Mach-1, a long-context RNA foundation model trained on a comprehensive compendium of full-length transcriptome data from 26 cancer cell lines, learns the sequence grammar governing transcriptome architecture through splice-aware tokenization and StripedHyena architecture, enabling accurate zero-shot prediction of isoform abundance, splicing patterns, and variant effects, while also generating biologically realistic and functionally viable synthetic transcripts that recapitulate human transcriptome structure.

Linking DNA sequence to genomic function remains one of the grand challenges in genetics and genomics. Here, we present a large-scale compendium of single-molecule transcriptome sequencing of diverse cancer cell lines, revealing their isoform diversity and specificity. We used this compendium to build Mach-1, an RNA foundation model that learns how the nucleotide sequence of unspliced pre-mRNA dictates transcriptome architecture—the relative abundances and molecular structures of mRNA isoforms. By using the Striped-Hyena architecture, Mach-1 handles extremely long sequence inputs at nucleotide resolution (64 kilobase pairs), allowing for quantitative, zero-shot prediction of all aspects of transcriptome architecture, spanning isoform abundance, structure, and variant-induced splicing changes. To test both the interpretive and generative capabilities of Mach-1, we experimentally validated its learned regulatory grammar and predictions through perturbation of RNA-binding proteins nominated by Mach-1 to impact targeted splicing, precise CRISPR editing of variants of uncertain significance that the model predicted to alter splicing, and de novo transcript synthesis and expression in human cells. Together, this release establishes a new foundation for sequence-to-transcript modeling. Mach-1’s representations can be extended and fine-tuned across a spectrum of biological contexts, from variant interpretation to RNA engineering.