acceptodds
Under review as a conference paper at ICLR 2027

Designing Cell-Type-Specific Regulatory DNA with Annealed Diffusion Beam Search

Abstract

Designing regulatory DNA with cell-type-specific activity is broadly relevant for cell engineering and gene therapy. An ideal AI-designed sequence should drive high expression in desired cell types, have minimal activity in undesired cell types, and adhere to the genome's natural regulatory grammar. Diffusion models have become powerful priors for DNA sequence generation. Inference-time search guided by predictive oracles can steer these models toward high specificity. However, activity scoring requires complete DNA sequences, which must be derived heuristically from partially masked intermediates during search. Positionwise maximum a posteriori (MAP) estimation provides an efficient heuristic for completing these intermediates and computing lookahead scores. Here, we report that these scores are unreliable early in denoising and their reliability improves non-uniformly as sequence context accumulates. Based on this observation, we developed Progress-Annealed Candidate Selection (PACS), a training-free method that schedules how strongly MAP scores influence selection within diffusion beam search. PACS standardizes scores within each parent's candidate pool and increases selection strength with the expected fraction of revealed nucleotides, allowing broader exploration early and a stronger preference for high-scoring sequences later. We provide a local KL-regularized interpretation of the selector, characterize its limiting cases, and establish large-pool consistency under stated assumptions. Across three human cell lines, PACS improves sequence fidelity by 12–16% relative to greedy selection while maintaining high predicted specificity. On an eight-cell-type immune benchmark, PACS outperforms most of the evaluated reinforcement learning, inference-time alignment, and iterative optimization methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
regulatory DNA design, masked diffusion models, inference-time search, beam search, cell-type specificity

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_mgPnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Progress-Annealed Candidate Selection (PACS), a training-free method for guiding masked diffusion models toward cell-type-specific regulatory DNA. Motivated by evidence that positionwise MAP lookahead scores provide unreliable candidate rankings early in denoising, PACS standardizes scores within each parent’s candidate pool and progressively increases selection strength as more nucleotides are revealed. The authors provide a local KL-regularized interpretation of the selector and establish large-pool consistency under stated assumptions.

Evaluation covers three human cell-line activity tasks and eight immune-cell accessibility tasks, using separately trained evaluation oracles, comparisons with greedy search and other design methods. PACS consistently improves motif-frequency, sequence-composition, and embedding fidelity relative to matched greedy search while retaining competitive predicted specificity.

Critical strengths
  1. The paper identifies that in diffusion search, MAP lookahead rankings are poorly informative early in denoising and improve as sequence context accumulates. PACS addresses this through a simple selection schedule, and its comparison with greedy search holds the comparable setups. This provides convincing evidence that changing candidate selection can improve the specificity–fidelity trade-off without retraining.

  2. PACS improves motif-frequency correlation, 3-mer correlation, and embedding fidelity over matched greedy search across all three cell lines and eight immune-cell targets, while maintaining competitive predicted specificity. Separately trained evaluation oracles and complementary motif-arrangement metrics strengthen the validation. This consistency supports the practical usefulness of the approach beyond a single design task.

Critical weaknesses
  • The broader comparison does not establish that the proposed selector is preferable to adequately resourced existing methods. PACS evaluates 2,048 candidates per parent and reports the best retained completion along each lineage, whereas the main SVDD baseline uses ten candidates and returns terminal outputs. Differences in search budget, candidate valuation, and output selection therefore accompany differences in method. Additional budget sweeps partially address this concern, but do not jointly control these factors.
Questions for authors
  • After accounting for Appendix E.3’s candidate-budget sweeps, which advantages over SVDD remain attributable to PACS’s selection rule? Are comparisons available using the same candidate valuation and output-reporting rule?
  • PACS improves sequence fidelity relative to greedy selection, but its specificity gains are target-dependent. Under what conditions should PACS be preferred over greedy selection? Can the main results establish whether PACS achieves better fidelity at comparable specificity, particularly for targets where specificity decreases?
3 comments
AuthorsOpenReview ↗

Response to Reviewer mgPn (1/2)

We thank the reviewer for their positive assessment and for recognizing the MAP-ranking diagnostic, the evidence for improving the specificity–fidelity trade-off without retraining, and the consistent fidelity gains across cell types. Our response adds (1) a comparison with SVDD-PM at matched oracle budgets and common terminal reporting and (2) evidence of improved fidelity at comparable specificity.

Re W1 & Q1: PACS improves the specificity–fidelity trade-off over SVDD at matched compute and terminal reporting

We agree that the main comparison combines differences in budget, valuation, and output reporting. To isolate the selection rule’s contribution, we compare SVDD-PM and PACS using the same oracle-evaluation budget and terminal outputs (last branch) for every method. We provide an extended comparison with SMC/TDS in our response to reviewer wxZH (W2).

