SPLAIRE: Improving Splice Site Usage Prediction with a Cell-Type-Optimized Model

In a preprint posted to bioRxiv on June 8, 2026, the lab reports SPLAIRE (SPLicing in AIRway Epithelium), a sequence-based deep learning model of RNA splicing trained on one of the largest paired RNA sequencing and genotyping datasets assembled to date: cultured human airway epithelial cells (HAECs) from more than 100 donors. The work, by co-first authors Matthew Runyan and Saumya Gupta with Yul Leshem, David Geller-McGrath, Congjian Liu, Aabida Saferali, Jennifer Dy, Predrag Radivojac, Yohannes Tesfaigzi, Peter Castaldi, and Ayan Paul, makes two contributions: a model that improves quantitative prediction of splice site usage over existing tools, and a systematic evaluation of the current generation of sequence-based splicing models.

Background

Alternative splicing affects more than 95% of human protein-coding genes and is a principal source of isoform and proteomic diversity; genetically altered splicing accounts for a substantial fraction of human disease. Deep learning models that predict splice sites from pre-mRNA sequence, including SpliceAI, Pangolin, and SpliceTransformer, report high aggregate performance on this task. The study examines the composition of the datasets underlying those metrics and shows that they are dominated by high-usage, constitutively spliced sites. Performance on low-usage splice sites, tissue-specific splice sites, and continuous quantification of usage is substantially weaker across all evaluated models.

The central quantity in the study is splice site usage (SSU): the fraction of reads at a position that support a splice event, ranging from 0 (never used) to 1 (constitutively used). SSU is quantified directly from RNA-seq with SpliSER and serves as the regression target for SPLAIRE. Its empirical distribution is strongly bimodal, with density concentrated near 0 and near 1, so aggregate metrics computed over such a distribution carry limited information about performance in the intermediate and low-usage ranges.

Data and Model Architecture

The training strategy prioritizes repeated observation of the same splice sites across individuals over broad coverage across tissues or species. RNA-seq from primary HAECs cultured in basal media provides the training data, with 10 donors held out for evaluation. A companion model, SPLAIRE-Var, is architecturally identical but is trained on pre-mRNA sequences containing each individual’s genetic variants rather than the reference genome.

The architecture follows the dilated residual convolutional design of SpliceAI and Pangolin. One-hot encoded sequence is projected to 32 channels by a 1×1 convolution and processed through 16 residual blocks arranged in four groups of four, with kernel sizes of 11, 11, 21, and 41 and dilation rates of 1, 4, 10, and 25 across successive groups. The resulting receptive field integrates sequence information up to 5,000 bp from each predicted position. The model produces two jointly trained outputs: a classification head that labels each central nucleotide as acceptor, donor, or neither, and a regression head that predicts continuous SSU. The parameter count is on the order of a few million, which permits fast training and inference and supports attribution analysis; genomic foundation models with hundreds of millions to billions of parameters underperform dedicated splicing models on this task.

Classification and Regression Performance

SPLAIRE was compared with SpliceAI, Pangolin, and SpliceTransformer on held-out, non-paralogous protein-coding genes from chromosomes 1, 3, 5, and 7 in five tissues: HAEC and GTEx lung, testis, brain cortex, and whole blood, with 10 held-out donors each.

For SSU regression, SPLAIRE achieved the highest R² in all five tissues (0.82 to 0.90), including R² = 0.90 on HAEC and 0.82 to 0.84 on GTEx tissues absent from its training data, compared with 0.78 to 0.82 for Pangolin, 0.73 to 0.79 for SpliceTransformer, and 0.70 to 0.80 for SpliceAI, whose classification probability was used as an SSU proxy since it has no regression head. Calibration analysis showed that SPLAIRE’s predictions tracked measured usage most closely, while SpliceAI systematically overpredicted SSU and Pangolin underpredicted it. On Pangolin’s own multi-tissue test set, SPLAIRE reached R² of 0.62 to 0.68 against 0.31 to 0.51 for Pangolin.

Classification results depended strongly on the choice of reference set. Against all GENCODE v45 splice sites, SpliceAI attained the highest AUPRC (0.96) with SPLAIRE at 0.95; SpliceAI was trained directly on these annotations. Restricted to MANE Select transcripts, SPLAIRE attained 0.95 while SpliceAI fell to 0.87. Against all splice sites observed in RNA-seq, including unannotated and low-usage junctions (28% to 51% of observed sites per tissue are absent from GENCODE), AUPRC fell to 0.68 to 0.77 for every model, with SPLAIRE highest in four of five tissues.

