Cancer drug response prediction with surrogate modeling-based graph neural architecture search

Area of Science:

• Bioinformatics
• Computational Biology
• Personalized Medicine

Background:

• Personalized medicine aims to tailor cancer treatments by understanding individual drug responses.
• Graph neural networks (GNNs) are powerful tools for bioinformatics but require extensive manual tuning for optimal performance.
• Developing effective GNN models for drug sensitivity prediction is challenging and time-consuming.

Purpose of the Study:

• To develop an automated framework, AutoCDRP, for predicting cancer drug response using GNNs.
• To leverage surrogate modeling for efficient GNN architecture search.
• To overcome the limitations of manual GNN model design and hyperparameter tuning.

Main Methods:

• Proposed AutoCDRP, a novel framework for automated cancer drug-response prediction.
• Utilized surrogate modeling to predict and evaluate GNN architectures within a defined search space.
• Employed a systematic approach to identify the most effective GNN architecture for drug sensitivity prediction.

Main Results:

• AutoCDRP efficiently identifies optimal GNN architectures for cancer drug-response prediction.
• The GNN architecture generated by AutoCDRP demonstrated superior performance compared to state-of-the-art methods on benchmark datasets.
• The identified optimal architecture consistently outperformed baseline models from the initial training epoch.

Conclusions:

• AutoCDRP automates the selection of optimal GNN architectures, accelerating the development of personalized cancer therapies.
• The framework significantly enhances the accuracy and efficiency of cancer drug-response prediction.
• AutoCDRP offers a promising solution for advancing personalized medicine in oncology.