acceptodds
Under review as a conference paper at ICLR 2027

Beyond Single Variants: A Benchmark for Regulatory Epistasis in Genomic Sequence Models

Abstract

Genomic sequence models, both supervised sequence-to-function predictors (Enformer, Borzoi, AlphaGenome) and self-supervised genome language models (Nucleotide Transformer, Evo 2), are increasingly used to interpret non-coding variation, yet they are benchmarked almost exclusively on variants tested one at a time against a reference background. Real haplotypes carry many variants per regulatory element and regulatory grammar is non-additive, so the practical question is whether models can compose variant effects. We assemble a benchmark from MPRA diplotype experiments that assay all four haplotypes (REF, A, B, AB) of thousands of variant pairs, together with  1.15M single-variant (variant × cell-type) measurements across three studies, and evaluate five models along a hierarchy of tasks: single-variant regression (T1), interaction regression (T2), non-additive pair detection (T3), and full-versus-additive prediction across 421 phased personal genomes (T4), each under a zero-shot and a frozen-encoder linear-probe readout. On Siraj et al. (2026), models rank confidently measured single-variant effects moderately well (zero-shot Spearman ρ up to 0.42; linear probe up to 0.73) but capture far less of the measured interaction signal (ρ ≤ 0.32; AUROC ≤ 0.68, against 0.62 for inter-variant distance alone). Linear probes recover substantial single-variant signals from gLM representations that their likelihoods do not expose, yet bring interaction detection at most to the level of the distance baseline. On personal genomes, full-haplotype predictions are no more accurate than additive reconstructions from each model's own marginal effects. Current models' interaction predictions thus add little beyond their marginal effects, a gap that matters for variant prioritization on real haplotypes and for sequence design, and that motivates interaction-aware training data, objectives and benchmarks.

Primary area:
datasets and benchmarks
Keywords:
genomic sequence models, regulatory epistasis, variant effect prediction, MPRA, non-coding variants, genome language models, benchmark, compositional generalization, linear probing, gene regulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_gkEyOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents a benchmark dataset for genetic variant epistasis from massively parallel reporter assay (MPRA) datasets. The motivation is to evaluate current genomic foundation models' ability to predict interactions between variants on the same haplotype, which is often overlooked by the current benchmarks. The proposed benchmark contains hierarchical predictive tasks: single-variant regression, interaction regression, non-additive pair detection and personal-genomic compositional gain. The authors test multiple pretrained foundation models on these tasks using zero-shot and linear probing. The results show a major performance gap between single-variant prediction and interaction prediction, and the weak interaction signals learned by the models offer little benefit to the gene expression prediction compared to additive reconstructions.

Critical strengths
  1. The proposed benchmark provides useful datasets and predictive tasks to evaluate the foundation models' capability in predicting complex and biologically relevant targets, namely epistatic variant interactions. The results establish a critical performance gap between single variant prediction and interaction prediction. Overall, the benchmark raises concerns about model capability and can facilitate the evaluation and tuning of future models.
  2. The hierarchical predictive tasks, especially T3 and T4, progressively test the model's capability and informativeness, which provides valuable insights into the models' predictive behavior.
Critical weaknesses
  1. As already stated in the discussion, the high-confidence interaction datasets are rather small. Though it is mainly caused by data availability, it makes statistical comparison less confident.
Questions for authors
  • How reliable is the distinction between positives and negatives in T3? Since negatives seem to be defined as pairs not significantly called non-additive, they may not sufficiently represent the negative space.
  • Since the baseline performance of T4 is already low, how much of the little performance gain compared to additive reconstruction is caused by the models' real failures to model epistatic interactions , and how much is by the task difficulty itself? Would metrics other than Spearman correlation show a different pattern?
5 comments
AuthorsOpenReview ↗

Response to Reviewer gkEy (Part 1/3)

We thank the reviewer for the positive assessment and thoughtful feedback. Please also see our General Response above, which details the harmonization of Siraj T2 interaction targets to REF/REF (all T2 results below use the corrected targets) and shared personal-genome (T4) analyses. Below, we address each point in detail.


[Weakness — Dataset Size & Statistical Confidence]: "As already stated in the discussion, the high-confidence interaction datasets are rather small. Though it is mainly caused by data availability, it makes statistical comparison less confident."

We agree that the high-confidence interaction sets are small, reflecting the current availability of published MPRA datasets that assay all four haplotypes of naturally occurring variant pairs in the same sequence context—to our knowledge, Siraj et al. (2026) and Abell et al. (2022) are the largest such datasets. We therefore designed the benchmark so that new four-haplotype datasets can be added with the same task definitions and splits as they become available.

To assess statistical confidence directly, we report 95% locus-group bootstrap confidence intervals for every T1–T3 comparison. Several key comparisons remain clearly resolved despite the sample size: on the matched set (Appendix C.2, Table C2), single-variant ρ\rho exceeds interaction ρ\rho with CIs excluding zero for Enformer and AlphaGenome zero-shot and for four of five probed models (e.g., probed AlphaGenome 0.710.71 vs. 0.390.39, difference 0.310.31 [0.17,0.46][0.17, 0.46]). In addition, as detailed under Question 1 below, our T3 conclusions hold across stratified negative definitions and 20 alternative labeling thresholds.


[Question 1 — Reliability of T3 Positives & Negatives]: "How reliable is the distinction between positives and negatives in T3? Since negatives seem to be defined as pairs not significantly called non-additive, they may not sufficiently represent the negative space."

We thank the reviewer for raising this point. The 2,021 negatives represent all candidate pairs not classified as non-additive by Siraj et al., rather than pairs confirmed to be additive. Recognizing this, we audited the calling procedure and stratified the negative space to test the robustness of our conclusions:

  1. Label validation: Re-running Siraj et al.'s calling code on our benchmark cell types reproduces all 143 positives, and 95%95\% of positives called from independent biological-replicate halves match the official labels. Excluding six negatives called positive only in cell types without full model coverage changes every AUROC by ≤0.001\le 0.001.
  2. Negative-set stratification: We divided the 2,021 negatives into four mutually exclusive groups, ordered from most to least positive-like (decreasing median measured ∣ε∣|\varepsilon|: 0.270.27, 0.160.16, and 0.050.05; not measurable for double-inactive elements):
    • Weak interaction signal (397397, 20%20\%): Nominal interaction evidence (p<0.05p < 0.05), but below the full calling thresholds.
    • Underpowered (1,1451{,}145, 57%57\%): No nominally significant interaction, but insufficient precision to exclude ∣ε∣≥0.25|\varepsilon| \ge 0.25.
    • Confirmed additive (5151, 2.5%2.5\%): The 95%95\% CI of ε\varepsilon lies within ±0.25\pm 0.25 in every active cell type, supporting additivity within Siraj et al.'s effect threshold.
    • Double-inactive elements (428428, 21%21\%): No active haplotypes in any benchmark cell type, preventing an interaction call under the original criteria.

