CellSpliceNet: Interpretable Multimodal Modeling ofAlternative Splicing Across Neurons in C. elegans

Abstract

Alternative splicing profoundly diversifies the transcriptome and proteome, but decoding its regulatory mechanisms remains a challenge. We introduce CellSpliceNet, an interpretable transformer-based multimodal deep learning framework designed to predict splicing outcomes across the neurons of C. elegans. By integrating four complementary data modalities, namely long-range genomic sequence, local regions of interest (ROIs) in the RNA sequence, secondary structure, and gene expression, CellSpliceNet captures the complex interplay of factors that influence splicing decisions within the cellular context. CellSpliceNet employs modality-specific transformer embeddings, incorporating structural representations guided by mutual information and scattering graph embeddings. To this end, a novel and carefully designed multimodal multi-head attention mechanism preserves the integrity of each modality while facilitating selective cross-modal interactions, notably allowing gene expression data to inform sequence and structural predictions. Attention-based pooling within each modality highlights biologically critical elements, such as canonical intron-exon splice boundaries and accessible single-stranded RNA loop structures within the exon. Quantitative comparisons with current state-of-the-art methods demonstrated CellSpliceNet s superior predictive accuracy (Spearman{rho} = 0.88) and high accuracy across diverse neuron subtypes. Furthermore, CellSpliceNet elucidates a hierarchical, neuron-specific splicing code by preferentially weighting upstream enhancer motifs (e.g., GGAAGAAC) and identifying neuron-class-specific splicing-factor signatures, including smu-1, unc-75, and hrp-1. Thus, CellSpliceNet not only advances the frontiers of alternative splicing predictive capabilities but also provides mechanistic insights into the multimodal regulation of alternative splicing.