acceptodds
Under review as a conference paper at ICLR 2027

From Sequence Uncertainty to Phenotype Hypotheses: Uncertainty-Conditioned Inference over Biomedical Knowledge Graphs

Abstract

Genomic foundation models provide variant-level evidence, but connecting that evidence to phenotype hypotheses requires accounting for uncertainty throughout structured reasoning. We introduce an uncertainty-conditioned path-energy framework that converts sequence predictions and calibrated predictive uncertainty into explicit distributions over graph-supported phenotype hypotheses, without patient-level phenotype supervision. An interpretable learned energy scores admissible gene–disease–phenotype routes, query uncertainty controls the target temperature, and path probabilities are aggregated into phenotype hypotheses. This formulation makes evidence quality an explicit determinant of hypothesis breadth while retaining traceable multi-path support. Exact normalization over enumerable candidate sets provides a reference for evaluating a Hastings-corrected learned revision operator. On AlzKB and Hetionet, phenotype-level entropy tracks upstream uncertainty, with Spearman correlations of 0.951 and 0.878, respectively; ablations identify the contribution of uncertainty conditioning. In masked-edge recovery, the model exceeds chance and degree-matched probability-permutation controls. With query genes held out from calibration and energy learning, mean recall@10 reaches 0.206 across three edge masks, exceeding DistMult by 0.108, with a roughly fivefold lower across-mask coefficient of variation. These results establish an uncertainty-aware interface from genomic predictions to graph-supported phenotype hypotheses that combines interpretable probability allocation, improved held-out association recovery, and reduced sensitivity to graph masking.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Genotype-to-phenotype prediction, Biomedical knowledge graphs, Genomic foundation models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6kjLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes an uncertainty-conditioned path-energy framework for mapping genomic sequence predictions and calibrated predictive uncertainty to distributions over phenotype hypotheses in biomedical knowledge graphs. The method assigns an interpretable energy to admissible gene–disease–phenotype paths, uses uncertainty-dependent temperature scaling to control distributional breadth, aggregates path probabilities into phenotype-level hypotheses, and implements a Hastings-corrected Metropolis–Hastings sampler whose stationary distribution can be checked against an exactly normalized target. The authors evaluate the framework on AlzKB and Hetionet. They report strong correlations between upstream uncertainty and phenotype-level entropy, verify that the sampler reproduces the exact target distribution, and evaluate masked disease–symptom edge recovery under gene-held-out settings. The proposed method achieves higher average recall@10 than DistMult across three graph masks and lower across-mask variability.

Critical strengths

The paper clearly separates the learned target distribution from the fidelity of the sampling operator. Exact enumeration allows the authors to verify the Hastings-corrected sampler directly, which is a useful methodological check.

The framework provides an interpretable mechanism for propagating upstream uncertainty into the breadth of graph-supported hypotheses. The use of path-level features and phenotype-level aggregation makes the output more inspectable than a black-box ranking score.

Critical weaknesses

The central uncertainty result is largely imposed by the model parameterization rather than demonstrated as calibrated phenotype uncertainty. The temperature is explicitly defined as an increasing function of the supplied uncertainty, and the paper proves that increasing temperature increases entropy for fixed utilities. The high uncertainty–entropy correlations therefore mainly confirm the designed coupling; they do not establish that the resulting phenotype distributions are calibrated against phenotype-level uncertainty or clinical outcomes.

The biological evaluation is based on a small and highly constrained setup. The AlzKB experiments use 982 variants from only 15 genes, with APP, PSEN1, and PSEN2 accounting for 94% of queries, and the energy is learned from only 121 definitive pathogenic/benign variants. The candidate supports are constructed from the same knowledge graphs used for inference, so the results do not establish generalization to unseen genes, diseases, phenotypes, or external patient-level data.

The evidence for improved recovery over DistMult is not robust across the reported splits. Under gene-held-out evaluation, the pooled recall@10 improvement is positive, but the improvement for mask seed 42 has a confidence interval containing zero. On the primary split, the full model and DistMult differ by only 0.013 recall@10 with an interval that also contains zero. Thus, the headline aggregate improvement is sensitive to how the graph is masked and aggregated.

The method’s contribution is difficult to isolate from the learned energy and candidate-support design. The full model includes route features, endpoint features, confidence interactions, and uncertainty-dependent temperature. Several fixed-temperature and component ablations achieve similar recovery, and the paper does not show that uncertainty conditioning itself improves held-out biological association recovery. The sampler, meanwhile, is not a computational improvement over direct categorical sampling because the candidate supports are enumerable.

Questions for authors

Can the authors evaluate phenotype-level probability calibration using independent phenotype or patient-level outcomes, rather than only entropy correlations and masked graph-edge recovery? Without such an evaluation, what is the intended interpretation of the output probabilities?

