Bridging VAE-Derived Latent Gene Representations and Graph Neural Networks for Improved Drug Response Prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Pharmacogenomics seeks to personalize medicine by predicting drug response from genetic variations.
• Current methods often struggle with high-dimensional genomic data, limiting personalized treatment strategies.

Purpose of the Study:

• To develop and evaluate a novel pharmacogenomics classification framework using Graph Convolutional Networks (GCNs).
• To compare the performance of GCNs against conventional Multi-Layer Perceptrons (MLPs) for predicting drug response.
• To assess the interpretability and robustness of GCNs in leveraging gene-gene interactions for enhanced accuracy.

Main Methods:

• A GCN model was developed, incorporating gene-gene interaction networks derived from Pearson Correlation Coefficients (PCC) of gene expression data.
• Models were trained and validated on the Genomics of Drug Sensitivity in Cancer (GDSC) dataset.
• Generalizability was tested using The Cancer Genome Atlas (TCGA) data, with performance evaluated across varying input dimensions.

Main Results:

• GCNs demonstrated superior performance compared to MLPs across different input sizes, achieving a peak F1 score of 0.84 ± 0.07 at PCC > 0.7 and 8192 input features.
• MLP performance degraded with increasing input size, highlighting the robustness of GCNs to high-dimensional genomic data.
• Odds ratio analysis confirmed a strong link between GCN model accuracy and biological relevance, indicating effective capture of meaningful genomic features.

Conclusions:

• GCN-based models offer a powerful and effective framework for pharmacogenomics classification.
• Leveraging biological priors, such as gene-gene interactions, significantly enhances the accuracy and interpretability of drug response prediction.
• This approach holds promise for advancing personalized medicine through more precise genomic data analysis.