- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes ToDNA, a test-time optimization framework for designing cell-type-specific regulatory DNA sequences. It first adapts the pretrained Evo 2 genomic foundation model to enhancer and promoter sequences through LoRA fine-tuning. It then optimizes individual candidate sequences by updating nucleotide logits using a straight-through estimator. These updates are guided by a sequence-to-activity oracle trained on multiple cell types and regularized using position-wise categorical KL divergence from a local reference distribution. The authors evaluate ToDNA on an enhancer design benchmark covering three cell types and a promoter design benchmark covering one cell type.
Overall, I find this work interesting, but it needs a more rigorous evaluation, clearer motivation, presentation, and a better-supported claim of originality. I look forward to seeing these concerns (see below) addressed in the discussion or a revised submission.
Critical strengths
- The test-time optimization method is practically useful because it can be applied to multiple cell-type targets without additional fine-tuning for each target.
- The method preserves sequence composition and transcription factor motif profiles reasonably well, providing some evidence that optimization retains biologically relevant sequence features.
Critical weaknesses
- The choice of Evo 2 (Brixi et al., 2026) as the backbone is not sufficiently justified. The paper does not compare it with other DNA foundation models, such as Nucleotide Transformer v2 (Dalla-Torre et al., 2025), Nucleotide Transformer v3 (Boshar et al., 2025), Caduceus (Schiff et al., 2024), or HyenaDNA (Nguyen et al., 2023). Comparisons should account for differences in these models’ training objectives and generation procedures.
- In Table 2, the reported standard deviations for “Base logits” (I assume unadapted Evo 2) and ToDNA, overlap across metrics. This alone does not establish whether the differences are statistically significant and leaves the benefit of adaptation unclear. The authors should report uncertainty on the differences and discuss whether the gains justify the fine-tuning cost.
- The optimization procedure appears closely related to FastSeqProp (Linder & Seelig, 2021). Both use straight-through estimators to optimize discrete sequences through a differentiable predictor, and support regularization. The paper should cite this work where relevant and explain which technical differences account for the proposed method’s contribution. These differences should be motivated and tested through ablations.
- Ledidi (Schreiber et al., 2025) is another directly relevant baseline. It uses Gumbel–Softmax optimization and explicitly penalizes the number of edits to an initial sequence. A comparison with Ledidi, together with an ablation replacing KL regularization with an edit penalty, would help establish the value of the proposed formulation. Its omission also leaves the state-of-the-art claim insufficiently supported.
- Several aspects of the evaluation need clarification:
o The benchmark appears incomplete relative to the Ctrl-DNA study from which it is adapted (Chen et al., 2025). In particular, activity in individual off-target cell types and results for the remaining promoter targets are not fully reported.
o Some reproduced baseline values appear lower than previously published results on the same benchmark. The authors should explain these discrepancies and identify any differences in evaluation settings.
o Relevant baselines are missing, including Ledidi (Schreiber et al., 2025), DNA-CRAFT (Awasthi et al., 2026), GPA (Sarkar et al., 2026), and SVDD (Li et al., 2025).
o Metrics assessing output diversity or mode collapse are missing.
o The design and evaluation predictors share the Enformer architecture and pretraining, which may lead to correlated prediction errors. Evaluation with additional predictors, such as Borzoi (Linder et al., 2025) or AlphaGenome (Avsec et al., 2026), would help assess whether the gains transfer beyond the current predictors, provided suitable outputs and sequence contexts are available.
- The evaluation covers only three enhancer targets and one promoter target, using a single MPRA-derived activity prediction setup. Additional cell types and complementary assay modalities, such as chromatin accessibility, would provide stronger evidence of generality.
- The results do not show consistent improvements across targets. Ctrl-DNA achieves higher enhancer specificity for HepG2 and K562, while promoter specificity remains negative for every method. The paper acknowledges these limitations, but its broader performance claims should reflect them.
Minor comments:
- Several claims in the introduction need correction or addition of relevant citations.
- The notation is dense in places, and several symbols appear before they are defined.
- Table 1 reports both motif Pearson and motif Spearman correlations. These appear to provide largely overlapping information in the reported results. The authors should explain the value of reporting both or retain only one.
- The motivation for the different ToDNA variants, including the reward formulations in Appendix C, should be explained before the variants are introduced.
- MinGap is mentioned but never introduced.
Questions for authors
- In Table 5, the mean-difference objective achieves the best MinGap for some targets, while directly optimizing MinGap does not consistently produce the best score on that metric. Can the authors explain this pattern? In particular, were the objectives compared using matched optimization budgets and comparable hyperparameter tuning? I also find it surprising that optimzing for Target only still yields positive MinGap scores raising concerns about the accuracy of the reported results.
