acceptodds
Under review as a conference paper at ICLR 2027

Base composition confounds classifier-free guidance in DNA diffusion models

Abstract

Designing short DNA sequences that act as enhancers selectively in a target cell type is a central goal in generative genomics. Diffusion models have recently shown promise for this task, typically using classifier-free guidance to steer generation toward the target cell type. We find that stronger classifier-free guidance can raise predicted chromatin accessibility, a proxy for regulatory activity, while shifting base composition away from that of real sequences and distorting motif frequencies. We introduce composition-neutral guidance (CNG), which subtracts the mean guidance logit update across positions for each base. This removes sequence-wide base preferences from the update without retraining. Applied to BaseDiff, our 21.6M-parameter convolutional discrete diffusion model, CNG reduces composition shifts and improves agreement with real motif frequencies and co-occurrence patterns with little change in predicted accessibility. We compare generators across guidance scales using their accessibility–quality trade-off curves. Across two datasets, BaseDiff sampled with CNG achieves larger trade-off areas than DNA-Diffusion and D3, despite using fewer than one quarter of their parameters.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
discrete diffusion, classifier-free guidance, guidance correction, regulatory DNA design, genomics

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_V4GNOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies sequence-wide base-composition shifts as one failure mode of strong classifier-free guidance in regulatory-DNA generation. It proposes composition-neutral guidance (CNG), which subtracts the per-base positional mean from the conditional-minus-unconditional logit update, and applies it to a compact convolutional discrete diffusion model, BaseDiff. Models are compared through accessibility-versus-motif-quality trade-off curves rather than at one guidance scale. Experiments evaluate composition, motif frequencies/co-occurrence, predicted accessibility, and several robustness checks.

Critical strengths
  1. The main finding is informative beyond a model leaderboard: increasing predicted accessibility can coincide with distorted base composition and motif statistics. CNG is a simple, training-free intervention targeting this mechanism. The same-network comparison in Fig. 3 and the composition/error analysis in Appendix J help isolate the effect from the backbone change; composition-matched forward-noise experiments provide an additional check.

  2. The empirical validation is substantially broader than the main table alone. It includes a second accessibility predictor, a second genomic insertion locus, a second dataset, three independent BaseDiff training runs, and comparisons with APG/CFG-Zero* and alternate composition corrections (Appendices C, E–G, K–M). On K562, Table 1 shows co-occurrence trade-off area improving from 0.364 to 0.423 with CNG. These checks strengthen the case for a useful computational finding.

Critical weaknesses
  1. The claimed portability of the correction remains under-tested. Other headline generators use ordinary CFG, while CNG and the additional correction comparisons are applied to BaseDiff. The results therefore establish a BaseDiff-plus-CNG system and a within-BaseDiff improvement, but do not show whether the intervention transfers to other discrete generators or how rankings change when competitors receive it. This matters to the broader contribution of CNG as a reusable sampler correction.

  2. The biological conclusion remains limited by the evaluation proxies. Both motif-quality measures originate from the same JASPAR scan, and accessibility comes from predictors; Appendix D itself shows that short-range statistical generators can achieve high realism. The second predictor/locus/dataset checks are valuable but do not establish improved cellular function. I regard the evidence as supporting reduced compositional distortion and better computational trade-offs, with functional claims requiring more direct validation.

Questions for authors
  1. Can CNG be applied to an independently trained competing discrete model, such as the D3 baseline, and evaluated at comparable accessibility? Does it improve quality and preserve accessibility, and how does the corrected competitor compare with BaseDiff-CNG?

  2. Appendix K acknowledges that zeroing the mean logit update preserves composition only approximately after softmax. Can the main-text statement that the correction changes only composition be qualified, and can effects on motif probabilities be separated from the exact algebraic property of zero-mean logits?

  3. How sensitive are the CNG gains to the motif library and co-occurrence pair-selection rules? If functional validation is not available, please keep conclusions explicitly at the computational level and distinguish matching motif statistics from preserving regulatory function.

2 comments
AuthorsOpenReview ↗

Author Response (1/2)

Thank you for recognizing the computational finding and the controlled comparisons. We tested CNG on D3 on both datasets and examined the sensitivity to the motif library and pair-selection rules.

Transfer to D3. We sampled the same trained D3 baseline with and without CNG, keeping its weights, conditioning and guidance scales fixed. DHS areas follow Table 1, and lentiMPRA areas use the shared accessibility range of Table 7.

DatasetCell typeMetricD3D3 + CNGBaseDiff-CNG
DHSGM12878Realism0.368±0.0070.368 \pm 0.0070.388±0.0070.388 \pm 0.0070.388±0.0060.388 \pm 0.006
DHSHepG2Realism0.373±0.0080.373 \pm 0.0080.387±0.0080.387 \pm 0.0080.396±0.0080.396 \pm 0.008
DHSK562Realism0.361±0.0070.361 \pm 0.0070.378±0.0050.378 \pm 0.0050.392±0.0060.392 \pm 0.006
DHSGM12878Co-occurrence0.388±0.0040.388 \pm 0.0040.400±0.0040.400 \pm 0.0040.405±0.0040.405 \pm 0.004
DHSHepG2Co-occurrence0.389±0.0050.389 \pm 0.0050.399±0.0050.399 \pm 0.0050.398±0.0040.398 \pm 0.004
DHSK562Co-occurrence0.374±0.0060.374 \pm 0.0060.398±0.0060.398 \pm 0.0060.423±0.0050.423 \pm 0.005
lentiMPRAK562Realism0.064±0.0030.064 \pm 0.0030.070±0.0020.070 \pm 0.0020.074±0.0010.074 \pm 0.001
lentiMPRAHepG2Realism0.101±0.0040.101 \pm 0.0040.110±0.0020.110 \pm 0.0020.120±0.0020.120 \pm 0.002
lentiMPRAWTC11Realism0.046±0.0010.046 \pm 0.0010.046±0.0010.046 \pm 0.0010.047±0.0010.047 \pm 0.001
lentiMPRAK562Co-occurrence0.054±0.0030.054 \pm 0.0030.067±0.0020.067 \pm 0.0020.074±0.0020.074 \pm 0.002
lentiMPRAHepG2Co-occurrence0.069±0.0060.069 \pm 0.0060.114±0.0030.114 \pm 0.0030.121±0.0030.121 \pm 0.003
lentiMPRAWTC11Co-occurrence0.042±0.0010.042 \pm 0.0010.049±0.0010.049 \pm 0.0010.052±0.0010.052 \pm 0.001

CNG increases D3's area in all twelve comparisons, including WTC11 realism before rounding. The lentiMPRA comparison holds the accessibility range fixed across ordinary D3, corrected D3, and BaseDiff-CNG. BaseDiff-CNG retains larger areas in most comparisons with 21.6M parameters, compared with D3's 92.4M. The corrected models are similar on DHS GM12878 realism and HepG2 co-occurrence.