Table gkEy-1. Siraj T3 AUROC across alternative negative sets (each vs. the 143 positives; zero-shot / linear probe):

Negative setnnEnformerBorzoiAlphaGenomeNTEvo 21/distance
All negatives (main text)2,0210.63 / 0.68*0.66 / 0.68*0.66 / 0.68*0.58 / 0.590.50 / 0.620.62 / 0.58
Weak interaction signal397 (20%)0.58 / 0.62*0.60 / 0.62*0.61* / 0.63*0.56 / 0.540.49 / 0.570.55 / 0.53
Underpowered1,145 (57%)0.60 / 0.66*0.63 / 0.67*0.64 / 0.67*0.59 / 0.580.50 / 0.610.62 / 0.59
Confirmed additive51 (2.5%)0.65 / 0.68*0.67 / 0.67*0.64 / 0.66*0.54 / 0.580.50 / 0.620.58 / 0.54
Double-inactive elements428 (21%)0.76* / 0.78*0.78* / 0.78*0.75* / 0.75*0.61 / 0.650.51 / 0.68*0.67 / 0.61

*Paired locus-group bootstrap 95% CI for (model −- 1/distance) excludes zero. The distance probe uses one-feature logistic regression on the same folds.

AuthorsOpenReview ↗

Response to Reviewer gkEy (Part 2/3)

  1. Robustness across negative definitions and thresholds: Probed S2F models significantly outperform 1/distance in every stratum, including against confirmed-additive negatives (AUROC 0.66–0.680.66\text{–}0.68 vs. 0.540.54 for the distance probe). Re-calling labels across 20 alternative threshold settings (q=0.01–0.2q = 0.01\text{–}0.2; ∣ε∣=0–0.5|\varepsilon| = 0\text{–}0.5), probed S2F models also consistently outrank both gLMs and 1/distance in every setting.

Revisions: Added these analyses to Appendix E.6 (Table E4), and updated Section 5.3 to state explicitly that T3 negatives are pairs not called non-additive rather than pairs confirmed to be additive.

[Question 2a — T4 Low Baseline Accuracy vs. Model Failure]: "Since the baseline performance of T4 is already low, how much of the little performance gain compared to additive reconstruction is caused by the models' real failures to model epistatic interactions, and how much is by the task difficulty itself?"

We agree that low baseline expression-prediction accuracy makes it difficult to separate model limitations from task difficulty in T4 alone. As detailed in General Response, Section 2 (Tables R1–R2), two new analyses narrow the interpretation:

  1. Model-predicted non-additivity (Table R1): Most models produce substantial non-additive differences across individuals (median q=0.18–0.63q = 0.18\text{–}0.63), yet these adjustments do not improve expression prediction over additive reconstruction (median Δρ=−0.002\Delta\rho = -0.002 to +0.002+0.002).
  2. Stratification by cis-genetic predictability (Table R2): Even among the top tertile of genes with stronger cis-genetic signal—where Enformer and AlphaGenome achieve clear baseline accuracy (ρfull=0.110–0.124\rho_{\text{full}} = 0.110\text{–}0.124)—joint haplotype scoring adds zero gain over summing single-variant effects (all high-minus-low Δρ\Delta\rho CIs include zero).

Because the available population data still cannot quantitatively separate model error from residual task noise, we clarify in Section 5.4 that T4 evaluates the downstream predictive utility of modeled interactions, whereas our primary conclusions about interaction accuracy rest on the controlled MPRA evaluations (T2/T3).


[Question 2b — Alternative Evaluation Metrics]: "Would metrics other than Spearman correlation show a different pattern?"

No—alternative metrics across all four tasks confirm the same conclusions (Table gkEy-2). We added Pearson correlation (T1–T2, T4), chance-normalized NDCG (T1 NDCG@100 for large-effect prioritization; T4 two-tailed NDCG across all 421 individuals), and precision@50 and recall@200 (T3; AP was already reported), using 95% locus-group bootstrap CIs for T1–T3 and gene-level bootstrap CIs for T4.

Table gkEy-2. Evaluation across additional metrics (ZS = zero-shot, LP = linear probe):

TaskMetricEnformerBorzoiAlphaGenomeNTEvo 21/distance
T1 (emVars)Spearman ZS / LP (paper)0.36 / 0.680.41 / 0.660.46 / 0.73−0.04 / 0.45−0.03 / 0.61–
Pearson ZS / LP0.30 / 0.630.31 / 0.590.37 / 0.67−0.03 / 0.400.00 / 0.56–
NDCG@100 ZS / LP0.22 / 0.430.20 / 0.350.33 / 0.500.03 / 0.180.01 / 0.31–
T2Spearman ZS / LP (paper)0.19 / 0.430.27 / 0.460.38 / 0.39−0.18 / 0.160.18 / 0.350.03 / −0.16
Pearson ZS / LP0.27 / 0.520.32 / 0.510.37 / 0.58−0.26 / 0.360.13 / 0.48−0.13 / −0.21
T3Precision@50 ZS / LP0.32 / 0.460.38 / 0.480.50 / 0.580.28 / 0.200.10 / 0.420.22 / 0.21
Recall@200 ZS / LP0.29 / 0.310.31 / 0.350.35 / 0.410.24 / 0.250.13 / 0.280.26 / 0.26
T4Median Δr\Delta r (full − additive)−0.0010.0030.0010.0000.000–
% genes Δr>0\Delta r > 048%55%54%48%50%–
Median ΔNDCG\Delta\text{NDCG} (full − additive)+0.002+0.006+0.0040.000−0.001–

NDCG is chance-normalized (0=0 = random, 1=1 = perfect). T1 NDCG@100 ranks variants by absolute effect; T4 NDCG is two-tailed (scoring both highest and lowest expressers) over all 421 individuals.

