Biological network topology features predict gene dependencies in cancer cell-lines

Motivation: Protein-protein interaction (PPI) networks have been shown to successfully predict essential proteins. However, such networks are derived generically from experiments on many thousands of different cells. Consequently, conventional PPI networks cannot capture the variation of genetic dependencies that exists across different cell types, let alone those that emerge as a result of the massive cell restructuring that occurs during carcinogenesis. Predicting cell-specific dependencies is of considerable therapeutic benefit, facilitating the use of drugs to inhibit those pro-teins on which the cancer cells have become specifically dependent. In order to go beyond the limitations of the generic PPI, we have attempted to personalise PPI networks to reflect cell-specific patterns of gene expression and mutation. By using twelve topological features of the resulting PPIs, together with matched gene dependency data from DepMap, we trained random-forest clas-sifiers (DependANT) to predict novel gene dependencies. Results: We found that DependANT improves the power of the baseline generic PPI models in predicting common gene dependencies, by up to 10.8% and is more sensitive than the baseline generic model when predicting genes on which only a small number of cell types are dependent. Availability: Software available at https://bitbucket.org/bioinformatics_lab_sussex/dependant2