Stratification by usage decile quantifies the dependence of performance on SSU. Classification AUPRC declined with decreasing usage for all models; in the lowest decile (SSU 0 to 0.1), SPLAIRE and SpliceAI performed best at approximately 0.23 to 0.30, while Pangolin and SpliceTransformer fell below 0.15. Regression error was lowest at the SSU extremes and highest at intermediate values. The authors conclude that splice site prediction is not a solved problem, and that reported aggregate metrics primarily reflect performance on constitutive sites.

Tissue-Specific Prediction

Pangolin and SpliceTransformer include tissue-specific output heads. To assess whether these heads capture tissue-specific splicing, the authors computed the tissue specificity index τ and its reverse on Pangolin’s four-tissue test set, identifying 1,746 tissue-specific sites (τ ≥ 0.5 or τ_rev ≥ 0.7, with cross-tissue SSU range ≥ 0.3). On these sites, the tissue-matched head of each model was usually outperformed by at least one of the same model’s heads trained for a different tissue. DeepLIFT-SHAP attributions computed at brain-specific sites across all four Pangolin tissue models were nearly identical (median pairwise Pearson r = 0.99), indicating that the four models extract the same sequence features and weight them equivalently. Tissue specificity in these architectures is therefore encoded almost entirely in the final 1×1 convolution that maps a shared representation to per-tissue outputs, rather than in the learned sequence features themselves.

Variant Effect Prediction

Variant effects were evaluated in two settings. In massively parallel reporter assays, predicted ΔPSI was computed as the alternate-minus-reference change averaged over the tested exon’s splice sites. On Vex-seq (2,055 variants across 110 alternatively spliced exons), SPLAIRE-Var achieved the highest Pearson correlation and AUPRC (0.60, 0.36), followed by SPLAIRE (0.59, 0.35); on MFASS (27,733 SNVs across 2,198 constitutive exons), Pangolin performed best (0.54, 0.35). Training on personal genotypes produced only marginal improvement over the reference-trained model. Across all models, accuracy was concentrated on variants within a few nucleotides of a splice junction and declined for deeper exonic and intronic positions, and all models overpredicted effects for neutral variants near junctions.

In the endogenous genomic context, the models were evaluated on fine-mapped sQTL credible sets from GTEx V8 across 49 tissues, classifying high-confidence causal variants (posterior inclusion probability greater than 0.9) against non-causal variants matched by gene expression and distance from the splice site. Median AUPRC ranged from 0.76 to 0.80 across models, with SpliceAI at 0.79, Pangolin at 0.78, and SPLAIRE, SPLAIRE-Var, and SpliceTransformer at 0.76. Tissue-matched heads conferred little or no advantage over alternative heads in this task. Applied to unresolved credible sets (maximum PIP below 0.9, 2 to 10 variants), SPLAIRE reduced the candidate list to a single variant in 53% of two-variant sets and to three or fewer variants in 46% of ten-variant sets, indicating practical utility for genetic fine-mapping.

Feature Attribution

DeepLIFT-SHAP attributions on held-out splice sites show that both output heads concentrate importance on the splice site dinucleotide motifs, with the regression head distributing attribution more broadly across the input window; low-usage sites show weaker attribution at the splice motifs and the polypyrimidine tract. Stratifying donor sites by the +1/+2 dinucleotide, SPLAIRE maintained R² above 0.76 on non-canonical donors. The attribution pattern at GC donors matched that of canonical GU donors, while the rare AU donors elicited a distinct recognition pattern, indicating that the model learned a separate representation for this junction class from a small number of training examples.

Conclusions

The study supports two conclusions. First, training deeply on many individuals of a single cell type improves quantitative splice site usage prediction, and a compact dilated CNN outperforms both a transformer-based splicing model and general-purpose genomic foundation models on this task. Second, commonly reported performance metrics depend strongly on the usage composition of the evaluation set: low-usage and tissue-specific splicing remain open problems, and variant effect prediction, while useful for narrowing fine-mapped credible sets, exhibits systematic error modes shared across models. The authors argue that usage-aware, task-specific evaluation is as important to further progress as architectural development.