Can the authors report performance on gene-held-out, disease-held-out, and phenotype-held-out splits that prevent the model from relying on graph neighborhoods closely related to the training queries? The current evaluation leaves held-out genes in the graph and uses a small, highly concentrated set of genes.

How does the method perform when the upstream sequence uncertainty is deliberately miscalibrated or perturbed? Since the uncertainty-to-temperature schedule is hand-specified, a robustness analysis would clarify whether the framework responds to uncertainty quality or merely to the numerical value of H.

Can the authors provide per-mask and per-gene statistical results, rather than relying primarily on pooled query-level averages? In particular, the non-significant improvement on one of the three masks and the wide variation in the number of reachable paths should be discussed more prominently.

Can the authors isolate the contributions of uncertainty conditioning, route features, and the learned energy through matched ablations evaluated on the same held-out recovery task? The current results show that the full framework works better in aggregate, but do not establish which component is responsible for the improvement.

Since the candidate supports are enumerable, please clarify the practical motivation for the Metropolis–Hastings sampler. What computational regime or graph size would make sampling preferable to exact normalization, and how does the method scale when enumeration is no longer feasible?

3 comments
AuthorsOpenReview ↗

Reply to Reviewer 6kjL

Thank you for the detailed review. We ran every analysis you asked for that our data allow (Tables R1–R8 in our general comment; gene-held-out, gene-cluster 95% bootstrap intervals).

W3 / Q4. Robustness of the DistMult comparison; per-mask and per-gene statistics. Per-mask and per-gene results for ten masks are in Table R2a–c (pooled difference +0.053 [−0.038, +0.086] query-weighted, +0.016 [−0.049, +0.091] gene-balanced). To separate mask effects from DistMult's training noise we then ran a confirmation test with the protocol fixed in advance: 30 fresh masks, DistMult averaged over three training seeds (Table R8 in the general comment):

recall@10, 30 fresh masksOurs (submitted model)DistMultOurs − DistMultOurs higher
Query-weighted0.1830.123+0.060 [−0.028, +0.093]24/30 masks (Wilcoxon p < 10⁻⁴)
Gene-balanced0.2520.239+0.013 [−0.043, +0.082]19/30 masks (p = 0.18)

Narrower claim. Query-weighted recall@10 is higher than DistMult's on most masks (24 of 30), an advantage concentrated in APP, PSEN1 and PSEN2 (0.191 / 0.197 / 0.110 vs 0.132 / 0.094 / 0.097); the gene-cluster interval includes zero and the other 12 genes split 6–6, so we do not claim an advantage across genes. Across-mask CV is lower for our method query-weighted (0.30 vs 0.39) and higher gene-balanced (0.16 vs 0.12), so we no longer claim lower variability. Both statements will replace the current sentence in the abstract and Section 5.7, with gene-balanced and per-gene results reported next to query-weighted ones.

Why the paper's conclusions do not rest on exceeding DistMult. DistMult is a reference for task difficulty: a trained triple scorer that receives neither the sequence prediction nor its uncertainty and returns a score, not a distribution over traceable paths. The contribution is the uncertainty-conditioned interface, and each part of it is supported independently of that comparison: (i) hypothesis breadth tracks upstream uncertainty within every gene with ≥5 variants and without the three large genes, on both graphs (Table R1); (ii) the target is exact and the sampler is verified against it; (iii) training under θ(H) improves held-out recovery over the same energy retrained at constant temperatures from 0.20 to 0.80, by 0.043–0.061 with intervals excluding zero (Table R3); (iv) recovery is above chance under both weightings on all 30 fresh masks (+0.118 [+0.060, +0.141]; +0.095 [+0.037, +0.167]) and under disease hold-out (Table R6). The method does this while remaining at least on par with the trained embedding.

W4 / Q5. What the contribution is, and which component is responsible for what. The novelty is not a new energy family or a new sampler; both are standard, as Sections 2 and 3.4 say. It is the interface: an upstream model's calibrated uncertainty about a query sets the concentration of an exactly normalized distribution over graph-supported phenotype hypotheses, and that distribution, not a ranking score, is the evaluated output. The matched ablations you asked for, on the same held-out recovery task (ten masks, identical queries; Tables R3 and R4), assign a role to each component:

ComponentTestEffect on recall@10Effect on breadth, ρ(H, entropy)
Uncertainty conditioning (the contribution)Energy retrained at constant θ ∈ {0.20, …, 0.80}, bracketing the mean of θ(H) (0.69)0.106–0.124 vs 0.167; gain +0.043 to +0.061, all intervals exclude zero0.19–0.64 vs 0.74
Sequence prediction pp-interaction features removed−0.027 [0.012, 0.040]0.53 vs 0.74
Learned energyReplaced by uniform path mass+0.042 [−0.075, +0.134], not resolved−0.20 vs 0.74
Confidence and route-mechanism featuresRemovedno loss (−0.006; −0.036, intervals include zero)0.68; 0.77

So uncertainty conditioning is responsible for the breadth behaviour and, when the energy is trained under it, for a recovery gain over matched constant temperatures; the submitted "fixed θ" ablation changed the temperature only at inference, which is why it showed no difference. Hotter constants (1.00, 1.25) come close (+0.022 [−0.003, +0.040]; +0.015 [−0.008, +0.033]), and we will state that. The sampler is not claimed as a contribution: it is the component that would be needed beyond enumerable supports, included here because it can be verified against the exact target.