Cell lineMetricSVDD-PMPACS
HepG2AlphaGenome MinGap ↑5.888 ± 0.1586.151 ± 0.083
HepG2Motif correlation ↑0.835 ± 0.0180.851 ± 0.024
HepG23-mer correlation ↑0.937 ± 0.0050.961 ± 0.004
HepG2NTv2 FBD ↓1.972 ± 0.0241.797 ± 0.055
HepG2Shannon entropy ↑1.980 ± 0.0011.981 ± 0.000
HepG2Mean edit distance ↑149.764 ± 0.056149.848 ± 0.026
HepG2Search time (s) ↓1703.033 ± 18.4541671.820 ± 8.644
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
K562AlphaGenome MinGap ↑6.038 ± 0.0656.251 ± 0.185
K562Motif correlation ↑0.854 ± 0.0020.867 ± 0.006
K5623-mer correlation ↑0.909 ± 0.0040.943 ± 0.003
K562NTv2 FBD ↓1.908 ± 0.0301.719 ± 0.014
K562Shannon entropy ↑1.980 ± 0.0001.980 ± 0.002
K562Mean edit distance ↑149.803 ± 0.028149.778 ± 0.108
K562Search time (s) ↓1702.700 ± 9.9741682.632 ± 15.601
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SK-N-SHAlphaGenome MinGap ↑5.136 ± 0.0594.938 ± 0.116
SK-N-SHMotif correlation ↑0.797 ± 0.0150.820 ± 0.026
SK-N-SH3-mer correlation ↑0.892 ± 0.0110.915 ± 0.015
SK-N-SHNTv2 FBD ↓2.446 ± 0.0912.186 ± 0.033
SK-N-SHShannon entropy ↑1.972 ± 0.0011.973 ± 0.001
SK-N-SHMean edit distance ↑149.224 ± 0.051149.306 ± 0.086
SK-N-SHSearch time (s) ↓1717.867 ± 21.7331675.205 ± 2.442

PACS reduces FBD relative to SVDD-PM by 8.9%, 9.9%, and 10.6% on HepG2, K562, and SK-N-SH, and improves motif-frequency and 3-mer correlations on all three targets. It increases specificity on HepG2 and K562 but decreases it on SK-N-SH, at comparable runtime. We have updated the main benchmark figure with these budget-matched results.

AuthorsOpenReview ↗

Response to Reviewer mgPn (2/2)

Re Q2: PACS offers better fidelity at comparable specificity

The intended application of PACS is to design regulatory DNA for targeted gene expression in gene therapies, which requires both high predicted specificity in the desired cell type and sequence fidelity relative to natural regulatory DNA. PACS is most useful when it improves both, as on HepG2 and K562 above. When a trade-off exists, PACS should be preferred if it meets the required specificity threshold and improved fidelity is valuable for the application. To examine this directly, we analyzed the SK-N-SH branching-factor sweep, where the main B=2048B = 2048 results favor greedy selection over PACS on specificity.

Branching factor BMetricGreedy SelectionPACS
32AlphaGenome MinGap ↑3.420 ± 0.0563.409 ± 0.076
32Motif correlation ↑0.788 ± 0.0320.799 ± 0.007
323-mer correlation ↑0.918 ± 0.0030.921 ± 0.002
32NTv2 FBD ↓2.061 ± 0.0351.977 ± 0.035
64AlphaGenome MinGap ↑3.900 ± 0.0923.922 ± 0.035
64Motif correlation ↑0.817 ± 0.0420.833 ± 0.018
643-mer correlation ↑0.922 ± 0.0070.937 ± 0.007
64NTv2 FBD ↓2.131 ± 0.0481.969 ± 0.008
2048AlphaGenome MinGap ↑5.387 ± 0.0365.125 ± 0.120
2048Motif correlation ↑0.784 ± 0.0130.819 ± 0.029
20483-mer correlation ↑0.897 ± 0.0100.915 ± 0.016
2048NTv2 FBD ↓2.403 ± 0.0582.126 ± 0.033

At B=32B = 32 and 64, mean MinGap differs by only −0.011 and +0.022, respectively, while PACS reduces FBD by 4.1% and 7.6% and also improves motif-frequency and 3-mer correlations. These results support better fidelity at comparable mean specificity and equal candidate budgets.

The revised comparisons show that fidelity gains persist against SVDD after matching the oracle budget and terminal reporting and identify specific examples with improved fidelity at comparable specificity. We thank the reviewer again for suggestions that strengthened the study.

Thank you for the additional experiments. The matched-budget comparisons address much of my concern about the original evaluation and strengthen my confidence in the results. I find the analysis of MAP ranking quite interesting, and PACS provides a well-motivated approach to acting on that observation. Together with the theoretical analysis and empirical evaluation, this makes a useful contribution. Overall, the rebuttal reinforces my positive assessment, and I will maintain my recommendation.

Reviewer_MGUoOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposed a beam-search-based diffusion model tunng methods for designing regulatory DNA. The method performed good on the existing regulatory DNA benchmark.

Critical strengths

Good introduction.

Critical weaknesses
  1. There are tons of missing baselines and related works, for example PepTune, TR2-D2, GLID2E, etc.. The authors are strongly suggested to follow more recent works and try to refer these methods since there may be something in common.
  2. One small benchmark is not enough to provet the effectiveness of the method. For example, for DRAKES, there are also protein-related benchmark to verify the model's capability.
  3. From machine learning perspective, there is no/near zero novelty. It is just applying an algorithm already exist to DNA design tasks, but not the SOTA performance. This brings no values to this field.
Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer MGUo (1/2)

We thank the reviewer for their time and suggestions regarding broader comparisons, evaluation scope, and methodological clarity. Our response addresses the three concerns with (1) comparisons with adapted PepTune, TR2-D2, and GLID²E; (2) clarification of the evaluation scope; and (3) a clearer statement of the methodological contribution.

Re W1: Comparisons with PepTune, TR2-D2, and GLID²E support PACS’s usefulness for cell-type-specific DNA design

We appreciate the suggestion and have evaluated all three methods on our benchmark. In addition to our existing discussion of PepTune in the Introduction and Related Work, we have expanded the Related Work section to discuss TR2-D2 and GLID²E.

