CellSpliceNet: Explainable Multimodal Transformers for Splicing in C. elegans Neurons

Alternative splicing diversifies the transcriptome, yet its regulation remains difficult to decode. We present CellSpliceNet, an interpretable transformer model that predicts splicing outcomes across C. elegans neurons by integrating four modalities: long-range genomic sequence, local RNA regions of interest (ROIs), predicted RNA secondary structure, and cell-type–specific gene expression. Modality-specific encoders—including graph-signal scattering for structure and expression—feed a multimodal multi-head attention module that preserves per-modality signals while enabling expression-informed interactions with sequence and structure. Attention pooling highlights salient biology (e.g., splice boundaries and single-stranded loop regions) and enables deep model interpretability. On held-out data, CellSpliceNet outperforms strong baselines and achieves Spearman $\rho=0.88$, with robust accuracy across neuron subtypes.