W1 / Q1. Calibration. Agreed, and this is not claimed. Proposition 1 covers path entropy at fixed utilities; that phenotype entropy still tracks H after aggregation is empirical and holds within every gene (median within-gene ρ 0.96 / 0.90) and without the three dominant genes (0.950 / 0.769; R1). It is not calibration: under a log score on held-out positives the conditioned distribution is not better than fixed temperature (−10.99 vs −8.32 bits). No patient-level data exist for these variants. The outputs are relative support among graph-supported hypotheses, as §3.3 states.

W2 / Q2. Hold-out structure. Held-out genes must stay as nodes or no path exists; all their labels are withheld from calibration and energy fitting. New disease hold-out (all symptom edges of 20% of diseases removed, five folds, combined with gene hold-out; R6): recall@10 0.108 vs chance 0.056, +0.052 [+0.025, +0.088]; not different from DistMult (+0.013 [−0.127, +0.032]). A phenotype with no remaining edges is unreachable for any graph-supported method, so phenotype hold-out is outside the method's scope; we say so.

Q3. Miscalibrated H (Table R5). Breadth follows the value supplied: ρ with the true H is 0.74 with the true signal, 0.68 with added noise (σ = 0.2), 0.17 when H is permuted across variants, −0.58 when inverted. Recall@10 is unchanged (0.166–0.182). The framework transmits H faithfully; it cannot repair a poor H, and on locked likely-label variants our upstream H flags sequence-call errors only weakly (AUROC 0.56). We add this as a limitation.

Q6. Why MH. As §3.4 states, there is no computational benefit on enumerable supports; the sampler is included so it can be verified against an exact target. Uncapped supports grow about 100× per gene hop (median 85, 4,980, 6.0×10⁵, 7.3×10⁷ paths for 0–3 hops; R7), so enumeration fails beyond one hop. Our current proposal is a global independence proposal and would itself need replacing by a local or amortized one in that regime; the Hastings correction and the exact-reference test carry over. We make no scalability claim.

Dear Reviewers,

Thank you again for your thoughtful comments. Please let us know whether our responses have addressed your concerns or if any further clarification is needed.

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

The paper introduces a query-conditioned probabilistic interface that merges predictive uncertainty into a structured distribution over phenotype hypotheses. The framework gives each path among gene, disease, and phenotype an interpretable linear energy score. The framework then obtains phenotype probabilities by summing over the paths connecting them. The authors demonstrate the model's superiority by applying it to two graphs, AlzKB and Hetionet, and report improved performance in mean recall and across-mask variability.

Critical strengths
  1. The paper is well organized and presents the model in a clear way. The overall results are consistent across both graphs.
  2. Each path receives an explicit score and probability. This makes the hypotheses interpretable and allows readers to check each path’s contribution directly.
Critical weaknesses
  1. My major concern is the novelty. The components presented in Sections 2 and 3, such as path energy and the Metropolis-Hastings sampler, are mostly standard and already exist in the literature. How is the paper different from temperature-conditional GFlowNets (Kim et al., 2024)?

  2. Most of the AlzKB variants come from APP, PSEN1, and PSEN2. The authors should provide more comprehensive results and analysis to show that the uncertainty–breadth correlation holds not only for these three genes.

Questions for authors
  1. How do you choose the constants for Theta(H)? What if you change the slope, and will the results change much? I think the paper needs more sensitivity analysis regarding this

  2. Since APP, PSEN1 and PSEN2 dominate the variants. can the authors report leave-one-gene-out or gene-balanced results

2 comments
AuthorsOpenReview ↗

Reply to Reviewer vWvr

Thank you for the constructive questions. We ran both requested analyses. Full tables (R1–R8) are in our general comment; intervals are gene-cluster 95% bootstrap.