Our design objective is to maximize predicted activity in a target cell type while minimizing off-target activity, as measured by MinGap. PepTune uses Pareto-based selection for peptide-SMILES. We adapted its Pareto reward for regulatory DNA design. For TR2-D2 and GLID²E, whose published DNA applications optimize activity in a single target cell type, we adapted the rewards to optimize MinGap and recalibrated TR2-D2’s temperature for the benchmark oracle’s reward scale. We used the same pretrained starting model, oracle, and sampler across methods and otherwise retained the published hyperparameters. The table compares design performance under a common objective. We note that TR2-D2 and GLID²E fine-tuned the generator for each objective, whereas PACS uses the frozen generator at inference time.

Cell lineMetricGLID²ETR2-D2PepTune adaptationPACS
HepG2AlphaGenome MinGap ↑−0.130 ± 0.0584.976 ± 0.1050.558 ± 0.0706.372 ± 0.061
HepG2Motif correlation ↑0.502 ± 0.0180.608 ± 0.0080.732 ± 0.0160.834 ± 0.036
HepG23-mer correlation ↑0.596 ± 0.0270.574 ± 0.0090.906 ± 0.0120.960 ± 0.008
HepG2Shannon entropy ↑1.919 ± 0.0081.924 ± 0.0031.981 ± 0.0001.981 ± 0.001
HepG2NTv2 FBD ↓2.969 ± 0.1853.075 ± 0.0751.778 ± 0.0521.789 ± 0.066
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
K562AlphaGenome MinGap ↑1.895 ± 0.0625.403 ± 0.0211.283 ± 0.1376.463 ± 0.162
K562Motif correlation ↑0.259 ± 0.0190.745 ± 0.0120.829 ± 0.0360.865 ± 0.003
K5623-mer correlation ↑−0.196 ± 0.0020.422 ± 0.0030.941 ± 0.0200.942 ± 0.005
K562Shannon entropy ↑1.515 ± 0.0021.852 ± 0.0031.977 ± 0.0021.980 ± 0.002
K562NTv2 FBD ↓34.954 ± 0.1254.499 ± 0.3011.741 ± 0.0551.760 ± 0.012
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SK-N-SHAlphaGenome MinGap ↑2.181 ± 0.0352.160 ± 0.1760.377 ± 0.0735.125 ± 0.120
SK-N-SHMotif correlation ↑0.391 ± 0.0240.501 ± 0.0080.810 ± 0.0280.819 ± 0.029
SK-N-SH3-mer correlation ↑0.283 ± 0.0070.530 ± 0.0040.941 ± 0.0100.915 ± 0.016
SK-N-SHShannon entropy ↑1.825 ± 0.0061.937 ± 0.0021.981 ± 0.0021.972 ± 0.001
SK-N-SHNTv2 FBD ↓6.610 ± 0.2133.982 ± 0.1831.819 ± 0.0492.126 ± 0.033

PACS achieves higher mean predicted specificity and motif-frequency correlation than all three methods on all targets. The PepTune adaptation achieves the lowest mean FBD on all three cell lines, with mean MinGap ranging from 0.377 to 1.283. These results support PACS’s favorable specificity–fidelity trade-off on the cell-type-specific design objective.

AuthorsOpenReview ↗

Response to Reviewer MGUo (2/2)

Re W2: The study evaluates two benchmarks spanning eleven targets and two assay modalities

We evaluated PACS on two benchmarks: three human cell-line activity tasks (MPRA measurements) and eight immune-cell accessibility tasks (chromatin-accessibility data) differing in both cellular context and measured property. Across all eleven targets, PACS improves motif-frequency, 3-mer, and embedding fidelity over matched greedy search. These experiments test whether the selection strategy transfers across regulatory-DNA tasks and data modalities.

Re W3: The contribution is progress-dependent selection for approximate MAP-valued search

Our central question is: How strongly should inference-time search trust an approximate endpoint score when its ranking reliability changes during generation?

We first show that MAP scores rank candidate continuations poorly early in denoising and more reliably as sequence context accumulates. PACS uses this observation by standardizing scores among siblings, and increasing selection strength as denoising progresses. We provide a local KL-regularized interpretation (Appendix C) and a large-pool consistency result. Controlled comparisons with greedy selection further reinforce that changing the selector alone reduces FBD by 11.5–16.0% across three cell lines while improving motif and composition fidelity. Across all eleven targets, these fidelity gains hold at the same oracle-evaluation budget as greedy search. Our methodological contribution is to use this measured change in MAP ranking reliability to guide how strongly the search favors higher-scoring candidates during denoising.

We have incorporated the additional comparisons and clarifications into the revised manuscript. We thank the reviewer again for their suggestions.

Thank the authors for the rebuttal. The rebuttal addressed most of my concerns. I decide to raise my score to 2. I still believe the scope of this paper, as well as the method novelty are not good enough, although the authors did a good job on the other fields. The limited insights to this community makes me keep this paper below acceptance threshold.

Reviewer_wxZHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the design of cell-type-specific regulatory DNA (enhancers). It performs inference-time beam search on top of a frozen, class-conditional masked discrete diffusion model, guided by a sequence-to-activity oracle. Because the oracle can only score complete sequences, each partially masked candidate is first completed position-by-position via MAP, and the score of the completed sequence is used as it look-ahead score.

The authors begin with a diagnostic. The within-group correlation between the MAP look-ahead score and the value of ancestral-sampling continuations rises from about 0 early in denoising to about 0.82 later on, indicating that the look-ahead score is unreliable early and more reliable late.

Building on this, they propose PACS (Progress-Annealed Candidate Selection), a training-free candidate selection rule.

