acceptodds
Withdrawn from ICLR 2027

An Uncertainty-aware Optimal Transport Framework for Cross-Population Polygenic Risk Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Polygenic risk scores (PRS) often lose predictive accuracy across ancestry groups because differences in allele frequencies, linkage disequilibrium (LD), and genetic architecture create distribution shifts between source and target populations. Existing cross-population PRS methods improve portability through Bayesian shrinkage and multi-ancestry modeling, but generally do not explicitly model structured correspondence between ancestry-specific genomic contexts. To address this limitation, we propose OT-PRS, an uncertainty-aware optimal transport framework for cross-population polygenic risk prediction. OT-PRS uses block-wise fused Gromov–Wasserstein transport to align source and target variants by jointly modeling SNP-level similarity and ancestry-specific LD geometry. The transported source effects define a target-aligned prior that is refined using target-population summary statistics, while transport entropy quantifies transfer uncertainty and enables uncertainty-aware score construction and individual-level reliability assessment. Experiments across multiple diseases and ancestry groups in UK Biobank demonstrate competitive performance against established PRS methods, and external validation in the All of Us Research Program further supports the portability of OT-PRS across cohorts. Our results highlight optimal transport as a flexible framework for improving the portability and reliability of polygenic risk prediction.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Optimal transport, Polygenic risk prediction, Uncertainty quantification, Statistical genetics

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_xJYLOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

OT-PRS aligns SNPs across populations using fused Gromov-Wasserstein transport, transfers source effects into a prior, and refines them with target GWAS statistics. Transport entropy supplies additional weighting and prediction features. Evaluation covers nine UK Biobank disease-ancestry tasks and covariate-adjusted discrimination in All of Us.

Critical strengths
  1. The transport objective and target-effect refinement provide an explicit, inspectable formulation of cross-population information transfer.

  2. The target-only and identity-prior ablations directly examine the distinctive transport component, while the paper appropriately distinguishes entropy from calibrated predictive uncertainty.

Critical weaknesses
  1. The primary evaluation does not establish discovery–evaluation independence. The UK Biobank analysis uses GWAS statistics and evaluation genotypes from the same resource, while cross-validation partitions only downstream phenotype-model fitting. Potential overlap is acknowledged, but exclusion of evaluation individuals from discovery is not established; consequently, the AUCs, pseudo-R² values, and rankings in Section 5.1 cannot establish independent genetic prediction.

  2. The external study does not resolve the central validation gap. Figure 5 reports only the full covariate-adjusted model in All of Us, without covariate-only, competing PRS, or matched-prior results. As Section 5.2 acknowledges, this does not isolate the genetic increment; the fitting and evaluation separation for the external regression also needs clarification before these AUCs establish the proposed method's predictive advantage.

  3. The matched ablation provides little evidence for the distinctive transport mechanism. Section 5.3.1 reports mean AUCs of 0.7691 for OT and 0.7690 for an identity prior under matched conditions. This does not show that cross-SNP redistribution adds meaningfully beyond target refinement and direct borrowing; the larger differences against feature-only or structure-only transport in Figure 6 do not resolve that question.

  4. Signed effect transfer lacks a sufficiently justified correspondence. The transport cost uses allele-frequency differences and unsigned LD geometry, whereas Equation 2 averages signed effects across different variants. Harmonizing the same variant across populations does not determine the appropriate relative sign between distinct variants, so lower alignment cost need not preserve useful tagging effects; this weakens the biological motivation for the proposed prior.

  5. The effect-size parameterization is insufficiently specified to assess the estimator. Equation 3 relates marginal effects to joint effects through the correlation matrix R, while Equation 5 scores effect-allele dosages. These expressions require compatible standardized or per-allele coefficient conventions and the corresponding genotype-variance conversion; without those definitions, it is unclear whether target refinement and downstream scoring estimate the same genetic quantity.

  6. The transport-derived uncertainty measure lacks empirical validation. The method uses transport entropy to quantify alignment ambiguity and attenuate SNP effects, but does not establish that higher entropy corresponds to less reliable effect transfer or greater prediction error. Appendix D.4 shows that most SNPs have highly diffuse alignments (median entropy 0.924), resulting in substantial attenuation of effects. Although the authors acknowledge that entropy is not calibrated predictive uncertainty, its usefulness as a reliability measure is not demonstrated.

Questions for authors
  1. Can the primary comparison and identity-prior ablation be evaluated with strictly disjoint GWAS discovery and test samples? In All of Us, what are the held-out genetic increments over covariates alone and the corresponding baseline PRS methods?

  2. Are the external regression coefficients fitted and evaluated on separate individuals? What effect-size conventions connect Equations 3 and 5, including any genotype-standard-deviation conversions?