Exact property of the correction. Let Δp,b=ℓc,p,b−ℓ∅,p,b\Delta_{p,b}=\ell_{c,p,b}-\ell_{\varnothing,p,b} and let Δˉb\bar\Delta_b be its mean over positions. CNG uses Δ~p,b=Δp,b−Δˉb\widetilde\Delta_{p,b}=\Delta_{p,b}-\bar\Delta_b. Each base's mean update is exactly zero, and variation across positions is unchanged in logit space. Softmax couples the four bases at each position, and the effect on sampled composition is approximate. Appendix K tests a probability-space variant of this correction, and it does not improve on the mean correction (CNG). We will qualify Section 3.3 to state this property precisely and distinguish it from effects on motif probabilities.

AuthorsOpenReview ↗

Author Response (2/2)

Motif library and pair selection. We rescanned the sequences with HOCOMOCO v14 CORE, which contains 1,595 motif matrices against JASPAR's 1,019, and recomputed the quality metrics. Below are area differences between BaseDiff-CNG and the same model with ordinary CFG.

Cell typeJASPAR realismHOCOMOCO realismJASPAR co-occurrenceHOCOMOCO co-occurrence
GM12878+0.0070±0.0043+0.0070 \pm 0.0043+0.0132±0.0040+0.0132 \pm 0.0040+0.0033±0.0035+0.0033 \pm 0.0035+0.0115±0.0053+0.0115 \pm 0.0053
HepG2+0.0061±0.0027+0.0061 \pm 0.0027+0.0075±0.0024+0.0075 \pm 0.0024−0.0001±0.0040-0.0001 \pm 0.0040+0.0216±0.0092+0.0216 \pm 0.0092
K562+0.0253±0.0047+0.0253 \pm 0.0047+0.0269±0.0046+0.0269 \pm 0.0046+0.0592±0.0076+0.0592 \pm 0.0076+0.1401±0.0135+0.1401 \pm 0.0135

We also varied the number of selected pairs, the motif-prevalence window, and the Jaccard redundancy threshold. Each variant reselects pairs from the real training data and recomputes the co-occurrence metric. The table gives CNG's co-occurrence-area improvement over ordinary CFG.

Pair-selection ruleGM12878HepG2K562
Submitted rule+0.0033±0.0036+0.0033 \pm 0.0036−0.0001±0.0040-0.0001 \pm 0.0040+0.0592±0.0077+0.0592 \pm 0.0077
Top 250 pairs+0.0017±0.0042+0.0017 \pm 0.0042−0.0040±0.0051-0.0040 \pm 0.0051+0.0689±0.0089+0.0689 \pm 0.0089
Top 1,000 pairs+0.0061±0.0039+0.0061 \pm 0.0039+0.0027±0.0028+0.0027 \pm 0.0028+0.0460±0.0069+0.0460 \pm 0.0069
Prevalence 2–40%+0.0046±0.0043+0.0046 \pm 0.0043+0.0009±0.0042+0.0009 \pm 0.0042+0.0636±0.0083+0.0636 \pm 0.0083
Prevalence 0.5–60%+0.0033±0.0037+0.0033 \pm 0.0037−0.0001±0.0039-0.0001 \pm 0.0039+0.0591±0.0076+0.0591 \pm 0.0076
Jaccard threshold 0.3+0.0001±0.0038+0.0001 \pm 0.0038+0.0012±0.0031+0.0012 \pm 0.0031+0.0376±0.0051+0.0376 \pm 0.0051
Jaccard threshold 0.7+0.0099±0.0085+0.0099 \pm 0.0085−0.0004±0.0045-0.0004 \pm 0.0045+0.0525±0.0098+0.0525 \pm 0.0098

The K562 improvement remains positive across all pair-selection variants. The effects on GM12878 and HepG2 are smaller, consistent with the original comparison.

Computational interpretation. We agree with your characterization of the evidence as reduced compositional distortion and improved computational trade-offs. Appendix D illustrates why motif agreement must be considered alongside predicted accessibility. On GM12878, the order-5 Markov model has realism of 0.897. Its predicted accessibility is 0.172, compared with 0.611 for real sequences.

Our study concerns building and evaluating DNA generators in silico. The cost of experimental characterization motivates careful computational comparison before selecting sequences for synthesis. Motif agreement and predicted accessibility describe specific properties of the generated sequences, while cellular enhancer function requires experimental validation. We will clarify this scope in the Introduction and Discussion and qualify wording that could imply preservation of regulatory function.

Reviewer_DdpSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how classifier-free guidance (CFG) affects the biological realism of generated regulatory DNA sequences, identifying base-composition shifts as an important source of degradation under strong guidance. The authors introduce BaseDiff, a compact 21.6M-parameter convolutional discrete diffusion model for cell-type-specific enhancer generation, and propose composition-neutral guidance (CNG), a training-free modification that subtracts the position-averaged, base-specific component of the CFG logit update. This correction aims to preserve cell-type-specific regulatory sequence patterns while preventing guidance from excessively amplifying global nucleotide preferences. The authors further propose evaluating generated sequences through accessibility–quality trade-off curves, combining predicted chromatin accessibility with motif-frequency realism and motif co-occurrence. Experiments on DNase hypersensitive-site sequences from GM12878, HepG2, and K562, together with an additional lentiMPRA dataset, show that BaseDiff-CNG generally achieves better accessibility–quality trade-offs than DNA-Diffusion, D3, and Dirichlet flow matching. Additional analyses examine alternative guidance corrections, diffusion processes, model architectures, accessibility predictors, genomic contexts, and sequence diversity. The results suggest that controlling base composition during guidance can improve motif-level agreement with real regulatory sequences while largely preserving predicted accessibility.

Critical strengths
  • The first critical strength is the identification and systematic characterization of base-composition bias as a confounding factor in classifier-free guidance for regulatory DNA generation. The paper addresses an important problem in sequence design: stronger guidance can increase predicted regulatory activity while simultaneously introducing nucleotide-composition shifts that distort transcription-factor motif distributions.

  • The second critical strength is the relatively comprehensive experimental evaluation of the accessibility-quality trade-off in regulatory DNA generation. Rather than comparing models at a single guidance scale or relying exclusively on predicted enhancer activity, the authors evaluate how regulatory sequence quality changes as guidance strength increases.

Critical weaknesses
  • The first major weakness is that the biological validity of the generated regulatory sequences remains insufficiently established because the evaluation relies entirely on computational predictions and motif-based statistics. Although the authors report improvements in predicted chromatin accessibility and motif realism, neither directly establishes that the generated sequences function as cell-type-specific enhancers in biological experiments. Accessibility is evaluated using Enformer or ChromBPNet after inserting a generated 200-bp sequence into a fixed genomic context, whereas enhancer activity depends on additional biological factors that may not be adequately captured by these predictors. The authors appropriately acknowledge the absence of experimental validation, but this limitation remains consequential given the manuscript's emphasis on biologically faithful regulatory DNA design. Furthermore, motif-frequency agreement and motif co-occurrence are relatively coarse summaries of regulatory sequence structure. They do not directly evaluate motif spacing, orientation, positional organization, cooperative transcription-factor interactions, or other aspects of cis-regulatory grammar. Appendix D also demonstrates that simple fifth-order Markov models can reproduce much of the motif-realism score without explicitly modeling regulatory grammar. Consequently, the reported improvements establish better agreement with selected computational proxies but provide limited evidence that CNG produces functionally superior enhancers. Experimental validation through MPRA or related reporter assays, or at least stronger independent computational evaluations of regulatory grammar and sequence function, would substantially strengthen the biological conclusions.

