- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes PEAC, a framework for activity-conditioned joint generation of promoter–enhancer pairs. PEAC consists of three stages: regulatory prior distillation from promoter and enhancer teacher models into a shared autoregressive Transformer, activity-conditioned initialization using STARR-seq-derived element-level supervision, and pair fusion with bidirectional cross-attention and conditional maximum mean discrepancy (MMD) alignment using FANTOM5 CAGE-derived pair-level data. The authors evaluate PEAC on K562-related datasets using single-element activity predictors, a CAGE-expression oracle, sequence realism and diversity metrics, ablation studies, constructed-pair baselines, and multiple training and generation seeds. Compared with a direct-concatenation baseline, PEAC achieves higher target-bin hit rates under the primary pair-activity oracle for low-, medium-, and high-activity conditions (79.1%, 91.7%, and 67.8%, respectively, versus 31.8%, 68.6%, and 11.6%). The authors also report competitive sequence statistics and closer matching to the reference activity distribution than several constructed-pair baselines. However, independent-oracle evaluation reveals substantially weaker low-activity controllability, and an external EPInformer diagnostic shows little difference between matched and shuffled pairs, although the same limitation is observed for reference pairs. Overall, the results provide encouraging evidence for oracle-based activity-conditioned pair generation, but whether PEAC learns meaningful promoter–enhancer functional compatibility or produces biologically functional sequences remains insufficiently established.
Critical strengths
-
The paper addresses a meaningful and relatively underexplored problem. Extending regulatory DNA generation from individual promoters or enhancers to activity-conditioned promoter–enhancer co-generation is a relevant and potentially valuable research direction, given that regulatory activity depends on the combination of these elements. The problem formulation is clear, and the autoregressive factorization provides a natural way to generate an enhancer conditioned on both the target activity and the generated promoter. The work therefore offers a useful formulation for studying joint regulatory sequence generation.
-
The evaluation covers multiple aspects of generation quality and includes useful robustness checks. Beyond evaluating activity controllability with the primary CAGE-expression oracle, the paper reports sequence-level realism and diversity diagnostics, component ablations, constructed-pair baselines, independent-oracle evaluations, and multi-seed experiments. These experiments provide a broader assessment than relying on a single activity metric or a single model configuration. In particular, the independent-oracle analysis and constructed-pair controls help characterize the limitations of the reported activity-control results and provide a useful starting point for evaluating future methods for promoter–enhancer co-generation.
Critical weaknesses
-
The central claim of pair-specific functional modeling is not sufficiently established. The main evidence for activity controllability relies on a CAGE-expression oracle trained on the same pair-level data distribution used for model development. Although the authors evaluate the generated samples using three additional oracles trained with leakage-controlled splits, the low-activity target-bin hit rate drops from 79.1% under the primary oracle to 35.1% under majority voting across the independent oracles. Furthermore, the EPInformer diagnostic shows nearly identical predicted activity and pair-attention scores for matched and shuffled PEAC pairs. While the similar results for reference pairs suggest that this diagnostic may have limited sensitivity to pair-specific compatibility, it does not provide positive evidence that PEAC captures such compatibility either. Taken together, the results support some degree of oracle-predicted activity control, but do not sufficiently establish that the model learns meaningful promoter–enhancer functional compatibility beyond marginal sequence properties and condition-dependent generation. This limitation directly affects the central motivation and biological interpretation of the proposed framework.
-
The contribution of pair fusion is difficult to isolate, and potential shortcuts in the training objectives remain insufficiently controlled. The pair-fusion encoder and auxiliary prediction heads receive explicit target-condition information, while generated and reference representations for the MMD objective are computed using the same jointly trained encoder. Consequently, the activity-prediction heads may partially exploit the supplied condition, and the MMD objective may align representations through condition information rather than sequence-derived functional features. The authors acknowledge this potential shortcut in Appendix A.7, but the reported ablations and external-oracle evaluations do not directly rule it out. Separately, although the direct-concatenation baseline uses the same pair-level training data and activity labels, its modeling and training pipeline differs from that of PEAC, making it difficult to attribute the observed improvements specifically to cross-attention pair fusion. A condition-blind evaluation and a more closely matched baseline that differs primarily in the pair-fusion mechanism would be needed to establish the specific contributions of these components. Without such evidence, the extent to which pair fusion and latent distribution alignment enforce sequence-dependent functional compatibility remains unclear, weakening the methodological claims supported by the experiments.
Questions for authors
-
Robustness of activity conditioning. The low-activity target-bin hit rate decreases from 79.1% under the primary CAGE-expression oracle to 35.1% under majority voting across the three leakage-controlled oracles (Appendix Tables 7–8). The authors attribute this discrepancy partly to differences in oracle calibration near the low/medium decision boundary. Can the authors clarify how much of the discrepancy can be explained by this boundary effect, and whether the low-activity conditioning result remains robust under continuous-score comparisons or a calibration-independent evaluation? More broadly, how should the primary-oracle results be interpreted in light of the substantially weaker agreement across independent oracles?
