RegJEDI: A Joint-Embedding Diffusion Language Model for Regulatory DNA Design

Regulatory DNA generators are commonly conditioned on cell-type labels or scalar activity measurements. Such targets identify a context or endpoint but under-specify the coordinated chromatin program a sequence should realize. We formulate regulatory design instead against measured multi-assay fingerprints that capture accessibility, histone state, and factor occupancy across cellular contexts. RegFP pairs cell-differential regulatory loci with these fingerprints under chromosome-held-out evaluation. RegJEDI is a fingerprint-conditioned masked diffusion language model that directly generates one sequence per target. During training, functional representation alignment (fREPA) aligns a condition-blind denoiser state with a frozen sequence-to-function teacher trained only on training chromosomes. The teacher is removed at inference and is distinct from the held-out evaluator. RegJEDI realizes requested fingerprints more faithfully than matched conditioning controls while preserving regulatory grammar and sequence realism; an independent accessibility model corroborates the intended cell-context specificity. These results establish measured functional fingerprints as a practical conditioning interface for direct regulatory DNA design.