  • The second major weakness is that the improvement attributable specifically to CNG is relatively modest and inconsistent across cell types, while the broader performance gains are substantially influenced by the BaseDiff architecture. Table 1 shows that BaseDiff without CNG already outperforms several larger baseline models on most accessibility–quality comparisons, suggesting that the convolutional backbone and associated architectural choices account for an important fraction of the overall improvement. The incremental benefit of CNG is considerably smaller than the comparison between BaseDiff-CNG and other generative frameworks. For example, on HepG2, the accessibility–co-occurrence trade-off area remains unchanged at 0.398 with and without CNG, while GM12878 improves only from 0.402 to 0.405. More importantly, Appendix Table 8 shows that at guidance scale w=6w=6, CNG actually reduces HepG2 motif co-occurrence from 0.780 to 0.764, despite improving motif realism. These results indicate that removing the global composition component does not consistently preserve higher-order motif relationships. Additionally, Appendix M demonstrates that using cell-type-specific base frequencies in the forward diffusion process recovers much of the same improvement without applying CNG, suggesting that the observed benefit is not unique to the proposed sampling correction. Although the training-free nature of CNG is valuable, the paper does not demonstrate its effectiveness when directly applied to independently developed DNA diffusion models such as DNA-Diffusion or D3. This limits the evidence for CNG as a generally applicable guidance strategy rather than a correction primarily validated within the proposed BaseDiff framework.

Questions for authors
  • Could the authors clarify whether the improvements in motif realism and co-occurrence reflect genuine recovery of regulatory grammar or primarily correction of low-order nucleotide statistics? The central motivation is that strong CFG distorts motif distributions by amplifying global base-composition preferences. However, both evaluation metrics are derived from PWM-based motif scans, which are themselves sensitive to nucleotide composition and local sequence statistics. Appendix D demonstrates that an order-5 Markov generator achieves motif-realism scores of approximately 0.884-0.900 and motif co-occurrence scores of approximately 0.859-0.889 across the evaluated cell types, despite lacking an explicit model of regulatory grammar. This raises the question of how much of the improvement obtained through CNG reflects biologically meaningful motif organization rather than better preservation of short-range sequence statistics. Could the authors evaluate motif frequencies and co-occurrence after controlling for GC content and dinucleotide or higher-order k-mer distributions? Additional analyses of motif spacing, orientation, and positional dependencies would help determine whether CNG preserves regulatory structure beyond composition-dependent motif occurrence. This distinction would materially affect the interpretation of the proposed method as improving biological realism rather than primarily correcting nucleotide-distribution artifacts.

  • Could the authors provide a more controlled assessment of the contribution of CNG independently of the BaseDiff architecture and demonstrate whether the correction generalizes to other diffusion models? Table 1 indicates that BaseDiff already achieves strong performance without CNG, and Appendix N suggests that replacing its convolutional backbone with transformer-based alternatives produces relatively large performance differences. In contrast, the additional improvements from CNG are smaller and vary across cell types and evaluation metrics. Since CNG operates directly on the conditional and unconditional logit predictions, its formulation appears potentially applicable to other discrete diffusion models without retraining. Applying the correction to independently developed models such as D3 would help establish whether the observed improvements reflect a general property of classifier-free guidance or depend on characteristics of the BaseDiff architecture and training procedure. It would also be useful to distinguish the performance gains attributable to the convolutional backbone, the diffusion objective, and the guidance correction through controlled comparisons. Such experiments would strengthen the claim that CNG constitutes a broadly useful methodological contribution.

  • Could the authors explain why CNG does not consistently improve motif co-occurrence, particularly for HepG2, and whether the removed composition component may also contain biologically meaningful regulatory information? The proposed correction assumes that the position-independent component of the CFG update primarily reflects undesirable nucleotide-composition bias, while the remaining position-dependent component captures biologically relevant motif preferences. However, this decomposition is not necessarily equivalent to separating nuisance information from regulatory information. Global base composition can itself be associated with biologically meaningful regulatory properties, including transcription-factor binding preferences and cell-type-specific sequence characteristics. Appendix Table 8 shows that at w=6, HepG2 motif co-occurrence decreases from 0.780 under standard CFG to 0.764 under CNG, despite improvements in motif realism and composition matching. Furthermore, Table 1 shows essentially no aggregate co-occurrence improvement for HepG2. Could the authors analyze which transcription-factor motifs and motif pairs improve or deteriorate after applying CNG, and whether these changes are associated with motif GC content or cell-type-specific regulatory functions? It would also be informative to examine partial composition correction, in which only a fraction of the position-averaged guidance component is removed. These analyses would clarify whether complete removal of the global component is generally appropriate or whether it can suppress biologically useful conditioning signals.

  • Could the authors provide stronger evidence that the generated sequences achieve biologically meaningful cell-type-specific enhancer activity beyond predictions from Enformer and ChromBPNet? The manuscript evaluates accessibility by inserting generated sequences into fixed genomic contexts and obtaining predictions from pretrained models. Although using two predictors and multiple insertion loci improves robustness, both approaches remain computational approximations of regulatory activity, and neither directly measures enhancer function. The lentiMPRA dataset provides experimentally measured transcriptional activity for real sequences, but the manuscript uses only categorical cell-type labels during training and evaluates generated sequences through ChromBPNet rather than a predictor trained directly on the measured reporter activities. Could the authors evaluate generated sequences using an independently trained MPRA activity predictor or assess their agreement with experimentally characterized regulatory motifs and sequence-function relationships? Ideally, a subset of generated sequences from BaseDiff with and without CNG could be tested experimentally under matched predicted accessibility conditions. This would help determine whether improved motif statistics translate into stronger or more selective regulatory activity, which is essential for establishing the practical value of the method in enhancer design.

  • Could the authors clarify the sensitivity of the accessibility–quality trade-off results to the evaluation metrics, reference points, and guidance-scale sampling strategy? The proposed trade-off-area metric is a useful way to compare models across guidance strengths, but its numerical value depends on the selected accessibility range, quality floor, and set of evaluated guidance scales. Table 1 computes areas using a quality floor of 0.5, whereas Appendix Table 7 reports that several lentiMPRA co-occurrence values fall below this threshold and therefore contribute nothing to the area. The authors also evaluate only five guidance scales, which may not capture the optimal trade-off curves equally well for models with different sensitivity to guidance. Although Appendix O provides comparisons over shared accessibility ranges, some improvements remain relatively small, particularly for HepG2 and GM12878. Could the authors report sensitivity analyses using alternative quality floors, denser guidance-scale grids, and normalized or fixed-accessibility comparisons? It would also be useful to report performance at matched accessibility levels rather than relying primarily on integrated trade-off areas. These analyses would help establish whether the reported model rankings reflect robust improvements in sequence quality or are partly influenced by the chosen evaluation protocol.

3 comments
AuthorsOpenReview ↗

Author Response (1/3)

Thank you for recognizing the composition-bias finding and the evaluation across guidance strengths. We have now tested CNG on D3 on both datasets and extended the evaluation to eleven further ENCODE cell types. We also examined composition-controlled comparisons, partial correction, and sensitivity to the evaluation protocol.

