acceptodds
Withdrawn from ICLR 2027

Phenotype-Associated Gene Discovery with AlphaGenome

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Identifying which genes are relevant to a phenotype is a prerequisite for editing the genome to treat disease, and the established tool, genome-wide association, has a known failure mode: where population structure correlates with the trait, a variant-level test cannot separate a causal locus from one that is merely ancestry-informative. We study this failure on a testbed built to make it unavoidable: a cohort from the 1000 Genomes Project relabelled by a phenotype, skin pigmentation, that coincides exactly with continental ancestry. For each candidate gene individually, we predict melanocyte RNA-seq with a frozen sequence-to-function model and train one classifier per gene on that signal alone; the balanced accuracy it reaches induces a ranking over genes. Over nine literature-curated pigmentation genes against thirty-three independently screened controls, the classifier's ranking recovers the panel at (, permutation null), where a variant-level association test on the same cohort and label reaches . The gap is largest where a candidate list is drawn: four of the classifier's five highest-ranked genes are panel genes, against two for the association test, and by the threshold recovering its th panel gene it admits control of where the two association estimators admit and —an advantage significant against both estimators over the top 15% of the ranking, a cap fixed in advance (, ). Over the whole ranking it persists but, at nine positives, is resolved only against the aggregating estimator. Since follow-up is spent only on the head of a candidate list, this means fewer ancestry-driven false leads per experiment, from a filter built into the representation rather than a model of the confound.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
gene prioritization, genome-wide association studies (GWAS), sequence-to-function models, AlphaGenome, skin pigmentation genetics

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_aM49OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a phenotype-associated gene ranking approach using frozen AlphaGenome representations. For each candidate gene, AlphaGenome predicts melanocyte RNA-seq tracks from individual haplotypes, and a separate CNN is trained to predict a binary pigmentation label from that gene's representation. The held-out balanced accuracy of each classifier is then used to rank genes.

The method is evaluated on 1,072 individuals from the 1000 Genomes Project, where African populations are labeled as strongly pigmented and European populations as weakly pigmented. On 9 literature-curated pigmentation genes and 33 screened controls, the proposed ranking achieves AUC 0.880 and outperforms variant-level GWAS rankings and several ablations, particularly near the top of the ranking. The authors argue that the AlphaGenome representation acts as a functional bottleneck that suppresses ancestry-associated variation without predicted transcriptional consequences.

Critical strengths
  1. The functional bottleneck perspective is interesting. Rather than performing association directly on genotype, the method passes sequence variation through frozen, cell-type-specific AlphaGenome predictions. The ablations provide some evidence that this representation suppresses ancestry information that unrestricted sequence retains broadly.

  2. The experimental protocol is relatively transparent. The paper reports top-of-ranking metrics, permutation tests, several representation ablations, and explicitly acknowledges important limitations such as the ancestry-derived phenotype, small positive set, and distinction between locus and causal gene.

Critical weaknesses
  1. The phenotype is not actually measured. The binary pigmentation label is deterministically constructed from continental ancestry. Therefore, the experiment cannot distinguish pigmentation-specific signal from other biological signals that differ between African and European populations.

  2. The per-gene score lacks a clear interpretation. A gene is scored by the held-out accuracy of a CNN predicting this ancestry-derived label from its AlphaGenome representation. High predictability does not imply phenotype-specific relevance or causal involvement of the gene.

  3. The gene ranking may depend strongly on the downstream CNN, but does not establish that the ranking is stable across classifiers, optimization or resampled individuals. It is therefore unclear whether the ranking reflects a robust property of the AlphaGenome representation.

  4. The benchmark may be unusually favorable to the proposed representation. The positives are known pigmentation genes, while controls are screened to exclude pigmentation, melanin, and melanocyte annotations. The proposed representation is specifically melanocyte RNA-seq.

  5. The empirical comparison is limited. The evaluation contains only 9 positive genes and 33 controls from one cohort. The comparison also lacks a strong modern gene-prioritization baseline under a non-circular evaluation protocol.

  6. The evidence supports locus prioritization more directly than gene discovery. A high-scoring gene-centered window does not identify the causal gene. The HERC2/OCA2 example illustrates this distinction.

  7. The GWAS comparison represents a deliberately pathological setting. Population structure is identical to the constructed label, and the GWAS baseline is consequently evaluated without population-structure correction, but it does not support a general claim that the proposed approach is superior to established association-based approaches.

