acceptodds
Withdrawn from ICLR 2027

Drug-Target-Disease Embeddings Improve the Accuracy of Drug Repurposing Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Drug repurposing offers a promising strategy to accelerate therapeutic development by identifying new indications for approved drugs. Recent work in this area has made extensive use of biomedical knowledge graphs (KGs). However, most computational models built on KGs predict drug-disease associations without explicitly incorporating information about molecular targets of drug molecules. We present a target-aware drug repurposing framework that leverages drug-target-disease (DTD’) triplet embeddings to explicitly capture target information. We generated 512-dimensional Node2Vec embeddings for DTD’ combinations labeled as either “positives” (fully connected DTD’ triplets in ROBOKOP KG) or “negatives” (random DTD’ combinations with no known DD’, i.e., “treats”, connections) and built multiple binary classification machine learning models. Multilayer Perceptron (MLP) models using DTD’ triplet embeddings outperformed matched drug-disease (DD’) doublet models across all classification and ranking metrics such as precision (0.61 vs 0.41, respectively), recall (0.76 vs 0.35), F1-score (0.68 vs 0.38) and enrichment factor (EF1%: 27.54 vs 15.56). These models also outperformed both XGBoost and Random Forest baselines, and Node2Vec embeddings yielded superior performance compared to embeddings from pretrained large language models such as PubMedBERT, and BioBERT. Our results confirm that explicit target information does improve the predictive accuracy of drug repurposing models built with KGs. We employed our models to generate novel repurposing hypotheses for approximately 22,000 human diseases annotated in the Mondo Disease Ontology. We present case studies of top-ranked drug repurposing candidates nominated by our approach for two neurodegenerative diseases (ALS and Huntington’s) supported by evidence in PubMed and ClinicalTrials.gov.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Drug repurposing, FDA-approved drugs, ROBOKOP Knowledge Graph, drug-target-disease triplets, machine learning, large language models (LLMs)

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_jQb9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a target-aware drug repurposing framework based on drug-target-disease triplets. Drug, target, and disease nodes are embedded using Node2Vec on the ROBOKOP knowledge graph, and their embeddings are concatenated and fed to an MLP classifier. The authors compare triplet-based models with drug-disease doublet models, several conventional classifiers, and text-based embeddings. They report improved classification and ranking performance when target information is included, and further use the trained model to generate drug repurposing hypotheses for ALS and Huntington’s disease.

Critical strengths

S1. Explicitly incorporating molecular targets into drug repurposing is well motivated and provides a more mechanistic formulation than direct drug-disease prediction.

S2. The paper evaluates multiple classifiers and embedding types and includes ranking-based metrics and case studies for downstream drug repurposing hypotheses.

Critical weaknesses

W1. Positive and negative triplets differ structurally by construction. Positive samples form complete drug-target-disease motifs, while most negatives are created by random entity replacement. The model may therefore exploit local KG connectivity rather than learn a generalizable therapeutic relationship.

W2. The evaluation is mainly transductive. Drug-disease pairs are split, but drugs, diseases, and targets can still appear in both training and test sets, with Node2Vec embeddings learned from the same graph. Generalization to unseen entities is therefore unclear.

W3. The baseline comparison is limited for a KG-based prediction study. Standard knowledge-graph embedding and heterogeneous graph methods, such as TransE, RotatE, ComplEx, R-GCN, or HGT, are not included.

Questions for authors

Q1. How does the method perform when drugs, diseases, or targets in the test set are completely unseen during embedding and classifier training?

Q2. Could the authors compare against standard KG link-prediction methods such as TransE, RotatE, ComplEx, or heterogeneous GNN baselines under the same split?

Q3. Can the authors provide a matched DTD′ versus DD′ experiment with identical training pairs, optimization budget, and classifier capacity, differing only in whether target information is included?

Q4. How much performance can be obtained using a simple graph heuristic, such as whether the drug and disease share a target or the number of common target paths? This would help quantify how much the learned model adds beyond local KG topology.

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

This paper introduces a target-aware framework to predict if a drug might treat a disease via a specific target. They basically pulled drug-target-disease triplets out of the ROBOKOP knowledge graph and supplemented the drug info with MeDIC.

Critical strengths
  1. the paper model drug–target–disease relationships yields a clear empirical improvement. The problem seesm to be interesting and may be help in AI-driven drug discovery.

  2. The paper also consider some LLM feature like node2vector to improve the performance

Critical weaknesses
  1. The topic is not new, drug-target-disease is a old problem in bioinformatics, like, and some related work is missing

Chen, H. and Li, J., 2019, May. Modeling relational drug-target-disease interactions via tensor factorization with multiple web sources. In The World Wide Web Conference.

It seems that the paper just use text emebeding + MLP as method.