Contribution of CNG and transfer beyond BaseDiff. The comparison in Figure 3 samples one trained BaseDiff model with and without CNG. It isolates the guidance correction while holding the architecture, objective, and weights fixed. Table 1 reports the two versions separately, and Appendix N compares backbones and objectives. The architecture and correction each contribute to the reported results.

The table below gives the paired difference between the two samplings, BaseDiff-CNG vs. BaseDiff, at the submitted quality floor of 0.5.

DatasetCell typerealism area marginppco-occurrence area marginpp
DHSGM12878+0.0070±0.0043+0.0070 \pm 0.00430.0010.001+0.0033±0.0035+0.0033 \pm 0.00350.0700.070
HepG2+0.0061±0.0027+0.0061 \pm 0.0027<10−4<10^{-4}−0.0001±0.0040-0.0001 \pm 0.00400.9490.949
K562+0.0253±0.0047+0.0253 \pm 0.0047<10−4<10^{-4}+0.0592±0.0076+0.0592 \pm 0.0076<10−4<10^{-4}
lentiMPRAK562+0.0074±0.0018+0.0074 \pm 0.0018<10−4<10^{-4}+0.0106±0.0029+0.0106 \pm 0.0029<10−4<10^{-4}
HepG2+0.0114±0.0029+0.0114 \pm 0.0029<10−4<10^{-4}+0.0458±0.0064+0.0458 \pm 0.0064<10−4<10^{-4}
WTC11+0.0012±0.0011+0.0012 \pm 0.00110.0300.030+0.0067±0.0013+0.0067 \pm 0.0013<10−4<10^{-4}

Eleven of the twelve margins are positive, and nine remain significant at the Bonferroni threshold for twelve tests, α=0.05/12\alpha = 0.05/12.

We additionally sampled the same trained D3 baseline with and without CNG. DHS areas follow Table 1, and lentiMPRA areas use Table 7's shared accessibility range.

DatasetCell typeMetricD3D3 + CNGBaseDiff-CNG
DHSGM12878Realism0.368±0.0070.368 \pm 0.0070.388±0.0070.388 \pm 0.0070.388±0.0060.388 \pm 0.006
DHSHepG2Realism0.373±0.0080.373 \pm 0.0080.387±0.0080.387 \pm 0.0080.396±0.0080.396 \pm 0.008
DHSK562Realism0.361±0.0070.361 \pm 0.0070.378±0.0050.378 \pm 0.0050.392±0.0060.392 \pm 0.006
DHSGM12878Co-occurrence0.388±0.0040.388 \pm 0.0040.400±0.0040.400 \pm 0.0040.405±0.0040.405 \pm 0.004
DHSHepG2Co-occurrence0.389±0.0050.389 \pm 0.0050.399±0.0050.399 \pm 0.0050.398±0.0040.398 \pm 0.004
DHSK562Co-occurrence0.374±0.0060.374 \pm 0.0060.398±0.0060.398 \pm 0.0060.423±0.0050.423 \pm 0.005
lentiMPRAK562Realism0.064±0.0030.064 \pm 0.0030.070±0.0020.070 \pm 0.0020.074±0.0010.074 \pm 0.001
lentiMPRAHepG2Realism0.101±0.0040.101 \pm 0.0040.110±0.0020.110 \pm 0.0020.120±0.0020.120 \pm 0.002
lentiMPRAWTC11Realism0.046±0.0010.046 \pm 0.0010.046±0.0010.046 \pm 0.0010.047±0.0010.047 \pm 0.001
lentiMPRAK562Co-occurrence0.054±0.0030.054 \pm 0.0030.067±0.0020.067 \pm 0.0020.074±0.0020.074 \pm 0.002
lentiMPRAHepG2Co-occurrence0.069±0.0060.069 \pm 0.0060.114±0.0030.114 \pm 0.0030.121±0.0030.121 \pm 0.003
lentiMPRAWTC11Co-occurrence0.042±0.0010.042 \pm 0.0010.049±0.0010.049 \pm 0.0010.052±0.0010.052 \pm 0.001

CNG increases D3's area in all twelve comparisons, including WTC11 realism before rounding. Across the eleven additional ENCODE cell types, BaseDiff-CNG also has larger realism areas than plain D3 in all eleven and larger co-occurrence areas in ten.

Appendix M provides a complementary intervention. Matching the forward-process noise to each cell type's composition reduces much of the same distortion through training. Agreement between the training and sampling interventions supports the interpretation that composition contributes to the guidance effect. CNG can be applied to an existing trained model without changing its weights.

Scope of the biological conclusions. Our study concerns the development and evaluation of diffusion models for DNA generation in silico. We identify a guidance-induced distortion in composition and motif statistics and evaluate a correction through accessibility–quality trade-offs. Experimental characterization of generated sequences requires substantial resources, which motivates careful computational comparison before selecting sequences for synthesis.

BaseDiff is trained on sequence and cell-type labels, without using the accessibility predictors as a training reward. We evaluate it with two predictors, two genomic insertion loci, and two datasets. The lentiMPRA experiment tests whether the guidance effect also occurs in sequences obtained from a different assay. A reporter-activity predictor would evaluate an additional predicted property, and conditioning on a continuous activity is a different task that the paper names as a limitation and we leave to future work. Our claims concern the reported motif statistics and predicted accessibility, and cellular enhancer function requires experimental validation. We will make this scope clearer in the Introduction and Discussion and qualify wording that could imply functional validation.

AuthorsOpenReview ↗

Author Response (2/3)

Composition controls and regulatory grammar. Correcting composition-dependent distortions in motif statistics is part of CNG's intended benefit. We included the Markov controls in Appendix D to show why motif agreement must be evaluated jointly with predicted accessibility. On GM12878, the order-5 Markov model reaches realism of 0.897. Its predicted accessibility is 0.172 compared with 0.611 for real sequences.

We reweighted ordinary CFG samples toward the real GC distribution and recomputed the accessibility–quality curves. This recovers much of CNG's improvement on GM12878 and HepG2. On K562, CNG retains larger realism and co-occurrence areas, by 0.008 and 0.017. We also used approximate dinucleotide balancing. The following are point estimates on K562 at w=6w=6.

Sampling or controlRealismCo-occurrence
Ordinary CFG0.7210.700
Ordinary CFG with dinucleotide balancing0.7570.738
CNG0.7800.820

For approximate dinucleotide balancing, we trained a logistic classifier to distinguish real and generated sequences from their 16 dinucleotide frequencies. We used five-fold cross-fitting and weighted each generated sequence by its predicted odds of belonging to the real set. We then recomputed motif realism and co-occurrence using these weights. The effective sample size was 564 from 5,000 sequences in this comparison.

CNG remains ahead of both controls on K562, where matching composition does not account for the whole gain. An added benefit is that CNG applies to every sample as it is generated, while reweighting and filtering approaches are less sample efficient (see Reviewer X1Kk comment 2).

We report the metrics used throughout our other evaluations, which are motif realism (from prior work) and motif co-occurrence which extends it to pairs. Spacing, orientation and physical cooperativity are further aspects of regulatory grammar. We will qualify broader grammar language in the paper accordingly.

HepG2 and partial correction. The HepG2 co-occurrence decrease at w=6w=6 is present in the reported results. Reweighting ordinary CFG samples toward the real GC distribution produces a similar decrease, from 0.780 to 0.766. CNG gives 0.764. The change involves many motifs and is associated with their GC content, and CNG brings motif frequencies closer to real on all three cell types.