AuthorsOpenReview ↗

Response to Reviewer gkEy (Part 3/3)

  • T1–T2: Pearson rr reproduces the Spearman results: the zero-shot ranking is unchanged (AlphaGenome>Borzoi>Enformer\text{AlphaGenome} > \text{Borzoi} > \text{Enformer}), probing improves every model on both tasks, and after probing, single-variant correlation still exceeds interaction correlation for all five models (T1 0.40–0.670.40\text{–}0.67 vs. T2 0.36–0.580.36\text{–}0.58). On T2, all probed models and zero-shot AlphaGenome and Borzoi significantly beat 1/distance.
  • T3: Precision and recall confirm improved top-of-list prioritization: zero-shot AlphaGenome reaches precision@50 =0.50= 0.50 versus 0.220.22 for 1/distance (6.6%6.6\% positive base rate), and recall@200 =0.35= 0.35 versus 0.260.26.
  • T4: Neither Pearson rr nor two-tailed NDCG shows a meaningful gain from full-haplotype over additive predictions (median Δr=−0.001\Delta r = -0.001 to +0.003+0.003, all 95% CIs include zero; median ΔNDCG=−0.001\Delta\text{NDCG} = -0.001 to +0.006+0.006), including on genes with stronger cis-genetic predictability (see General Response, Section 2).

Revisions: Added these results to Appendix Table A3 (T1–T3) and Table F4 (T4).

Reviewer_9RF4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The authors present a new benchmark to evaluate if genome foundation models (DNA LMs and sequence to function models) can predict non-additive regulatory variant effects as measured by four-haplotype MPRA and phased personal genomes. They evaluate five models in total. The main finding is that models can predict marginal effects moderately well but perform weak on interactions beyond additive effects. The models, especially with probing instead of zero-shot evaluations, do outperform very simple distance based baseline, but only modestly.

Critical strengths

The benchmark is timely and the capability of genome foundation model to model interaction effects is an important research question, as this is a setting where conventional additive genomic methods have limitations and genome foundation models could provide substantial value if they performed well in this setting.

As such, the main findings, if robust, of the study, that current models mostly fail to exploit variant interactions is an important limitation and should motivate future research in the field, e.g. interaction-aware training of genome foundation models.

Critical weaknesses

As stated by the authors themselves in the limitations, there is uncertainty about the interpretation of the MPRA derived interactions, as they may reflect global nonlinearities. While I applaud the authors for stating this clearly in the limitations sections, it seems to me that most of the findings hinge on the validity of this assumption, and it is thus something that needs to be tested explicitly.

Section 4.5 is referenced but does not exist.

Appendix C lacks results stated in the main text. Same for Appendix E.

Kircher dataset counts differ between pp. 4 and 12.

Questions for authors

How much measured epistasis remains after accounting for global reporter nonlinearities?

The linear probes are trained to predict single-variant effects and then applied to variant combinations without interaction-specific training. Would the conclusions of the study changes if the probes were trained on interactions?

Would interaction-supervised probes change the conclusions about representation quality?

How informative is T4 given low baseline accuracy?

Please clarify if there is overlap between Kircher data and model training/evaluation.

Are the authors planning to publicly release the benchmark?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 9RF4 (Part 1/4)

Response to Reviewer 9RF4 (Part 1/4)

We thank the reviewer for recognizing the timeliness of our benchmark and for the detailed, constructive feedback. Please also see our General Response above, which details the harmonization of Siraj T2 interaction targets to REF/REF (all T2 results below use the corrected targets) and shared personal-genome (T4) analyses. Below, we address each weakness and question point by point.


[Weakness 1 — Global Reporter Nonlinearities]: "As stated by the authors themselves in the limitations, there is uncertainty about the interpretation of the MPRA derived interactions, as they may reflect global nonlinearities... it seems to me that most of the findings hinge on the validity of this assumption, and it is thus something that needs to be tested explicitly."

We agree this must be tested directly rather than only noted as a limitation. As detailed under Question 1 below, noise-calibrated sign-epistasis tests and cross-validated polynomial corrections show that global reporter nonlinearity explains almost none of the measured epistasis in Siraj and does not change any model comparison or ranking in either MPRA dataset.


[Weaknesses 2–4 — Manuscript Consistency & Presentation]: "Section 4.5 is referenced but does not exist. Appendix C lacks results stated in the main text. Same for Appendix E. Kircher dataset counts differ between pp. 4 and 12."

Thank you for catching these presentation issues; we have corrected all three in the revision:

  1. Section 4.5 reference: Corrected to point to Section 4.3 and Appendix C.2, which describe the matched-set comparison.
  2. Appendix C and Appendix E: Appendix C now includes the gLM layer sweep (Table C1), the matched-set results behind Fig. 2B–C (Table C2), and the pair-supervised probe results (Table C3). Appendix E now includes every T2–T3 result cited in Section 5.3 (Figures E1–E4, Tables E1–E5).
  3. Kircher dataset counts (pp. 4 and 12): Reconciled both pages to the exact same counts: 19,933 harmonized measurements (17,663 SNVs) across 14 elements, of which 12,466 SNVs in 10 elements are scored by every model under both readouts (5,332 significant at FDR <0.05< 0.05). All Kircher results use this 12,466-SNV set.

[Question 1 — Global Reporter Nonlinearities]: "How much measured epistasis remains after accounting for global reporter nonlinearities?" In Siraj, nearly all of it: after removing the fitted global component, the corrected εclean\varepsilon_{\text{clean}} correlates with the original ε\varepsilon at Spearman ρ=0.97\rho = 0.97, and 98.6%98.6\% of T2 pairs retain their sign. We tested this in both MPRA datasets using three complementary analyses (Appendix E.7):

