Multi-modal contrastive drug synergy prediction model guided by single modality

Tongyong Luo, Zheng Zhang, Chen Xiangan, Zhi LiJournal of Cheminformatics 2025 · DOI

Abstract

Drug combinations offer a potent strategy against cancer, yet identifying synergistic pairs remains a significant challenge due to the vast search space and high costs of traditional screening. We present MCDSP (Multi-modal Contrastive Drug Synergy Prediction), a novel deep learning framework that leverages multi-modal contrastive learning guided by single-modality prediction tasks. By integrating drug chemical structures with biological network embeddings and aligning these disparate data types, MCDSP reduces prediction bias and achieves state-of-the-art performance on large-scale datasets. Our experiments demonstrate that MCDSP is particularly robust in predicting synergy for unseen drug combinations and cell lines, making it a valuable tool for clinical application.

Introduction

In cancer treatment, monotherapy often leads to drug resistance and high toxicity. Combining drugs can overcome these hurdles, but finding the right "synergistic" pairs—where the combined effect is greater than the sum of its parts—is like finding a needle in a haystack. Traditional wet-lab screening is too slow and expensive to test the millions of possible combinations.

While AI models have stepped in to help, most struggle with multimodal integration. They often treat different types of data (like a drug’s chemical structure vs. its position in a protein interaction network) with equal weight. This "blind fusion" can introduce noise if one data type is less relevant for a specific case, leading to biased or inaccurate predictions.

The MCDSP Framework

The authors introduce MCDSP, which processes data through four distinct modules designed to optimize the integration of biological and chemical information:

  1. Feature Extraction: The model looks at two "modalities." The first is biomolecular features (molecular fingerprints for drugs and gene expression for cell lines). The second is entity embeddings, which use a Relational Graph Convolutional Network (RGCN) to learn the "neighborhood" of drugs and proteins within a massive biological network.
  2. Highway Network Module: This acts as a sophisticated gatekeeper, regulating how information from these two modalities is transformed and combined to prevent the "vanishing gradient" problem during training.
  3. Single Modality Guided Contrastive Learning (SMGCL): This is the core innovation. The model trains individual classifiers for each modality. It then compares their performance: if one modality predicts correctly while the other fails, contrastive learning "pulls" the ineffective features toward the effective ones. This aligns the two different views of the data toward a consistent, accurate prediction.
  4. Prediction Module: Finally, the optimized features are concatenated and fed into a Multi-Layer Perceptron (MLP) to produce the final synergy score.
The MCDSP workflow, illustrating the path from raw features to the final synergy prediction.

Key Insights and Alignment Strategy

The primary strength of MCDSP lies in its ability to reduce prediction bias. Most multimodal models assume all data is equally reliable. By using single-modality performance to guide the contrastive learning process, MCDSP effectively filters noise and focuses on the most predictive components of each data type.

The alignment strategy ensures that even if one data source is weak for a specific drug-cell line pair, the model can "correct" it using the other source. This creates much clearer decision boundaries between synergistic and antagonistic samples.

An overview of the contrastive learning task, showing how features are aligned in the joint feature space.

Experimental Results

MCDSP was trained and evaluated on two major benchmarks: DrugCombDB (63,010 combinations) and OncologyScreen (4,176 combinations). It consistently outperformed baseline models like DeepSynergy and HypergraphSynergy:

  • Accuracy: On the large DrugCombDB dataset, it achieved an accuracy (ACC) of 0.7824 and an ROC-AUC of 0.8594.
  • Robustness: The model proved highly effective at predicting synergy for unseen drug combinations and cell lines, which is critical for real-world clinical applications.
  • Ablation Studies: Removing either the biomolecular features or the entity embeddings significantly degraded performance, proving that both "what" a drug is (structure) and "where" it acts (network position) are essential.

Conclusion

MCDSP provides a sophisticated approach to multimodal integration in bioinformatics. While it requires a certain amount of data to reach full potential—as seen by its slightly lower performance on very small datasets—it sets a new benchmark for drug synergy prediction on large-scale cancer datasets. Future work will focus on making the model more data-efficient for niche clinical applications.

Summary updated

Abstract

Compared to monotherapy, drug combinations exhibit stronger efficacy, fewer side effects, and lower drug resistance in cancer treatment. However, traditional wet-lab methods for screening synergistic drug combinations are both costly and inefficient. Lately, the development of various drug synergy methods has been promoted by the emergence of multiple drug synergy databases. Many of these methods use multimodal data and achieve good results. However, if various modalities of data is given equal consideration without taking into account the differences in features between the two modalities, this may lead to less effective multi-modal learning. We propose a multi-modal contrastive learning method for drug synergy prediction, named MCDSP. Specifically, MCDSP extracts entity embedding features of drugs and cell lines from heterogeneous graphs, while leveraging molecular fingerprints and gene expression features as biomolecular features for drugs and cell lines. These two different types of features serve as two types of modality information. Under the guided of single modality prediction tasks, we evaluated the relevant information of each modality. Through contrastive learning, the prediction bias of the two modalities are reduced, which obtain improved quality of multi-modal feature. Experiments show that MCDSP outperforms baseline methods on large datasets, and it performs well in handling unknown drug combinations and cell lines. MCDSP has demonstrated significant effectiveness in predicting drug synergy.
BibTeX
@article{luo2025multi,
  author = {Tongyong Luo and Zheng Zhang and Chen Xiangan and Zhi Li},
  title = {{Multi-modal contrastive drug synergy prediction model guided by single modality}},
  journal = {Journal of Cheminformatics},
  year = {2025},
  doi = {10.1186/s13321-025-01087-0},
}
Figure