W2 / Q2. Dominance of APP, PSEN1, PSEN2. The uncertainty–breadth association does not depend on them (Table R1):

ρ(H, phenotype entropy)AlzKB, θ(H)AlzKB, fixed θHetionet, θ(H)Hetionet, fixed θ
All variants0.9510.5820.8780.495
Excluding the three genes0.950 (61 variants, 12 genes)0.5090.769 (64 variants, 15 genes)0.218
Lowest value when one of the three is left out0.9340.4950.8700.442
Within gene, median [min, max], 8 genes with ≥5 variants0.959 [0.901, 0.987]0.7130.900 [0.602, 0.931]0.782
Gene-balanced (equal weight per gene)0.9350.6220.7820.430

For recovery, the energy and calibrators were already refitted with each gene group left out. Gene-balanced and per-gene recovery are in Tables R2 and R8. We also extended the recovery evaluation to ten masks (Table R2) and to 30 further masks in a confirmation test with the protocol fixed in advance, DistMult averaged over three training seeds (Table R8). On the 30 fresh masks query-weighted recall@10 is 0.183 versus 0.123 for DistMult, higher on 24 of 30 masks; gene-balanced it is 0.252 versus 0.239 (+0.013 [−0.043, +0.082]). We therefore state the comparison more narrowly: the advantage is consistent across masks and concentrated in the three large genes, it is not claimed across genes, and we no longer claim lower across-mask variability. The paper's conclusions do not rest on exceeding DistMult, which serves as a reference for task difficulty and uses neither the sequence prediction nor its uncertainty: the breadth result (table above), the gain over constant-temperature training (Q1 below) and recovery above chance under both weightings (+0.118 [+0.060, +0.141]; +0.095 [+0.037, +0.167]) each hold on their own.

Q1. Choice of θ(H) and sensitivity. The schedule 0.35 + 0.90H was fixed in the study specification before the reported experiments and was not tuned on any reported metric. We now retrain the energy for every schedule a + bH (R3, ten masks, gene-held-out):

recall@10ρ(H, entropy)
Paper schedule (0.35, 0.90)0.1670.74
11 other schedules, a ∈ {0.20, 0.35, 0.50}, b ∈ {0.45, 0.90, 1.35, 1.80}0.147–0.1830.62–0.80
Constant temperature (b = 0), θ ∈ {0.20, 0.35, 0.50, 0.60, 0.80}0.106–0.1240.19–0.64
Constant temperature, θ ∈ {1.00, 1.25}0.145–0.1520.58–0.66

Results are stable across slopes and intercepts. Any uncertainty-dependent schedule recovers more than constant temperatures up to 0.80 (paper schedule vs these: +0.043 to +0.061, all intervals exclude zero); very hot constants come close (+0.022 [−0.003, +0.040], +0.015 [−0.008, +0.033]).

W1. Novelty and temperature-conditional GFlowNets. The components are standard and we say so. Differences from Kim et al. (2024): (i) there the temperature is a user-chosen exploration setting over one reward; here it is supplied per query by another model's calibrated uncertainty, so hypothesis breadth becomes a function of evidence quality; (ii) we do not amortize: supports are enumerable, the target is normalized exactly, and sampler error is measured (JS 0.003) instead of assumed; (iii) the energy is learned from 121 variant labels through an aggregated readout, not given as a reward. A temperature-conditional GFlowNet is the natural sampler once enumeration fails, and our exact-reference test is what would validate it. The R3 comparison above is the evidence that this conditioning adds value.

Thank you for your candid and detailed response. All of my concerns have been addressed, and I have increased my score accordingly.

Reviewer_c5GeOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how uncertainty in genomic predictions can be reflected in distributions over graph-supported phenotype hypotheses. It learns scores for gene–disease–phenotype paths from variant pathogenicity labels, uses an uncertainty-dependent temperature to control path probabilities, and aggregates them into phenotype scores. Experiments on AlzKB and Hetionet examine the relationship between uncertainty and hypothesis breadth, check the sampler against the exact distribution, and evaluate recovery of masked graph associations. The recovery experiments also include a setting where query genes are excluded from sequence calibration and energy fitting.

Critical strengths
  • I do not have much experience in this area but I think the problem is interesting and well-motivated.

  • The phenotype scores can be traced to explicit graph paths, making their construction inspectable.

  • The paper includes gene-held-out splits and an explicit leakage audit for the masked-edge experiments, which shows care in the evaluation design.

  • The code is publicly available.

Critical weaknesses
  • I wonder if the technical novelty is limited. Learned path features and temperature-controlled distributions have precedents in Lao et al. (2011) and Kim et al. (2024), while the sampler uses standard Metropolis–Hastings. The contribution seems mainly to lie in combining these ideas for phenotype hypotheses, and the paper does not convincingly establish the added value of this particular combination.

  • The practical benefit of uncertainty conditioning remains unclear. The association between uncertainty and broader distributions is largely built in: Θ(H)\Theta(H) increases with HH by design, and Proposition 1 already implies higher entropy at higher temperature. No variant-level phenotype data are used to check whether the broader distributions are more accurate. The recovery comparisons also do not show an improvement over fixed temperature. The gap between demonstrating the model’s behavior and demonstrating its usefulness therefore remains.

