acceptodds
Under review as a conference paper at ICLR 2027

TARGEN: A STRUCTURAL SCORE BEYOND MOLECULAR EVIDENCE FOR CANCER TARGET GENE PREDICTION

Abstract

Cancer-associated gene prediction requires integrating molecular measurements with biological relations. We study whether learning the predictive contributions of a candidate’s connection partners improves ranking beyond molecular and functional scores and their neighborhood propagation. We develop TARGEN, a regularized additive model that combines multi-omics data, optional pretrained GeneRAIN representations, and functional gene sets. Alongside direct scores and their propagation through protein–protein interaction and functional relations, TARGEN learns node coefficients from training labels and aggregates them into a structural score. For the linear special case with fixed relation operators and scale, we derive a stationary-point identity under L2 regularization: the structural score is a weighted sum of training residuals, with weights determined by shared-neighborhood similarity. This identity explains how training labels contribute to a candidate’s structural score. In five repetitions of ten-fold cross-validation, TARGEN achieves pan-cancer average precision (AP) of 0.921 and cancer-specific macro-average AP of 0.841. It obtains the highest mean AP among the evaluated configurations for pan-cancer prediction and in 15 of 16 cancer types. Compared with a control that fixes node coefficients to zero, the full model improves pan-cancer AP by 0.278 percentage points and cancer-specific macro-average AP by 1.663 percentage points. Each model repeats selection and fitting independently; mean AP improves in all 16 cancer types. Public patient-derived organoid data further support growth dependency on the prescreened candidate GRB2 in organoids from 34 of 49 patients with esophageal adenocarcinoma. Together, the score decomposition, linear analysis, and paired ablation explain the structural score and support its complementary predictive value in the evaluated tasks. candidate’s structural score. In five repetitions of ten-fold cross-validation, TARGEN achieves pan-cancer average precision (AP) of 0.921 and cancer-specific macro-average AP of 0.841. It obtains the highest mean AP among the evaluated configurations for pan-cancer prediction and in 15 of 16 cancer types. Compared with a control that fixes node coefficients to zero, the full model improves pancancer AP by 0.278 percentage points and cancer-specific macro-average AP by 1.663 percentage points. Each model repeats selection and fitting independently; mean AP improves in all 16 cancer types. Public patient-derived organoid data further support growth dependency on the prescreened candidate GRB2 in organoids from 34 of 49 patients with esophageal adenocarcinoma. Together, the score decomposition, linear analysis, and paired ablation explain the structural score and support its complementary predictive value in the evaluated tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
graph-based learning, pretrained representations, model interpretability, cancer gene prediction, multi-omics integration, biological networks

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

This paper proposes TARGEN for cancer-gene prioritization. It combines molecular and functional evidence, graph-based propagation, and a learned connection-partner structural score to rank candidate genes. The study evaluates pan-cancer and cancer-specific settings, reports ablations and network-rewiring analyses, derives a stationary-point interpretation of the structural component, and examines selected candidates using organoid dependency evidence.

Critical strengths
  1. The paper evaluates the added structural branch with an s = 0 ablation, label-retention controls, and degree-preserving network rewiring. Table 2 reports the highest mean AP in 15 of 16 cancer-specific tasks, which is a meaningful result within the evaluated benchmark.
  2. The cancer-specific evaluation and prescreened organoid analysis provide biological context. The authors also explicitly acknowledge transductive evaluation, unequal baseline resources, and the limited scope of the organoid follow-up.
Critical weaknesses
  1. The independent methodological novelty remains uncertain. Eqs. (9)–(10) relate the linear structural score to supervised residuals through a shared-neighborhood kernel under specified assumptions. The s = 0 ablation demonstrates a contribution within TARGEN, but does not distinguish this mechanism from resource-matched supervised graph-kernel or residual-propagation alternatives. Full-system comparisons also use different molecular and network resources.
  2. The favorable benchmark results do not yet establish generalization beyond the chosen labels and splits. Table 1 reports statistically significant gains of 0.278 percentage points in pan-cancer AP and 1.663 points in cancer-specific macro-average AP, with 14 of 16 individual cancers significant under the reported tests. However, the cancer-specific positives are DriverDB-derived computational candidates, 1,030 negative genes are shared across cancers, and evaluation is transductive. These design choices limit what can be inferred about robustness to alternative labels or unseen settings.
Questions for authors
  1. With identical molecular features, relation operators, gene coverage, and tuning budgets, how does the structural branch compare with supervised graph-kernel, label-propagation, and residual-correction alternatives?
  2. How sensitive are the results to alternative negative definitions and splits that separate related gene families or temporal evidence sources?
  3. Can the prescreened organoid candidates be compared with baseline-ranked candidates under the same selection and validation protocol?
2 comments
AuthorsOpenReview ↗