Questions for authors
  1. What is the intended statistical or causal interpretation of per-gene classifier accuracy? Why should Y-predictability from an AlphaGenome track be interpreted as phenotype-specific gene relevance rather than label information retained at that locus? How stable is the ranking across alternative classifiers, seeds, and bootstrap resampling?

  2. Can the authors evaluate on hard matched controls, such as genes with substantial melanocyte activity and ancestry differentiation but no established pigmentation role? This would help determine whether the method identifies phenotype relevance rather than simply distinguishing melanocyte/pigmentation genes from generic controls.

  3. Can the method be evaluated on a measured phenotype where ancestry and phenotype are correlated but not deterministically identical? I consider this important for establishing the central claim beyond the constructed stress test.

  4. Can the authors provide stronger gene-prioritization/functional-genomics baselines under a non-circular evaluation protocol? If this is impossible because existing methods use information involved in constructing the positive panel, please clarify the scope of the comparative claim.

Reviewer_MqxAOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a method for identifying genes associated with a phenotype using genetic sequencing data and AlphaGenome. The experiments use DNA sequences from 1,072 individuals from African and European populations, with population membership used as a proxy for skin pigmentation rather than actual pigmentation measurements.

For each candidate gene, the authors use AlphaGenome to predict RNA-seq signals in melanocytes (pigment-producing cells) from the DNA sequences of all individuals. They then train a separate neural network for each gene to predict whether an individual belongs to the African or European group, using only the predicted RNA-seq signal around that gene. Genes that allow more accurate classification receive higher scores.

The authors evaluate whether this procedure ranks nine known pigmentation genes above 33 control genes. The method achieves a gene-ranking AUC of 0.880, compared with 0.724 for the best-SNP GWAS baseline.

Critical strengths
  1. The idea of using AlphaGenome-predicted RNA-seq in a phenotype-relevant cell type for gene prioritization is interesting.

  2. The ablation experiments provide useful evidence about the role of the representation.

Critical weaknesses
  1. The main concern is the use of ancestry as a substitute for the phenotype. The authors train their classifiers to predict population membership, without using actual pigmentation measurements. Consequently, a gene receives a high score when its predicted RNA-seq signal is informative about ancestry, regardless of whether that signal is related to pigmentation. The main component connecting this ancestry prediction task to pigmentation is the use of AlphaGenome-predicted RNA-seq in melanocytes. Although this representation improves the ranking of known pigmentation genes, it does not establish that the method separates pigmentation-related variation from other ancestry-associated variation. This is particularly important because disentangling these signals is a central motivation of the paper. A more convincing validation would use measured pigmentation phenotypes and demonstrate that the AlphaGenome representation improves the identification of relevant genes.

  2. The choice of control genes raises questions about the evaluation. The authors select 33 randomly sampled genes screened for known pigmentation associations. However, the classification target is ancestry, and the controls are not selected or matched according to their ancestry informativeness or predicted melanocyte RNA-seq properties. It is therefore unclear whether the observed separation reflects pigmentation relevance or differences in how much ancestry-related information the two groups retain after the AlphaGenome transformation. This matters because the reported ranking performance is the main evidence supporting the proposed approach.

Questions for authors
  1. The current evaluation uses ancestry-derived labels rather than measured pigmentation. Could the authors clarify what evidence supports interpreting the resulting gene rankings as phenotype-specific rather than ancestry-specific? Is evaluation with measured phenotype labels feasible, or could additional experimental settings help assess this distinction?

  2. Could the authors provide further analysis of the 33 control genes, particularly whether their ancestry informativeness or predicted melanocyte RNA-seq activity differs systematically from that of the pigmentation genes? This would help assess whether these differences contribute to the reported ranking advantage.

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

The paper investigates whether predicted functional information can improve candidate-gene ranking when phenotype labels are determined by population ancestry. It assigns pigmentation labels from African and European sample membership, without individual-level pigmentation measurements. A frozen AlphaGenome model predicts melanocyte RNA-seq tracks from each individual’s two haplotypes. Separate CNNs classify the labels for each candidate gene, and genes are ranked by test-set balanced accuracy.