We tested removal fractions α=0,0.25,0.5,0.75,1\alpha=0,0.25,0.5,0.75,1, where zero gives ordinary CFG and one gives CNG. The following are point estimates at the same w=6w=6 used in your question.

Fraction removedGM12878 realismGM12878 co-occurrenceHepG2 realismHepG2 co-occurrenceK562 realismK562 co-occurrence
00.69420.82130.68850.78000.72150.6998
0.250.70920.82680.69210.77550.74340.7286
0.50.72140.83030.70070.77350.75610.7533
0.750.72600.83410.70460.76850.76830.7771
10.72150.83070.70890.76390.77960.8202

On GM12878 and K562 both metrics improve. HepG2 shows a trade-off between realism and co-occurrence. The largest gain is on K562, whose GC content lies furthest from the average over cell types, and the smallest on HepG2, which is already close to it. CNG also improves the trade-offs in most of the eleven additional ENCODE cell types. Global composition can contain regulatory information. CNG retains the conditional prediction and removes a position-constant term from the additional guidance update. We will clarify in Section 3.3 that the decomposition is defined in logit space and its effects on motif statistics are evaluated empirically.

AuthorsOpenReview ↗

Author Response (3/3)

Quality floor. We repeated the area calculations at floors of 00, 0.30.3, 0.50.5, 0.60.6 and 0.70.7 while retaining each table's accessibility range. On lentiMPRA, BaseDiff-CNG has the largest area in all six comparisons at every tested floor. Below are its margins over the best remaining Table 7 method.

Cell typeMetricFloor 0Floor 0.3Floor 0.5Floor 0.6Floor 0.7
K562Realism+0.0074±0.0018+0.0074 \pm 0.0018+0.0074±0.0018+0.0074 \pm 0.0018+0.0074±0.0018+0.0074 \pm 0.0018+0.0074±0.0018+0.0074 \pm 0.0018+0.0074±0.0018+0.0074 \pm 0.0018
HepG2Realism+0.0114±0.0029+0.0114 \pm 0.0029+0.0114±0.0029+0.0114 \pm 0.0029+0.0114±0.0029+0.0114 \pm 0.0029+0.0114±0.0029+0.0114 \pm 0.0029+0.0114±0.0029+0.0114 \pm 0.0029
WTC11Realism+0.0012±0.0011+0.0012 \pm 0.0011+0.0012±0.0011+0.0012 \pm 0.0011+0.0012±0.0011+0.0012 \pm 0.0011+0.0012±0.0011+0.0012 \pm 0.0011+0.0012±0.0011+0.0012 \pm 0.0011
K562Co-occurrence+0.0106±0.0029+0.0106 \pm 0.0029+0.0106±0.0029+0.0106 \pm 0.0029+0.0106±0.0029+0.0106 \pm 0.0029+0.0106±0.0029+0.0106 \pm 0.0029+0.0106±0.0029+0.0106 \pm 0.0029
HepG2Co-occurrence+0.0458±0.0064+0.0458 \pm 0.0064+0.0458±0.0064+0.0458 \pm 0.0064+0.0458±0.0064+0.0458 \pm 0.0064+0.0439±0.0055+0.0439 \pm 0.0055+0.0405±0.0045+0.0405 \pm 0.0045
WTC11Co-occurrence+0.0067±0.0013+0.0067 \pm 0.0013+0.0067±0.0013+0.0067 \pm 0.0013+0.0067±0.0013+0.0067 \pm 0.0013+0.0067±0.0013+0.0067 \pm 0.0013+0.0067±0.0013+0.0067 \pm 0.0013

Table 7 computes each area over the accessibility range shared by all methods, where lowering the floor adds the same area to every method and leaves the margins unchanged. The exception is HepG2 co-occurrence at 0.60.6 and 0.70.7, where more points get clipped because the floor is too high. The same check on DHS preserves the main pattern of larger K562 gains and similar HepG2 co-occurrence areas with and without CNG.

Guidance grid density. We added w=0.5,3,5w=0.5,3,5 for BaseDiff with and without CNG, giving eight scales instead of five. At the submitted floor, the K562 realism-area improvement changes from 0.0253 to 0.0225. The co-occurrence improvement changes from 0.0592 to 0.0534. The denser grid raises each method by a similar amount, which leaves the comparison between them unchanged. Every method's area rises with the number of points, which is why Table 1 and Table 7 continue to use the same five scales.

Quality at matched accessibility. We also interpolated the BaseDiff curves at the predicted accessibility of real training sequences for each cell type. The table reports point estimates, with linear interpolation between adjacent evaluated guidance scales.

Cell typeAccessibilityPlain CFG realismCNG realismPlain CFG co-occurrenceCNG co-occurrence
GM128780.6110.9030.9070.9110.914
HepG20.7600.8860.8930.8950.894
K5620.4370.9160.9230.9150.930

The largest gains remain on K562 when predicted accessibility is held fixed at the real level.

We hope the additional experiments and clarifications address your concerns. If so, we would be grateful if you would reconsider your score.

Reviewer_X1KkOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper identifies an important issue in classifier-free guidance (CFG) for DNA diffusion models: stronger guidance can amplify biases in base composition, leading to distorted motif distributions even as predicted chromatin accessibility improves. To address this issue, the authors propose composition-neutral guidance (CNG), a training-free correction that removes the position-averaged guidance update for each nucleotide. They also introduce BaseDiff, a compact convolutional discrete diffusion model, and evaluate DNA generators using accessibility–quality trade-off curves. Experiments across two datasets show that BaseDiff with CNG reduces composition shifts and improves motif realism and co-occurrence while largely preserving predicted accessibility, outperforming several existing baselines.

Critical strengths
  1. The paper identifies a meaningful issue in guided DNA generation and provides a clear analysis of how base composition shifts can distort motif statistics under stronger CFG.
  2. CNG is simple, well motivated, and requires no additional training. The analysis of GC content and motif-level errors provides useful evidence supporting the proposed correction.
  3. BaseDiff achieves strong results with substantially fewer parameters than existing diffusion models. The experiments are fairly comprehensive, covering multiple baselines, two datasets, different accessibility predictors, and several useful ablations.
Critical weaknesses
  1. While CNG is presented as a general training-free correction, it is only evaluated on BaseDiff. Applying it to other diffusion models such as DNA-Diffusion or D3 would provide stronger evidence of its general applicability.
  2. Although the authors compare several composition correction strategies, it remains unclear whether CNG provides benefits beyond matching the generated sequences to the real base composition. For example, ordinary CFG samples could be filtered to match the target cell's base composition and compared with CNG at similar predicted accessibility, while accounting for the additional sampling cost.
  3. The evaluation covers only four distinct cell types across the two datasets. Testing more diverse cell types, such as additional ENCODE DHS cell types or the multi-cell-type Drosophila brain dataset used by D3, would provide stronger evidence for the generality of the observed bias and the effectiveness of CNG.
Questions for authors

How sensitive are the motif quality results to the scanning procedure? For example, do the improvements hold with an alternative motif collection such as HOCOMOCO or different scanning thresholds?

