Deep neural profiling reveals RAP1GAP2 as a latent regulator of tumor tnvasion in Oropharyngeal Carcinoma

Abstract

Background: Conventional differential gene expression (DGE) analysis inadequately captures the complex molecular changes that drive the progression of oropharyngeal carcinoma (OC). Variational Autoencoder (VAE) offers a deep learning approach to uncover hidden patterns in high-dimensional transcriptomic data, potentially revealing novel molecular regulators.

Methods: Gene expression datasets were combined from multiple databases, and a VAE model was trained to compress the data into a low-dimensional latent space. Integrated Gradients, an automated attribution technique, was utilized to determine the contribution of each gene to each latent node (biological representation). Genes that consistently had high attribution scores across all latent dimensions were selected as potential regulators (driver genes). Pathway enrichment analysis and classification analyses unveiled the biological significance of these genes.

Results: The VAE learned latent features that are biologically important, and Integrated Gradients revealed a group of genes that have a substantial impact on these features. RAP1GAP2 was consistently one of the top contributors across all 50 latent variables, which is noteworthy. RAP1GAP2 had the highest latent-space importance and strong discriminative power for distinguishing OC, with a performance of 0.769. This occurred despite the lack of substantial differential expression in tumors relative to normal samples. Biological interpretation suggests that RAP1GAP2, a protein that activates Rap1 GTPase, may facilitate tumor invasion by deactivating Rap1 and modulating MAPK signaling and Golgi-mediated secretion.

Conclusion: Our deep learning framework identified RAP1GAP2 as a hidden driver in oropharyngeal carcinoma. This demonstrates how VAE and Integrated Gradients may discover molecular regulators overlooked by alternative approaches. This method delivers novel insights into the biology of OC tumors that could benefit future research and therapeutic approaches.

Joy Prokash Debnath

Graduate Research Assistant

My research interests include computational biology and machine learning.