Across nine preselected pigmentation-associated genes and 33 controls, the method achieves a ranking AUC of 0.880, versus 0.724 for the Šidák-corrected association baseline. Their top-five lists contain four and two known associated genes, respectively. The primary metric is pAUC at FPR ≤ 0.15: 0.499 for the classifier versus 0.155 and 0.089 for the two association baselines (Table 1). The experiment measures recovery of known genes; it does not validate newly ranked candidates.

Critical strengths
  1. The raw-sequence ablation distinguishes classification accuracy from gene-ranking utility. In Section 4.3 and Figure 2, raw DNA yields mean balanced accuracies of 0.979 and 0.963 for known pigmentation genes and controls. Predicted RNA-seq yields lower accuracies of 0.918 and 0.778, but better separates the groups. Accurate prediction of ancestry-derived labels is therefore insufficient for useful candidate prioritization. Appendix E also reports ranking AUCs of 0.8586–0.9007 across four training seeds on the fixed split.

  2. The method enriches the top of the ranking for known pigmentation genes. In Appendix I, Table 4, it includes one control before recovering the fifth known gene, compared with six and twenty controls for the association baselines. This matters when only a few candidates can be validated. Controls lack pigmentation annotations under the authors’ screening criteria; they are not experimentally established negatives. The result supports known-gene enrichment, rather than a measured reduction in unproductive experiments.

Critical weaknesses
  1. Direct controls are needed to distinguish target-cell-type information from gene properties. Because labels are determined by ancestry, the key question is why the functional representation preserves more ancestry-discriminative information for known pigmentation genes than for controls. Differences in expression levels, predicted-track variability, or local population differentiation could explain the separation. The raw-sequence ablation does not establish melanocyte specificity, and Appendix C’s variant-count and linkage-disequilibrium checks do not directly test these alternatives. Other-cell-type tracks and property-matched controls would help determine what drives the ranking improvement.

  2. The statistical strength of the comparison depends on the null and the fixed gene panel. Section 3.5, Table 1, and Appendix F explicitly distinguish the two tests. At c = 0.15, label-reassignment p-values are 0.0073 and 0.0057, whereas method-exchange p-values are 0.1198 and 0.0716. Label reassignment is informative about enrichment relative to arbitrary nine-gene sets, but the interpretation of this result as comparative method superiority on a fixed, curated panel needs stronger justification. The four-seed check and leave-one-gene-out analysis support robustness within this panel; they do not establish performance on another panel.

Questions for authors
  1. Cell-type controls: How does ranking performance change when tracks from other biologically justified cell types are used under the same pipeline? Is the melanocyte advantage consistent across repeated runs?

  2. Property-matched controls: Does the ranking advantage persist with controls matched on expression or predicted-track properties? Please define these properties and explain the matching procedure.

  3. Statistical interpretation: Given the fixed, curated panel, why does the label-reassignment test provide the appropriate basis for claiming superiority over the association baselines? How should that claim be interpreted alongside the nonsignificant method-exchange results at c = 0.15?

  4. Metric correction: Sections 3.1, 3.4, and 4.2 and Figure 2 use the test-set majority proportion of 0.7099 as an accuracy baseline, including comparisons with balanced accuracy. Under the definition in Table 7, an always-majority classifier has balanced accuracy 0.5, as Figure 3 also states. Please correct the reference line and comparisons involving balanced accuracy.

Reviewer_dZtJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper extends the AlphaGenome-based classification pipeline of De Souza et al. (2026) from a single-gene comparison to a systematic gene-ranking protocol, and compares it with variant-level association (GWAS) on a testbed where population structure and phenotype coincide (Genomes Project individuals labelled weak/strong pigmentation by continental ancestry). The authors pass two phased haplotypes per individual through the AlphaGenome model to extract melanocyte RNA-seq profiles, and train one CNN per gene to predict pigmentation proxy derived from continental ancestry. Across nine known pigmentation genes and 33 screened controls, the proposed protocol recovers the panel at AUC = 0.880 (p < 0.001, permutation null test) compared with 0.724 for a variant-level association test (GWAS). The authors conduct ablations that compare raw DNA sequences, scalar expression summaries, and unaligned profiles, suggesting that the reduction of false positives comes from a filter built into the representation of AlphaGenome rather than from explicitly modeling the confound.