3 comments
AuthorsOpenReview ↗

Author Response (1/2)

Thank you for the positive assessment and the suggestions. We tested CNG on D3 on both datasets, extended the evaluation to eleven further ENCODE cell types, and evaluated filtering and motif-scan sensitivity.

Transfer to another model. We sampled the same trained D3 baseline with and without CNG, using identical weights and the same five guidance scales. The table uses the submitted quality floor of 0.5. DHS areas follow Table 1, and lentiMPRA areas use Table 7's shared accessibility range.

DatasetCell typeMetricD3D3 + CNGBaseDiff-CNG
DHSGM12878Realism0.368±0.0070.368 \pm 0.0070.388±0.0070.388 \pm 0.0070.388±0.0060.388 \pm 0.006
DHSHepG2Realism0.373±0.0080.373 \pm 0.0080.387±0.0080.387 \pm 0.0080.396±0.0080.396 \pm 0.008
DHSK562Realism0.361±0.0070.361 \pm 0.0070.378±0.0050.378 \pm 0.0050.392±0.0060.392 \pm 0.006
DHSGM12878Co-occurrence0.388±0.0040.388 \pm 0.0040.400±0.0040.400 \pm 0.0040.405±0.0040.405 \pm 0.004
DHSHepG2Co-occurrence0.389±0.0050.389 \pm 0.0050.399±0.0050.399 \pm 0.0050.398±0.0040.398 \pm 0.004
DHSK562Co-occurrence0.374±0.0060.374 \pm 0.0060.398±0.0060.398 \pm 0.0060.423±0.0050.423 \pm 0.005
lentiMPRAK562Realism0.064±0.0030.064 \pm 0.0030.070±0.0020.070 \pm 0.0020.074±0.0010.074 \pm 0.001
lentiMPRAHepG2Realism0.101±0.0040.101 \pm 0.0040.110±0.0020.110 \pm 0.0020.120±0.0020.120 \pm 0.002
lentiMPRAWTC11Realism0.046±0.0010.046 \pm 0.0010.046±0.0010.046 \pm 0.0010.047±0.0010.047 \pm 0.001
lentiMPRAK562Co-occurrence0.054±0.0030.054 \pm 0.0030.067±0.0020.067 \pm 0.0020.074±0.0020.074 \pm 0.002
lentiMPRAHepG2Co-occurrence0.069±0.0060.069 \pm 0.0060.114±0.0030.114 \pm 0.0030.121±0.0030.121 \pm 0.003
lentiMPRAWTC11Co-occurrence0.042±0.0010.042 \pm 0.0010.049±0.0010.049 \pm 0.0010.052±0.0010.052 \pm 0.001

CNG increases D3's area in all twelve comparisons, including WTC11 realism before rounding. BaseDiff-CNG remains competitive after correcting D3, with larger areas in most comparisons and fewer than one quarter of its parameters.

Additional cell types. We evaluated eleven further ENCODE cell types from the same DHS index, using the same sequence construction and filtering as the original dataset. Ten are additional cell types, and hESCT0 was included in training but not previously evaluated. The table reports areas over the accessibility range shared by the three methods, with 95%95\% bootstrap half-widths. This additional analysis uses a quality floor of zero.

Cell typeD3 realismBaseDiff realismBaseDiff-CNG realismD3 co-occurrenceBaseDiff co-occurrenceBaseDiff-CNG co-occurrence
NHDF Ad0.454±0.0090.454 \pm 0.0090.464±0.0090.464 \pm 0.0090.488±0.0080.488 \pm 0.0080.508±0.0080.508 \pm 0.0080.525±0.0030.525 \pm 0.0030.532±0.0030.532 \pm 0.003
DE0.060±0.0010.060 \pm 0.0010.059±0.0010.059 \pm 0.0010.070±0.0020.070 \pm 0.0020.056±0.0010.056 \pm 0.0010.054±0.0010.054 \pm 0.0010.060±0.0020.060 \pm 0.002
HAEpiC0.324±0.0050.324 \pm 0.0050.330±0.0050.330 \pm 0.0050.344±0.0040.344 \pm 0.0040.352±0.0050.352 \pm 0.0050.327±0.0060.327 \pm 0.0060.341±0.0060.341 \pm 0.006
NHLF0.358±0.0060.358 \pm 0.0060.377±0.0050.377 \pm 0.0050.388±0.0040.388 \pm 0.0040.380±0.0050.380 \pm 0.0050.386±0.0040.386 \pm 0.0040.387±0.0040.387 \pm 0.004
M059J0.153±0.0030.153 \pm 0.0030.166±0.0020.166 \pm 0.0020.167±0.0020.167 \pm 0.0020.153±0.0020.153 \pm 0.0020.166±0.0020.166 \pm 0.0020.164±0.0010.164 \pm 0.001
Trophoblast0.232±0.0070.232 \pm 0.0070.276±0.0040.276 \pm 0.0040.283±0.0040.283 \pm 0.0040.228±0.0050.228 \pm 0.0050.260±0.0030.260 \pm 0.0030.270±0.0040.270 \pm 0.004
MCF10a ER SRC 24h Tam0.200±0.0060.200 \pm 0.0060.216±0.0030.216 \pm 0.0030.217±0.0030.217 \pm 0.0030.196±0.0060.196 \pm 0.0060.210±0.0020.210 \pm 0.0020.211±0.0020.211 \pm 0.002
HEEpiC0.285±0.0050.285 \pm 0.0050.298±0.0030.298 \pm 0.0030.303±0.0030.303 \pm 0.0030.272±0.0060.272 \pm 0.0060.304±0.0030.304 \pm 0.0030.308±0.0030.308 \pm 0.003
hESCT00.222±0.0060.222 \pm 0.0060.259±0.0030.259 \pm 0.0030.261±0.0030.261 \pm 0.0030.210±0.0060.210 \pm 0.0060.271±0.0030.271 \pm 0.0030.284±0.0040.284 \pm 0.004
HCPEpiC0.083±0.0060.083 \pm 0.0060.092±0.0010.092 \pm 0.0010.094±0.0010.094 \pm 0.0010.090±0.0070.090 \pm 0.0070.094±0.0010.094 \pm 0.0010.094±0.0010.094 \pm 0.001
EH0.015±0.0040.015 \pm 0.0040.015±0.0000.015 \pm 0.0000.016±0.0000.016 \pm 0.0000.015±0.0040.015 \pm 0.0040.015±0.0000.015 \pm 0.0000.016±0.0000.016 \pm 0.000

BaseDiff-CNG has larger realism areas than D3 in all eleven and larger co-occurrence areas in ten. CNG has a positive effect relative to plain BaseDiff across most cell types and is neutral at the others. This extends the evaluation beyond the original four distinct cell types. Cross-species transfer to the fly-brain dataset remains a useful further test.

AuthorsOpenReview ↗

Author Response (2/2)

Filtering ordinary CFG samples. We retained plain-CFG sequences whose GC content was within two percentage points of the real reference mean for that cell type. We then recomputed both motif quality and predicted accessibility on the retained sequences at each guidance scale.