  1. Sign epistasis (model-free test beyond monotone readouts): Any monotone global readout curve g(⋅)g(\cdot) (such as assay saturation) can compress or stretch effect sizes, but it cannot reverse the direction of a variant's effect across backgrounds. We tested whether either variant in a pair has a statistically significant effect (∣z∣>1.96|z| > 1.96) in opposite directions on the REF versus ALT background of the other variant, and compared this rate against an additive-projected noise null (removing ε\varepsilon from each pair to obtain Yadd=Y−ε4[1,−1,−1,1]Y_{\text{add}} = Y - \frac{\varepsilon}{4}[1, -1, -1, 1] and resampling per-haplotype measurement noise across 1,000 simulations). In Siraj, 20.3%20.3\% of the 143 T2 pairs show significant sign epistasis, compared with 0.7%0.7\% at the 95th percentile of the noise null (p=0.001p = 0.001). Across all 8,311 pair–cell-type units, the observed rate (0.85%0.85\%) remains 5×5\times the null 95th percentile (0.17%0.17\%).
  2. Estimating the global-nonlinearity component directly: If apparent epistasis ε\varepsilon were merely an artifact of a smooth global reporter curve g(⋅)g(\cdot) distorting additive effects, then ε\varepsilon would be predictable from each locus's baseline activity level (yˉ\bar{y}) and single-variant effect sizes (MA,MBM_A, M_B)—for example, saturation produces larger negative ε\varepsilon when baseline activity is already high and both variants push activity in the same direction. To capture any such global curvature without assuming a fixed parametric shape, we trained a degree-2 polynomial ridge model per cell type predicting ε\varepsilon from [yˉ,  MA+MB,  MAMB,  MA2+MB2][\bar{y},\; M_A + M_B,\; M_A M_B,\; M_A^2 + M_B^2] under 5-fold locus-disjoint cross-validation. Global activity and effect-size terms explain almost none of the measured epistasis: out-of-fold R2R^2 is ≤0\le 0 in every individual Siraj cell type and only 0.040.04 on the 143 T2 pairs (meaning at least 96%96\% of T2 epistasis variance is not explained by global reporter nonlinearity). Importantly, this test has statistical power: in simulations with a planted sigmoid readout of realistic magnitude, it detects the global component in 61–75%61\text{–}75\% of replicates.
AuthorsOpenReview ↗

Response to Reviewer 9RF4 (Part 2/4)

  1. Benchmark robustness after correction: Subtracting the out-of-fold global prediction (εclean=ε−ε^global\varepsilon_{\text{clean}} = \varepsilon - \hat{\varepsilon}_{\text{global}}) preserves model performance and rankings (Table 9RF4-1):
    • T2: Δρ\Delta\rho ranges from −0.049-0.049 to +0.015+0.015, with only one of 12 paired CIs excluding zero (Borzoi zero-shot). Model rankings are unchanged in both zero-shot and linear-probe settings, and AlphaGenome remains the strongest zero-shot model.
    • T3: Because Siraj's published calls cannot be recomputed on εclean\varepsilon_{\text{clean}}, we applied the same significance rule (BH, 5% FDR) before and after correction. It calls 127 positive pairs before and 123 after, and T3 AUROC changes by ≤0.03\le 0.03. Table 9RF4-1. Siraj performance before and after global-nonlinearity correction: | Setting | Model | T2 ρ\rho: original →\to corrected | Δρ\Delta\rho [95% CI] | T3 AUROC: original →\to corrected | | :--- | :--- | :---: | :---: | :---: | | Zero-shot | AlphaGenome | 0.375 →\to 0.383 | +0.007 [−0.034, 0.048] | 0.709 →\to 0.716 | | | Borzoi | 0.274 →\to 0.225 | −0.049 [−0.093, −0.008] | 0.678 →\to 0.680 | | | Enformer | 0.190 →\to 0.199 | +0.009 [−0.031, 0.053] | 0.653 →\to 0.659 | | | Evo 2 | 0.176 →\to 0.190 | +0.015 [−0.030, 0.060] | 0.501 →\to 0.497 | | | NT | −0.180 →\to −0.217 | −0.038 [−0.094, 0.007] | 0.632 →\to 0.632 | | | 1/distance | 0.030 →\to 0.017 | −0.012 [−0.044, 0.015] | 0.693 →\to 0.695 | | Linear probe | Borzoi | 0.465 →\to 0.447 | −0.017 [−0.055, 0.016] | 0.709 →\to 0.720 | | | Enformer | 0.426 →\to 0.395 | −0.031 [−0.073, 0.001] | 0.716 →\to 0.721 | | | AlphaGenome | 0.393 →\to 0.381 | −0.012 [−0.060, 0.030] | 0.728 →\to 0.742 | | | Evo 2 | 0.350 →\to 0.322 | −0.028 [−0.079, 0.014] | 0.657 →\to 0.683 | | | NT | 0.160 →\to 0.141 | −0.018 [−0.064, 0.020] | 0.623 →\to 0.620 | | | 1/distance | −0.158 →\to −0.153 | +0.005 [−0.037, 0.044] | 0.670 →\to 0.671 | T2: Spearman ρ\rho on the 143 published non-additive pairs against ε\varepsilon and εclean\varepsilon_{\text{clean}}; 95% CIs from a paired locus bootstrap. T3: AUROC for ranking all 2,164 candidate pairs under a uniform BH 5% FDR rule (127 positives before, 123 after correction; original T3 values therefore differ from Table 1, which uses Siraj's published labels).
  2. Abell: Abell does show a detectable global component (out-of-fold R2=0.135R^2 = 0.135 vs. noise-null mean 0.0540.054; p=0.02p = 0.02), consistent with possible saturation. Nevertheless, interaction prediction remains near zero before and after correction (∣ρ∣≤0.06|\rho| \le 0.06; ∣Δρ∣≤0.02|\Delta\rho| \le 0.02 for all models and distance), so no conclusion depends on it.
  3. Why not a parametric global-epistasis model (e.g., MAVE-NN)? Parametric global-epistasis models such as MAVE-NN (Tareen et al., 2022) rely on dense combinatorial mutagenesis where each variant is observed across many genetic backgrounds to identify a shared global readout curve g(⋅)g(\cdot). Here, each locus has only four genotypes (REF, A, B, AB), leaving only a single degree of freedom per locus after fitting baseline activity and the two main effects; under measurement noise, multi-parameter sigmoid readouts become poorly identified and bias curvature estimates even on purely additive data. We therefore rely on the noise-calibrated tests above (alongside noise-calibrated one-parameter curvature checks in Appendix E.7). Revisions: We added the correction protocol, noise-calibrated nulls, power analysis, and rescored results to Appendix E.7 (Table E5), and updated Limitations accordingly.
AuthorsOpenReview ↗

Response to Reviewer 9RF4 (Part 3/4)

[Questions 2 & 3 — Interaction-Supervised Probes & Representation Quality]: "The linear probes are trained to predict single-variant effects and then applied to variant combinations without interaction-specific training. Would the conclusions of the study change if the probes were trained on interactions? Would interaction-supervised probes change the conclusions about representation quality?"

