Integration of massively parallel splicing assays across 43 cell lines with deep learning models enables both accurate prediction of cell type-specific alternative splicing patterns and generative design of synthetic sequences with programmable splicing activity, including sequences that selectively target cancer cells harboring splicing factor mutations.

Programmable control of gene expression in specific cell types is essential for both basic discovery and therapeutic intervention, yet current strategies lack scalability across diverse cellular contexts. Here, we introduce SPICE (Splicing Proportions In Cell types), an integrated experimental and computational framework that harnesses alternative RNA splicing as a programmable modality for cell type-specific gene regulation. To power SPICE, we constructed a massively parallel reporter assay (MPRA) comprising 46,372 human-derived sequences and profiled exon skipping across 43 cell lines spanning 10 lineages, uncovering widespread cell type-specific exon skipping. Using this data, we trained deep learning models that both predict splicing in unseen contexts and generate synthetic sequences with programmed, cell type-specific splicing patterns. Leveraging these models, we further engineered sequences that selectively splice in cells harboring oncogenic splicing factor mutations, demonstrating translational potential. SPICE provides a generalizable strategy for dissecting splicing regulation and engineering alternative splicing as a gene expression regulatory layer for research and therapeutic applications.

We introduce SPICE, an integrated framework that couples large-scale splicing assays with generative design to uncover regulatory principles and design programmable, cell-specific gene expression for research and therapeutic applications.