Methodological Contribution and Generalization Results

Thank you for your constructive comments on our methodological contribution and generalization evaluation. We have revised the main text accordingly. TARGEN integrates direct molecular and functional scores, their propagation, and label-trained connection-partner contributions. Full AP is 0.921 versus 0.869 for molecular-only and 0.906 for direct molecular and gene-set scoring; the isolated structural gain across 16 cancers is 1.109 percentage points (Tables I.1 and J.2). We appreciate the suggestion of a same-feature, same-graph kernel comparison, which remains a separate control beyond these reported experiments; Equation (10) concerns the fixed-operator linear setting. We have also added the CGMap comparison using shared numerical molecular inputs, labels, and folds, giving AP 0.9208 for TARGEN and 0.8580 for CGMap-CPDB (Table 3). Each method retains its graph and functional resources. DriverDB supplies target-cancer positives; reference negatives exclude positives across all 118 projects, leaving 1,030 common-coverage genes; all methods share labels and folds. NCG6 evaluation gives macro AP 0.713 and a structural gain of 1.189 percentage points (Table J.3). With disjoint families in outer testing and inner validation, macro AP is 0.8581 for TARGEN, 0.8491 for its additive ablation, 0.7652 for DISFusion, 0.7688 for deepCDG, and 0.7553 for EMOGI (Table 5). The structural gain is 0.901 percentage points, with a 95% interval of [0.532, 1.261] and one-sided p<0.05; Tables J.6 and J.7 also report class-count-matched random comparisons. Under the common organoid protocol, macro P@100 is 0.50%, 2.00%, 2.50%, and 2.50% for TARGEN, DISFusion, deepCDG, and EMOGI, respectively (Table H.6). Original-prescreening EAC replication counts are 1/4, 2/9, 3/5, and 2/4, respectively (Table H.7). Tables H.8 to H.12 retain matched backgrounds, Top200 results, all five contexts, and rank overlap, with source-data links in the reviewer guide.

Thank you for the additional experiments and clarifications. I appreciate the authors' efforts, but I remain unconvinced by the rebuttal. My main concern is still the methodological contribution. The additional ablations show some benefit from the structural component, but they do not sufficiently clarify what is fundamentally new compared with existing graph-based approaches. The reported structural gains are also relatively modest, making it difficult to judge the importance of this component.

Overall, the rebuttal provides more experimental evidence, but my main concerns about novelty and generalization remain. I will therefore keep my original rating.

Reviewer_up4SOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TARGEN, a transductive, regularized additive model for cancer-associated target-gene ranking. The method combines multi-omics features, optionally selected frozen GeneRAIN components, and MSigDB C2/C5 gene-set features. In addition to direct feature scores and their propagation through six protein–protein interaction networks and gene-set relations, it learns a coefficient for each gene and aggregates neighboring coefficients into a structural score. For a linear special case with fixed operators, the authors derive a stationarity identity expressing that score as a sum of training residuals weighted by shared-neighborhood similarity. In five repetitions of ten-fold cross-validation, TARGEN obtains pan-cancer AP 0.921 and macro-average AP 0.841 across 16 cancer-specific tasks. Relative to a refitted s = 0 control, the reported improvements are 0.278 and 1.663 AP percentage points, respectively. The study also includes reduced-evidence and graph-rewiring controls and a public organoid analysis in which GRB2 shows dependency in 34/49 esophageal adenocarcinoma patients.

