Dates: 2023
Principal Investigator: Dr. Omer Al-Khanbashi
Description: Cancer has been known as one of the leading causes of death worldwide with nearly 10 million deaths reported in 2020. Despite much research has been conducted in the direction of finding effective treatments for curing cancer and progress has been made during recent decades, cancer has remained one of the most difficult diseases to be completely cured. It is due to its incredibly high complexity in terms of types and mechanisms. However, it is common that the initialization and progression of each cancer are primarily involved in a set of genes, called cancer drivers. Hence, uncovering cancer drivers will not only help improve our understanding of the biological mechanism underlying cancer development and progression but also develop novel treatments. In the cancer driver prediction task, computational methods have shown promising results in discovering cancer drivers, especially once they are equipped with an effective multi-omics data integration paradigm, as highlighted in recent studies. However, the performance of the existing methods is still far below the realistic expectation due to the lack of effective solutions for data integration, context exploitation, and a small number of positive examples for model training. In this proposal, we introduce a novel context-based multi-modal graph neural network for predicting cancer driver genes that overcome the three mentioned issues. Precisely, the method is based on our assumption that a gene with a higher context dissimilarity between its characterization in healthy and cancerous contexts is more likely to be a cancer driver compared with the one with a lower dissimilarity. For such, we design the network architecture of our method with a two-module graph neural network that takes two corresponding distinct graphs constructed from healthy and cancerous tissues as input. This, on the one hand, allows different types of data to be effectively integrated through graph structures, on the other hand, enables the network to learn the representations of each gene in both contexts simultaneously, to eventually measure the context dissimilarity of the gene.