Effective expression analysis using gene interaction matrices and convolutional neural networks

Artificial intelligence recently experienced a renaissance with the advancement of convolutional neural networks (CNNs). CNNs require spatially meaningful matrices (e.g., image data) with recurring patterns, limiting its applicability to high-throughput omics data. We present GIM, a simple, CNN-ready framework for omics data to detect both individual and network-level entities of biological importance. Using gene expression data, we show that GIM-CNNs can outperform comparable neural networks in performance and their design facilitates network-level interpretability. GIM-CNNs provide a means to discover novel disease-relevant factors beyond individual genes and their expression, factors that are likely missed by standard differential gene expression approaches.