acceptodds
Withdrawn from ICLR 2027

GeneSyn-HGNN: Structured Context-Conditioned Interaction Learning for Drug Synergy Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learning the effect of an entity pair under a high-dimensional context requires structured entity representations and a scorer that captures pair-context dependence. Drug synergy prediction exemplifies this setting, as the same drug pair can exhibit different effects across cellular states. Directly modeling this dependence with an unrestricted interaction over two drug embeddings and raw gene expression, however, requires a large number of decoder coefficients. We introduce the Gene-aware Drug Synergy Hypergraph Neural Network (GeneSyn-HGNN), which integrates motif-guided multi-relational hypergraph encoding and a structured cell-conditioned decoder. Its central Gene-Synergy Kernel (GSK) models a parameter-efficient interaction between the gated drug pair and the encoded cell representation. Two complementary pathways support this interaction: Cell-Conditioned Gating (CCG) modulates both drug representations according to the encoded cell state, and Gene-Selective Attention (GSA) applies a gated-pair query to expression-scaled gene keys to produce a pair-dependent additive score. We analyze the decoder's expressiveness and coefficient complexity and derive a conditional decoder-level capacity bound under fixed, sample-independent upstream feature maps. Experiments on O'Neil and NCI-ALMANAC show that GeneSyn-HGNN achieves the highest listed AUC and AUPR values and maintains strong performance under reduced training fractions. These findings support structured cell conditioning as an effective and parameter-efficient approach to modeling drug-pair-cell interactions.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Drug synergy prediction, hypergraph neural networks, motif-guided representation learning, context-conditioned decoding, gene expression, precision oncology

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_CiVXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes GeneSyn-HGNN for drug synergy prediction. It integrates a motif-guided multi-relational hypergraph encoder and a structured cell-conditioned decoder. The decoder consists of cell-conditioned gating, gene-synergy kernel, and gene-selective attention. Experiments on O’Neil and NCI-ALMANAC datasets demonstrate it has some improvements over the selected baselines.

Critical strengths

None

Critical weaknesses
  1. Unjustified problem simplification and baseline choices. The drug synergy prediction problem is simplified as a binary classification problem. Recent baselines are reviewed but not included in the performance comparison, such as TensoGraph (March 2026, Open access with code released).
  2. Lack of novelty. The proposed method is incremental and it doesn't provide sufficient value to either graph learning or drug synergy prediction research community.
Questions for authors
  1. Why do you formulate the drug synergy prediction as a binary classification problem? Is it useful in real applications?
  2. Why baselines like TensoGraph is reviewed yet omitted in the performance comparison? New researches like Cosynllm is also missing.
  3. Are there any theoretical support on the design of GSK? How do we further advance the research of drug synergy prediction and graph learning based on the findings presented in the paper?
Reviewer_kQ47OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

GeneSyn-HGNN predicts drug synergy using an MRHGNN-based hypergraph encoder with additional motif relations and a cell-conditioned decoder. The decoder combines Cell-Conditioned Gating (CCG), a diagonal drug-pair-cell interaction termed the Gene-Synergy Kernel (GSK), and Gene-Selective Attention (GSA). The paper reports improvements on O'Neil and NCI-ALMANAC under a transductive within-cell-line protocol, together with component ablations, reduced-data experiments, and theoretical analyses of fixed-feature decoder classes.

Critical strengths
  1. Consistent empirical gains from GSK. The reported GSK ablations show substantial AUC reductions on both datasets when this component is removed: 0.9692 to 0.9164 on O'Neil and 0.9232 to 0.8446 on NCI-ALMANAC (§4.3). These results provide evidence that the additional interaction branch is useful within the tested architecture.

  2. Encouraging performance with limited training data. The learning-curve experiment uses nested training subsets while keeping validation and test partitions fixed. On O'Neil, the full model trained with 40% of the training data exceeds the test AUC of the model without GSK trained on the entire training set (§4.4, Figure 5). This suggests potential value when labeled combinations are scarce under the evaluated protocol.