Cell typeMetricPlain CFGGC-filtered CFGCNG
GM12878Realism0.381±0.0020.381 \pm 0.0020.363±0.0050.363 \pm 0.0050.388±0.0020.388 \pm 0.002
HepG2Realism0.390±0.0020.390 \pm 0.0020.358±0.0070.358 \pm 0.0070.396±0.0030.396 \pm 0.003
K562Realism0.367±0.0020.367 \pm 0.0020.372±0.0040.372 \pm 0.0040.392±0.0020.392 \pm 0.002
GM12878Co-occurrence0.402±0.0030.402 \pm 0.0030.385±0.0080.385 \pm 0.0080.405±0.0030.405 \pm 0.003
HepG2Co-occurrence0.398±0.0050.398 \pm 0.0050.366±0.0110.366 \pm 0.0110.398±0.0040.398 \pm 0.004
K562Co-occurrence0.364±0.0050.364 \pm 0.0050.382±0.0110.382 \pm 0.0110.423±0.0050.423 \pm 0.005

Filtering improves K562 but falls short of CNG and reduces the areas on the other two cell types. It retains 407–1,305 of 5,000 sequences across cells and guidance scales, or 8–26% of the pool. The sampling cost is approximately 4–12 generated sequences per retained sequence. CNG applies during sampling and retains every generated sequence.

We also tried reweighting plain-CFG samples toward the real GC distribution. This recovers much of the improvement on GM12878 and HepG2, while on K562 CNG retains larger realism and co-occurrence areas, by 0.008 and 0.017. These comparisons show that composition explains an important part of the problem, and that CNG has a practical advantage in sampling efficiency.

Motif library. We repeated the analysis with HOCOMOCO v14 CORE, which contains 1,595 motif matrices against JASPAR's 1,019. The following are area differences between BaseDiff-CNG and the same BaseDiff model with ordinary CFG.

Cell typeJASPAR realismHOCOMOCO realismJASPAR co-occurrenceHOCOMOCO co-occurrence
GM12878+0.0070±0.0043+0.0070 \pm 0.0043+0.0132±0.0040+0.0132 \pm 0.0040+0.0033±0.0035+0.0033 \pm 0.0035+0.0115±0.0053+0.0115 \pm 0.0053
HepG2+0.0061±0.0027+0.0061 \pm 0.0027+0.0075±0.0024+0.0075 \pm 0.0024−0.0001±0.0040-0.0001 \pm 0.0040+0.0216±0.0092+0.0216 \pm 0.0092
K562+0.0253±0.0047+0.0253 \pm 0.0047+0.0269±0.0046+0.0269 \pm 0.0046+0.0592±0.0076+0.0592 \pm 0.0076+0.1401±0.0135+0.1401 \pm 0.0135

The realism gains hold under the alternative library, and the largest co-occurrence gain remains on K562.

Scanning threshold. We also rescanned the JASPAR library at p=10−3p=10^{-3} and p=10−5p=10^{-5}, around the paper's p=10−4p=10^{-4}, and recomputed both areas over all five guidance scales (w=0,1,2,4,6w = 0, 1, 2, 4, 6). The margins below are BaseDiff-CNG over the next best baseline (excluding BaseDiff with plain CFG), at a quality floor of zero.

ThresholdCellrealism area marginppco-occurrence area marginpp
10−310^{-3}GM12878+0.0255±0.0041+0.0255 \pm 0.0041<10−4<10^{-4}+0.0179±0.0027+0.0179 \pm 0.0027<10−4<10^{-4}
HepG2+0.0336±0.0047+0.0336 \pm 0.0047<10−4<10^{-4}+0.0212±0.0029+0.0212 \pm 0.0029<10−4<10^{-4}
K562+0.0433±0.0042+0.0433 \pm 0.0042<10−4<10^{-4}+0.0392±0.0039+0.0392 \pm 0.0039<10−4<10^{-4}
10−410^{-4}GM12878+0.0279±0.0048+0.0279 \pm 0.0048<10−4<10^{-4}+0.0247±0.0055+0.0247 \pm 0.0055<10−4<10^{-4}
HepG2+0.0326±0.0061+0.0326 \pm 0.0061<10−4<10^{-4}+0.0227±0.0075+0.0227 \pm 0.0075<10−4<10^{-4}
K562+0.0379±0.0059+0.0379 \pm 0.0059<10−4<10^{-4}+0.0565±0.0079+0.0565 \pm 0.0079<10−4<10^{-4}
10−510^{-5}GM12878+0.0266±0.0061+0.0266 \pm 0.0061<10−4<10^{-4}−0.0298±0.0706-0.0298 \pm 0.07060.4080.408
HepG2+0.0356±0.0070+0.0356 \pm 0.0070<10−4<10^{-4}+0.0140±0.0273+0.0140 \pm 0.02730.3130.313
K562+0.0321±0.0084+0.0321 \pm 0.0084<10−4<10^{-4}+0.0063±0.0376+0.0063 \pm 0.03760.7440.744

The realism margin holds at all three thresholds and all three cell types. Co-occurrence is more sensitive to the threshold, with less precise comparisons at 10−510^{-5} as motif hits become sparse.

Thank you for the detailed response and additional experiments, which have adequately addressed my concerns. I will maintain my positive rating and wish the authors the best of luck!

Reviewer_Ya2ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies DNA generation for a target cell type. It finds that strong classifier-free guidance can raise predicted accessibility while changing the overall mix of A, C, G, and T. This change can also affect motif patterns. The authors propose composition-neutral guidance (CNG). At each sampling step, it removes the average guidance update for each base across sequence positions. It needs no retraining. The authors test CNG with BaseDiff on two datasets. They report motif patterns closer to real data, with little change in predicted accessibility.

Critical strengths
  1. The paper points out a clear problem with strong guidance. A higher prediction score may come with larger changes in base composition and motif patterns. CNG provides a simple way to reduce this problem. It only changes sampling and needs no retraining.

  2. The experiments include two datasets and two accessibility prediction models. The authors also test a second DNA insertion location and compare several guidance rules. The comparison with and without CNG uses the same trained BaseDiff model. This helps show the effect of the correction.

Critical weaknesses
  1. CNG is tested with BaseDiff, while the other generators and alternative backbones use plain guidance. Its value across other models remains unclear. The model ranking may change if methods such as D3 also use CNG. Testing this would help show how widely the correction can be used.

  2. The area score uses a quality floor of 0.5. In Table 7, four points fall below this floor and contribute nothing to the area. I did not find a check using other floors. It is therefore unclear whether the model ranking is stable under this choice, especially on the second dataset.

Questions for authors
  1. Can the authors test CNG on D3 or one of the alternative backbones? How much does it improve those models? Does BaseDiff still perform better when the other models use the same correction?

  2. Why is the quality floor set to 0.5? Do the model rankings stay the same with other reasonable floors? Appendix O changes the accessibility range but does not test this choice.

1 comment
AuthorsOpenReview ↗

Author Response

Thank you for the positive assessment and the two questions. We tested CNG on D3 on both datasets and repeated the area calculations with alternative quality floors.

Transfer to D3. We sampled our trained D3 baseline with and without CNG, keeping its weights, conditioning and guidance scales fixed. The following trade-off areas use the submitted quality floor of 0.5. DHS areas follow Table 1, and lentiMPRA areas use the shared accessibility range of Table 7.