Critical strengths
  1. Clear decomposition and a verifiable analytical connection. The separation of direct evidence, propagated evidence, and learned connection-partner coefficients (Section 3.3; Figure 1) makes the source of each contribution explicit. Equation (10), while restricted to the stated fixed-operator linear setting, provides a useful exact characterization of how training residuals contribute to the structural score, rather than presenting the extra graph branch as an unexplained heuristic.
  2. Substantial within-paper empirical controls. The authors use repeated, paired cross-validation; refit the s = 0 control; report a cancer-specific macro-average AP gain of 1.663 points with a 95% interval of [1.291, 2.035] (Table 1); and investigate 16 omics/label-retention conditions and 30 degree-preserving PPI rewirings (Figure 2; Appendices F and I). These experiments offer more targeted evidence for a useful network-dependent component than aggregate leaderboard scores alone.
Critical weaknesses
  1. The central contribution is not yet adequately isolated or differentiated from related supervised graph corrections. In the full model, setting s = 0 removes both the additive structural term and the dependence of the bounded nonlinear correction on s (Equations (5)–(6); Section 4.3), so the 0.278-point pan-cancer gain does not isolate Gs itself. The cleaner linear-model analysis does isolate the structural term, but its residual identity follows from the first-order condition of an L2-regularized linear model and does not establish a new property of the full nonlinear predictor. The paper discusses supervised connection learning and residual propagation (Section 2.3), yet the main comparisons do not directly establish gains over matched-feature and matched-graph kernel/residual-propagation alternatives. Given the small incremental pan-cancer improvement, this distinction is consequential to the method’s ICLR-level significance.
  2. Benchmark ascertainment and unequal comparator inputs limit interpretation of the large absolute AP numbers. For the pan-cancer labels, positive genes are drawn from curated cancer-gene evidence, whereas negatives are explicitly filtered to exclude several cancer/disease annotations (Appendix C.1). Appendix G.4 reports 27 positives among 995 genes in the lowest annotation-count tertile versus 628 among 980 in the highest, illustrating a pronounced association between annotation availability and the benchmark labels. The paper appropriately includes count-based and count-stratified controls, but these do not establish generalization to equally characterized unknown candidates or independently ascertained labels. Moreover, pan-cancer methods can differ in label coverage and omics/network inputs (Table 3; Appendix C.1), while the cancer-specific comparison uses the same labels and splits but retains method-specific modalities and network resources (Appendix D.4). Consequently, strong relative AP does not, on its own, identify the causal value of the proposed scoring mechanism. The organoid results are hypothesis-generating rather than a ranking-enrichment validation: among top-100 non-common-essential candidates measured in the organoid data, Appendix H.3 reports only 1 supported gene out of 92 measured.
Questions for authors
  1. Could the authors provide a factorial full-model ablation that separately disables (i) Gs in the additive logit and (ii) the s-dependent inputs to the bounded correction, reoptimizing each variant on identical folds? How does the isolated structural branch compare with supervised adjacency/kernel and residual-propagation baselines given the same molecular features, candidate universe, and relation operators?
  2. To what extent do the gains persist under a benchmark that balances or matches positives and negatives by annotation count, network degree, and molecular measurement availability, or under an independent label source? Appendix G.4 gives informative marginal strata, but the severe class-composition differences across these strata make a matched comparison particularly important.
  3. Could the authors disentangle unequal gene coverage and inputs from architectural differences in the pan-cancer comparison (Appendix C.1) and clarify the independent evidential value of the organoid analysis, including how the prior dependency prescreen was conducted and whether the top-ranked list shows enrichment relative to a matched background (Appendix H.1–H.4)? Answers to these points could materially change the assessment of the work’s substantive contribution.
1 comment
AuthorsOpenReview ↗

Structural Scoring and Evaluation Protocols

Thank you for your careful review and the helpful suggestion to separate the two node-dependent pathways. We have revised the main text accordingly. We used the original s=0 control to assess the combined contribution of node coefficients through additive scoring and correction; removing only Gs retains their correction pathway. Following your suggestion, we completed a four-cell ablation using identical folds and independent selection and refitting, while retaining other correction inputs (Table J.1). Adding Gs increases pan-cancer AP by 0.320 percentage points with node inputs to correction disabled and 0.159 points with them retained; the 16-cancer macro-average gain is 1.109 points (Tables J.1 and J.2). The full predictor exceeds direct molecular and gene-set scoring by 1.481 AP percentage points, and real PPI connections outperform all 30 degree-preserving rewirings (Table I.1; Figure 2). We appreciate the proposed resource-matched comparisons; graph-kernel and residual-propagation comparisons remain separate controls beyond these reported experiments, and Equation (10) concerns the fixed-operator linear setting. Cancer-specific methods share 13,627 covered genes, labels, and splits; pan-cancer coverage appears in Table C.1. TARGEN fits preprocessing on training genes, selects hyperparameters within inner validation, and rebuilds preprocessing during outer refitting. Joint matching yields 321 pairs: AP is 0.758 for TARGEN and 0.573 for the count predictor, with gains of 0.059 and 0.050 percentage points over additive-ablation and zero-node controls (Tables J.4 and J.5). The common organoid comparison gives macro P@100 of 0.50% for TARGEN, 2.00% for DISFusion, and 2.50% for deepCDG and EMOGI (Table H.6). The matched 46-target TARGEN subset has 0.00% dependency versus 0.43% in the weighted background in both CRC and EAC (Table H.8). Appendix H details the full 200-candidate, five-context analysis, prescreening, overlap exclusions, and patient aggregation.