-
Evidence for pair-specific functional compatibility. PEAC is motivated by the hypothesis that promoter–enhancer co-design benefits from modeling functional interactions between the two elements. However, EPInformer assigns nearly identical predicted activity and pair-attention scores to matched and shuffled PEAC pairs, with similar results for reference pairs (Appendix Table 10). What evidence supports the claim that PEAC learns meaningful promoter–enhancer compatibility, rather than primarily learning marginal sequence distributions and target activity labels? In particular, can the authors clarify what the current experiments establish about pair-specific functional compatibility, as distinct from the generation of plausible individual elements and the control of oracle-predicted pair activity?
-
Potential condition-information shortcuts. The pair-fusion encoder and auxiliary prediction heads receive explicit target-condition information, and the same jointly trained encoder is used to compute generated and reference representations in the MMD objective. Consequently, the activity-prediction heads may exploit the supplied condition, while the MMD loss may align representations partly through condition information rather than sequence-derived functional features. What evidence establishes that these objectives impose meaningful constraints on the generated sequences themselves? For example, have the authors evaluated condition-blind or sequence-only representations, or tested whether the predicted activity and latent representations change appropriately when promoter–enhancer sequences are shuffled while the target condition is held fixed? If such controls are unavailable, how should the claimed contributions of pair fusion and latent distribution alignment be interpreted?
-
Attribution of improvements to pair fusion. The direct-concatenation baseline differs from PEAC in its training and modeling pipeline, making it difficult to isolate the contribution of explicit pair-level modeling. Although the ablation study reports degradation when pair fusion is removed, can the authors clarify exactly which architectural and training components are changed in this ablation and which are held fixed? In particular, how much of the improvement can be attributed specifically to cross-attention between promoter and enhancer representations, rather than differences in supervision, optimization, or model capacity? A parameter-matched autoregressive baseline trained on the same pair data and activity labels, but without cross-attention pair fusion, would help establish the incremental value of the proposed mechanism.
3 comments
Question 1: We found that the reported 79.1% low-bin hit rate came from a stale score summary. On the identical 1,000 saved low-conditioned pairs used for disjoint-oracle rescoring, the primary oracle gives 48.7% and majority voting gives 35.1%. Comparing inconsistent result artifacts accounted for 30.4 percentage points of the original 44.0-point gap. The remaining same-sample difference is 13.6 points. The primary oracle predicts Bergman K562 ExP STARR-seq log₂(RNA/DNA) pair-reporter activity. CAGE-derived supervision is used during initialization.
The generator and evaluators use different condition edges, and the primary and disjoint oracles also differ in binning. Predictions close to the low/medium boundary can therefore change bins. The current analysis does not estimate what fraction of the remaining 13.6 points is attributable to the boundary alone.
We fit a monotone calibration map separately on each disjoint oracle’s validation predictions and froze it before scoring the generated pairs. The calibrated PEAC means for low, medium, and high targets were −4.964, −4.945, and −1.952. Continuous target MAEs were 1.720, 1.060, and 0.861 for PEAC and 1.637, 0.984, and 2.150 for direct concatenation. The low and medium PEAC distributions nearly coincide, and PEAC improves continuous target error for the high condition. A threshold-free comparison on raw oracle scores found that a randomly drawn high-conditioned pair scored above a randomly drawn low-conditioned pair with probability 0.987–0.989 across the three disjoint evaluators.
The reconciled primary-oracle hit rates describe performance under that evaluator’s decision rule. The strongest cross-oracle result is the upward shift for high-conditioned samples. A robust low-versus-medium distinction is unsupported by the continuous analysis.
Question 2: Our design goal is for the two generated elements to yield a combined reporter activity near the requested target. It does not require the original partners to outperform alternative partners. We recognize that phrases such as “better matched” can imply a stronger goal and will clarify the wording.
In Table 10, EPInformer predicted essentially the same mean activity and pair attention after swapping enhancers in both generated and reference pairs. The reference control shows that weak sensitivity to shuffling is shared by both groups under EPInformer. The test leaves open whether alternative partners have similar effects in this setting or whether the evaluator has limited sensitivity to the swaps. We assess activity target attainment separately with sequence-only reporter oracles.