  3. What justifies transferring effect signs across distinct SNPs under an unsigned alignment cost, and is the resulting score stable under consistent alternative effect-allele coding?

Reviewer_ji8AOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an optimal transport (OT)-based PRS approach for underrepresented ancestry groups. It accounts for block-wise LD structure, applies OT within each block to align source and target populations, and incorporates transport uncertainty into prediction. Improving PRS performance in underrepresented populations is important, and the paper presents an interesting idea. However, several limitations weaken the empirical evidence, and I do not consider the paper suitable for publication in its current form.

Critical strengths

• An interesting application of OT to PRS construction for underrepresented ancestry groups.

Critical weaknesses

• Training-data contamination: In UKBB, it seems that the target summary statistics include all samples in the target dataset, thereby inflating prediction accuracy. Due to this issue, all results from the UKBB analysis are questionable. • Low performance gain: Section 5.3.1 shows that the OT prior produces nearly identical results to those obtained using target-only estimation or the identity prior. The maximum gain is < 0.005 (Figure 6). • It is possible that some heterogeneity across ancestry groups is due to effect-size heterogeneity, but OT only considers differences in genetic structure (MAF and LD).

Questions for authors

• Since the UKBB test individuals were included in the training summary statistics, the authors should either obtain summary statistics that exclude the test individuals or compare methods in All of Us. • In the All of Us data analysis, it seems that covariates are included in the prediction model, so the prediction performance is not adjusted for them. Therefore, the predictive performance attributable to PRS is not isolated. The authors should evaluate delta PRS in All of Us.

Reviewer_1yLfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

OT-PRS uses entropy-regularized fused Gromov-Wasserstein transport to align source and target SNPs using allele frequencies and LD structure. Transported source effects define a prior for target-informed MAP refinement. Entropy-derived weighting and individual uncertainty features subsequently modify prediction. Evaluation uses UK Biobank and an external All of Us cohort.

Critical strengths
  1. Cross-population PRS transfer is an important problem, and the method provides a concrete combination of structured alignment and target-informed refinement. The regularized quadratic refinement has a well-defined unique solution under the stated positive-definiteness conditions.
  2. The manuscript explicitly acknowledges potential discovery-evaluation overlap in UK Biobank and includes external evaluation with frozen SNP effects. It also provides a useful matched comparison of OT, identity-prior, and target-only refinement.
Critical weaknesses
  1. The experiments do not convincingly establish independent comparative benefit from optimal transport.
    • Discovery-evaluation overlap. Appendix C.1 acknowledges that the five-fold split does not itself separate participants used to estimate the Pan-UK Biobank GWAS summary statistics. It also notes that potential discovery-evaluation overlap limits interpretation as independent out-of-sample prediction.
    • All of Us. This cohort helps, but only OT-PRS is reported there. Its covariate-adjusted AUCs do not establish incremental discrimination over covariates (as Section 5.2 acknowledges) or superiority over competing PRS methods in that cohort.
    • Matched ablation. Section 5.3.1 excludes uncertainty components. Mean AUC is 0.7691 for the OT prior versus 0.7690 for the identity prior and 0.7677 for target-only estimation. The OT prior matches or exceeds the identity prior in seven of nine tasks, but the per-task differences shown in Figure 6A are small. No matched identity-prior comparison is reported for the full model with uncertainty components.
    • The aggregate evidence for the central transport step is weak.
  2. Transporting signed effects across different SNPs needs stronger justification. The structural cost uses 1−∣r∣21-|r|^2 (Section 3.1; Equation 8), which removes LD sign. Meanwhile, the transported prior in Equation 2 averages signed effects through a nonnegative coupling. Harmonizing alleles for the same SNP does not by itself justify combining effects from different SNPs without accounting for their relative signed relationship. The paper needs assumptions or evidence showing when this operation preserves useful effect direction and avoids cancellation or coding sensitivity.
  3. The entropy statistics establish diffuse alignment under the fitted configuration (median normalized entropy 0.924; Appendix D.4), but they do not by themselves establish poor prediction. Likewise, the paper explicitly states that transport entropy and the individual uncertainty score are not calibrated predictive variances, posterior precisions, or confidence intervals (Section 3.4; Appendix A.4). Lack of calibration is therefore not my principal objection.
Questions for authors
  1. Can the authors provide independent external comparisons containing covariate-only models and the strongest competing PRS methods, with a clear separation of GWAS discovery, tuning, and evaluation?
  2. What assumptions justify Equation 2 for signed effects transported between different variants? How does it behave under equivalent allele recodings and sign-sensitive LD controls?
  3. Does OT improve upon identity-prior refinement across independent cohorts and repeated evaluations when preprocessing, refinement, and downstream predictor specifications are matched?

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

Related papers

Discussion (0)

Sign in to comment.