acceptodds
Withdrawn from ICLR 2027

Recovery over Agreement: Cross-View Masked Graph Learning for Drug Repositioning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Computational drug repositioning seeks new indications for existing drugs, yet drug–disease association (DDA) prediction is hampered by sparse labels and by evidence scattered across heterogeneous entity attributes and auxiliary relations. Existing multi-view methods encode attribute- and relation-derived graphs independently and align them only through agreement objectives, which reward redundancy rather than complementarity; they typically leave supervised edges visible to relation message passing, and multi-task variants force DDA, drug–protein, and disease–protein prediction onto a single fused representation. We propose Cross-view Attribute–Relation Exchange (CARE), which replaces cross-view agreement with cross-view reconstruction. Under a single node- and edge-masking process applied only to the relation encoder, the relation view reconstructs masked entity representations. In contrast, the attribute view ranks masked associations above type-compatible unknown pairs, so each view must supply what the other withholds and target edges never enter relation message passing. A type-aware contrastive objective keeps entity semantics consistent across views, and task-conditioned gates let drug–protein and disease–protein supervision strengthen the shared encoders without imposing a common fusion on the target task. On three benchmarks, CARE consistently outperforms existing methods, with the largest gains in recall, indicating that our approach helps recover associations that sparse relational evidence alone misses and showing the benefit of coupling multi-task learning with reciprocal cross-view reconstruction for sparse DDA prediction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
computational drug repositioning, drug–disease association prediction, multi-view graph learning, heterogeneous biomedical networks, cross-view reciprocal supervision, contrastive alignment, multi-task learning, task-conditioned fusion

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_j5yYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies drug–disease association (DDA) prediction, addressing the limitations of existing multi-view methods that primarily align attribute and relational representations without explicitly leveraging their complementary information. To address this problem, proposing CARE, a cross-view masked graph learning framework that employs reciprocal reconstruction between attribute and relation views, type-aware contrastive alignment, and task-conditioned multi-task learning with drug–protein (DPI) and disease–protein (DPA) associations. Experiments on three benchmark datasets using 10-fold cross-validation demonstrate that CARE consistently outperforms 12 baseline methods across six evaluation metrics. Ablation studies further show the contributions of reciprocal reconstruction, contrastive alignment, and auxiliary relation supervision, while case studies provide supporting clinical and literature evidence for top-ranked drug candidates.

Critical strengths

The paper clearly explained the methodological strengths of this study through both experiments and logical reasoning. The necessary experiments are also well demonstrated using benchmark datasets.

Critical weaknesses

The performance scores on these benchmarks are already close to 100%, raising concerns that the models may be competing on already saturated benchmarks. While demonstrating the effectiveness of the proposed method using benchmark datasets is reasonable, additional comparisons on more recent external benchmarks or newly constructed experimental settings seem necessary.

Questions for authors
  1. In general, the number of proteins is much larger than that of drugs and diseases. Even in datasets such as MSI and Hetionet, dense PPI information can make it difficult to identify meaningful drug–protein–disease relationships. Although the proposed method achieves nearly perfect scores on the current benchmark datasets, it is unclear whether similar performance can be achieved on external datasets or larger KGs with highly imbalanced degree distributions. The slight performance improvements on the benchmarks are acknowledged, but whether the proposed methodology remains effective on more complex, heterogeneous biomedical KGs is still questionable.
  2. Is there a case study showing how the two views actually complement each other's missing information? Typically, cross-attention or embedding fusion can enhance shared information, while this study emphasizes that the two views can recover complementary information from each other. An embedding-level case study demonstrating this mutual recovery would better support the main claim.
  3. The code could not be found in the supplementary materials. Providing the code would be helpful for reproducibility. And main architecture figure seems to illustrate the framework somewhat differently from the methodological description. A clearer illustration would help readers better understand the proposed approach.
Reviewer_U9XSOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets a drug repositioning problem (drug-disease association prediction). The authors argue that existing methods have the following three issues: first, rewarding redundancy: the two views are trained to agree rather than to complement each other; second, visible training labels: the associations the model is trained to predict remain in the relation graph during training, so the model can partly read its targets instead of inferring them; and third, one fusion for all tasks: the attribute and relation views are combined in the same way for each task but each task may need a different mix.

To solve these issues, they propose CARE which introduces the following: cross-view recovery: each view must recover information hidden from the other, which according to the authors, makes the two views more complementary; masking: the training target edges are hidden from the relation graph, so the model cannot see the associations it is being trained to predict; and a multi-task strategy: relation-specific gates decides how much attribute and relation evidence each task uses.

To validate CARE, they use 10 folds cross validation to compare it against 12 methods on 3 datasets, reporting six metrics. CARE ranks first on every metric and dataset. They also present ablations that remove each reconstruction loss, the contrastive alignment, and the auxiliary relations and tasks. Finally, they present a case study on breast cancer (with lung cancer and Alzheimer’s disease in the appendix), in which they report that the top-ranked candidates are supported by clinical or preclinical evidence.

Critical strengths
  1. The paper proposes an interesting way where the two views (attribute and relation) to interact before fusing them: each view must recover information hidden from the other. This cross-view reconstruction, together with the use of an attribute graph and a relation graph as the two views, could be extended to other representation learnings settings with multiple views.