  • The recovery evaluation is heavily concentrated: 921 of 964 queries come from three genes, and variants from the same gene share the recovery targets. The stability comparison also depends on the averaging scheme: giving each gene equal weight yields an across-mask coefficient of variation of about 0.16 for the method and 0.12 for DistMult. The average recovery advantage remains, but the robustness claim needs this qualification.

Questions for authors

Please check weakness.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer c5Ge

Thank you for the careful reading. We ran new experiments for each of your three points. All new tables (R1–R8) are in our general comment; intervals are gene-cluster 95% bootstrap.

W3. Gene concentration, weighting and stability. We evaluated this directly. First, seven more masks under the submitted protocol (ten in total; Table R2), reported query-weighted and gene-balanced; your gene-balanced CV values (0.16 vs 0.12 on the three submitted masks) are reproduced there. Second, a confirmation test with the protocol fixed in advance: 30 fresh masks, DistMult averaged over three training seeds (Table R8 in the general comment):

recall@10, 30 fresh masksOurs (submitted model)DistMultOurs − DistMultOurs higher
Query-weighted0.1830.123+0.060 [−0.028, +0.093]24/30 masks (Wilcoxon p < 10⁻⁴)
Gene-balanced0.2520.239+0.013 [−0.043, +0.082]19/30 masks (p = 0.18)

Narrower claim. Query-weighted recall@10 is higher than DistMult's on most masks (24 of 30), an advantage concentrated in APP, PSEN1 and PSEN2 (0.191 / 0.197 / 0.110 vs 0.132 / 0.094 / 0.097); the gene-cluster interval includes zero and the other 12 genes split 6–6, so we do not claim an advantage across genes. Across-mask CV is lower for our method query-weighted (0.30 vs 0.39) and higher gene-balanced (0.16 vs 0.12), so we no longer claim lower variability. Both statements will replace the current sentence in the abstract and Section 5.7, with gene-balanced and per-gene results reported next to query-weighted ones.

Why the paper's conclusions do not rest on exceeding DistMult. DistMult is a reference for task difficulty: a trained triple scorer that receives neither the sequence prediction nor its uncertainty and returns a score, not a distribution over traceable paths. The contribution is the uncertainty-conditioned interface, and each part of it is supported independently of that comparison: (i) hypothesis breadth tracks upstream uncertainty within every gene with ≥5 variants and without the three large genes, on both graphs (Table R1); (ii) the target is exact and the sampler is verified against it; (iii) training under θ(H) improves held-out recovery over the same energy retrained at constant temperatures from 0.20 to 0.80, by 0.043–0.061 with intervals excluding zero (Table R3); (iv) recovery is above chance under both weightings on all 30 fresh masks (+0.118 [+0.060, +0.141]; +0.095 [+0.037, +0.167]) and under disease hold-out (Table R6). The method does this while remaining at least on par with the trained embedding.

W2. Practical benefit of uncertainty conditioning.

  • Built in? Proposition 1 concerns path entropy at fixed utilities. The reported quantity is phenotype entropy after aggregation, with H-dependent utilities, so its behaviour is empirical. It holds within every gene with ≥5 variants (median within-gene ρ 0.96 on AlzKB, 0.90 on Hetionet) and without APP/PSEN1/PSEN2 (ρ 0.950 and 0.769; fixed temperature: 0.51 and 0.22) (Table R1).
  • Useful? Our "fixed temperature" ablation only swapped the temperature at inference. We now retrain the energy at constant temperature (R3, ten masks, gene-held-out). Constants from 0.20 to 0.80, which bracket the mean of θ(H) (0.69), reach recall@10 0.106–0.124 versus 0.167 with θ(H); the paired gains are +0.043 to +0.061 and every interval excludes zero. Hotter constants (1.00, 1.25) close most of the gap (+0.022 [−0.003, +0.040]; +0.015 [−0.008, +0.033]). So conditioning improves recovery over matched constant temperatures, and we state the hot-constant caveat.
  • Accuracy of broader distributions. No variant-level phenotype data exist for these variants, and under a log score on held-out associations the conditioned distribution is not better than fixed temperature. We keep the statement that outputs are relative support, not calibrated probabilities.

W1. Novelty. The components are standard, as §2 and §3.4 say. What Lao et al. and Kim et al. do not provide is the object: a per-query, exactly normalized distribution over graph-supported phenotype hypotheses whose concentration is set by an upstream model's calibrated uncertainty about that query, with sampler error measured against the exact target. PRA trains a relation ranker from path features and has no uncertainty input; in temperature-conditional GFlowNets the temperature is a user-chosen exploration setting. The added value of the combination is the R3 result above.

Dear Reviewers,

Thank you again for your thoughtful comments. Please let us know whether our responses have addressed your concerns or if any further clarification is needed.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General new results to all reviewers

We thank the reviewers. This comment collects every new result; the individual replies refer to the tables here as R1–R8.