No, adding interaction supervision does not change our conclusions, and it rules out "probe mismatch" as the reason for limited interaction performance.

Setup. Keeping pretrained representations frozen, we used the same ridge readout s(X)=w⊤h(X)s(X) = w^\top h(X) under the same locus-grouped five-fold splits (holding out ∼20%\sim 20\% of pairs and all pairs from their loci per fold). To the single-variant training set (h(A)−h(REF)→yA−yREFh(A) - h(\text{REF}) \to y_A - y_{\text{REF}}), we added double-variant training examples (h(AB)−h(REF)→yAB−yREFh(AB) - h(\text{REF}) \to y_{AB} - y_{\text{REF}}) amounting to 0%0\%, 4%4\%, 8%8\%, 16%16\%, 40%40\%, 64%64\%, or 80%80\% of all pairs (80%80\% = all training-fold pairs; 0%0\% = original single-only probe). Interaction prediction is evaluated on unseen loci via ε^=s(AB)−s(A)−s(B)+s(REF)\hat{\varepsilon} = s(AB) - s(A) - s(B) + s(\text{REF}) against measured ε\varepsilon (under the corrected REF/REF convention).

Results (Tables 9RF4-2 and 9RF4-3). Pair supervision produced no improvement in held-out interaction prediction. On Siraj, every model stayed within 0.0030.003 of its single-only baseline at every supervision level; on Abell, correlations stayed near zero (∣ρ∣≤0.06|\rho| \le 0.06, max change 0.0070.007). For every model on both datasets, the 95% CI of Δρ\Delta\rho (80%−0%80\% - 0\%) includes zero. Training on pairs labeled directly with ε\varepsilon gives the same null gain, and a probe trained on ε\varepsilon alone without single-variant examples performs worse (ρ=0.04–0.23\rho = 0.04\text{–}0.23 on Siraj).

Table 9RF4-2. Siraj T2 Spearman ρ\rho vs. pair supervision (143 held-out pairs; 2,000 locus-level bootstraps):

Pairs added (% of all pairs)EnformerBorzoiAlphaGenomeNTEvo 2
0% (single-only)0.4260.4650.3930.1600.350
4%0.4250.4650.3920.1610.349
8%0.4260.4650.3930.1610.349
16%0.4260.4650.3940.1600.350
40%0.4240.4660.3940.1610.349
64%0.4240.4650.3960.1600.350
80% (all training-fold pairs)0.4240.4650.3940.1600.349
Δρ\Delta\rho (80% − 0%) [95% CI]−0.002 [−0.007, +0.004]+0.001 [−0.003, +0.005]+0.000 [−0.025, +0.023]−0.000 [−0.005, +0.005]−0.001 [−0.005, +0.003]

Table 9RF4-3. Abell T2 Spearman ρ\rho vs. pair supervision (1,673 records across 304 loci; same splits as Table 9RF4-2):

Pairs added (% of all pairs)EnformerBorzoiAlphaGenomeNTEvo 2
0% (single-only)−0.058−0.014−0.0040.0190.020
4%−0.058−0.0140.0020.0190.020
8%−0.059−0.0150.0020.0190.019
16%−0.058−0.0150.0010.0180.018
40%−0.058−0.0190.0020.0180.018
64%−0.059−0.020−0.0020.0170.017
80% (all training-fold pairs)−0.060−0.0210.0020.0160.017
Δρ\Delta\rho (80% − 0%) [95% CI]−0.003 [−0.009, +0.004]−0.007 [−0.017, +0.002]+0.006 [−0.003, +0.016]−0.003 [−0.012, +0.006]−0.003 [−0.013, +0.007]

Implications for representation quality. Together, the two probe designs separate single-to-pair generalization (original probes) from linear decodability of interactions (pair-supervised probes):

  1. Single-variant probing did not underestimate the representations: If frozen embeddings contained additional, linearly decodable interaction signal missed by single-variant training, adding pair supervision would have improved held-out ρ\rho. Instead, ρ\rho is invariant across all supervision levels, showing that the linearly accessible interaction signal is already aligned with the single-variant activity direction (which is also why ε\varepsilon-only training performs worse without single-variant anchoring).
  2. Scope of the representation claim: Our conclusion about representation quality does not change, at the available data scale, a linear readout cannot extract more interaction signal from these frozen embeddings than single-variant probing already recovers. We note explicitly that this applies to linear ridge readouts on frozen embeddings; whether nonlinear probes, substantially larger pair-training sets, or encoder fine-tuning can extract non-linearly encoded interactions remains an open question.

Revisions: We updated Section 5.1 and Appendix C.3 (Table C3) to distinguish single-to-pair generalization from pair-supervised probing and to report the full supervision sweep.

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

The paper introduces a benchmark for evaluating the performance of genomic sequence models in predicting regulatory epistasis. Using publicly available MPRA data and personal genome data, the authors evaluate whether models learn non linear variant interactions by benchmarking a variety of sequence to function and DNA language models. The authors find that models generally struggle to predict the effect of interactions compared to predicting single variant effects

Critical strengths

The paper identifies a useful gap in evaluation of DNA models. Systematically benchmarking models against experimentally measured non-additive regulatory effects is a useful contribution, and has not yet been explored extensively in one unified benchmark. This provides a potential foundation for future work and offers a useful failure mode of current DNA models.

The paper evaluates multiple genomic models (ie both supervised sequence to function models and unsupervised sequence only models) across MPRA datasets and personal-genome expression data, using both zero-shot and linear-probe settings. This is useful for providing robust insights and strengthens the claims of the paper.

Critical weaknesses

The limitations of the datasets make it difficult to draw strong conclusions. MPRA activity doesn't directly correspond to the models' output tracks and previous studies have shown that models struggle significantly to predict gene expression differences across individuals, with correlations often close to zero [1]. This makes it difficult to tell whether the poor results reflect an inability to model regulatory interactions or broader limitations of the prediction tasks. I appreciate that better datasets may not be available, but this does limit the strength of the conclusions and utility of the benchmark.

