CellSpliceNet

Alternative splicing lets a single gene produce many transcripts, and nervous systems exploit it heavily: neighbouring neurons of different classes can include or skip the same cassette exon at very different rates. Predicting that rate — the percentage-spliced-in (PSI) value — requires reasoning simultaneously about sequence motifs, the RNA's folded structure, and which splicing factors a particular cell happens to express. Most splicing predictors see only one of those signals.

CellSpliceNet is an interpretable multimodal transformer that predicts cell-type-resolved PSI in Caenorhabditis elegans. It was developed by Smita Krishnaswamy's group at Yale University with collaborators at Vanderbilt University and New York University, and posted as a preprint in June 2025. The model fuses four views of each splicing event — the long-range genomic sequence, a local region of interest (ROI) in the RNA, the predicted secondary structure, and the neuron's gene expression profile — through a masked cross-modal attention scheme adapted from the Zorro framework, so that each modality keeps its own representation while still being able to condition on the others.

Where RNA language models such as SpliceBERT learn splice-site grammar from sequence alone, CellSpliceNet is built around the observation that the same sequence produces different splicing outcomes in different cells, and that the difference is carried by the expression modality.

#Key Features

• Four-modality fusion: Genomic sequence, RNA region of interest, secondary structure, and gene expression are embedded by modality-specific encoders and combined by a multimodal multi-head attention block that permits selective, not indiscriminate, cross-modal mixing.
• Cell context as input: Neuron-specific splicing-factor expression is encoded as a graph and injected into the model, allowing one trained network to serve every neuron class rather than requiring a separate model per cell type.
• Attention-based interpretability: Pooling weights are read out directly, and they concentrate on intron–exon boundaries, single-stranded loop nucleotides inside the exon, and a small set of high-impact splicing factors including smu-1, unc-75, and hrp-1.
• Graph scattering embeddings: Secondary structures and splicing-factor co-expression networks are converted to graphs and encoded with the geometric scattering transform, giving fixed-length, permutation-invariant descriptors.
• Released checkpoint and dataset: A pretrained CellSpliceNet.pth checkpoint, preprocessing scripts, and the curated dataset are public, alongside a training configuration for human GTEx data.

#Technical Details

Full-length RNA sequences are tokenised into patches carrying both base identity and exon/intron/flanking annotations, avoiding the four-token vocabulary problem of nucleotide-level models while retaining long-range context. Secondary structures come from ViennaRNA and are encoded with geometric scattering; the expression modality builds a mutual-information graph over 243 curated splicing factors from CeNGEN transcriptomes covering 50 isolated neuron types, with counts imputed by MAGIC. Exon annotations are drawn from WormBase. The concatenated modality embeddings feed a multilayer perceptron trained end-to-end under squared-error loss on a single A100 GPU, with a 65/15/20 train/validation/test split, k-fold cross-validation, and ten independent seeds.

On held-out data the full model reaches a Spearman correlation of 0.88, against 0.74 without the sequence modality, 0.81 without the ROI, 0.82 without structure, and 0.84 without expression. It beats a general-purpose vision transformer baseline by 0.07 in Spearman and Pearson correlation and by 0.11 in R², with wider margins over the domain-specific predictors SpliceTransformer, SpliceFinder, SpliceAI, and Pangolin. Accuracy varies by neuron class: ASI (0.97), IL1 (0.96), and AIM (0.95) are predicted best, DVB (0.66) worst, with most classes above 0.80.

#Applications

The model is aimed at researchers studying neuron-specific post-transcriptional regulation, where it can nominate the sequence elements and splicing factors that drive a given exon's inclusion in a given cell class. Its attention maps make it usable as a hypothesis generator for antisense oligonucleotide design, for prioritising candidate splice-altering variants, and for charting cell-type-specific splicing networks. The shipped GTEx configuration extends the same architecture to human tissue data.

#Impact

CellSpliceNet is a demonstration that cellular context, supplied as an expression modality, is worth as much as sequence for splicing prediction — its largest single-modality ablation loss after sequence itself. The masking architecture is generic and transferable to other multimodal genomics problems that pair sequence with epigenome or expression tracks. Its scope is nonetheless narrow: training and evaluation are confined to the C. elegans nervous system, the work remains a preprint awaiting peer review, and the code carries a Yale non-commercial licence that requires a separate agreement for commercial use.