acceptodds
Under review as a conference paper at ICLR 2027

A Bidirectional Generative DNA Model with Motif-Inpainting Policy Optimization

Abstract

Early DNA foundation models adopted BERT-style training, achieving strong performance on sequence understanding tasks but offering limited support for generation. Recent autoregressive models enable DNA generation, yet their left-to-right factorization does not naturally reflect the bidirectional dependencies underlying gene regulation. We present D3LM, which unifies both capabilities through masked diffusion. D3LM retains the Nucleotide Transformer (NT) architecture while reformulating its training objective as discrete masked diffusion. It retains NT's sequence understanding capabilities while generating regulatory sequences that more closely resemble real DNA than those produced by the evaluated autoregressive baselines. Building on D3LM's bidirectional generation capabilities, we introduce Motif-Inpainting Policy Optimization (MIPO) for cell-specific regulatory sequence design. MIPO anchors regulatory motifs within high-scoring sequences and inpaints their local neighborhoods while preserving the outer context, using functional feedback to select improved candidates for reinforcement learning. Our results highlight masked diffusion as a promising framework for unified DNA understanding, generation, and functional design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
ai4sciecne, diffusion model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fDS9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces D3LM, a bidirectional generative DNA model that retains the Nucleotide Transformer architecture and replaces fixed-rate masked language modeling with a masked-diffusion objective. It supports both iterative sequence generation and representation learning for downstream genomic tasks. The authors also propose Motif-Inpainting Policy Optimization (MIPO), which uses diffusion RL by inserting cell-associated motifs into high-scoring sequences and regenerating their local neighborhoods. The authors evaluate unconditional generation, 18 genomic classification tasks, and cell-specific regulatory design, showing competitive results against the baselines.

Critical strengths
  • The paper explores masked diffusion as a framework for capturing bidirectional context in DNA sequences, achieving competitive performance on both sequence generation and downstream classification tasks. This demonstrates the potential of a single DNA model to support both generation and representation learning.

  • The paper introduces MIPO, a motif-guided RL strategy for functional DNA design. MIPO explores around promising sequences by inserting regulatory motifs and inpainting their local neighborhoods while preserving the surrounding context. Under matched scoring budgets and shared optimizer hyperparameters, it substantially improves predicted specificity in HepG2 and moderately improves SK-N-SH. The shuffled-motif ablation further supports the contribution of motif structure beyond additional proposal generation or nucleotide composition.

Critical weaknesses
  • MIPO selects high-scoring parent sequences and accepts inpainted proposals using the same reward model that supplies RL feedback and evaluates the final generated sequences. Consequently, the reported gains may reflect increasingly effective optimization of that predictor rather than improved biological function. Validation with an independent predictor or experimental assay is needed to assess whether these improvements generalize beyond the training reward model.

  • The specific benefit of bidirectional diffusion modeling is insufficiently demonstrated. D3LM retains the NT architecture and primarily changes its training objective, but the paper lacks a controlled comparison against an autoregressive variant using the same backbone. A matched NT baseline would also help isolate the contribution of diffusion training. Furthermore, the introduction claims that left-to-right autoregressive modeling cannot adequately capture downstream regulatory relationships, but the experiments do not directly test this limitation. This motivation requires stronger evidence.

Questions for authors
  • Could the authors report scores for ordinary generated sequences, selected parents, and their corresponding inpainted sequences? This would separate the gain from parent selection from the additional gain due to motif insertion. A training control that reuses 1. all genereted sequence + motif insertion, and 2. selected high-scoring parents without motif insertion would also help establish whether MIPO’s improvement over d1 comes from its motif-guided exploration or simply from emphasizing high-reward sequences.

  • Although the paper includes published NT-v2 classification results, could the authors add an NT baseline continued on the same data with the same training budget while retaining its original masked-language-modeling objective? Comparing downstream performance and generation quality, with an explicit sampling procedure for NT, would help isolate the contribution of diffusion training.

  • How does D3LM differ from D3 [1] and the Dirichlet Diffusion Score Model [2]? Comparisons using matched datasets, sequence lengths, and training budgets would help establish whether D3LM offers better generation quality or functional controllability.