  • Selection rule: For each parent, several child candidates are drawn; the scores within the pool are standardized, and one candidate is sampled to keep according to a Boltzmann distribution, replacing the original argmax.
  • Annealing scheme: The inverse temperature increases smoothly with denoising progress (clean fraction), so selection is near-uniform early and approaches greedy late.
  • Theory: This selector is the optimal solution to a local KL-regularized objective, and as the candidate pool grows it converges to a proposal distribution that is locally score-tilted.

Contributions

  1. A progress-coupled candidate selection rule, PACS.
  2. A diagnosis of the reliability of MAP look-ahead, together with a KL interpretation of the selector and a large-pool consistency result.
  3. Empirical evaluation on three human cell lines and eight immune cell types.

Validation

  • Using homology-aware data splits, the design oracle and the evaluation oracle are trained separately, and evaluation considers both predictive specificity (MinGap) and sequence fidelity (motif frequency, 3mer, embedding-space FBD, and motif layout).
  • Relative to greedy selection under the same configuration, FBD drops by 12%–16% across the three cell lines and by 5.7%–36.2% on immune cells, while predictive specificity is largely maintained. The abstract further claims that it outperforms most RL, inference-time alignment, and iterative-optimization baselines.
  • The schedule ablation in the appendix shows that a fixed temperature is close to greedy, wheraes raising the temperature with progress markedly improves fidelity.
Critical strengths
  1. The method is simple, and controlled comparisons support its empirical value. PACS and greedy are matched on backbone, reward, N, S, B, and output rule, changing only the candidate selection rule (D.4), with low additional selection overhead. Across three cell lines and eight immune-cell tasks, PACS improves the motif, 3-mer, and FBD metrics relative to greedy while maintaining competitive predictive specificity. The candidate-budget sweep further supports its practical value across multiple settings.

  2. The experiments have fairly broad coverage, and the boundaries of the evidence are disclosed transparently. E.2 compares multiple schedules and fixed-strength selectors; the multiple evaluation models, homology-aware splits, and the separation between design and evaluation oracles strengthen the credibility of the evaluation. The authors also clearly state the scope of applicability of the local theory, the descriptive nature of the ablations, and the uncertainty of the early-rollout reference. These are important strengths of the paper.

Critical weaknesses

W1: There remains an evidence gap between the key ablation and the main experiments that affects judgments about contribution.

E.2 does provide a valuable comparison of selectors, but it uses B=32 and final-state completion, whereas the main experiments use B=2048 and a lineage-best archive. Both the candidate-pool size and the output rule may affect the relative performance of the selectors, so the existing ablation cannot yet directly show how large the gains in the main experiments are relative to simple random selection schemes. The results for fixed κ∈{10,20,40} support PACS over those settings, but weaker fixed selection or an "early-uniform, late-strong-selection" scheme remains a simple explanation worth comparing against.

What this affects is the judgment of the incremental contribution of the schedule design. The most helpful addition would be to validate a small number of key controls under the main-experiment setting on one representative task. This might show that a clean fraction or a specific smoothing function is optimal; if a simple warm-up is equally effective, that would still support the idea of delayed strong selection, but the claims about the contribution of the specific schedule would need to be adjusted accordingly.

W2: The cross-method advantage needs to be interpreted by combining the configurations, budgets, and multi-metric trade-offs in the main tables and the appendices.

In D.4, SVDD-PM uses M=10, SMC/TDS use 128 particles, and PACS uses B=2048 per parent; these are not directly comparable budget units. The larger-candidate-budget results in E.3 also change some of the specificity rankings: for example, on HepG2 at B=2048, the MinGap of SMC/TDS is 6.81/7.05, higher than the 6.37 of PACS in the main table, but this comes with worse fidelity and a clear drop in diversity. These results can support PACS's overall trade-off advantage, but the relationship between the two implementations and output rules must be made explicit before the sources of the cross-method differences can be judged.

I suggest using the existing results to present, in a concentrated way, the specificity, fidelity, diversity, and cost of comparable configurations, and stating in the summary that the specificity change relative to greedy is task-dependent.

Questions for authors
  1. A limited supplementary experiment addressing W1: On one representative task, using B=2048 and the lineage-best output rule, could you compare PACS against a reasonably tuned fixed-strength selection and a simple "early-uniform, late-strong-selection" scheme, and briefly describe the tuning budget? If you already have corresponding results, please provide them directly. The aim is to judge the incremental gain under the main setting, without needing to expand into a full selector sweep.
  2. A configuration clarification addressing W2: What are the differences among SMC/TDS/SVDD in the main table and E.3 in terms of candidate generation, value estimation, and output rule? Using the existing budgets and multi-metric results, please explain which comparisons support PACS's advantage and the trade-offs those advantages correspond to.

These two points are the questions most likely to change my assessment.

5 comments
AuthorsOpenReview ↗

Response to Reviewer wxZH (1/2)

We thank the reviewer for the constructive feedback and for highlighting several strengths of our method. Our response and the revised manuscript add: (1) selector controls at B=2048B = 2048 with identical lineage-best reporting, including constant strengths and early-uniform/late-strong switches; and (2) systematic comparison of specificity, fidelity, diversity, and runtime at matched oracle budgets and common terminal reporting.