- Can the authors report evaluation metrics for the starting sequences sampled from the adapted prior, before any test-time optimization? This would provide a reference point for motif and composition metrics and help separate the contribution of optimization from that of the adapted prior.
- How is the number of edits determined? Does the method use an early-stopping criterion, or does it always run for a fixed number of iterations? Can the authors vary the KL regularization strength and iteration count to characterize the trade-off between sequence fidelity and predicted specificity?
- Can the authors compare KL regularization with a minimum-edit loss, as used in Ledidi (Schreiber et al., 2025)?
- What is the source of the enhancer and promoter sequences used to fine-tune Evo 2?
- Could the authors use a DNA diffusion model, such as D3 (Sarkar et al., 2024) or DNA-CRAFT (Awasthi et al., 2026), as the proposal model instead of Evo 2? What changes to the optimization framework would this require?
5 comments
Response to Reviewer njXC
Q4: The benchmark appears incomplete relative to the Ctrl-DNA study from which it is adapted (Chen et al., 2025). Some reproduced baseline values appear lower than previously published results on the same benchmark.
**A4:**Thank you for pointing this out. We agree that the original presentation was incomplete. We have completed the results for the two omitted promoter targets, JURKAT and THP-1. As shown in Table 3, our model ToDNA continues to hold an advantage across most metrics. Regarding the lower baseline scores, our experiments adapt the benchmark to a common evaluation protocol. We use separately trained Design and Evaluation predictors, different sequence-group partitions and starting pools, and three runs with 256 candidates locked before independent evaluation. (Chen et al., 2025) report five runs and metrics over sequences generated in the final round. These differences limit direct numerical comparisons with the published results.
Table 3. The results for JURKAT and THP-1. Bold marks the highest mean in each row.
| Promoter target | Metric | Ctrl-DNA | ToDNA |
|---|---|---|---|
| JURKAT | Mean specificity margin ↑ | 0.069 ± 0.014 | 0.039 ± 0.002 |
| JURKAT | MinGap ↑ | −0.035 ± 0.019 | −0.075 ± 0.002 |
| JURKAT | Motif Pearson ↑ | 0.136 ± 0.054 | 0.857 ± 0.011 |
| JURKAT | Motif Spearman ↑ | 0.717 ± 0.038 | 0.884 ± 0.004 |
| JURKAT | 3-mer Pearson ↑ | 0.449 ± 0.060 | 0.891 ± 0.005 |
| JURKAT | Positional entropy (bits) ↑ | 1.872 ± 0.011 | 1.779 ± 0.002 |
| JURKAT | Mean pairwise Hamming distance ↑ | 0.710 ± 0.003 | 0.680 ± 0.001 |
| THP-1 | Mean specificity margin ↑ | 0.215 ± 0.013 | 0.186 ± 0.001 |
| THP-1 | MinGap ↑ | 0.148 ± 0.013 | 0.127 ± 0.002 |
| THP-1 | Motif Pearson ↑ | 0.470 ± 0.174 | 0.868 ± 0.018 |
| THP-1 | Motif Spearman ↑ | 0.789 ± 0.054 | 0.699 ± 0.061 |
| THP-1 | 3-mer Pearson ↑ | 0.409 ± 0.107 | 0.718 ± 0.151 |
| THP-1 | Positional entropy (bits) ↑ | 1.557 ± 0.112 | 1.928 ± 0.017 |
| THP-1 | Mean pairwise Hamming distance ↑ | 0.595 ± 0.058 | 0.729 ± 0.006 |
Q5: Relevant baselines are missing, including Ledidi (Schreiber et al., 2025), DNA-CRAFT (Awasthi et al., 2026), GPA (Sarkar et al., 2026), and SVDD (Li et al., 2025).