  • How does D3LM with MIPO compare with existing RL-based DNA design methods, such as TACO [3], Ctrl-DNA [4], and GO-CRE [5]? A shared evaluation setting with matched scoring budgets, and clear reporting of motif priors and inference-time filtering, would make the relative benefits easier to assess.

References

[1] Sarkar et al. Designing DNA With Tunable Regulatory Activity Using Discrete Diffusion. bioRxiv, initial preprint 2024. https://doi.org/10.1101/2024.05.23.595630

[2] Avdeyev et al. Dirichlet Diffusion Score Model for Biological Sequence Generation. ICML, 2023. https://proceedings.mlr.press/v202/avdeyev23a.html

[3] Yang et al. Regulatory DNA Sequence Design with Reinforcement Learning. ICLR, 2025. https://arxiv.org/abs/2503.07981

[4] Chen et al. Ctrl-DNA: Controllable Cell-Type-Specific Regulatory DNA Design via Constrained RL. NeurIPS, 2025. https://arxiv.org/abs/2505.20578

[5] Ma et al. Reconstructing sequence-grammar trajectories enables interpretable and tunable cis-regulatory element design. bioRxiv, 2026. https://doi.org/10.64898/2026.07.10.737719

Reviewer_rdfwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

D3LM combines DNA representation learning and generation through masked diffusion, while MIPO uses motif-guided inpainting to improve policy optimization for regulatory design. Experiments evaluate generation fidelity, 18 classification tasks, and three cell types, showing competitive generation and representation performance and improved predictor-assessed specificity in two cell types.

Critical strengths

-MIPO puts bidirectional inpainting to concrete use in regulatory design. With equal scoring budgets, multiple seeds, and composition-preserving shuffled-motif controls, the HepG2 gain is convincing as an optimization result (Tables 2–3) and inference needs no motif hints or reward filtering.

  • Evaluation spans generation and representation learning, with useful ablations on initialization, tokenization, and sampling. Notably, scratch initialization favors generation while pretrained initialization favors classification, informative for adapting existing genomic models.
Critical weaknesses
  • Functional validation uses the same Malinois predictor that selected SFT data and supplied RL rewards (Appendix H). The gains thus show better optimization of that predictor, not transferable regulatory activity versus predictor exploitation. Independent evaluation matters most for the large HepG2 gain.

  • The evidence doesn't isolate the paradigm's claimed advantage: generation baselines differ in training length, budget, and sampling, and classification gains add data and compute without a matched continued-MLM control. The claim that autoregressive factorization cannot model downstream enhancers is also too strong - a left-to-right factorization can represent such dependencies in principle. What's established is the practical benefit of flexible conditioning and inpainting, not a fundamental expressive limit.

Questions for authors
  • Do MIPO's gains hold under an independent predictor or measured activity? Please also report sequence diversity and similarity to training data.
  • How does D3LM compare to continued MLM with identical initialization, data, and updates, and to an autoregressive control matched on training length and budget? These comparisons would materially change my assessment of the diffusion contribution.
Reviewer_8kuEOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents D3LM, a masked-diffusion DNA model intended to support both sequence understanding and generation. It also introduces Motif-Inpainting Policy Optimization for designing regulatory sequences through motif-centered inpainting and functional feedback.

Critical strengths

None

Critical weaknesses
  1. The methodological novelty appears limited. Masked diffusion has already been explored for DNA generation, and the policy optimization component appears to apply an existing approach to motif-centered design. The paper does not make a sufficiently clear case for what new technical insight this combination provides.

  2. The empirical results do not convincingly establish an advantage over existing approaches. In unconditional generation, the model underperforms the reported baseline, and the evaluation omits strong reference models such as Evo and Evo 2. Together with the limited methodological advance, this leaves the practical benefit insufficiently supported.

Overall, I recommend rejection because the technical novelty is limited and the empirical evidence does not sufficiently establish the benefit of the approach.

Questions for authors

Which parts of the model and policy optimization procedure are technically new relative to prior work? How should the unconditional-generation results be interpreted, and what comparisons with Evo or Evo 2 are available under comparable conditions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.