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

Pharmacogenomics aims to predict drug response based on genetic variations, facilitating personalized medicine. In this study, we develop a pharmacogenomics classification framework using Graph Convolutional Networks (GCNs) and compare its performance with conventional multi-layer perceptrons (MLPs). The GCN model integrates gene-gene interactions to enhance classification accuracy and interpretability.We train and validate our models using the Genomics of Drug Sensitivity in Cancer (GDSC) dataset and evaluate their generalizability using The Cancer Genome Atlas (TCGA) data. In the GCN model, graph edges are constructed based on the Pearson correlation coefficient (PCC) of gene expression values, allowing biologically relevant relationships to be incorporated into the learning process. Our results show that GCNs outperform MLPs across different input sizes, with the highest F1 score of 0.84 ± 0.07 achieved at PCC > 0.7 with an input size of 8192. In contrast, MLP performance declined as input size increased, indicating that GCNs are more robust to high-dimensional data. Furthermore, odds ratio analysis reveals a strong correlation between model accuracy and biological relevance, suggesting that GCNs capture meaningful genomic features.In conclusion, GCN-based models provide an effective approach for pharmacogenomics classification by leveraging biological priors.