R1. The uncertainty–breadth association is not driven by APP/PSEN1/PSEN2

Spearman ρ between H and phenotype entropy (AlzKB: 982 queries; Hetionet: 985). Within gene: the 8 genes with ≥5 variants, median [min, max]. Gene-balanced: 20 variants resampled per gene, 500 draws.

GraphModelAll variantsExcluding the 3 genesLowest when one of the 3 is left outWithin geneGene-balanced
AlzKBFull θ(H)0.9510.950 (61 variants, 12 genes)0.9340.959 [0.901, 0.987]0.935 [0.919, 0.948]
AlzKBFixed θ0.5820.509 (61 variants, 12 genes)0.4950.713 [0.293, 0.896]0.622 [0.544, 0.702]
HetionetFull θ(H)0.8780.769 (64 variants, 15 genes)0.8700.900 [0.602, 0.931]0.782 [0.749, 0.812]
HetionetFixed θ0.4950.218 (64 variants, 15 genes)0.4420.782 [−0.195, 0.945]0.430 [0.358, 0.500]

R2. Gene-held-out recall@10 on ten edge masks

R2a. Per mask (first three rows: masks of the submission).

Mask seedOurs QWDistMult QWOurs − DistMult, QW [95% CI]Ours GBDistMult GBChance QW
420.1820.151+0.032 [−0.016, +0.042]0.2390.1900.063
70.2130.057+0.156 [+0.043, +0.265]0.3120.2330.066
1230.2220.085+0.137 [+0.040, +0.228]0.2370.1920.065
10.2120.221−0.009 [−0.133, +0.058]0.2180.2730.064
20.0560.068−0.012 [−0.178, −0.004]0.1570.2730.064
30.1230.212−0.089 [−0.122, −0.059]0.2190.2590.064
40.2370.060+0.177 [+0.093, +0.246]0.3610.2810.064
50.1120.130−0.017 [−0.106, +0.045]0.2620.2040.064
60.0810.101−0.020 [−0.193, +0.104]0.1880.2370.063
80.2330.053+0.180 [+0.126, +0.251]0.2740.1620.064

R2b. Summary (mean over masks; SD and CV across masks).

MasksWeightingOurs mean (SD; CV)DistMult mean (SD; CV)Ours − DistMult [95% CI]Masks ours higherOurs − chance [95% CI]
3 (submission)QW0.206 (0.021; 0.10)0.097 (0.048; 0.49)+0.108 [+0.024, +0.160]3/3+0.141 [+0.108, +0.146]
3 (submission)GB0.263 (0.043; 0.16)0.205 (0.024; 0.12)+0.058 [−0.010, +0.136]3/3+0.102 [+0.035, +0.182]
7 (new)QW0.151 (0.075; 0.50)0.121 (0.071; 0.59)+0.030 [−0.059, +0.054]2/7+0.087 [+0.029, +0.124]
7 (new)GB0.240 (0.067; 0.28)0.241 (0.044; 0.18)−0.001 [−0.074, +0.079]3/7+0.084 [+0.033, +0.142]
10 (all)QW0.167 (0.068; 0.40)0.114 (0.063; 0.55)+0.053 [−0.038, +0.086]5/10+0.103 [+0.057, +0.130]
10 (all)GB0.247 (0.059; 0.24)0.230 (0.041; 0.18)+0.016 [−0.049, +0.091]6/10+0.090 [+0.039, +0.155]

R2c. Per gene (mean over the ten masks; variants of one gene share candidates and held-out targets).

GeneQueries per maskOursFixed θ (inference)DistMultOurs − DistMult
APP3720.1780.1750.153+0.025
PSEN13340.1670.1660.074+0.093
PSEN22150.1060.1100.046+0.061
APOE120.4730.4730.582−0.109
HFE80.1300.1420.216−0.086
MPO80.3130.3020.257+0.056
NOS360.9200.9201.000−0.080
CSF1R20.0000.0000.126−0.126
ADAM1010.2170.1690.292−0.074
GRN10.1560.1560.051+0.104
MAPT10.1330.1330.068+0.065
PCDH11X10.4720.4100.090+0.382
PLAU10.3790.3670.185+0.194
PLCD110.0000.0000.192−0.192
VCP10.0600.0600.125−0.065

R3. Schedule sensitivity with the energy retrained for every θ(H) = a + bH (ten masks)

b = 0 rows are constant-temperature models trained and evaluated at that temperature. (The submission's "fixed θ" ablation changed the temperature at inference only.) The mean of the paper's θ(H) over evaluation queries is 0.69. "Paper − this" is the paired recall@10 difference to (0.35, 0.90).