We also tested partner preference with a frozen evaluator that can detect some non-additive effects in measured Bergman data. On 54,545 held-out combinations it achieved RMSE 1.1508, compared with 1.2191 for an additive promoter-plus-enhancer model on the log-activity scale. Its predicted and measured difference-in-differences correlated at Pearson r = 0.373. On generated pairs, we fixed the promoter and swapped enhancers within the same requested bin and nearby activity quantiles. The full model’s mean matched-minus-crossed prediction was 0.00193 (95% training-seed-clustered bootstrap interval −0.00220 to 0.00527), with a matched win rate of 50.12%. The full-minus-no-fusion contrast was 0.00140 (95% interval −0.02366 to 0.02975).
Question 3: We tested three independently trained checkpoints each for full PEAC, no pair fusion, and no MMD. We permuted 10-bp blocks within one element at a time while holding the scalar target, activity bin, and within-bin quantile fixed. In the converse control we held both DNA sequences fixed and changed only the condition. These within-element perturbations probe sequence sensitivity. The activity-matched enhancer exchanges in response 2 address partner preference.
For full PEAC, the mean absolute change in the internal activity prediction was 0.0025 activity units after sequence perturbation and 4.670 after changing only the condition. The corresponding latent RMS changes were 0.0099 and 1.818. The bin head remained 100% accurate after sequence perturbation. These results indicate that the internal heads and the MMD representation depend strongly on the supplied condition under this audit.
We also evaluated the same sequences with a separately trained, frozen Bergman predictor that accepts only DNA. Its full-model scores had mean Spearman correlation 0.621 ± 0.187 with the requested activity across training seeds. Sequence perturbation changed its score by 0.396 ± 0.024 activity units and increased target MAE by 0.022 ± 0.008. Thus, generated sequences carry activity information detectable by a sequence-only predictor. Full PEAC did not consistently outperform the no-fusion or no-MMD variants on these frozen-evaluator outcomes. We interpret both objectives as condition-aware auxiliary regularizers, with their incremental sequence-level contribution still unresolved.
Question 4: The closest isolation test is the matched no-pair-fusion variant. It retains the same Bergman pair rows and labels, condition embeddings, language-model and auxiliary losses, and final-stage optimization schedule of six epochs, batch size 8, learning rate 5 × 10⁻⁵, and up to 7,200 updates. Both variants have 18,095,018 nominal training-time parameters and the same 7,408,912-parameter causal rollout architecture. The ablation disables bidirectional promoter–enhancer state exchange and the pair token’s attention to the sequence states. Its disabled weights remain in the parameter count. Both variants retain promoter-first autoregression, so the enhancer can depend on the promoter during generation. The ablation tests the disabled fusion paths together, and the changes cannot be attributed to either path individually.
Across three independent training seeds, after averaging the three generation seeds within each checkpoint, the paired full-minus-no-fusion changes in low, medium, and high hit rates were −13.1 ± 19.9, +6.9 ± 7.3, and +6.3 ± 7.1 percentage points (mean ± SD). In these runs, fusion raised the mean medium and high rates and lowered the mean low rate. The seed variability and partner-swap results limit stronger attribution.
The direct-concatenation Transformer uses the same pair table and update count and has 7,408,656 rollout parameters. It uses a different training pipeline, batch size 16, and learning rate 3 × 10⁻⁴. It provides a conventional autoregressive benchmark. A conventional baseline matched on the training and example budgets, optimization, supervision, and seeds would be needed to isolate the incremental effect against that model family.
We thank the reviewer for raising these questions. The additional analyses support a narrower conclusion: PEAC jointly generates promoter and enhancer sequences with plausible element-level statistics and can shift sequence-only predictions of pair-reporter activity, with the clearest evidence in the high-activity regime. We have revised the manuscript for clarification.
Weakness 1: Thank you for highlighting this distinction. PEAC aims to jointly generate promoter–enhancer sequences with a requested pair-level activity; it does not assume or claim partner-specific compatibility or interactions beyond multiplicative effects. Accordingly, matched pairs outperforming shuffled pairs is not a necessary criterion for our generation objective. We have clarified this scope in the manuscript and revised wording that could imply biological compatibility. We also explicitly discuss evaluator dependence: low-condition accuracy decreases from 79.1% to 35.1%, whereas the high-condition advantage remains substantial under majority voting across the additional oracles (78.2% versus 12.0% for direct concatenation). The revised discussion treats EPInformer as an inconclusive diagnostic for partner specificity and acknowledges that the additional oracles cannot eliminate shared dataset bias.
Weakness 2: We agree that the existing comparisons do not fully isolate the contribution of pair fusion or exclude condition-based shortcuts. We have revised the methodological interpretation accordingly: cross-attention enables information exchange between element representations, and MMD provides an auxiliary distributional constraint; neither is presented as proof of sequence-dependent functional compatibility. We have also clarified that the direct-concatenation comparison establishes the performance difference between the evaluated pipelines. Condition-blind auxiliary objectives and a baseline matched in initialization, losses, and training budget remain necessary controls for stronger component-level conclusions, and this limitation is now stated explicitly.