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

The authors propose computing a structural gene score based on its neighboring genes in PPI and gene set graphs. Their overall objective consists of an additive linear function that adds the linearly transformed omics + geneRAIN features and the propagated scores (G_u and G_s) from the precalculated graph matrix. By weighting the 6 PPIs and gene sets at 50% each, they ensure the PPIs and gene sets contribute equally to the precalculated graph. When evaluated on cancer specific and pan cancer targets, they show that their model outperforms previous benchmarks. Through an ablation, they emphasize the importance of the structural score in driving better predictions. Through patient-derived organoids, they show that knocking out their predicted cancer genes caused significant tumor growth impairment, presenting strong clinical evidence.

Critical strengths
  1. The authors’ choice of using a decoupled model architecture and the careful ablation of the new structural score provides clear evidence of the score’s influence on model predictions and improves model explainability, which was presented as one of the main research gaps.
  2. The authors also presented further empirical evidence using real-world organoid specimens to show clinical applicability: knocking out the predicted cancer genes slowed cancer growth.
Critical weaknesses
  1. The lack of definitions for key terms, such as those used in computing the structural score Gs, makes the manuscript difficult to follow and evaluate carefully. For example, the authors need to clarify clearly the feature dimension of beta and theta in the computation of u. Moreover, because this structural score is the main novelty of the work, the authors should provide intuitive explanations of what a “structural score” and “propagated score” are and how they relate to gene interactions in the 7 networks.
  2. Building on the weakness above, depending on how the X features were transformed by matrix u, the resulting Gu and Gs could have very different meanings, and it remains unclear whether the computations truly transformed the matrices into the desired representations.
Questions for authors
  1. While Equation 6 clearly introduces c(v,a) as a correction term bounded by a tanh activation, the main text fails to briefly explain the intuitive purpose of this term. How exactly does this bounded correction contribute to the model's overall additive predictions?
  2. While the experiment conducted in Section 4.6 is much appreciated, these 4 candidate genes have just been tested with two cancer types, CRC and EAC. Since the TARGEN prediction results from Table 2 show good predictive scores for 15 out of 16 cancer types tested, were the 4 candidate genes also tested with those other cancer types aside from CRC and EAC? And if so, were the results comparable to the ones shown in the paper?
1 comment
AuthorsOpenReview ↗

Method Clarifications and Organoid Analysis

Thank you for your helpful suggestions on clarifying the method and organoid analysis. We have revised the main text accordingly. Section 3.3 now defines β∈ℝd, θ∈ℝM, s∈ℝN, and u=Xβ+Aθ∈ℝN. With or without GeneRAIN, d is 96 or 64 for pan-cancer prediction and 36 or 4 for cancer-specific prediction; M counts training-selected gene sets. On aligned gene axes, Gu aggregates connected genes' molecular and functional scores, while Gs aggregates their label-trained node coefficients. The correction provides bounded nonlinear logit adjustments; AP is 0.921 with correction and 0.919 with all channels disabled (Table G.3). The organoid reanalysis covers 200 candidates across five cancers and 143 patients; the four primary genes underwent earlier NextGen prescreening before Sanger outcome analysis. We aggregate published BAGEL2 calls and scaled LFCs by patient, requiring at least five measured patients, dependency in at least half, and median scaled LFC below -0.500. GRB2 in EAC meets this criterion in the four-candidate analysis, with dependency in 34/49 patients; Tables H.4 and H.5 retain all cohorts and 20 candidate-context comparisons. Cancer dependencies vary with lineage and genotype (Tsherniak et al., 2017; Behan et al., 2019; Herranz-Ors et al., 2026).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.