A5: Thank you for pointing this out. We completed a direct comparison with Ledidi using the same starting sequences, predictors, and search-query budget. As shown in Table 4, ToDNA achieved a higher mean specificity score on four of the six tasks, while Ledidi performed better on the K562 and THP-1 promoter tasks. Complete matched comparisons with DNA-CRAFT, GPA, and SVDD have not yet been completed due to time constraints**.**
Table 4. Comparison with Ledidi. The metric is normalized mean specificity margin on the Evaluation predictor (↑). Bold marks the higher mean in each row.
| Task | Ledidi | ToDNA |
|---|---|---|
| HepG2 enhancer | 0.333 ± 0.006 | 0.349 ± 0.006 |
| K562 enhancer | 0.543 ± 0.017 | 0.546 ± 0.011 |
| SK-N-SH enhancer | 0.214 ± 0.009 | 0.247 ± 0.004 |
| JURKAT promoter | 0.085 ± 0.013 | 0.110 ± 0.015 |
| K562 promoter | −0.115 ± 0.003 | −0.120 ± 0.005 |
| THP-1 promoter | 0.190 ± 0.005 | 0.184 ± 0.004 |
Q6: Design and evaluation predictors share the Enformer architecture and pretraining, which may lead to correlated prediction errors. Evaluation with additional predictors, such as Borzoi (Linder et al., 2025) or AlphaGenome (Avsec et al., 2026), would help assess whether the gains transfer beyond the current predictors, provided suitable outputs and sequence contexts are available.
**A6:**Thank you for highlighting this concern. We conducted an exploratory evaluation with AlphaGenome, obtaining ATAC, CAGE, and RNA-seq predictions for ToDNA and Ledidi candidates across three enhancer targets. As shown in Table 5, we interpret them as exploratory cross-predictor assessments and do not claim that they establish transfer of cell-specific gains.
Table 5. Exploratory AlphaGenome evaluation of ToDNA and Ledidi(Schreiber et al., 2025). All scores are expressed in ×10⁻³. “Design target” identifies the intended target of sequence optimization, not a cell-specific AlphaGenome output track. Bold indicates the higher mean signal within each target and output.
| Design target | Method | ATAC | CAGE | RNA-seq |
|---|---|---|---|---|
| HepG2 | Ledidi | 56.365 ± 18.993 | 5.495 ± 0.246 | 0.363 ± 0.086 |
| HepG2 | ToDNA | 53.356 ± 6.052 | 5.063 ± 1.020 | 0.372 ± 0.111 |
| K562 | Ledidi | 58.009 ± 10.490 | 13.035 ± 3.378 | 1.563 ± 0.775 |
| K562 | ToDNA | 47.609 ± 1.571 | 10.504 ± 0.287 | 0.971 ± 0.278 |
| SK-N-SH | Ledidi | 41.289 ± 20.951 | 11.312 ± 4.690 | 1.732 ± 2.065 |
| SK-N-SH | ToDNA | 23.045 ± 8.036 | 6.147 ± 0.830 | 0.831 ± 0.364 |
Response to Reviewer njXC
Q7: The evaluation covers only three enhancer targets and one promoter target, using a single MPRA-derived activity prediction setup. Additional cell types and complementary assay modalities, such as chromatin accessibility, would provide stronger evidence of generality.
A7: Thank you for this suggestion. In addition to the four displayed tasks, we have evaluated the JURKAT and THP-1 promoter targets. As shown in Table 4 on A5, ToDNA achieves a higher mean specificity score on the JURKAT and THP-1 promoter targets. This demonstrates that our model also performs well on tasks involving other cell types.
Q8: The mean-difference objective achieves the best MinGap for some targets, whereas directly optimizing MinGap does not consistently yield the highest score. Could the authors explain this pattern and clarify whether optimization budgets and hyperparameter tuning were comparable? The positive MinGap scores obtained by the Target-only objective also require clarification to address concerns about the accuracy of the reported results.
A8: Thank you for raising this point. Our objective design follows the Ctrl-DNA (Chen et al., 2025) formulation, which measures specificity as normalized target activity minus mean normalized off-target activity. We included Target-only and smooth-MinGap variants to examine alternative reward formulations.
Q9: Could the authors report evaluation metrics for the starting sequences sampled from the adapted prior, before test-time optimization? This would provide a reference for motif and composition metrics and help distinguish the contribution of optimization from that of the adapted prior.
A9: Thank you for your suggestions. We evaluated both the full candidate pool sampled from the adapted prior and the selected starting sequences before test-time optimization. As shown in Table 6, we observe that test-time optimization can generate more natural sequences with the desired properties.