AlphaGenome is evaluated using only 16 kb of sequence context, compared with 524 kb for Borzoi. AlphaGenome's original publication shows that prediction performance can deteriorate with shorter inference contexts, since the model is only trained at 1MB sequence lengths. Similar issues with UNet models have also been explored in [2]. Hence, it is unclear whether the reported performance reflects a limitation of the model or degradation in performance due to the evaluation context length not matching the training context length.

[1] Huang, C. et al. (2023). Personal transcriptome variation is poorly explained by current genomic deep learning models. Nature Genetics, 55, 2056–2059

[2] Boshar, S et al (2025). A foundational model for joint sequence-function multi-species modeling at scale for long-range genomic prediction

Questions for authors

Why was AlphaGenome evaluated at 16 kb rather than a longer supported context? Can the authors demonstrate that the results for this are not due to a mismatch between evaluation context length and the training context length of the model?

6 comments
AuthorsOpenReview ↗

Response to Reviewer ugaP (1/3)

We thank the reviewer for recognizing the value of our unified benchmark and for the constructive feedback. Please also see our General Response above, which details the harmonization of Siraj T2 interaction targets to REF/REF (all T2 results below use the corrected targets) and shared personal-genome (T4) analyses. Below, we address each weakness and question point by point.


[Weakness 1a — MPRA vs. Model Output Tracks]: "The limitations of the datasets make it difficult to draw strong conclusions. MPRA activity doesn't directly correspond to the models' output tracks"

We agree that episomal MPRA reporter activity differs from the endogenous genomic measurements used to train S2F models. To test whether readout mismatch explains the interaction gap, we conducted four controls:

  1. Same assay, same variants (in the submission): Our main T1–T2 comparison already holds the assay fixed: single-variant is computed on the 286 constituent variants of the same 143 pairs, in the same oligos and cell type (Fig. 2B–C, Appendix C.2). On this matched set, AlphaGenome reaches =0.54 on single variants versus 0.38 on interactions under the zero-shot setting, and 0.71 versus 0.39 with the linear probe (difference 0.31 [0.17,0.46]). A readout mismatch that affected both tasks equally would not create this gap.

  2. Linear probing bridges the track-to-MPRA readout gap (in the submission): To remove the zero-shot track-mapping bottleneck altogether, we also evaluate linear probes trained on frozen model embeddings directly against measured MPRA activity. Probing substantially boosts single-variant prediction (T1 =0.66–0.73 for S2F models), confirming that model representations encode strong MPRA-relevant regulatory signal; yet interaction prediction (T2 =0.39–0.46) and interaction detection (T3 AUROC 0.68 vs. 0.62 for 1/distance) remain limited even after learning a direct mapping to MPRA activity.

  3. Cell-type-matching control: Comparing zero-shot predictions from experimentally matched cell-type tracks against mismatched cell-type tracks on the same variants and oligos shows that endogenous S2F tracks capture cell-type-specific MPRA signal on both T1 and T2 (Table ugaP-1; averaged across HCT116, HepG2, K562, and SK-N-SH; 1,000 locus-group bootstraps; A549 excluded as n=1):

Table ugaP-1. Advantage of cell-type-matched over mismatched tracks ( Spearman , 95% CI):

ModelT1 (n=55,356)T2 (n=142)
Enformer+0.08 [0.08, 0.08]+0.05 [−0.11, 0.21]
Borzoi+0.06 [0.06, 0.07]+0.10 [−0.08, 0.25]
AlphaGenome+0.09 [0.09, 0.10]+0.27 [0.09, 0.45]

Matched tracks outperform mismatched tracks on T1 across all three S2F models despite strong cross-cell-type correlation in measured MPRA effects (=0.76–0.92). On T2, AlphaGenome achieves mean =0.43 with matched tracks versus 0.16 with mismatched tracks (=+0.27 [0.09,0.45], positive in HCT116, HepG2, and SK-N-SH; K562 is -0.21 [-0.53,0.14]), demonstrating that endogenous model tracks capture cell-type-specific MPRA interaction signal.

AuthorsOpenReview ↗

Response to Reviewer ugaP (2/3)

  1. Output-track modality sweep: Evaluating six predefined combinations of accessibility, histone, and transcription (CAGE/RNA-seq) tracks in a zero-shot setting shows that while transcription tracks improve specific cases (e.g., AlphaGenome in HepG2 rises from 0.55 to 0.71), transcription is the strongest single modality in only 2 of 10 available model–cell-type combinations, and no modality family systematically dominates (Table ugaP-2; 1,000 locus-group bootstraps):

Table ugaP-2. Siraj T2 Spearman across output-track families and cell types (95% CI):

ModelCell type (n)AccessibilityHistoneTranscriptionAcc. + Histone (orig.)Acc. + Transcr.All modalities
EnformerHCT116 (25)0.07 [−0.41, 0.47]0.02 [−0.44, 0.48]n/a0.03 [−0.40, 0.47]n/a0.03 [−0.41, 0.46]
HepG2 (34)0.42 [0.12, 0.68]0.18 [−0.20, 0.52]0.02 [−0.27, 0.34]0.26 [−0.13, 0.56]0.32 [−0.01, 0.57]0.21 [−0.17, 0.50]
K562 (45)0.08 [−0.24, 0.38]0.04 [−0.26, 0.35]0.23 [−0.06, 0.47]0.15 [−0.19, 0.44]0.25 [−0.06, 0.52]0.15 [−0.14, 0.44]
SK-N-SH (38)0.15 [−0.22, 0.50]0.51 [0.21, 0.73]n/a0.26 [−0.12, 0.59]n/a0.26 [−0.14, 0.60]
BorzoiHCT116 (25)0.40 [−0.04, 0.74]0.02 [−0.40, 0.41]0.12 [−0.33, 0.53]0.24 [−0.22, 0.58]0.31 [−0.15, 0.68]0.23 [−0.20, 0.61]
HepG2 (34)0.27 [−0.05, 0.56]−0.05 [−0.38, 0.32]0.13 [−0.21, 0.51]0.18 [−0.13, 0.50]0.23 [−0.09, 0.55]0.18 [−0.13, 0.51]
K562 (45)0.26 [−0.04, 0.52]0.13 [−0.19, 0.43]0.22 [−0.10, 0.50]0.17 [−0.13, 0.45]0.34 [0.06, 0.59]0.30 [−0.01, 0.58]
SK-N-SH (38)0.13 [−0.24, 0.50]0.38 [0.12, 0.60]0.33 [0.06, 0.57]0.37 [0.00, 0.65]0.30 [−0.09, 0.62]0.42 [0.05, 0.69]
AlphaGenomeHCT116 (25)0.52 [0.17, 0.78]0.21 [−0.20, 0.58]0.19 [−0.22, 0.56]0.49 [0.15, 0.77]0.48 [0.10, 0.77]0.44 [0.07, 0.75]
HepG2 (34)0.51 [0.16, 0.75]0.58 [0.26, 0.83]0.62 [0.33, 0.81]0.55 [0.23, 0.77]0.70 [0.48, 0.86]0.71 [0.47, 0.86]
K562 (45)0.24 [−0.09, 0.55]0.21 [−0.11, 0.51]0.09 [−0.22, 0.43]0.22 [−0.09, 0.52]0.23 [−0.08, 0.52]0.23 [−0.07, 0.51]
SK-N-SH (38)0.34 [0.09, 0.60]0.47 [0.16, 0.72]0.29 [−0.10, 0.61]0.45 [0.13, 0.69]0.44 [0.07, 0.69]0.45 [0.11, 0.71]