Critical weaknesses
  1. The motivation and methodological advance are not established. The introduction surveys existing architectures and then introduces the proposed components without identifying a concrete limitation that requires this decomposition. Existing fusion, attention, and cell-conditioned scorers already model dependence on cellular context. The parameter comparison against an unrestricted O(d²G) tensor does not establish an advantage over these practical alternatives. Within the base bilinear term and for observed cell lines, CCG's gate can be absorbed into the learned cell-specific scaling vector (§3.4, Eqs. 6–7). GSK's diagonal scoring operation is closely related to DistMult [1], although GSK generates its relation vector from encoded cell expression. GSA uses standard query-key attention (§3.6). The manuscript does not explain what new modeling principle emerges from combining these operations, or why this particular decomposition is appropriate for drug synergy.

  2. The evidence does not establish the proposed design's specific benefits. Removing GSK demonstrates the utility of an additional scoring branch, but does not distinguish expression-informed structure from additional cell-specific modeling capacity. Replacing GSK's expression-derived vector with freely learned per-cell parameters is not evaluated. A comparison with a comparably sized low-rank scorer using the same encoder is also missing. This distinction matters because evaluation uses only observed cell lines and the decoder retains cell-indexed parameters. The theoretical analyses do not resolve it: Appendix E compares CCG with a cell-independent reference, although the implemented base decoder already conditions on the cell; Proposition F.2 assumes fixed, sample-independent feature maps and does not characterize the jointly trained model or its GSK ablation. The authors acknowledge these limitations, but the resulting evidence still does not connect the claimed advantages to the proposed mechanisms.

Questions for authors
  1. What specific limitation of existing expression-conditioned scorers motivates the three-component decoder? Beyond architectural integration, what is the central new insight relative to diagonal trilinear scoring and established feature-conditioning mechanisms?
  2. Does generating GSK's interaction vector from expression offer an advantage over learning an additional free vector for each observed cell line? Were these alternatives compared with the same encoder and comparable tuning budgets?
Reviewer_bXUjOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers drug synergy prediction, where the same drug pair can be synergistic in one cell line and not in another. The key idea is that cellular context should enter the final drug-pair score through a few structured, parameter-efficient components rather than an unrestricted drug × drug × gene interaction. GeneSyn-HGNN encodes drugs with two relational hypergraphs: one built from six drug–protein motifs and one from training synergy labels. Its decoder sums a cell-scaled bilinear score inherited from MRHGNN with three added components:

Critical strengths

The model architecture is described clearly. Each component (motif hypergraph, CCG, GSK, GSA) has a well-defined role, explained with equations and an overview figure.

Critical weaknesses
  1. There are unconvincing baseline results. Each test set is half positive and half negative, so random guessing already gives an AUPR of about 0.5. A model with high AUC ranks most positives above most negatives, so its AUPR should also be high. With AUC 0.87, AUPR should be larger than 0.5. Yet HypertranSynergy reports 0.49, which is worser than random. Similarly, MHGSynergy and HypergraphSynergy report AUPR 0.67–0.68 with AUC 0.93. These same models also report ACC of 0.90–0.92, which does not fit such low AUPR.
  2. The motif and biological claims are not supported by the evidence. Changing which motifs are used moves AUC by at most 0.002. There is no run without motifs and no comparison to MRHGNN’s original four motifs, so the value of the six-motif hypergraph is unclear. The gene attributions come from one hand-picked drug pair and one checkpoint, with no check that the rankings are stable across folds or seeds.
  3. GSK is not really efficient, and the ablation doesn’t prove it matters. The paper compares GSK (131K coefficients) to a full 64M-coefficient tensor, but none of the compared methods use such a tensor. Real methods use an MLP over concatenated features, a per-cell vector as in MRHGNN, or low-rank/Tucker decoders. With only 37 cell lines in O’Neil, GSK effectively produces just 37 vectors. A per-cell vector like MRHGNN’s could learn these with about 9.5K parameters. Removing GSK makes the model worse than MRHGNN (0.916 vs 0.94), even though it still contains MRHGNN’s decoder. So the other added parts may be hurting, and GSK may just be compensating for that.
Questions for authors
  1. Can you provide an example of a prediction on a balanced dataset (1,0) and the AUC ~ 0.8 but AUPR < 0.5?
  2. Can the authors should compare GSK against a simple learned vector per cell line?

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

Related papers

Discussion (0)

Sign in to comment.