Table 6. Characteristics of HepG2 candidate sequences before test-time optimization. Bold indicates the better mean according to the stated direction.
| Metric | Before TTO | After TTO |
|---|---|---|
| Design mean specificity margin ↑ | 0.001 ± 0.001 | 0.030 ± 0.003 |
| Motif Pearson ↑ | 0.800 ± 0.026 | 0.818 ± 0.037 |
| 3-mer Pearson ↑ | 0.947 ± 0.012 | 0.921 ± 0.029 |
| Mean GC content (%) | 46.619 ± 0.269 | 54.068 ± 0.491 |
| Sequences with GC content between 30% and 70% (%) ↑ | 91.667 ± 1.502 | 95.573 ± 1.966 |
| Sequences with maximum homopolymer length ≤ 6 (%) ↑ | 83.377 ± 2.215 | 90.104 ± 1.626 |
| Mean maximum homopolymer length (bp) ↓ | 5.621 ± 0.122 | 5.193 ± 0.255 |
Q10: How is the number of edits determined? Does the method use early stopping or a fixed number of iterations? Could the authors vary KL regularization strength and iteration count to characterize the trade-off between sequence fidelity and predicted specificity?
A10: Thank you for highlighting this concern. We assume edits are measured as the Hamming distance between each final discrete sequence and its own starting sequence. The edit count is a computational metric. We implement each trajectory run for a fixed number of iterations without early stopping, while KL regularization softly constrains deviation from the reference distribution. In the matched HepG2 comparison below Table 7, increasing the KL weight from 0 to 1 reduced edits while maintaining a slightly higher mean Design-predicted specificity.
Table 7. Matched KL comparison on HepG2. Both settings use the same Design predictor, paired pools and starts. Bold indicates the best displayed mean.
| KL weight | Iterations | Edits ↓ | Design ΔR ↑ |
|---|---|---|---|
| 0.000 | 10 | 77.520 ± 1.392 | 0.383 ± 0.015 |
| 1.000 | 10 | 64.494 ± 1.539 | 0.391 ± 0.012 |
Q11: Could the authors compare KL regularization with a minimum-edit loss, as used in Ledidi (Schreiber et al., 2025)?
A11: Thank you for raising this point. We performed a regularizer ablation by replacing KL with a differentiable expected-edit penalty relative to the starting sequence, using the same starting pool and optimization budget. The edit penalty discourages nucleotide substitutions, while KL penalizes deviations from the full reference nucleotide distribution. As shown in Table 8, KL achieved a higher mean Design reward gain on all three enhancer targets at the tested regularization strengths.
Table 8. KL regularization versus an expected-edit penalty within ToDNA. The metric is Design reward gain relative to each retained candidate’s own starting sequence (↑). Bold marks the higher mean.
| Enhancer target | Edit penalty | KL regularization |
|---|---|---|
| HepG2 | 0.009 ± 0.001 | 0.024 ± 0.001 |
| K562 | 0.106 ± 0.005 | 0.120 ± 0.005 |
| SK-N-SH | 0.010 ± 0.001 | 0.021 ± 0.001 |
Response to Reviewer njXC
Thank you for your detailed and insightful comments. We address each concern point by point below.
Q1: Why choose Evo2 over other DNA foundation models?
A1: Thank you for pointing this out. We agree that comparisons must respect model-specific objectives. Other DNA foundation models, including Nucleotide Transformer v2 (Dalla-Torre et al., 2025), Nucleotide Transformer v3 (Boshar et al., 2025), Caduceus (Schiff et al., 2024), and HyenaDNA (Nguyen et al., 2023), learn reusable representations or sequence priors through pretraining on DNA. To broaden our comparison across architectures and generative paradigms, we additionally included D3 (DNA Discrete Diffusion; Sarkar et al., 2024) as an alternative proposal generator. We compared D3 proposal-only with D3 proposals followed by ToDNA test-time optimization. As shown in Table 1, we observe that adding ToDNA increased the mean evaluation reward. This indicates that our strategy can generalize to other DNA foundation model architectures.
Table 1. D3 model on three enhancer tasks. The metric is the final normalized evaluation reward (×10⁻³) ↑. Bold marks the best mean in each row.
Q2: Do the differences between ToDNA and Base logits in Table 2 justify the cost of adaptation?
A2: Thank you for highlighting this issue. We clarify that “Base logits” uses unadapted Evo2 for the initial logits and KL reference, while retaining the same adapted-model candidate pool. Therefore, this comparison evaluates adaptation in initialization and regularization. We observe that adaptation took about 40 minutes for enhancers and 39 minutes for promoters on one GPU each, once per element class.
Q3: How does ToDNA compare with FastSeqProp (Linder & Seelig, 2021) and Ledidi (Schreiber et al., 2025)?
A3: Thank you for pointing this out. We compared all three methods directly. As shown in Table 2, we observe that the ToDNA model continues to hold an advantage across most DNA design tasks.
Table 2. Comparison with Fast SeqProp and Ledidi. The metric is normalized mean margin on the evaluation predictor (↑). Bold marks the highest mean in each row.