BIONIC: Biological Network Integration using Convolutions

Abstract

Biological networks constructed from varied data, including protein-protein interactions, gene expression data, and genetic interactions can be used to map cellular function, but each data type has individual limitations such as bias and incompleteness. Network integration promises to address these limitations by combining and automatically weighting input information to obtain a more accurate and comprehensive representation of the underlying biology. However, existing network integration methods may fail to adequately scale to the number of nodes and networks present in genome-scale data, may perform poorly, and may not handle partial network overlap. To address these issues, we developed a deep learning-based network integration algorithm that incorporates a graph convolutional network (GCN) framework to effectively learn dependencies between any input network. Our method, BIONIC (Biological Network Integration using Convolutions), learns features which contain substantially more functional information compared to existing approaches, linking genes that share diverse functional relationships, including co-complex and shared bioprocess annotation. BIONIC can integrate networks in a fully unsupervised manner if functional gene annotations are not available, and it can also leverage available annotations in a semi-supervised manner. BIONIC is scalable in both size and quantity of the input networks, making it feasible to integrate numerous networks on the scale of the human genome. To demonstrate the utility of BIONIC in identifying novel biology, we predicted essential gene chemical-genetic interactions from a small set of diagnostic non-essential gene profiles in yeast, and experimentally validated these predictions. BIONIC correctly predicted many chemical-genetic interactions, and it correctly predicted genes that are required for proper {beta}-1,6-glucan synthesis as significant interactions with the bioactive compound pseudojervine.