AI Breakthrough in Cancer Therapy: Pisces Predicts Effective Drug Combinations with Unmatched Precision

In a pioneering leap for cancer treatment and precision medicine, researchers at the University of Washington have developed Pisces. This multi-modal, AI-based model predicts synergistic drug combinations with unprecedented accuracy.

Pisces overcomes a key limitation in oncology: the limited experimental data for the countless possible drug combinations. By augmenting each drug pair up to 64 times using eight distinct data types, Pisces trains smarter, faster, and with more accuracy than previous models.

“We’re not just predicting drug synergies; we’re identifying clinically relevant cancer pathways and optimizing treatments,” said Dr. Sheng Wang, the study’s senior author.

How Pisces Works

Pisces integrates 8 diverse data modalities per drug:

SMILES (Simplified Molecular Input Line Entry System) sequences
Molecular graphs
3D drug structures
Drug targets
Textual drug descriptions
Side effects
Drug response data
Drug ontology

By creating multiple pairwise combinations, Pisces simulates a vast and diverse training dataset, treating each pair of drug features as a separate instance, without the need for complete datasets across all drugs.

Key Achievements

64x dataset expansion via data augmentation
Outperformed all existing models on synergy prediction across three settings:
- Vanilla cross-validation
- Unseen drug combinations
- Unseen cancer cell lines
Identified a breast cancer drug-sensitive pathway validated through survival data
Successfully extended to 3-drug predictions trained only on 2-drug data
Accurate tumor volume predictions using patient-derived xenografts (Pearson r = 0.47)
Reliable drug-drug interaction (DDI) prediction, including novel, unrecorded interactions

Clinical Impact

Pisces not only predicts synergy, but it also reveals clinically actionable targets. By integrating prediction results with a genetic interaction network, Pisces uncovered a 19-gene pathway involved in breast cancer drug sensitivity. This pathway correlated with significantly longer survival in ER+ patients, offering a potential biomarker for prognosis and personalized treatment planning.

Real-World Use Cases

Personalized medicine: Predict patient-specific drug combinations
Drug repurposing: Find new use cases for existing drugs
Precision oncology trials: Prioritize combinations for testing
Adverse drug interaction avoidance: Enhance safety profiles

Future Applications

The team aims to extend Pisces with:

Self-supervised learning
Interpretation tools for clinical transparency
Integration of 3-drug+ combinations
Multi-task learning across synergy, toxicity, and side effects

“Pisces is more than a prediction engine—it’s a discovery platform,” said co-author Dr. Hoifung Poon of Microsoft Research.

Reference

Xu, H., Lin, J., Woicik, A., Liu, Z., Ma, J., Zhang, S., … & Wang, S. (2022). Pisces: A multi-modal data augmentation approach for drug combination synergy prediction. Cell Genomics. DOI: 10.1016/j.xgen.2025.100892