Associate Professor  Qiang  He
Northeastern University,  China

Title: Driver Gene-Based Synergistic Prediction of Drug Combinations by Domain Adaptive Meta-Learning

Abstract:

Predicting the synergistic effects of drug combinations for individual patients is essential for advancing precision clinical cancer therapy. Traditional approaches, which often rely on cell line data, may fail to adequately capture the complex genetic variability of cancer patients and the diversity among cancer types. To address this limitation, we develop a novel Domain Adaptive Meta-Learning (DAmeta) method that integrates domain adaptive networks with meta-learning to enhance the personalized prediction of drug combination synergy across diverse cancer types. Our method leverages the heterogeneity of driver genes to learn specific parameters for each cell line task through a meta-learning network. Furthermore, it aligns the distributional discrepancies between cell line and patient data using a domain adaptive network, thereby enhancing the accuracy of predictions across different biological conditions. Specifically, we initially employ a graph neural network to identify driver genes, then utilize a dynamic weight generation network to swiftly update the task embedding using a few labeled drug synergy samples. Furthermore, the incorporation of a domain obfuscator iteratively aligns the feature distributions between cell line and patient data, enhancing the translational applicability of our model from experimental settings to clinical environments. Our findings indicate that DAmeta outperforms contemporary state-of-the-art approaches, demonstrating superior efficacy in both patient-derived and cell-line data scenarios. Through a feature attribution analysis, we further highlight the pivotal role of driver genes in enhancing the prediction accuracy of drug combination synergy.

Biography: