Spliformer-v2

Roughly 15% of pathogenic mutations act by disrupting pre-mRNA splicing, and the splicing code itself differs between tissues — an intron retained in liver may be spliced out in cortex. Sequence-based predictors such as SpliceAI and its open reimplementation OpenSpliceAI score splicing effects without tissue context, while tissue-aware models like Pangolin and SpliceTransformer collapse the entire central nervous system into a single "brain" label. That granularity is a poor match for neurological disease genetics, where the relevant unit is a specific cortical region or the spinal cord.

Spliformer-v2, released as a bioRxiv preprint in July 2025 by Ming Zhang's group at Tongji University, addresses that gap. It fine-tunes the SegmentNT genome-annotation model — which reuses the DNA encoder of Nucleotide Transformer v2-500m-multi-species and swaps the masked-language-modeling head for a 1D U-Net segmentation module — on paired whole-genome sequencing and RNA-seq from 18 human tissues: 11 brain regions plus cervical spinal cord, and six peripheral tissues. Rather than classifying whether a position is a splice site, it regresses continuous donor and acceptor usage at single-nucleotide resolution, so a variant is scored by how much it shifts quantitative site usage in a given tissue.

The model succeeds Spliformer, the same group's attention-based splicing predictor used to link splicing variants to amyotrophic lateral sclerosis (ALS). Where v1 predicted splicing from reference sequence alone, v2 takes genotypes directly and distinguishes heterozygous from homozygous effects.

#Key Features

• Brain-region resolution: Separate models for amygdala, anterior cingulate cortex, caudate, cerebellum, frontal cortex, hippocampus, hypothalamus, nucleus accumbens, putamen, substantia nigra, anterior lateral motor cortex, and cervical spinal cord, alongside blood, heart, kidney, liver, lung, and skeletal muscle.
• Haplotype-aware input: Each prediction consumes two haplotype sequences concatenated with special tokens, letting the model separate heterozygous from homozygous variant effects rather than treating every allele as homozygous.
• Usage regression, not site classification: Outputs continuous splice-site usage, trained against SpliSER-quantified junction usage from RNA-seq, giving delta-usage scores that are directly interpretable as a change in isoform fraction.
• VCF-in, VCF-out inference: A command-line tool annotates variants with acceptor and donor usage gain/loss scores and their distances from the variant, for GRCh37 or GRCh38.

#Technical Details

Each per-tissue model holds approximately 559 million parameters: a 24-layer BERT-style encoder over 6-mer tokens, topped by a U-Net head with two downsampling and two upsampling convolutional blocks (1,024 and 2,048 filters). Inputs are 720 bp windows centered on a splice site, with training loss computed only over the central 501 bp; the short window reflects both the tokenizer's multiple-of-six requirement and the observation that most splice-altering SNVs act within ±250 bp. Training data came from GTEx (20 controls, 16 tissues) and TargetALS (10 ALS cases and 10 controls, motor cortex and cervical spinal cord), yielding about 170,000 splice sites and 200,000 sequences per tissue and 3,014,910 SNVs overall, with chromosomes 1, 3, 5, 7, and 9 held out and paralogs excluded from the test set. Each tissue model trained for 20 epochs on a single A100 with AdamW and MSE loss.

On held-out chromosomes, Spliformer-v2 reported R² of 0.83–0.89 and Spearman ρ of 0.88–0.92 across tissues, against 0.10–0.41 and 0.52–0.64 for SpliceTransformer, with AUPRC of 0.90–0.95 versus 0.45–0.61 at a 0.5 usage cutoff. The comparison is favorable to the newer model by construction: SpliceTransformer accepts only one haplotype and has no brain-subregion models, so its general brain predictions stood in for every CNS tissue.

#Applications

The tool targets interpretation of noncoding and synonymous variants in neurological disease. Applied to ClinVar, predicted splicing variants were consistently enriched among pathogenic over benign variants across all 18 tissues, with anterior lateral motor cortex showing the highest pathogenic fraction among CNS regions. Applied to GWAS summary statistics, it nominated 34 of 2,158 ALS-associated SNVs and 129 of 20,141 Pan-UK Biobank neurological disease SNVs as tissue-specific splicing variants, including candidate mechanisms at CREB3, SCFD1, MAPT, and TOMM40.

#Impact

Spliformer-v2 is the first splicing predictor trained at the resolution of individual human brain regions and spinal cord, giving neurogenetics groups a way to ask where in the CNS a candidate variant might act. Its limitations are substantial and stated plainly by the authors: 20 individuals per tissue, no phasing information (so compound heterozygous effects within a window cannot be modeled), a 720 bp receptive field that excludes distal regulatory elements, and disease findings that are computational nominations awaiting minigene or RNA-seq validation. The preprint has not yet been peer reviewed, and while the GitHub repository ships the inference pipeline and GENCODE-derived annotations, the per-tissue checkpoints it loads are not distributed there and the code carries no license file.