Revisions: We added the track-family sweep (Appendix E.4, Table E2) and cell-type-matching control (Appendix E.5, Table E3), while explicitly discussing episomal-versus-endogenous readout differences in Limitations.


[Weakness 1b — Personal-Genome Expression Prediction (T4)]: "...previous studies have shown that models struggle significantly to predict gene expression differences across individuals, with correlations often close to zero [1]. This makes it difficult to tell whether the poor results reflect an inability to model regulatory interactions or broader limitations of the prediction tasks."

We agree that low baseline accuracy on personal genomes (Huang et al., 2023) requires careful controls. As shown in General Response, Section 2 (Tables R1–R2), two new analyses across 421 GEUVADIS individuals and 300 genes rule out trivial explanations for 0:

  1. Models are not simply outputting identical full and additive scores (Table R1): Most models produce sizable non-additive differences across individuals (median q=SD(full-additive)/SD(full)=0.18–0.63), yet these adjustments yield zero correlation gain (median =-0.002 to +0.002).
  2. Zero gain persists on the most cis-predictable genes (Table R2): Stratifying genes into tertiles by held-out PrediXcan accuracy from Huang et al. (2023) shows that even in the top tertile—where Enformer and AlphaGenome reach clear baseline accuracy (full=0.110–0.124)—joint haplotype scoring adds zero gain over summing single-variant effects (High-Low=-0.004 to +0.002; all 95% CIs include zero).

Revisions: We incorporated Tables R1–R2 into Section 5.4 and Appendix F, clarifying how T4 complements our controlled MPRA evaluations (T2/T3).

AuthorsOpenReview ↗

Response to Reviewer ugaP (3/3)


[Weakness 2 & Question 1 — AlphaGenome Context Length]: "AlphaGenome is evaluated using only 16 kb of sequence context, compared with 524 kb for Borzoi... Why was AlphaGenome evaluated at 16 kb rather than a longer supported context? Can the authors demonstrate that the results for this are not due to a mismatch between evaluation context length and the training context length of the model?"

We initially used 16 kb windows for computational efficiency across thousands of MPRA haplotypes, as Siraj pairs sit within 150 bp of each other. First, we note that with the REF/REF-harmonized Siraj labels, AlphaGenome at 16 kb already achieves the strongest zero-shot T2 performance (=0.38), outperforming Borzoi at 524 kb (=0.27) and Enformer at 196 kb (=0.19). We also cited Borshar et al. in the manuscript.

Second, to directly test sensitivity to context length, we re-evaluated AlphaGenome across five supported context lengths up to its full 1 Mb training length, holding all variants, tracks, pooling windows, and scoring procedures constant (Table ugaP-3):

Table ugaP-3. AlphaGenome zero-shot performance across input context lengths:

Metric16 kb131 kb262 kb524 kb1 Mb (Full Training Length)
**Single-variant ** (emVars across all 2,164 candidate pairs, n=1,803)0.600.610.610.620.62
**Siraj T2 Spearman ** (143 pairs)0.380.400.400.400.38
Siraj T3 AUROC (2,164 pairs)0.660.670.670.670.67
Siraj T3 AP0.240.250.250.250.25

Across a 64-fold increase in context length (16 kb 1 Mb), performance is virtually invariant: T2 Spearman changes by +0.005 (95% CI [-0.05,0.06]), T3 AUROC changes by +0.014 ([-0.003,0.03]), and T2 interaction predictions at 16 kb and 1 Mb correlate at =0.91. Thus, shorter inference context does not degrade AlphaGenome's performance on proximal MPRA variant pairs.

Revisions: We added this ablation to Appendix B.1 (Table B1) and clarified per-model context lengths in Appendix B and Limitations.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers (Part 1/2)

We thank all three reviewers for their constructive feedback and recognition of the value of our work. Before addressing individual comments below, we summarize (1) an important target-harmonization correction to Siraj T2 and (2) new personal-genome (T4) analyses addressing a shared question raised by all three reviewers.

GR1. Harmonization of Siraj T2 Interaction Targets

Correction. Our original T2 analysis used the precomputed interaction estimates (ε\varepsilon) released by Siraj et al. (Appendix A). For each pair, those released estimates designate the least-active of the four haplotypes as the baseline rather than using the fixed genomic REF/REF convention used by our models, reversing the sign of ε\varepsilon when the least-active haplotype is ALT/REF or REF/ALT (64 of 143 pairs; 44.8%). We harmonized the targets and verified them against: (i) Siraj et al.’s preprocessing code, (ii) the synthesized oligo sequences, and (iii) ε\varepsilon re-estimated directly from raw barcode counts (which changes T2 ρ\rho by at most 0.010.01).

Scope. The correction affects only Siraj T2; Siraj T3 evaluates ∣Imodel∣|I_{\text{model}}| against the original non-additivity labels, while Abell and T4 remain unchanged. The updated table incorporates these corrections.

Updated Siraj T2 Results

n = 143 pairs; Spearman ρ\rho with 95% locus-group bootstrap CIs (2,000 resamples).