  2. The training design avoid label leakage by construction: the links the model is trained to predict are always removed from the relational graph.

Critical weaknesses
  1. The work’s motivation relies on three issues (rewarding redundancy, visible training labels, and one fusion for all tasks), but they are presented only briefly and not fully tested; therefore, the main thesis is not well supported by the experiments. For the first issue, the ablations show that reconstruction improved the performance but not why: the claim that reconstruction makes the two views complementary is not measured. For the second issue, no ablation puts the target edges back into the relation graph, so the gaining of hiding them is not measured. For the third one, there is no experiment showing the benefit of the gates, the ablations only showed that adding DPA and DPI tasks helps. Also, prior work was not supported by specific examples. For example, which existing methods leave the training labels in the relation graph.
Questions for authors
  1. What are the exact features for drugs, diseases and proteins, and from which sources and which versions were they extracted? The introduction mentions “chemical fingerprints, disease phenotypes, protein sequences, and observed associations”, but this describes the field in general, not CARE’s data. The paper presents everolimus as a temporal validation. CARE ranks everolimus as first candidate for breast cancer and it was approved after the 2011 training labels. DrugBank 6.0 (2024) is cited in line 326, but it does not say which version the features and auxiliary associations come from. Would you please clarify which versions were used?

  2. How were the evidence tiers in Table 9 assigned? Which sources were consulted, who assigned the tiers? Did the people who assigned them know which model produced each drug?

  3. How are negative pairs selected at test time?

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

This paper proposes CARE for drug–disease association (DDA) prediction. It combines an adaptive attribute graph with a typed relation graph over drugs, diseases, and proteins. Node and edge masking affect only the relation encoder: the relation branch reconstructs masked nodes’ initial attribute embeddings, while the attribute branch ranks masked training associations above sampled unknown pairs. Type-aware contrastive alignment and task-conditioned gates support joint DDA, drug–protein interaction (DPI), and disease–protein association (DPA) prediction. Evaluation includes three DDA benchmarks, comparisons with 12 baselines, ablations, sensitivity analyses, and retrospective case studies.

Critical strengths

(1) The figures clearly illustrate the model architecture, and the tables present the comparisons and ablation results in an organized, readable way. The experiments cover three datasets and 12 baselines, complemented by ablation studies, sensitivity analyses, and retrospective case studies, providing useful empirical evidence within the tested setting. (2) The masking protocol is explicitly described. Both directions of each masked positive edge are removed from relation message passing, and the attribute-only link decoder provides an inspectable separation between that prediction branch and the relation graph. (3) The auxiliary ablations distinguish the use of DPI/DPA relations from explicit auxiliary prediction losses. Shared hyperparameters across datasets and the sensitivity analyses also provide useful evidence about the tested configuration.

Critical weaknesses

The CARE framework proposed by the authors is merely a concatenation of existing methods such as masked reconstruction and contrastive learning, with no substantive innovation. The ablation study simply removes the reconstruction loss without controlling variables, so it cannot show whether the performance gain comes from the "cross-view mutual reconstruction" design itself or merely from adding an extra loss term. Redundant alignment, training-objective visibility, and task interference are all common problems in representation learning. The paper does not clearly describe what role CARE plays in drug-disease association (DDA) prediction, nor does it directly verify whether CARE helps sparse entities or reduces negative transfer. The experiments also cover only three transductive DDA benchmarks. A retrospective case study can neither establish cold-start generalization nor demonstrate that the method transfers to other biomedical link prediction tasks. Recall and F1 show clear improvements, but without a clear explanation of how the threshold is chosen, and without uncertainty estimates based on paired comparisons, it is hard to judge how meaningful these gains are. The paper needs to show whether, when all methods are evaluated under the same protocol, these gains actually improve the prioritization of candidates. The paper reports no results on runtime, memory usage, or scalability. Exact attribute-graph construction and full type-wise contrastive learning involve quadratic all-pairs computation, so their cost should be considered separately from sparse-graph message passing. Without a comparison against strong baselines in training time and peak memory usage, it is difficult to judge whether the limited ranking improvements are worth the additional computational overhead.

Questions for authors

(1) What is the precise methodological advance over GraphMAE/HGMAE-style reconstruction, DRMAHGC, and MTP-DDA? Can you include DRMAHGC and controls that retain comparable node/link supervision, encoder capacity, inputs, and tuning budgets? (2) Does reciprocal recovery still help against those controls? What evidence shows increased complementarity? Please separately test supervised-training-edge exclusion and task-specific versus shared gates, always keeping validation/test edges excluded. (3) Can you provide fold/seed-level paired differences against DIAM, explain whether its results were rerun or quoted, and document evaluation pairs, thresholds, and the unit used to compute standard deviations? (4) Do the gains persist for low-degree entities, held-out drugs/diseases, and ranking over the full unknown-pair space? What evidence supports the claimed transfer beyond DDA? (5) What exact attributes and auxiliary-relation snapshots are used? Are association-derived inputs reconstructed within each fold, and do all case-study inputs predate the claimed temporal cutoff? (6) What are training time, inference time, peak memory, and parameter count relative to the strongest baselines, and how do affinity construction and contrastive learning scale with graph size?

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

Related papers

Discussion (0)

Sign in to comment.