Re W1 & Q1: PACS improves the specificity–fidelity trade-off over fixed-strength and stepped selection at the main configuration

We agree that the B=32B = 32 ablation does not directly establish the schedule’s contribution under the main B=2048B = 2048, lineage-best configuration. We therefore compared PACS (p=2p=2, κ=40\kappa =40) against greedy selection, five constant strengths κ∈1,5,10,20,40\kappa \in \\{1, 5, 10, 20, 40\\}, and two switch selectors (uniform until clean-fraction progress reaches 0.25 or 0.50, then κ=40\kappa = 40). All these comparisons use B=2048B = 2048, identical search parameters, and the same lineage-best output rule, in order to provide direct comparability. Moreover, sibling standardization is retained in all stochastic selectors. PACS’s hyperparameters (p=2p = 2, κmax⁡=40\kappa_{\max} = 40) are fixed once across all cell lines and tasks without per-target tuning and we perform a full grid on each cell type for the constant-κ (five values) and switch (two thresholds). In these comparisons, PACS consistently improves the specificity–fidelity trade-off over alternative configurations:

Comparison with constant-strength selection. PACS achieves better fidelity (evident from lower mean FBD values) than every tested constant on all three targets. It also achieves higher mean MinGap than all constants on HepG2 and K562. Stronger constants reach higher specificity on SK-N-SH but at worse fidelity. Constant-strength selection therefore does not reproduce PACS’s fidelity gains under the main configuration.

Comparison with delayed selection. The 0.25 switch matches or slightly exceeds PACS’s specificity but sacrifices fidelity across all three cell lines (PACS reduces FBD by 4.0%, 15.6%, and 13.3% across the three cell lines). The 0.50 switch achieves better fidelity but at the cost of specificity (MinGap drops by 0.36, 0.18, and 0.40). Discrete switching thus trades one property for the other depending on the threshold. PACS’s smooth schedule navigates this trade-off, delivering a favourable balance between optimization strength and sequence fidelity without requiring a switch point.

Cell typeSelectorAlphaGenome MinGap ↑NTv2 FBD ↓
HepG2PACS6.372 ± 0.0611.789 ± 0.066
HepG2Greedy6.291 ± 0.0532.039 ± 0.047
HepG2Constant κ=1\kappa = 12.854 ± 0.2362.760 ± 0.229
HepG2Constant κ=5\kappa = 55.376 ± 0.1112.155 ± 0.016
HepG2Constant κ=10\kappa = 105.933 ± 0.0402.042 ± 0.034
HepG2Constant κ=20\kappa = 206.101 ± 0.0742.024 ± 0.066
HepG2Constant κ=40\kappa = 406.180 ± 0.0482.037 ± 0.025
HepG2Switch at 0.256.372 ± 0.0421.863 ± 0.039
HepG2Switch at 0.506.009 ± 0.1201.741 ± 0.088
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
K562PACS6.463 ± 0.1621.760 ± 0.012
K562Greedy6.389 ± 0.0742.095 ± 0.050
K562Constant κ=1\kappa = 12.918 ± 0.0705.397 ± 0.106
K562Constant κ=5\kappa = 55.881 ± 0.0612.032 ± 0.035
K562Constant κ=10\kappa = 106.155 ± 0.1122.031 ± 0.070
K562Constant κ=20\kappa = 206.411 ± 0.0042.037 ± 0.081
K562Constant κ=40\kappa = 406.431 ± 0.0912.031 ± 0.046
K562Switch at 0.256.519 ± 0.0312.086 ± 0.481
K562Switch at 0.506.286 ± 0.1321.700 ± 0.010
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SK-N-SHPACS5.125 ± 0.1202.126 ± 0.033
SK-N-SHGreedy5.387 ± 0.0362.403 ± 0.058
SK-N-SHConstant κ=1\kappa = 12.196 ± 0.0462.897 ± 0.083
SK-N-SHConstant κ=5\kappa = 54.250 ± 0.0962.264 ± 0.086
SK-N-SHConstant κ=10\kappa = 104.863 ± 0.0102.296 ± 0.042
SK-N-SHConstant κ=20\kappa = 205.193 ± 0.0602.337 ± 0.029
SK-N-SHConstant κ=40\kappa = 405.389 ± 0.0632.425 ± 0.050
SK-N-SHSwitch at 0.255.298 ± 0.0442.451 ± 0.416
SK-N-SHSwitch at 0.504.724 ± 0.0981.930 ± 0.031
AuthorsOpenReview ↗

Response to Reviewer wxZH (2/2)

Re W2 & Q2: PACS improves specificity-fidelity-diversity trade-off across methods at matched compute

We thank the reviewer for suggesting that the main (Fig. 1) and appendix (D4) results need to be interpreted together. The main table (Fig. 1) uses the following configurations: SVDD-PM evaluates 10 candidates per chain, SMC/TDS use 128 particles, and PACS evaluates 2,048 siblings per parent. To make the comparison explicit, we equalize oracle sequence evaluations per run across all methods, apply the same terminal-output rule, and report specificity, fidelity, diversity, and search time together. Experiments were conducted on NVIDIA H100 PCIe GPUs with 80 GB of memory.

Comparison with matched-resource SVDD. PACS improves fidelity with an FBD reduction relative to SVDD of 8.9%, 9.9%, and 10.6% for HepG2, K562, and SK-N-SH, and improves motif-frequency and 3-mer correlations for all three cell types. PACS increases specificity on HepG2 and K562 but decreases it on SK-N-SH, at comparable runtime.

Comparison with SMC and TDS. SMC and TDS maintain a global particle population with resampling across all particles, allowing high-scoring trajectories to overwrite others, causing mode collapse at high budgets (Shannon entropy 0.19–0.40, edit distance 15–32 nt) and low fidelity. PACS retains high diversity throughout. The results therefore reflect that PACS navigates the specificity–fidelity–diversity trade-off more effectively.

