To go beyond prediction, we built Melange, a generative model that designs new sequences with programmed cell type-specific splicing. We experimentally validated that Melange can design sequences that splice only in neural-lineage cells, and not cells from other lineages! (7/10)

Using this dataset, we built Soma, a deep learning model that predicts how any RNA sequence will splice across cell types.

We drew inspiration from ChromBPNet (@anshulkundaje.bsky.social) to encode cell type-specific information directly from gene expression data. (6/10)

Good ML models require high-quality experimental data, but large-scale, cell type-resolved data on RNA splicing is limited. Hence, we built SPICE (Splicing Proportions In Cell types), an integrated experimental + AI framework for the generative design of cell type-specific RNA sequences. (4/10)