  1. The entire paper look like a homework project, the abstract is hard to follow. The proposed algorithm consists primarily of standard Node2Vec embeddings concatenated and passed to a conventional multilayer perceptron. Its other components, including focal loss, Bayesian hyperparameter optimization, and fold ensembling, are also well known techniques. I don't see any new idea or deep inight about the drug-target-disase.

  2. The baselines are too weak, like XGB or random forest. The reported case studies offer literature-based rationales, partly filtered by LLMs whose reliability is not quantitatively evaluated, rather than independent confirmation of novel predictions.

Questions for authors

1 The entire paper look like a homework project, the abstract is hard to follow. The proposed algorithm just combine some existing technique. what is the new insight here?

  1. The baselines are too weak, like XGB or random forest. The reported case studies offer literature-based rationales, partly filtered by LLMs whose reliability is not quantitatively evaluated, rather than independent confirmation of novel predictions.

  2. In Table 4, it seems that the proposed method has almost same performance to XGBoost, then why need so many llm-based emebding, or deep learning. It is not very convicing.

Reviewer_YUKrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies drug repurposing by addressing the limited incorporation of molecular target information in existing knowledge graph (KG)-based drug–disease association prediction models. The paper proposes a target-aware framework that concatenates Node2Vec embeddings of drugs, targets, and diseases extracted from the ROBOKOP KG and employs a multilayer perceptron (MLP) to predict therapeutic relationships. To demonstrate the contribution of target information, the paper compares drug–target–disease (DTD') triplet models with drug–disease (DD') doublet models using an 80/20 drug–disease pair-level train/test split, achieving improved AUROC (0.98 vs. 0.91), AUC-PR (0.75 vs. 0.33), and EF@1% (27.54 vs. 15.56). Further evaluate different classifiers and embedding methods, showing that the Node2Vec-based MLP outperforms XGBoost, Random Forest, and biomedical language model embedding baselines on most reported metrics. The framework is applied to approximately 194 billion drug–target–disease combinations, with LLM-assisted literature validation used to present potential repurposing candidates for ALS and Huntington's disease.

Critical strengths

This study attempt to directly incorporate drug-target-disease information into model training, which is a meaningful approach to drug repurposing prediction. The comparison between DD' and DTD' models provides experimental results that including target information can improve prediction performance across multiple evaluation metrics. The authors also evaluated different classifiers and embedding methods, which helps demonstrate the effectiveness of their proposed framework.

Critical weaknesses

This paper’s main claim is that target information is useful for drug repurposing prediction, supported by experimental results comparing DD' and DTD' embeddings. However, this has already been extensively investigated in previous studies[1-4], making the findings not particularly informative. The study follow a conventional negative sampling approach without carefully designing the negative selection strategy from a biological perspective. Therefore, the methodological novelty is further limited.

I also have concerns about the negative sampling strategy. It would be helpful to compare whether using contraindications as negative samples is effective and how it differs from using only randomly generated negatives. However, it is also necessary to consider whether contraindications can actually be regarded as biologically opposite to indications.

[References]

  • [1] Bang, Dongmin, et al. "Biomedical knowledge graph learning for drug repurposing by extending guilt-by-association to multiple layers." Nature Communications 14.1 (2023): 3570.
  • [2] Dong, Xialan, and Weifan Zheng. "Emerging technologies for drug repurposing: Harnessing the potential of text and graph embedding approaches." Artificial Intelligence Chemistry 2.1 (2024): 100060.
  • [3] Pushpakom, Sudeep, et al. "Drug repurposing: progress, challenges and recommendations." Nature reviews Drug discovery 18.1 (2019): 41-58.
  • [4] Luo, Yunan, et al. "A network integration approach for drug-target interaction prediction and computational drug repositioning from heterogeneous information." Nature communications 8.1 (2017): 573.
Questions for authors
  1. Previous studies have already extensively investigated how incorporating target information can improve drug repurposing prediction. I would like to see an analysis of the effectiveness of negative triplet selection from a biological perspective, such as ATC classes or side effects for drugs and pathways for targets.
  2. A large number of positive triplets can be generated, and many of them may contain false positives. I wonder how the authors controlled for these potential false positives.
  3. Do indication and contraindication drugs for the same disease actually represent biologically opposite relationships during training? For example, indication drugs may have disease-specific therapeutic effects, while contraindicated drugs may be more general-purpose drugs, such as steroids. In such cases, it seems difficult to assume that the same logic applies consistently across all diseases.
  4. Was the initial 80/20 drug–disease pair-level split performed randomly, and was class stratification considered? The authors mention stratification in the cross-validation procedure, but it is unclear whether this was also applied to the initial train/test split. Was the dataset in Table 1 randomly split? Was stratification considered?

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

Related papers

Discussion (0)

Sign in to comment.