A simple statistical test to infer the causality of target/phenotype correlation from small molecule phenotypic screens

Motivation: Cell-based phenotypic screens using small molecule inhibitors is an important technology for early drug discovery if the relationship between the disease-related cellular phenotype and inhibitors' biological targets can be determined. However, chemical inhibitors are rightfully believed to be less specific than perturbation by biological agents, such as antibody and small inference RNA. Therefore, it is often a challenge in small molecule phenotypic screening to infer the causality between a particular cellular phenotype and the inactivation of the responsible protein due to the off-target effect of the inhibitors. Results: In this article, we present a Roche in-house effort of screening 746 structurally diverse compounds for their cytotoxicity in HeLa cells measured by high content imaging technology. These compounds were also systematically profiled for the targeted and off-target binding affinity to a panel of 25 pre-selected protein kinases in a cell-free system. In an effort to search for the kinases whose activities are crucial for cell survival, we found that the simple association method such as the chi-square test yields a large number of false positives because the observed cytotoxic phenotype is likely to be the result of promiscuous action of less specific inhibitors instead of true consequence of inactivation of single relevant target. We demonstrated that a stratified categorical data analysis technique such as the Cochran-Mantel-Haenszel test is an effective approach to extract the meaningful biological connection from the spurious correlation resulted from confounding covariates. This study indicates that, empowered by appropriate statistical adjustment, small molecule inhibitor perturbation remains a powerful tool to pin down the relevant biomarker for drug safety and efficacy research.