HiRAND: A novel GCN semi-supervised deep learning-based framework for classification and feature selection in drug...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Predicting drug response from transcriptome data is crucial but faces challenges due to limited labeled data and the curse of dimensionality.
• Existing methods often yield poor predictions and fail to identify robust biomarkers.

Purpose of the Study:

• To develop an interpretable predictive model for drug response using limited labeled transcriptome data.
• To leverage unlabeled data and data augmentation to overcome data scarcity.

Main Methods:

• Introduced a novel Hierarchical Graph Random Neural Networks (HiRAND) framework.
• Integrated gene and sample information using graph convolutional networks (GCN).
• Employed data augmentation and consistency regularization to optimize predictions with limited labeled and abundant unlabeled data.

Main Results:

• HiRAND demonstrated superior performance compared to existing methods on simulation and multiple drug response datasets.
• Achieved the best prediction performance for the drug vorinostat among 62 tested drugs.
• Identified key genes, including ribosomal protein-related genes, crucial for vorinostat response.

Conclusions:

• HiRAND offers an efficient framework for enhancing drug response prediction accuracy, especially with limited labeled data.
• The model's interpretability aids in identifying key biomarkers for drug efficacy.
• Highlights the potential of graph neural networks and data augmentation in precision medicine.