Challenges and opportunities for oncology drug repurposing informed by synthetic lethality

Data supporting: Challenges and opportunities for oncology drug repurposing informed by synthetic lethality Michael Vermeulen1, Andrew W. Craig3,4, Tomas Babak1,2 1. Department of Biology, Queen’s University, Kingston, Ontario, Canada2. Leapfrog Bio, San Mateo, CA, USA3. Department of Biomedical and Molecular Sciences, Queen’s University, Kingston, Ontario, Canada4. Cancer Biology and Genetics Division, Sinclair Cancer Research Institute, Kingston, Ontario, Canada DescriptionThis repository contains the full dataset supporting the analyses in the manuscript, "Challenges and opportunities for oncology drug repurposing informed by synthetic lethality," published in npj Systems Biology and Applications. The data includes processed tables, pharmacogenomic statistics, and the outputs from random forest (RF) models used to identify a network of high-confidence synthetic lethal (SL) interactions across various cancer types. The primary goal of this work was to develop a robust computational pipeline to predict SL pairs and validate them using pharmacogenomic data, thereby identifying promising targets for oncology drug repurposing. This dataset provides the necessary files to reproduce the key findings of the study and serves as a resource for further investigation into synthetic lethality and cancer therapeutics. The repository is organized into the following components: Model Outputs: Results from the Boruta feature selection algorithm and random forest models, including feature importance scores for each predicted genetic interaction. Pharmacogenomic Statistics: Results from the analysis of PRISM drug sensitivity data, including effect sizes and correlations linking genetic mutations to drug response. Network Files: Data files used to generate the interactive pan-cancer genetic interaction networks, with edge weights corresponding to model-derived importance scores. MethodsThe dataset was generated through a multi-step computational pipeline. Initially, cell line data (CRISPR-Cas9 knockout scores, mutation status, microsatellite instability) were acquired and preprocessed from the DepMap portal (Public 23Q2 and 24Q2). Cancer driver genes were defined according to Bailey et al. (2018). A genetic interaction network was constructed using a feature selection workflow based on the Boruta algorithm, which employs a random forest classifier to identify significant associations between gene mutations and CRISPR knockout fitness. A stringent consensus approach using three different importance algorithms was implemented to ensure the reliability of selected features. Predicted pan-cancer SL interactions were then subjected to statistical refinement using a Type II ANOVA model to control for confounding variables such as cancer type and cell line growth properties. The biological and clinical relevance of the predicted SL pairs was assessed through several analyses. Mutual exclusivity was evaluated using tumor mutation data from The Cancer Genome Atlas (TCGA). Protein-protein interaction (PPI) enrichment was analyzed using the STRING database. Finally, a pharmacogenomic analysis was performed by integrating the predicted SL network with PRISM drug sensitivity data to validate interactions and identify potential drug repurposing opportunities.