DatasetCell typeMetricD3D3 + CNGBaseDiff-CNG
DHSGM12878Realism0.368±0.0070.368 \pm 0.0070.388±0.0070.388 \pm 0.0070.388±0.0060.388 \pm 0.006
DHSHepG2Realism0.373±0.0080.373 \pm 0.0080.387±0.0080.387 \pm 0.0080.396±0.0080.396 \pm 0.008
DHSK562Realism0.361±0.0070.361 \pm 0.0070.378±0.0050.378 \pm 0.0050.392±0.0060.392 \pm 0.006
DHSGM12878Co-occurrence0.388±0.0040.388 \pm 0.0040.400±0.0040.400 \pm 0.0040.405±0.0040.405 \pm 0.004
DHSHepG2Co-occurrence0.389±0.0050.389 \pm 0.0050.399±0.0050.399 \pm 0.0050.398±0.0040.398 \pm 0.004
DHSK562Co-occurrence0.374±0.0060.374 \pm 0.0060.398±0.0060.398 \pm 0.0060.423±0.0050.423 \pm 0.005
lentiMPRAK562Realism0.064±0.0030.064 \pm 0.0030.070±0.0020.070 \pm 0.0020.074±0.0010.074 \pm 0.001
lentiMPRAHepG2Realism0.101±0.0040.101 \pm 0.0040.110±0.0020.110 \pm 0.0020.120±0.0020.120 \pm 0.002
lentiMPRAWTC11Realism0.046±0.0010.046 \pm 0.0010.046±0.0010.046 \pm 0.0010.047±0.0010.047 \pm 0.001
lentiMPRAK562Co-occurrence0.054±0.0030.054 \pm 0.0030.067±0.0020.067 \pm 0.0020.074±0.0020.074 \pm 0.002
lentiMPRAHepG2Co-occurrence0.069±0.0060.069 \pm 0.0060.114±0.0030.114 \pm 0.0030.121±0.0030.121 \pm 0.003
lentiMPRAWTC11Co-occurrence0.042±0.0010.042 \pm 0.0010.049±0.0010.049 \pm 0.0010.052±0.0010.052 \pm 0.001

CNG increases D3's area in all twelve comparisons, including the WTC11 realism difference obscured by rounding. BaseDiff-CNG retains larger areas in most comparisons with 21.6M parameters, compared with D3's 92.4M. The corrected models are similar on DHS GM12878 realism and HepG2 co-occurrence.

Choice of quality floor. We used 0.5 as a common reference below all DHS quality values. As you note, four lentiMPRA co-occurrence points fall below it. We repeated the calculations at floors of 00, 0.30.3, 0.50.5, 0.60.6 and 0.70.7, keeping the accessibility ranges fixed. Below are BaseDiff-CNG's area differences from the best remaining method in Table 7.

Cell typeMetricFloor 0Floor 0.3Floor 0.5Floor 0.6Floor 0.7
K562Realism+0.0074±0.0018+0.0074 \pm 0.0018+0.0074±0.0018+0.0074 \pm 0.0018+0.0074±0.0018+0.0074 \pm 0.0018+0.0074±0.0018+0.0074 \pm 0.0018+0.0074±0.0018+0.0074 \pm 0.0018
HepG2Realism+0.0114±0.0029+0.0114 \pm 0.0029+0.0114±0.0029+0.0114 \pm 0.0029+0.0114±0.0029+0.0114 \pm 0.0029+0.0114±0.0029+0.0114 \pm 0.0029+0.0114±0.0029+0.0114 \pm 0.0029
WTC11Realism+0.0012±0.0011+0.0012 \pm 0.0011+0.0012±0.0011+0.0012 \pm 0.0011+0.0012±0.0011+0.0012 \pm 0.0011+0.0012±0.0011+0.0012 \pm 0.0011+0.0012±0.0011+0.0012 \pm 0.0011
K562Co-occurrence+0.0106±0.0029+0.0106 \pm 0.0029+0.0106±0.0029+0.0106 \pm 0.0029+0.0106±0.0029+0.0106 \pm 0.0029+0.0106±0.0029+0.0106 \pm 0.0029+0.0106±0.0029+0.0106 \pm 0.0029
HepG2Co-occurrence+0.0458±0.0064+0.0458 \pm 0.0064+0.0458±0.0064+0.0458 \pm 0.0064+0.0458±0.0064+0.0458 \pm 0.0064+0.0439±0.0055+0.0439 \pm 0.0055+0.0405±0.0045+0.0405 \pm 0.0045
WTC11Co-occurrence+0.0067±0.0013+0.0067 \pm 0.0013+0.0067±0.0013+0.0067 \pm 0.0013+0.0067±0.0013+0.0067 \pm 0.0013+0.0067±0.0013+0.0067 \pm 0.0013+0.0067±0.0013+0.0067 \pm 0.0013

BaseDiff-CNG has the largest area in all six lentiMPRA comparisons at every tested floor. Table 7 computes each area over the accessibility range shared by all methods, where lowering the floor adds the same area to every method and leaves the margins unchanged. The exception is HepG2 co-occurrence at 0.60.6 and 0.70.7, where more points get clipped because the floor is too high. The same check on DHS preserves the main findings, including the larger K562 gains and the similar HepG2 co-occurrence areas with and without CNG. We retain the submitted convention and report the floor sensitivity separately.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Note to all reviewers and the area chair

We thank the reviewers for their careful reading of the paper. We were glad that the base-composition finding was recognized as our main contribution, that the evaluation across guidance strengths and the controlled comparisons were noted, and that the work was rated as clear and easy to follow.

The main questions were: whether the correction (CNG) is specific to our model, how sensitive the quality metrics are to the motif scan, whether the biological claims are too strong, and how robust the trade-off protocol is. We plan four sets of additions for the camera-ready version, all computed on the same trained models, data, and guidance scales as the submission, along with revisions to the wording of our claims.

1. Does CNG transfer? (Reviewers Ya2Z, X1Kk, DdpS, V4GN) A new appendix applies CNG to the D3 transformer on both datasets. It improves D3's trade-off area in all twelve dataset/cell-type/metric combinations. The same appendix extends to eleven additional ENCODE cell types, where BaseDiff-CNG has the larger realism area in all eleven and the larger co-occurrence area in ten.

2. Sensitivity to the motif scan. (Reviewers X1Kk, V4GN) A new appendix varies the scan: HOCOMOCO v14 CORE in place of JASPAR, thresholds of 10−310^{-3} and 10−510^{-5} alongside the submission's 10−410^{-4}, and the three constants defining co-occurrence pair selection.

3. Composition controls. (Reviewers X1Kk, DdpS) Appendix J will add controls that filter or reweight standard classifier-free-guidance samples toward the real base composition, along with a comparison across partial correction strengths. Matching composition recovers much of the improvement on two of the three cell types but not the third, which is the one whose GC content is furthest from the cross-cell-type average.

4. Robustness of the trade-off protocol. (Reviewers Ya2Z, DdpS) Appendix O will add checks across five quality floors, a denser guidance grid, and quality at matched predicted accessibility. The ordering of the results does not change at any floor or on either grid.

Scoping the claims. We will reiterate in the Discussion that our conclusions are computational, and will add the same statement to the Introduction. Section 3.3 will state precisely that CNG is exact in logit space while its effect on sampled composition is approximate.

Details and full tables are in our responses to the individual reviews.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.