abrecall@10 QW (SD over masks)recall@10 GBPaper − this [95% CI]Masks paper higherρ(H, entropy)
0.200.450.148 (0.062)0.229+0.019 [−0.006, +0.043]8/100.767
0.200.900.183 (0.071)0.246−0.015 [−0.023, +0.002]3/100.733
0.201.350.183 (0.064)0.251−0.016 [−0.067, +0.007]2/100.777
0.201.800.173 (0.073)0.251−0.006 [−0.050, +0.017]5/100.755
0.350.450.157 (0.067)0.238+0.010 [−0.004, +0.018]7/100.796
0.350.900.167 (0.068)0.247paper schedule–0.736
0.351.350.166 (0.073)0.249+0.002 [−0.018, +0.009]6/100.798
0.351.800.182 (0.077)0.252−0.014 [−0.042, +0.013]4/100.801
0.500.450.151 (0.068)0.237+0.016 [−0.005, +0.032]8/100.746
0.500.900.148 (0.072)0.244+0.019 [+0.000, +0.038]8/100.737
0.501.350.147 (0.069)0.245+0.020 [+0.008, +0.028]8/100.722
0.501.800.171 (0.070)0.254−0.004 [−0.027, +0.012]5/100.620
0.200.000.108 (0.035)0.214+0.059 [+0.028, +0.070]7/100.291
0.350.000.106 (0.045)0.214+0.061 [+0.017, +0.085]8/100.190
0.500.000.109 (0.040)0.212+0.058 [+0.012, +0.078]8/100.510
0.600.000.113 (0.034)0.206+0.054 [+0.031, +0.073]9/100.461
0.800.000.124 (0.045)0.218+0.043 [+0.011, +0.069]8/100.635
1.000.000.145 (0.066)0.239+0.022 [−0.003, +0.040]8/100.575
1.250.000.152 (0.070)0.246+0.015 [−0.008, +0.033]7/100.659

R4. Component ablations on the same held-out recovery task (ten masks, identical queries)

Feature ablations zero the standardized features at inference, as in Table 1 of the submission.

Modelrecall@10 QWrecall@10 GBrecall@25Full − this, QW [95% CI]Full − this, GB [95% CI]
Full model0.1670.2470.293––
− confidence interactions (1−H features)0.1730.2540.307−0.006 [−0.010, +0.004]−0.008 [−0.021, +0.005]
− route-mechanism features0.2030.2560.326−0.036 [−0.054, +0.004]−0.009 [−0.024, +0.004]
− sequence-prediction interactions (p features)0.1400.2120.274+0.027 [+0.012, +0.040]+0.035 [−0.009, +0.087]
Fixed θ = 0.80 at inference0.1660.2390.290+0.001 [−0.003, +0.003]+0.008 [−0.000, +0.019]
Fixed θ − confidence interactions0.1790.2470.316−0.012 [−0.017, +0.004]+0.000 [−0.018, +0.019]
No energy (uniform path mass)0.1250.2150.255+0.042 [−0.075, +0.134]+0.032 [−0.032, +0.113]
Personalized PageRank0.1100.2160.170+0.057 [−0.046, +0.141]+0.031 [−0.058, +0.123]
DistMult0.1140.2300.259+0.053 [−0.032, +0.086]+0.016 [−0.052, +0.094]
Symptom popularity0.1150.2450.251+0.052 [−0.072, +0.153]+0.002 [−0.069, +0.093]
Chance (analytic)0.0640.1570.151+0.103 [+0.057, +0.129]+0.090 [+0.034, +0.155]

R5. Deliberately corrupted uncertainty at inference (energy fitted with the true H; ten masks)

Uncertainty suppliedρ(true H, entropy)Median within-gene ρρ(supplied H, entropy)recall@10 QWTrue − this, recall@10 [95% CI]
True H (paper)0.7360.8980.7360.167–
H + N(0, 0.2²), clipped0.6750.8540.7660.166+0.001 [+0.001, +0.002]
H permuted across variants0.1710.2760.6640.170−0.003 [−0.013, +0.002]
1 − H−0.582−0.8990.5820.182−0.015 [−0.028, +0.001]
H ≡ 0.50.2220.786–0.168−0.001 [−0.006, +0.006]

Two limits we report alongside. Log score (mean log₂ of the normalized candidate mass on held-out positives; higher is better): full θ(H) -10.99, fixed θ at inference -8.32, uniform path mass -7.62; the conditioned distribution is not better calibrated by this measure. Upstream signal on the 399 locked likely-label variants (never used in any fit): sequence-call error rate by H quartile Q1 low 0.09, Q2 0.05, Q3 0.19, Q4 high 0.10; AUROC of H for detecting an upstream error 0.56.

R6. Disease hold-out combined with gene hold-out

The 133 symptom-bearing diseases are split into five folds; for each fold all disease→symptom edges of the held-out diseases are removed, the energy is refitted without the query gene, and the held-out diseases' symptoms must be recovered through gene relays. 2020 query evaluations in 15 genes; 724 of 2813 held-out (gene, symptom) pairs remain reachable.