Cell lineMetricSMCTDSSVDD-PMPACS
HepG2AlphaGenome MinGap ↑6.806 ± 0.1457.048 ± 0.6775.888 ± 0.1586.151 ± 0.083
HepG2Motif correlation ↑0.455 ± 0.0940.476 ± 0.0510.835 ± 0.0180.851 ± 0.024
HepG23-mer correlation ↑0.678 ± 0.0220.599 ± 0.0290.937 ± 0.0050.961 ± 0.004
HepG2NTv2 FBD ↓5.010 ± 0.5614.496 ± 0.2001.972 ± 0.0241.797 ± 0.055
HepG2Shannon entropy ↑0.189 ± 0.0770.315 ± 0.0871.980 ± 0.0011.981 ± 0.000
HepG2Mean edit distance ↑14.693 ± 6.43424.285 ± 7.288149.764 ± 0.056149.848 ± 0.026
HepG2Search time (s) ↓1707.000 ± 13.7201723.067 ± 14.5241703.033 ± 18.4541671.820 ± 8.644
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
K562AlphaGenome MinGap ↑6.320 ± 0.3205.401 ± 1.7696.038 ± 0.0656.251 ± 0.185
K562Motif correlation ↑0.391 ± 0.0660.354 ± 0.1420.854 ± 0.0020.867 ± 0.006
K5623-mer correlation ↑0.550 ± 0.0960.529 ± 0.1390.909 ± 0.0040.943 ± 0.003
K562NTv2 FBD ↓7.111 ± 3.9495.755 ± 0.1711.908 ± 0.0301.719 ± 0.014
K562Shannon entropy ↑0.216 ± 0.0830.223 ± 0.1111.980 ± 0.0001.980 ± 0.002
K562Mean edit distance ↑16.036 ± 6.05517.841 ± 8.857149.803 ± 0.028149.778 ± 0.108
K562Search time (s) ↓1697.633 ± 5.3981718.100 ± 19.4801702.700 ± 9.9741682.632 ± 15.601
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SK-N-SHAlphaGenome MinGap ↑5.152 ± 0.0743.719 ± 1.7465.136 ± 0.0594.938 ± 0.116
SK-N-SHMotif correlation ↑0.392 ± 0.0150.401 ± 0.0630.797 ± 0.0150.820 ± 0.026
SK-N-SH3-mer correlation ↑0.660 ± 0.1140.648 ± 0.0500.892 ± 0.0110.915 ± 0.015
SK-N-SHNTv2 FBD ↓5.552 ± 0.6746.327 ± 0.9102.446 ± 0.0912.186 ± 0.033
SK-N-SHShannon entropy ↑0.396 ± 0.2130.250 ± 0.2601.972 ± 0.0011.973 ± 0.001
SK-N-SHMean edit distance ↑32.179 ± 18.13921.085 ± 23.923149.224 ± 0.051149.306 ± 0.086
SK-N-SHSearch time (s) ↓1705.400 ± 15.8531705.133 ± 5.0861717.867 ± 21.7331675.205 ± 2.442

In summary, these additional experiments establish PACS’s contribution beyond constant and delayed selection at the main configuration and clarify the cross-method trade-offs at matched budgets. We have updated Figure 1 in the revised manuscript and thank the reviewer again for suggestions that strengthened the study.

Thank you again for your feedback. We addressed both concerns by comparing PACS with fixed and stepped schedules and by evaluating the baselines at matched oracle budgets. We’d be happy to provide any further clarification before the discussion period ends.

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

PACS is a training-free selection rule for masked discrete diffusion beam search. It completes partial sequences with position-wise MAP, standardizes those oracle scores inside each sibling pool, and samples a survivor from a Boltzmann distribution whose inverse temperature grows with the clean fraction. The claim is that MAP rankings are unreliable early and trustworthy late, so annealing selection strength improves regulatory-DNA fidelity without giving up specificity.

Critical strengths
  1. The motivating measurement is concrete. Within-sibling Spearman correlation between MAP scores and ancestral-rollout values is weak early in denoising and rises as nucleotides are revealed (Figure 2).

  2. Sibling standardization is the right fix if raw oracle scales differ across steps and parents, and the selector is exactly the solution of a local KL-regularized objective. The limiting cases (uniform at κ = 0, greedy as κ → ∞) are clean.

  3. Against the greedy control, fidelity gains are consistent. At the main cell-line setting, Nucleotide Transformer Fréchet distance drops by 12.3%, 16.0%, and 11.5% on HepG2, K562, and SK-N-SH, with higher JASPAR and 3-mer correlations. On eight immune cell types, the same three fidelity metrics improve on every target (FBD reductions 5.7–36.2%), and positional entropy does not collapse.

  4. Design and evaluation oracles are separated (Enformer for search; AlphaGenome for the reported MinGap), and the benchmark includes RL, guidance, and iterative optimizers on a shared objective.

Critical weaknesses
  1. Specificity is mixed, which the abstract compresses. PACS has the best HepG2 MinGap, is 6.463 versus Ledidi’s 6.475 on K562, and loses to greedy on SK-N-SH (5.125 versus 5.387). “Outperforms most of the evaluated methods” is accurate, and it is also the strongest specificity claim the data support.

  2. The 12–16% fidelity number is an FBD reduction on two of the three cell lines; SK-N-SH is 11.5%. At large branching factors the fidelity gap over greedy narrows, so much of the gain is avoiding greedy selection while the MAP signal is noise. A fixed intermediate temperature might also achieve that. The schedule ablation is target-dependent, and p = 2, κ_max = 40 is a chosen operating point.

  3. Every number is an oracle score. The limitations section correctly says these sequences can absorb Enformer or AlphaGenome biases. There is no MPRA or accessibility measurement of the designs.

