 0 to 125 (step size = 2), we summed Otari’s predicted isoform effects for each unique variant, gene, and tissue combination (for the matched tissue model scores) and filtered variants according to the threshold. The accuracy of alignment between predicted and observed aFC directions was computed at each threshold, with variability estimated using 1,000 bootstrap replicates (sampling variants with replacement to generate bootstrap samples of the same size as the original dataset). This analysis spanned ten tissues and brain regions, including brain cortex, hippocampus, cerebellum, liver, whole blood, pancreas, heart, lung, kidney, and colon.

Transcript-specific sQTL effects

For each sQTL (n = 7,412 variants), transcripts were split into overlapping (exons intersecting or bordering the sQTL spliced region) and non-overlapping sets. We computed the predicted effect on transcript abundance for each set within the relevant matched tissue, enabling a direct comparison of sQTL impact across isoforms with one-sided independent t-tests followed by Benjamini-Hochberg correction. Only single-nucleotide variants within genic regions, 2 kb of the TSS, or 2kb of the TES were retained. More statistical details can be found in the Figure 3 legend.

Estimating tissue-specific burden

For each QTL-overlapping isoform in a sample of n = 15,914 QTLs, we identified the most perturbed graph node, defined by the largest L2 norm distance between reference and alternative node attributes, and extracted the top 10 disrupted features. Feature contributions were weighted by the sum of predicted absolute QTL effects in the matched tissue and normalized by variant count to enable cross-tissue comparison. The top 750 features in each tissue variant set were retained and annotated to functional categories based on their model: Sei (transcriptional) or Seqweaver (post-transcriptional). ConvSplice-derived features were excluded due to the small feature space. Variants shared between the eQTL and sQTL tissue-matched sets were removed in this analysis, and min–max normalization was applied to Sei and Seqweaver feature counts separately across tissues. For this analysis, single-nucleotide variants within genic regions, 2 kb of the TSS, or 2kb of the TES were retained for both eQTL and sQTL sets. Mean z-scores for eQTLs = 0.35/0.65 (s.d. = 0.25). Mean z-scores for sQTLs = 0.92/0.08 (s.d. = 0.07).

Benchmarking on sQTL causality

Using SuSiE fine-mapping, variants with PIP ≥0.9 were labeled positives and variants with PIP ≤0.1 were labeled negatives; ambiguous variants were excluded. We performed tissue-matched evaluations in six tissues with available Otari models: Brain Cortex, Liver, Whole Blood, Lung, Heart Atrial Appendage, and Colon Sigmoid. For each variant–gene–tissue combination, Otari’s score was computed as the sum of absolute, tissue-specific isoform effect sizes across all transcripts of the gene (i.e., an isoform-resolved burden measure). For baseline methods, we followed the authors' recommendations. We used each tool’s canonical variant score (the maximum absolute Δ score for SpliceAI, the MTSplice tissue-specific predicted scores, and the maximum absolute Δ score from Pangolin), taking the maximum across affected splice sites per variant.

Isoform dysregulation in autism spectrum disorder

Datasets

The SPARK⁴³ cohort provides comprehensive phenotypic, clinical, and genetic data from autistic individuals and their families. We analyzed the genetic data of 3,507 probands and 2,206 non-autistic siblings from SPARK with available trios (mother, father, child) and whole genome sequencing (WGS). Any child with complete trio WGS and passing QC was included in these analyses. Measurements were taken from distinct samples. In total, we identified 241,276 de novo variants (DNVs) in SPARK probands and 150,668 in siblings occurring within genic regions, 2kb of the TSS, or 2kb of the TES. Within Satterstrom genes, we compared 1,200 DNVs from probands with 698 DNVs from siblings. We excluded ‘testis’ and ‘ovary’ tissues from subsequent analyses due to known sex-related biases in autism.⁷⁵

De novo variant calling

DNVs were identified using a pipeline from prior work⁷⁶ and applied to WGS data from Variant Call Format (VCF) files provided by the SPARK consortium. Briefly, the pipeline, which was optimized for WGS, uses HAT⁷¹ alongside variant calls from GATK HaplotypeCaller⁷² to detect variants present in the child (genotype 0/1 or 1/1) but absent in both parents (genotype 0/0). Variants were filtered based on quality metrics including read depth, genotype quality score, and genomic location. Regions with recent repeats, low complexity, or centromeric content were excluded. We used default thresholds of minimum depth = 10 and genotype quality (GQ) ≥ 20. Individuals with abnormally high DNV counts (>3 standard deviations above the mean across trios) were removed from further analysis, and non-singleton DNVs (those found in more than one family) were excluded. This approach yielded an average of 70.92 DNVs per proband (SEM = 0.25) and 70.44 per sibling (SEM = 0.31) in SPARK.

Assessing average and tissue-specific impact

We applied additional filtering to exclude indels and variants on chromosomes X, Y, and M. Otari was applied to the remaining SNPs from both probands and siblings. DNVs in SPARK were restricted to Satterstrom genes (genic regions, within 2kb of the TSS, or within 2kb of the TES). For each variant, we computed the maximum absolute effect across brain tissues for each isoform (bags of isoforms) to compare effect sizes between probands (n = 1,200 variants) and siblings (n = 698 variants) using a one-sided independent t test (p = 7.51 × 10⁻⁸). To assess tissue specificity, we computed the mean Otari-predicted effect per tissue per variant (bags of variants) in probands and siblings and then computed the log fold changes of the means. We compared the mean log fold change across brain tissues to that of non-brain tissues using a one-sided independent t test (p = 3.11 × 10⁻⁹). Variant effect sizes were Z score normalized to the combined distribution of benign ClinVar variants and all SPARK sibling de novo variants passing filtering criteria (n = 78,369 variants, n = 304,851 transcript-variant combinations). In analyses of Alzheimer’s disease–associated and schizophrenia-associated variants, de novo variants from non-autistic siblings were used as control variants. More statistical details can be found in the Figure 5 legend.

Microexons analysis

We applied Otari to de novo variants identified in SPARK probands (n = 57,929 DNVs) and siblings (n = 35,642 DNVs), restricting the analysis to variants mapping to brain-expressed genes (n = 13,774 genes). Brain-expressed genes were selected using the expression table from GTEx v7 (gene median transcripts per million per tissue). The genes selected were those whose expression in brain tissue was at least five times higher than the median expression across all tissues. Variants were located within genic regions or within 2kb of the TSS. Exons were classified as either microexons (3–27 nt) or long exons (>27 nt). We flagged proband and sibling variants within 650 nt of microexon splice sites (n = 650 and 355 transcripts, respectively), as well as proband and sibling variants within 650 nt of long exon splice sites (n = 63,786 and 38,794 transcripts, respectively). Otari-predicted effects were compared between these sets; effects were retained for the top impacted brain tissue per variant. One-sided independent t-tests were used for hypothesis testing, with Benjamini-Hochberg correction applied for multiple comparisons (FDR values: 0.044/6.005 × 10-4/7.800 × 10-3). Visualization was performed to determine if data met assumptions of the statistical method and showed no major deviations. More statistical details can be found in the Figure 5 legend.

Published: January 16, 2026