Graph Convolutional Neural Networks Applications to Cancer Cellular Network Analysis
Nora Kimbrough-Perry1, Konstantinos Krampis1,2
1Mathematics and Statistics – Hunter College
2Biological Sciences – Hunter College, Weill Cornell Medicine
Graph neural networks (GNNs) and graph convolutional neural networks (GCNNs) have the potential to solve and improve solutions to problems in natural language processing, computational chemistry, the computational biology of cancer research, and other fields. A graph is a mathematical object consisting of nodes connected by edges. This type of structure can be used to model, among other things, networks of cells—including protein-protein interactions, gene-protein interactions, and epigenetic interactions. Once nodes are represented in an embedding space, GNNs can begin to make predictions about graphs. Graph-level tasks may be performed; these include graph classification and various types of graph comparison. GCNs and GCNNs may also be used to predict the position of a node in a graph, the label of a node, relationships between nodes, and whole graph labels. GCNs employ key characteristics of ordinary convolutional neural networks such as the creation of feature maps and graph pooling, as well as other graph-specific characteristics. This work draws on data from The Cancer Genome Atlas to expand the scope of what machine learning and GNNs can address, specifically graphs pertaining to the computational biology of cancer and gene networks.
