Locality-Controlled Out-of-Distribution Guidance for Robust Regulatory DNA Sequence Design

What it claims

LOCOGEN reframes sampling-time guidance for regulatory DNA diffusion as a problem of where guidance acts, not how strong it is. The paper (NeurIPS 2026, Wellcome Sanger Institute) first argues that single-reward evaluation hides reward hacking, then introduces a three-axis evaluation: property (predicted activity), novelty (authenticity), and plausibility (compositional deviation from natural grammar, measured as |ΔGini|). It identifies an "empty corner" of this space where all three hold but no baseline lands, and reports LOCOGEN reaches it.

How it works

Two modules on a frozen MDLM (masked discrete diffusion) backbone.

flowchart TD
  A[Training sequences + activity labels] --> B[Masked separator Sϕ]
  B -->|soft mask M| C[Invariant xinv]
  B -->|soft mask M| D[Environment xenv]
  C --> E[Predictor hψ, V-REx across k-means envs]
  D --> F[Frozen random encoder gω]
  E --> G[Frozen separator + centroids saved]
  G --> H[Sampling on frozen MDLM]
  H -->|motif positions| I[Protected: no OOD]
  H -->|background positions| J[OOD damping applied]

Module 1 learns a soft mask M that splits each sequence into an invariant part and an environment part. The invariant part must predict activity, and must do so equally well across three k-means-inferred environments (variance-risk-extrapolation, V-REx). The environment encoder is randomly initialized and frozen to stop it collapsing. Loss combines sufficiency, invariance, a density target ρ, and fused-lasso TV smoothing for contiguous blocks. The approach adapts invariant-environment decomposition from chemical graphs (GIL, Li et al. 2022).

Module 2 extends MOOD's OOD control from a single global coefficient γ to a per-position one. OOD damping flattens the background nucleotide distribution toward uniform, but only at positions marked environment and only in the latter half of sampling. Motif positions keep their learned distribution.

Evidence

Balanced score (P·N·Pl)^(1/3), each axis min-max scaled against an ideal reference (property max = pred-activity 8.0, novelty max = authenticity 1.0, plausibility min = |ΔGini| 0): LOCOGEN 0.758, DNA-Diff 0.642, MDLM+CG 0.591, SVDD 0.576, DRAKES 0.454, MDLM 0.233. Tested on HepG2, K562, SK-N-SH cell lines. The paper also claims to expose two concealed reward-hacking modes in prior methods: predictor-specific motif concentration and global compositional drift.

How to use it

Code at https://github.com/Lotfollahi-lab/LocoGen. The method needs a pretrained unconditional MDLM over fixed-length DNA and an activity guidance model; the separator and centroids are trained once and frozen, so no backbone retraining is required per objective.

Limits

The fetched text is truncated mid-sentence in the sampling section, so the full inference procedure and ablations are not visible here. Baseline positioning is the authors' own; DNA-Diffusion is framed as a non-optimizing reference rather than a direct competitor, which affects how its 0.642 compares. All results are on three specific cell lines, in silico, with no wet-lab validation reported in this excerpt. No licence is stated on the page.