Questions for authors
  1. Does a constant κ, tuned once on a validation cell type, recover most of the FBD gain? If so, how much of PACS is the progress schedule versus sibling standardization plus stochastic selection?

  2. Can you report wall-clock oracle calls to reach a fixed MinGap, not only quality at matched branching factor?

3 comments
AuthorsOpenReview ↗

Response to Reviewer qSiq (1/2)

We thank the reviewer for the constructive feedback and for highlighting several strengths of our method and validation. With this response, we provide (1) more precise specificity–fidelity claims; (2) the requested comparison with an intermediate constant temperature at the main branching factor; (3) oracle-evaluation counts and search times at fixed specificity thresholds; and (4) a clarification of how we mitigate oracle bias in the evaluations.

Re W1: We show that PACS consistently improves fidelity

The key strength of PACS is its consistent improvement in sequence fidelity across all tested conditions. We have revised the abstract to state the performance of our method more clearly by replacing its last two sentences with: “Across three human cell lines, PACS reduces NTv2 Fréchet biological distance by 11.5–16.0% relative to matched greedy selection, while improving motif-frequency and 3-mer correlations and retaining competitive predicted specificity. Finally, in an application-relevant benchmark across eight immune cell types, PACS improves motif-frequency, sequence-composition, and embedding fidelity, with cell-type-dependent changes in predicted specificity."

Re W2 & Q1: Additional experiments confirm improved fidelity of PACS compared to constant-strength selectors

To isolate the effect of the progress schedule from sibling standardization and stochastic Boltzmann sampling, we evaluated five constant selection strengths κ∈1,5,10,20,40\kappa \in \\{1, 5, 10, 20, 40\\} at B=2048B = 2048. These constant-κ selectors used identical search parameters and retain within-sibling standardization and Boltzmann sampling, differing from PACS only by replacing the progress schedule with a fixed κ. Supporting the added value of our approach, controlling the selection strength with PACS achieves consistent fidelity gains, including a reduction of the Fréchet Biological Distance (FBD) by 11.6%, 13.3%, and 6.1% relative to the constant selection strength with the best FBD across the three cell lines (HepG2, K562, and SK-N-SH). PACS also achieved higher mean MinGap than all constants for HepG2 and K562. These results confirm that the progress-annealed schedule itself drives the improved specificity–fidelity trade-off.

Cell typeSelectorAlphaGenome MinGap ↑NTv2 FBD ↓
HepG2PACS, p=2p = 2, κmax⁡=40\kappa_{\max} = 406.372 ± 0.0611.789 ± 0.066
HepG2Greedy6.291 ± 0.0532.039 ± 0.047
HepG2Constant κ=1\kappa = 12.854 ± 0.2362.760 ± 0.229
HepG2Constant κ=5\kappa = 55.376 ± 0.1112.155 ± 0.016
HepG2Constant κ=10\kappa = 105.933 ± 0.0402.042 ± 0.034
HepG2Constant κ=20\kappa = 206.101 ± 0.0742.024 ± 0.066
HepG2Constant κ=40\kappa = 406.180 ± 0.0482.037 ± 0.025
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
K562PACS, p=2p = 2, κmax⁡=40\kappa_{\max} = 406.463 ± 0.1621.760 ± 0.012
K562Greedy6.389 ± 0.0742.095 ± 0.050
K562Constant κ=1\kappa = 12.918 ± 0.0705.397 ± 0.106
K562Constant κ=5\kappa = 55.881 ± 0.0612.032 ± 0.035
K562Constant κ=10\kappa = 106.155 ± 0.1122.031 ± 0.070
K562Constant κ=20\kappa = 206.411 ± 0.0042.037 ± 0.081
K562Constant κ=40\kappa = 406.431 ± 0.0912.031 ± 0.046
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SK-N-SHPACS, p=2p = 2, κmax⁡=40\kappa_{\max} = 405.125 ± 0.1202.126 ± 0.033
SK-N-SHGreedy5.387 ± 0.0362.403 ± 0.058
SK-N-SHConstant κ=1\kappa = 12.196 ± 0.0462.897 ± 0.083
SK-N-SHConstant κ=5\kappa = 54.250 ± 0.0962.264 ± 0.086
SK-N-SHConstant κ=10\kappa = 104.863 ± 0.0102.296 ± 0.042
SK-N-SHConstant κ=20\kappa = 205.193 ± 0.0602.337 ± 0.029
SK-N-SHConstant κ=40\kappa = 405.389 ± 0.0632.425 ± 0.050
AuthorsOpenReview ↗

Response to Reviewer qSiq (2/2)

Re Q2: Additional experiments confirm improved fidelity of PACS at matched specificity thresholds.

We evaluated PACS and greedy selection across B ∈ {1, 2, 4, 8, 16, 32, 64, 128, 512, 1024, 2048}, reporting oracle sequence evaluations, search time, and achieved specificity and fidelity at the smallest B that reaches each MinGap threshold. In these comparisons, PACS consistently achieves improved fidelity (evident from lower mean FBD values) for every combination of cell line and specificity threshold. We also assessed runtime performance and found that PACS and greedy selection reach all performance thresholds at the same budget and time for HepG2, while PACS reaches the target with half the oracle evaluations and time for K562 and with twice the oracle evaluations and time for SK-N-SH. The average fidelity improvements are: HepG2 16.1%, K562 48.1%, and SK-N-SH 7.1%.

