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

Conditional generators for regulatory DNA design typically condition on a coarse label or a single activity score, collapsing a high-dimensional target onto one axis. We instead condition directly on a multi-assay functional fingerprint, treating design as a continuous, structured conditional generation problem, and release RegFP, a benchmark for it. RegJEDI is an absorbing-state discrete diffusion language model over DNA with a goal-separated denoiser: a goal-blind prefix processes the masked sequence and diffusion time, while a conditioned suffix injects the target fingerprint $P \in \mathbb{R}^{C \times A}$ via adaLN-Zero modulation and cross-attended profile tokens, with no reward model, reranking, or classifier guidance. We additionally introduce fREPA, a representation-alignment auxiliary loss that aligns intermediate features to a frozen target encoder during training only, removed at sampling. Against eight baselines spanning label-conditioned diffusion, reward-optimized generation, and fingerprint-conditioned controls, evaluated on held-out targets by an independent judge, RegJEDI achieves the best overall target fidelity and leads on most evaluation axes. Compared against an identically fingerprint-conditioned control at matched compute, fREPA alone yields a consistent, statistically reliable gain across most cell contexts — a cheap, general training-time lever for conditional generation on structured, high-dimensional targets with no added sampling cost. Evaluation is by an in-silico surrogate rather than wet-lab validation, and the mechanism behind the gain remains underdetermined.