Modelrecall@10 QW [95% CI]recall@10 GB [95% CI]recall@25Full − this, QW [95% CI]Full − this, GB [95% CI]
Full θ(H)0.108 [0.080, 0.249]0.151 [0.092, 0.211]0.238––
Fixed θ at inference0.112 [0.085, 0.201]0.154 [0.097, 0.216]0.234−0.004 [−0.007, −0.002]−0.002 [−0.004, −0.001]
DistMult0.095 [0.057, 0.379]0.213 [0.131, 0.305]0.232+0.013 [−0.127, +0.032]−0.062 [−0.148, +0.019]
Uniform path mass0.072 [0.012, 0.239]0.141 [0.074, 0.218]0.162+0.037 [−0.045, +0.073]+0.011 [−0.052, +0.082]
Personalized PageRank0.053 [0.011, 0.243]0.127 [0.061, 0.204]0.135+0.056 [+0.010, +0.076]+0.024 [−0.049, +0.096]
Symptom popularity0.068 [0.015, 0.355]0.170 [0.080, 0.266]0.195+0.040 [−0.108, +0.069]−0.018 [−0.098, +0.063]
Chance (analytic)0.056 [0.044, 0.166]0.094 [0.060, 0.133]0.139+0.052 [+0.025, +0.088]+0.057 [+0.013, +0.107]

R7. Size of the admissible support versus number of gene hops

Hetionet, 18 study genes, no route cap; relay walks gene→…→gene→disease→symptom.

SupportMedian paths per geneMinMax
Submission (≤1 hop, 24 route prefixes)70413967
0 gene hops850620
exactly 1 gene hop, uncapped4,980051,832
exactly 2 gene hops604,5807,4905,275,656
exactly 3 gene hops73,043,5761,094,217957,342,572

R8. Confirmation test on 30 fresh masks (protocol fixed in advance)

The protocol was written down before any of these masks was run: mask seeds 101–130; DistMult trained with three seeds per mask and averaged; pre-declared criterion for a DistMult advantage = gene-cluster 95% interval of the query-weighted paired difference above zero. The criterion is not met. The query-weighted advantage is consistent across masks but comes from the three high-count genes. "No route-mechanism features" is the R4 ablation, selected on the ten masks of R2 and tested here once.

R8a. recall@10, mean (SD; CV) over the 30 masks.

ModelQuery-weightedGene-balanced
Submitted model θ(H)0.183 (0.055; 0.30)0.252 (0.041; 0.16)
Ablation: no route-mechanism features0.210 (0.057; 0.27)0.257 (0.044; 0.17)
Fixed θ at inference0.186 (0.053; 0.29)0.239 (0.041; 0.17)
DistMult, mean of 3 training seeds0.123 (0.048; 0.39)0.239 (0.029; 0.12)
DistMult, single training seed0.133 (0.083; 0.62)0.239 (0.031; 0.13)
Symptom popularity0.098 (0.024; 0.25)0.235 (0.027; 0.11)
Chance (analytic)0.065 (0.001; 0.02)0.158 (0.011; 0.07)

R8b. Paired comparisons. Intervals express uncertainty across genes; mask counts and Wilcoxon signed-rank p use the 30 per-mask differences.

ComparisonQW diff. [95% CI]QW masks higherQW Wilcoxon pGB diff. [95% CI]GB masks higherGB Wilcoxon p
Submitted model − DistMult+0.060 [−0.028, +0.093]24/30< 10⁻⁴+0.013 [−0.043, +0.082]19/300.18
No-mechanism ablation − DistMult+0.086 [−0.035, +0.130]28/30< 10⁻⁴+0.018 [−0.041, +0.084]19/300.07
No-mechanism ablation − submitted model+0.027 [−0.005, +0.038]24/30< 10⁻⁴+0.005 [−0.008, +0.019]19/300.11
Submitted model − chance+0.118 [+0.060, +0.141]30/30< 10⁻⁴+0.095 [+0.037, +0.167]30/30< 10⁻⁴
No-mechanism ablation − chance+0.145 [+0.087, +0.166]30/30< 10⁻⁴+0.100 [+0.040, +0.163]30/30< 10⁻⁴

R8c. Per gene (mean over the 30 masks).

GeneQueries per maskSubmitted modelNo-mechanism ablationDistMult (mean of 3)
APP3720.1910.2050.132
PSEN13340.1970.2350.094
PSEN22150.1100.1490.097
APOE120.4410.4450.497
HFE80.1040.0720.234
MPO80.3230.3050.320
NOS360.9650.9660.998
CSF1R20.0000.0000.129
ADAM1010.2310.2450.200
GRN10.1200.1880.072
MAPT10.1550.1440.054
PCDH11X10.5350.4960.166
PLAU10.4090.4050.349
PLCD110.0000.0000.136
VCP10.0290.0290.135

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.