Genomic sequencing has allowed identification of omic features that predict drug sensitivity and their mechanisms of action. However, findings from these associative models have the potential to be more powerful with the availability of data from genome-wide, forward perturbation approaches in large panels of cancer cell lines. Using data generated by RNA interference (RNAi) and CRISPR-Cas9 screens, we can begin to construct drug-gene networks to better understand the mechanisms of drugs. Here, we aim to investigate how two types of complementary functional screening data, gene dependency profiles and drug sensitivity profiles, may reveal cancer cell line drug sensitivities and mechanisms of action. We leverage publicly available drug sensitivity data from the Cancer Therapeutics Response Portal (CTRP), and vulnerability datasets from DepMap. We show how gene vulnerability profiles can be utilized to identify possible unknown drug mechanism of action (MOA) and to characterize cancer drugs with high versus low specificity.
This project first utilizes multiple linear regression to reveal the relationship between pan cancer drug sensitivity with gene essentiality. This initial approach reveals three orders of drug MOAs that have the strongest drug specificity: MDM2 target drugs, EGFR target drugs, and PI3K target drugs, respectively. However, besides drugs with these three target pathways, the majority of CTRP drugs do not demonstrate strong associations between drug targets and gene dependency even though they are labeled with known gene targets. In order to further understand the characteristics of these drugs, we used principal component analysis of the drug-gene association matrix formed by previous regression analysis to discover cancer drug clusters and their association with genes. We discovered cancer drug clusters with known MOAs, but the result also revealed unknown biological functional drug-gene relations. We then designed a Kolmogorov-Smirnov test to analyze the concordance of enrichment of gene targets and genes revealed to be associated with cancer drugs in our regression analysis. By this test we confirmed that there is an enrichment of genes in our analysis that do not correspond to putative drug targets. To further delve into these patterns, we utilized a supervised approach to confirm the unknown MOAs revealed in previous results. A top-down approach of K means clustering was used to find groups of cancer drugs that is further input to Partial Least Squares Discriminant Analysis (PLS DA) to ascertain cancer drug clustering that may require further experimental testing to confirm their putative targets. Therefore, this project provides a novel integrative method that combines gene vulnerability and drug sensitivity data to understand cancer drug specificity. It also motivates further experimental studies to identify characteristics of cancer drugs revealed with unknown MOAs.
Authors: Estelle (Ning) Yao1, Katherine Sheu2, Thomas Graeber3
1University of California, Los Angeles
2David Geffen School of Medicine
3Department of Molecular & Medical Pharmacology