Pisces: A multi-modal data augmentation approach for drug combination synergy prediction

Area of Science:

• Computational biology
• Machine learning
• Drug discovery

Background:

• Drug combination therapy offers potential for overcoming cancer resistance and enhancing treatment efficacy.
• Predicting effective drug combinations using machine learning necessitates extensive training datasets, which are often limited.
• Sparse data presents a significant challenge in developing accurate predictive models for drug synergy.

Purpose of the Study:

• To introduce Pisces, a novel machine learning approach designed to predict drug combination synergy.
• To address the challenge of sparse training data in drug combination prediction through data augmentation.
• To improve the accuracy and applicability of machine learning models in predicting drug synergy and drug-drug interactions.

Main Methods:

• Developed Pisces, a machine learning framework utilizing data augmentation by creating multiple views of drug combinations.
• Integrated eight different drug modalities to generate 64 augmented views for each drug combination.
• Processed augmented views as separate instances to effectively handle missing modalities and sparse datasets.

Main Results:

• Achieved state-of-the-art performance in predicting drug synergy using both cell-line and xenograft models.
• Demonstrated high accuracy in predicting drug-drug interactions.
• Identified a novel breast cancer drug-sensitive pathway in BRCA cell lines by interpreting model predictions with a genetic interaction network.

Conclusions:

• Pisces effectively predicts drug synergy and drug-drug interactions through a data augmentation strategy.
• The Pisces approach circumvents the issue of missing modalities, making it versatile for various biological applications.
• The findings highlight the potential of Pisces in advancing personalized cancer treatment and drug discovery.