Critical strengths
  1. Systematic gene-level evaluation. The paper extends the AlphaGenome-based classification pipeline of De Souza et al. (2026) to gene prioritization, evaluating nine literature-supported pigmentation genes against 33 screened controls. This moves beyond the previous single-gene comparison and allows the authors to assess whether the approach consistently prioritizes known pigmentation genes. The inclusion of both overall ranking quality and early-retrieval metrics is relevant to practical candidate gene prioritization, where only a small number of top-ranked genes may be selected for experimental follow-up.

  2. Informative empirical comparisons and ablations. The paper compares the proposed ranking against gene-based GWAS statistics using the same cohort and labels, and examines several alternative input representations. These experiments provide useful evidence that AlphaGenome-predicted RNA-seq profiles may improve the separation of known pigmentation genes from screened controls in this specific setting. The contrast between raw DNA, predicted RNA-seq profiles, and scalar expression summaries is particularly informative because it raises a testable hypothesis about the role of the representation.

Critical weaknesses
  1. Limited methodological novelty and insufficient support for the broader claims.

The core AlphaGenome-to-CNN pipeline is inherited from De Souza et al. (2026), with the main extensions involving gene-wise ranking, two-haplotype input, alignment, and evaluation. More importantly, the central claims appear stronger than the evidence consistently supports. At the pre-declared pAUC cap, neither GWAS comparison is significant under the method-exchange null (p = 0.1198 and 0.0716), although both are significant under label reassignment. The claimed mechanistic explanation is also not conclusively established: each ablation comparison loses significance under the primary null when a particular gene is removed (Appendix I). Nevertheless, Section 5 suggests that the representation's filtering property generalizes beyond pigmentation. Thus, the strongest empirical evidence is limited to this particular cohort, while the broader claims rely on less conclusive results.

  1. Limited evidence for phenotype-specific gene discovery and concerns about construct validity.

The target labels are deterministically derived from continental ancestry rather than measured pigmentation. High classification accuracy therefore demonstrates ancestry-informative signals, not necessarily pigmentation relevance or causality. Moreover, the positive and control gene sets are characterized using different evidence: the nine panel genes are literature-curated, whereas the 33 randomly sampled controls are checked against annotation databases. As Appendix B acknowledges, one control has no curated annotations, and three biologically borderline controls are retained. Since AUC, pAUC, and P@5 measure separation between these sets, their interpretation depends on the validity of this panel/control distinction. Without stronger evidence that the controls are biologically unrelated to pigmentation, it remains unclear how much of the observed separation reflects pigmentation-specific biology.

Questions for authors
  1. Choice of permutation null for claimed method superiority (clarification). Looking at Tables 1 and 5, the CNN achieves a higher pAUC than both GWAS baselines. However, at the pre-declared threshold c=0.15c = 0.15, the advantage over the Šidák minimum and mean-χ2\chi^2 estimators is significant under the label-reassignment null (p=0.0073p = 0.0073 and 0.00570.0057, respectively), but non-significant under the method-exchange null (p=0.1198p = 0.1198 and 0.07160.0716, respectively). The Šidák comparison remains non-significant under method exchange across all seven reported caps in Table 5. Since label reassignment fixes both rankings and shuffles the panel/control labels, whereas method exchange swaps the two methods' ranks within genes, could the authors clarify why label reassignment is chosen as the primary test for claiming superiority between the two methods? How is the exchangeability assumption justified given the literature-curated positive panel and randomly sampled, annotation-screened controls? Furthermore, should the Abstract explicitly qualify its significance claim by naming the primary null, as the Table 1 caption does?

  2. Uncertainty in per-gene scores and ranking stability (clarification). Each gene is ranked by balanced accuracy on only 162 test individuals (115 strong, 47 weak). In Table 7, EGF (control) and DDB1 (panel) differ by just 0.00046 in balanced accuracy, much less than the effect of changing a single prediction (0.0043–0.0106). This small difference determines whether DDB1 contributes to the primary pAUC at (c=0.15): reversing their order would increase pAUC from 0.4994 to approximately 0.5208 (=0.703030/1.35). Although Table 3 examines robustness across training seeds, it does not address test-sample variability. Could the authors assess the stability of per-gene rankings and pAUC under family-aware bootstrapping or repeated data splits? In particular, how robust is the reported advantage over GWAS to such sampling variation?

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

Related papers

Discussion (0)

Sign in to comment.