Leveraging Variational Autoencoders to Identify Genes Involved with Cancer Metastasis

Identifying genetic drivers of cancer metastasis within high-dimensional gene expression data is challenging. This study uses a variational autoencoder (VAE) to encode complex gene expression patterns into a low-dimensional space, capturing intricate relationships often missed by traditional methods. Trained on pan-cancer data, the VAE, combined with a constrained optimization framework and co-occurrence network analysis, accurately identified critical genes such as ADAR, GLI1, and METTL3, and correlated gene pairs such as ERG and YAP1. These findings a lign with $\mathbf{c}$ urrent 1 iterature, showcasing the method’s potential. This approach offers candidate genes for further experimental analysis, paving the way for targeted treatment strategies.