Predicting Clinical Anticancer Drug Response of Patients by Using Domain Alignment and Prototypical Learning.

Area of Science:

• Computational biology
• Genomics
• Pharmacogenomics

Background:

• Accurate anticancer drug response prediction is essential for personalized cancer treatment.
• Existing models struggle due to scarce patient data and differing distributions between cell line and patient datasets.
• Current transfer learning methods are limited by cell line data anomalies and fail to leverage unlabeled patient data.

Purpose of the Study:

• To develop a robust model, DAPL, for predicting patient anticancer drug response.
• To address the challenges of data scarcity and domain shift in anticancer drug response prediction.

Main Methods:

• DAPL extracts domain-invariant features from cell line (CCLE, GDSC) and patient (TCGA, PDTC) data using multiple Variational Autoencoders (VAEs).
• Drug features are extracted using Graph Neural Networks (GNNs).
• Prototypical learning combines these features to train a classifier for improved patient response prediction.

Main Results:

• The DAPL model demonstrated superior performance in predicting patient anticancer drug response.
• DAPL outperformed existing state-of-the-art methods in cross-domain prediction tasks.
• The model effectively handles domain shift and data anomalies.

Conclusions:

• DAPL offers a promising approach for accurate patient anticancer drug response prediction.
• The method enhances personalized medicine by enabling more effective treatment selection.
• Leveraging domain-invariant features and prototypical learning improves model robustness and accuracy.