ModelZero-shot (original manuscript)Zero-shot (updated) [95% CI]Linear probe (original manuscript)Linear probe (updated) [95% CI]
Enformer0.110.19 [0.01, 0.35]0.150.43 [0.26, 0.56]
Borzoi0.140.27 [0.10, 0.43]0.320.46 [0.30, 0.60]
AlphaGenome0.070.38 [0.22, 0.52]0.180.39 [0.24, 0.53]
NT0.00−0.18 [−0.38, 0.02]0.060.16 [0.00, 0.31]
Evo 20.190.18 [0.02, 0.32]0.190.35 [0.16, 0.51]
1/distance0.150.03 [−0.17, 0.22]−0.22−0.16 [−0.33, 0.01]

Updated Interpretation

  1. Revised zero-shot ranking. AlphaGenome now leads Siraj T2 (ρ=0.38\rho = 0.38), followed by Borzoi (0.270.27) and Enformer (0.190.19), matching their T1 ordering. The distance baseline falls to ρ=0.03\rho = 0.03, and all three S2F models outperform it, indicating that their predictions capture directional interaction information beyond genomic proximity.
  2. Main conclusions remain unchanged.
    • Interaction prediction remains more challenging than single-variant prediction: Probed S2F models achieve T1 ρ=0.66–0.73\rho = 0.66\text{–}0.73 on significant variants, versus T2 ρ=0.39–0.46\rho = 0.39\text{–}0.46.
    • Interaction detection remains limited: T3 AUROC ≤0.68\le 0.68 versus 0.620.62 for distance.
    • Other tasks are unaffected: Abell T2/T3 and T4 (median Δρ≈0\Delta\rho \approx 0) remain unchanged.
AuthorsOpenReview ↗

General Response to All Reviewers (Part 2/2)

2. Personal-Genome Prediction (T4) and Baseline Expression Accuracy

All three reviewers asked how low baseline accuracy in personal-genome expression prediction (Huang et al., 2023) affects the interpretation of T4. To test whether model-predicted interactions provide incremental value when baseline predictability is stronger, we conducted two additional analyses across 421 GEUVADIS individuals and 300 genes:

(1) Model-predicted non-additivity. We quantified non-additivity as q=SD(full−additive)/SD(full)q = \text{SD}(\text{full} - \text{additive}) / \text{SD}(\text{full}), where SD\text{SD} is the standard deviation across individuals. Several models produce substantial non-additive components (median q=0.18–0.63q = 0.18\text{–}0.63), yet these provide no detectable improvement in expression prediction over additive reconstruction (median Δρ=−0.002\Delta\rho = -0.002 to +0.002+0.002; Table R1).

Table R1. Predicted non-additivity and zero-shot compositional gain.

ModelRelative non-additivity (qq)Median Δρ\Delta\rho (full − additive)
Enformer0.31−0.002
Borzoi0.42+0.002
AlphaGenome0.18+0.001
Nucleotide Transformer0.040.000
Evo 20.63−0.001

Values are medians across 300 genes. qq measures the relative magnitude of non-additive prediction residuals across individuals; Δρ\Delta\rho measures the difference in Spearman correlation with observed expression. Non-additive outputs do not necessarily represent accurate biological interactions.

(2) Stratification by cis-genetic predictability. We stratified the 300 genes into tertiles (100 genes each) using held-out Pearson correlation of PrediXcan predictions released by Huang et al. (2023) with observed GEUVADIS gene expression.

While baseline ρfull\rho_{\text{full}} increases in the high-predictability tertile for Enformer (0.017→0.1240.017 \to 0.124) and AlphaGenome (0.051→0.1100.051 \to 0.110), compositional gain Δρ\Delta\rho remains indistinguishable from zero across all tertiles (Table R2).

Table R2. T4 performance by cis-genetic predictability.

Modelρfull\rho_{\text{full}} Lowρfull\rho_{\text{full}} Midρfull\rho_{\text{full}} HighΔρ\Delta\rho LowΔρ\Delta\rho MidΔρ\Delta\rho HighΔρ\Delta\rho High − Low [95% CI]
Enformer0.0170.0090.124−0.001+0.001−0.005−0.004 [−0.013, 0.003]
Borzoi0.0160.0300.0420.000+0.006+0.002+0.002 [−0.009, 0.011]
AlphaGenome0.0510.0690.110+0.0010.000+0.0010.000 [−0.005, 0.004]
Nucleotide Transformer−0.0030.0140.0120.0000.0000.0000.000 [−0.001, 0.001]
Evo 20.0050.0380.003−0.002−0.002−0.001+0.001 [−0.016, 0.014]

100 genes per tertile; median held-out PrediXcan Pearson r=0.14r = 0.14, 0.230.23, and 0.370.37. ρfull\rho_{\text{full}} is the median Spearman correlation with observed expression; Δρ=ρfull−ρadditive\Delta\rho = \rho_{\text{full}} - \rho_{\text{additive}}. CIs use 2,000 gene-level bootstraps. All Δρ\Delta\rho CIs include zero. High-versus-low increases in ρfull\rho_{\text{full}} are supported for Enformer (+0.107+0.107 [0.020,0.152][0.020, 0.152]) and AlphaGenome (+0.058+0.058 [0.011,0.123][0.011, 0.123]).

Interpretation and revisions. Together, Tables R1 and R2 rule out two alternative explanations for why Δρ≈0\Delta\rho \approx 0 in T4:

  • Models are not simply outputting identical full and additive scores: Most models produce substantial non-additive differences across individuals (q=0.18–0.63q = 0.18\text{–}0.63; Table R1), but those differences fail to improve agreement with measured expression.
  • The zero gain is not caused by noisy, unpredictable genes: Even on the top tertile of genes with the strongest cis-genetic signal—where Enformer and AlphaGenome achieve clear baseline accuracy (ρ=0.11–0.12\rho = 0.11\text{–}0.12; Table R2)—joint haplotype scoring still adds zero improvement over summing single-variant effects (Δρ≈0\Delta\rho \approx 0).

Together with our controlled MPRA evaluations (T2/T3, where ground-truth interactions are directly measured), T4 shows that on natural human haplotypes, current models' non-additive adjustments add no predictive value even on the most cis-predictable genes. We have added Tables R1–R2 to Section 5.4 and Appendix F.

Area_Chair_zNRtOpenReview ↗

Please respond

Hi reviewers,

Thanks for your work so far on this paper. Please respond to the authors before the end of the discussion period if you have not done so already, even if to acknowledge their response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.