AbSplice2

Rare variants that disrupt splicing are a substantial cause of Mendelian disease, but whether a variant matters depends on where and when the affected splice site is used. Splicing is strongly tissue-specific, and inclusion levels of many exons shift over organ development — in brain and testis, up to 20% of cassette exons are developmentally regulated. Splicing annotations used by variant-effect predictors are derived almost entirely from adult tissue, so variants whose effect is confined to embryogenesis or childhood are systematically underestimated.

AbSplice2 is the second generation of the AbSplice framework, developed in the Gagneur lab at the Technical University of Munich with Helmholtz Munich, the University of Oxford, and the Broad Institute, and posted as a preprint in July 2025. Like the original AbSplice (Nature Genetics, 2023), it estimates the probability that a rare variant causes an aberrant splicing event in a given tissue by combining sequence-based deep learning predictions with SpliceMaps — tissue-specific tables of splice site usage measured from RNA-seq.

The three rebuilt components are a more accurate ground truth from the aberrant splicing caller FRASER2, replacement of SpliceAI with the directional gain and loss scores from Pangolin, and a continuous rather than binarized representation of splice site usage. Together they double the precision of the original model, while SpliceMaps derived from developmental RNA-seq let it score a variant separately at each of 15 stages from four weeks post conception to adulthood.

#Key Features

• Developmental resolution: SpliceMaps for 7 organs — brain, cerebellum, heart, kidney, liver, ovary, and testis — at 15 timepoints from embryogenesis to adulthood let the model report how a variant's predicted effect changes with age.
• Directional deep learning features: Pangolin's signed splice-site-usage gain and loss scores, plus the position it predicts is affected, replace SpliceAI's maximum delta score for a further 9% gain in average precision.
• Continuous splice site usage: Replacing the binary threshold on split-read support with a learned sigmoid-shaped effect improved average precision by about 41% and removed abrupt prediction shifts caused by small fluctuations near the old cutoff.
• Refined ground truth: Benchmarking against FRASER2 outlier calls, which reduce false positives by about an order of magnitude relative to FRASER, doubled recall without materially changing the learned feature contributions.
• Ready-to-use predictions: Precomputed scores for every single-nucleotide variant in protein-coding genes across all 49 GTEx tissues are on Zenodo, and a web interface scores SNVs and indels at chosen tissues and timepoints.

#Technical Details

The predictor is a glass-box explainable boosting classifier (an InterpretML gradient-boosted generalized additive model), trained with sample-grouped cross-validation on rare variants linked to FRASER2 splicing outliers across 49 GTEx tissues. Its features are the MMSplice + SpliceMap score, the MMSplice + SpliceMap + reference-Ψ score, Pangolin's gain and loss scores, and continuous splice site usage — the median split-read count for the affected donor or acceptor. Developmental SpliceMaps come from RNA-seq across 7 tissues and 15 stages; because only 1–5 samples exist per tissue-timepoint pair, adjacent timepoints are pooled with a sliding window, supported by downsampling experiments in GTEx. AbSplice2-RNA, which additionally consumes splicing outliers called from a clinically accessible tissue such as blood or fibroblasts, adds about 21% average precision. Gains replicate on 230 BrainGVEx individuals with paired genomes and frontal cortex RNA-seq, and substituting AbSplice2 into the aberrant underexpression predictor AbExp improved its average precision by about 7%.

#Applications

The model annotates VCFs through a Snakemake workflow shipped with the code, or scores one-off variants including indels through the web interface. Its intended use is rare disease diagnostics: prioritizing splice-disrupting variants in the tissue and developmental window relevant to a patient's phenotype rather than taking a maximum score across all tissues, which the authors show reduces precision. Applied to 64,736 rare disease participants in the UK National Genomics Research Library, it surfaced 26 variants in confidently disease-linked genes with stronger predicted effects during development than in adulthood, including a candidate new diagnosis in FGFR1. SpliceMaps can be rebuilt from a handful of RNA-seq samples, extending the framework to tissues the released maps miss.

#Impact

Adding a developmental axis to splicing prediction identifies a variant class that adult-only annotation misses: 18,206 variants in brain are predicted to disrupt splicing early in development but not in the adult, and such variants are 6.8-fold enriched in the most loss-of-function-intolerant decile of genes. Precomputed scores, developmental SpliceMaps, trained model files, and a GPL-3.0 codebase make the results directly reusable, though the developmental maps rest on small sample sizes covering only seven organs, the framework cannot recover an event the underlying sequence models miss, and the preprint has not yet been peer reviewed. The authors position further progress as coming from sequence-based models such as Borzoi and AlphaGenome, or self-supervised RNA models like SpliceBERT, trained on broader tissue and developmental panels.