Cell typeMinGap thresholdMethodBAlphaGenome MinGap ↑NTv2 FBD ↓Oracle sequence evaluationsSearch time, s
HepG23PACS163.070 ± 0.0561.823 ± 0.08565,53615.686 ± 0.128
HepG23Greedy163.121 ± 0.0432.399 ± 0.27365,53615.676 ± 0.119
HepG24PACS644.547 ± 0.1341.883 ± 0.070262,14453.986 ± 0.401
HepG24Greedy644.556 ± 0.0572.183 ± 0.071262,14454.552 ± 0.256
HepG25PACS1285.281 ± 0.1211.798 ± 0.058524,288106.741 ± 1.274
HepG25Greedy1285.231 ± 0.0392.008 ± 0.060524,288106.279 ± 0.750
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
K5623PACS163.611 ± 0.0601.936 ± 0.07965,53615.727 ± 0.110
K5623Greedy163.402 ± 0.1034.750 ± 0.52665,53616.368 ± 1.151
K5624PACS324.310 ± 0.0722.017 ± 0.163131,07228.415 ± 0.089
K5624Greedy324.064 ± 0.0374.161 ± 0.040131,07228.484 ± 0.073
K5625PACS645.011 ± 0.2501.759 ± 0.112262,14454.431 ± 0.136
K5625Greedy1285.431 ± 0.0882.654 ± 0.274524,288107.327 ± 0.332
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SK-N-SH3PACS323.409 ± 0.0761.977 ± 0.035131,07228.353 ± 0.074
SK-N-SH3Greedy163.075 ± 0.0142.089 ± 0.02565,53625.422 ± 17.057
SK-N-SH4PACS1284.245 ± 0.0672.009 ± 0.099524,288106.803 ± 0.930
SK-N-SH4Greedy1284.387 ± 0.0792.186 ± 0.069524,288106.354 ± 0.707
SK-N-SH5PACS10245.044 ± 0.1732.095 ± 0.1134,194,304836.125 ± 0.955
SK-N-SH5Greedy10245.165 ± 0.1652.270 ± 0.0314,194,304842.847 ± 5.398

Re W3: Using different oracles during search and evaluation supports the robustness of our approach

To establish the robustness of our approach across different sequence-to-activity models, we use different oracles during search (Enformer) and evaluation (AlphaGenome). We further partition the data using homology-aware clustering and train design and evaluation oracles on separate splits. Search uses the Enformer design oracle, while final evaluation uses the AlphaGenome model trained on the held-out split. Prior work has experimentally validated model-designed regulatory elements (Gosai et al., Nature 2024; DaSilva et al., Nature Genetics 2026), supporting our oracle-driven regulatory DNA design approach.

In summary, we now provide additional results in support of the methodological value, rigor, and reproducibility of PACS; we show improved fidelity at comparable search cost; and we have incorporated the new ablations and revised claims into the revised manuscript. We thank the reviewer again for suggestions that strengthened this study.

Thank you again for your feedback. We’ve addressed both of your questions above, including the fixed-strength comparisons and the search cost at matched specificity. We’d be happy to discuss either point further before the discussion closes.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their constructive feedback and their interest in our manuscript. At the heart of our study is the observation that lookahead scores in training-free enhancer design gradually become informative as more of the DNA sequence is unmasked. This motivates our PACS method for systematically lending more confidence to these scores as the search progresses.

Through additional validation and benchmarking experiments in response to the reviewers’ comments, we were able to substantiate the key contribution of our method. PACS designs DNA sequences with improved fidelity relative to matched greedy search (i.e., sequences that are more faithful to the human genome’s regulatory grammar), while providing competitive predicted cell-type specificity in two rigorous benchmarks on biologically relevant datasets. This result is practically important because high-fidelity, cell-type-specific DNA sequences are required for many biological applications (e.g., cell-type-specific expression of a gene therapy). Our work also further contributes by presenting a theoretically grounded interpretation of our solution with well-defined limiting cases.

Below, we summarize the main questions raised by the reviewers and how we responded to each through additional experiments and manuscript revisions, as detailed in our individual responses.

  • Does the PACS progress schedule improve on simpler selection rules? (qSiq, wxZH)
    We tested five constant selection strengths and two early-uniform, late-strong schedules at the main branching factor (B=2048B=2048). For all three cell lines, PACS showed enhanced fidelity, with reductions in Fréchet biological distance (FBD) of 6.1–13.3% relative to the best tested constant selection strength. Early switching sacrificed fidelity and increased FBD variance over tenfold on K562 and SK-N-SH, while late switching improved fidelity at the cost of specificity. PACS provides a stable, tuning-free instantiation of progress-coupled selection using a single fixed schedule across all targets.

  • Do these fidelity gains persist at comparable runtime and specificity? (qSiq, wxZH, mgPn)
    With equal oracle-evaluation budgets and the same terminal-output rule, PACS continued to provide enhanced fidelity, with reductions in FBD of 8.9–10.6% relative to SVDD-PM at comparable runtime. Predicted specificity increased for HepG2 and K562 and decreased for SK-N-SH. Against greedy search, PACS achieved better fidelity at every tested specificity threshold, with average FBD reductions of 7–48% across thresholds for each cell line and runtime varying by target and threshold.

  • How well do these results generalize? (MGUo, qSiq)
    Across three cell-line activity tasks and eight immune-cell chromatin accessibility tasks, PACS improved fidelity measures including motif, 3-mer, and embedding fidelity relative to matched greedy search. We also compared PACS with adapted versions of PepTune, TR2-D2, and GLID²E. In these additional comparisons, PACS achieved the highest predicted specificity in all three cell lines. PepTune achieved the lowest FBD, but at substantially lower specificity.

Finally, we revised the manuscript to incorporate the new results, ensure that the validation and benchmarking results are adequately presented and discussed, and make a small number of textual changes in response to points raised by the reviewers, as detailed in our individual responses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.