Abstract
Drug combination therapy is promising for cancer treatment through simultaneously reducing resistance and improving efficacy. Machine learning approaches to drug combination response prediction can prioritize experiments and discover new combinations, but require lots of training data in order to fit the nonlinearity of synergistic effect. Here, we propose Pisces, a novel machine learning approach for drug combination synergy prediction. The key idea of Pisces is to augment the sparse drug combination dataset by creating multiple views for each drug combination based on its different modalities. We combined eight different modalities of a single drug to create 64 augmented views for a pair of drugs, effectively expanding the size of the original data 64 times. Pisces obtained state-of-the-art results on cell-line-based drug synergy prediction, xenograft-based drug synergy prediction, and drug-drug interaction prediction. By interpreting Pisces’s predictions using a genetic interaction network, we further identified a breast cancer drug-sensitive pathway from BRCA cell lines in GDSC. We validated this pathway on an independent TCGA-BRCA tumor dataset and found that patients with this pathway activated had substantially longer survival time. Collectively, Pisces effectively predicts drug synergy and drug-drug interactions through augmenting the original dataset 64 times, and can be broadly applied to various biological applications that involve a pair of drugs.
Introduction
In oncology, drug combination therapy is a cornerstone of treatment, aimed at reducing drug resistance and improving therapeutic efficacy. "Synergy" occurs when the combined effect of drugs is greater than the sum of their individual effects. However, the vast space of potential drug combinations makes exhaustive laboratory testing prohibitively expensive. Computational models offer a scalable alternative to prioritize candidates for experimental validation.
Representing drugs digitally presents a significant challenge. Two primary modalities are commonly used: SMILES strings (text-based sequences) and molecular graphs (mathematical maps of atoms and bonds). While SMILES are well-suited for sequence-based models like Transformers, they often struggle with 3D spatial relationships. Conversely, molecular graphs capture structural topology but can be difficult to embed effectively for large or complex molecules. Existing models typically rely on a single modality or simple concatenation, often failing to capture the full chemical essence of the drugs.
The Pisces Framework
Pisces addresses these limitations through a novel multi-modal data augmentation approach. The core innovation is the use of cross-modal contrastive learning to integrate SMILES and molecular graph views.
For every drug pair, Pisces generates four "augmented views" by combining different modalities: (SMILES, SMILES), (Graph, Graph), (SMILES, Graph), and (Graph, SMILES). The model employs a Transformer for SMILES sequences and a DeeperGCN for molecular graphs. Using an InfoNCE contrastive loss, the framework forces these different views of the same drug pair to converge in a shared embedding space while pushing apart views of different pairs. This process enables the model to learn high-quality, modality-invariant representations. To incorporate biological context, Pisces also embeds cancer cell line data by analyzing gene expression within a protein-protein interaction (PPI) network.
Experimental Results and Benchmarking
The authors evaluated Pisces using the GDSC-Combo dataset, which includes over 100,000 drug combinations across 125 cancer cell lines. This dataset is notably imbalanced, with only 5% of combinations exhibiting synergy, providing a rigorous benchmark for predictive accuracy.
As shown in Figure 3, Pisces consistently outperformed existing state-of-the-art methods such as DeepDDS and GraphSynergy. In the "vanilla" cross-validation setting, Pisces achieved a KAPPA score 21.05% higher than the nearest competitor. Crucially, the model maintained superior performance in stratified settings—where it predicted synergy for entirely unseen drug combinations or cell lines—demonstrating its ability to generalize beyond simple pattern memorization.
Discussion and Conclusion
The success of Pisces highlights the complementary nature of chemical modalities. By leveraging both SMILES and graphs, the model overcomes the "large molecule" problem where graph-based encoders often struggle with long-range dependencies. Ablation studies confirmed that the contrastive learning component is vital; removing it led to a significant drop in performance.
While Pisces represents a major advancement, future work could expand the framework to include 3D molecular structures or textual clinical descriptions. Additionally, transitioning from in vitro cell line predictions to in vivo clinical outcomes remains a critical frontier. Ultimately, Pisces provides a robust, interpretable framework for drug synergy prediction that can be broadly applied to various biological tasks involving drug pairs.
Summary updated