Negative Binomial Mixed Models for Identifying Oncogenic dependencies through analysis of RNAi Screening data

Background: Loss-of-function RNAi screening has been extensively used to identify cancer dependencies,including oncogene addiction. A few computational methods such as ATARiS, DEMETER, and RSA havebeen presented to compute gene-level scores by handling off-target effects in RNAi screening data. However,these methods often result in low statistical power in low sample-size settings. This paper presents a newstatistical approach to tackle the off-target effects in order to provide higher statistical power compared tothe mentioned methods.Methods: We applied DRIVE project data and the CCLE data to detect gene drivers in pan-cancer and breastcancer, by thoroughly scrutinizing all shRNA data. In the proposed method, we first removed the effect ofbatches including pool and thermodynamic stability of shRNAs using an empirical Bayesian model availablein the SVA package in R. Then negative binomial mixed effect models were performed on ranks of theselogFC in each cell line.Results: Among ۶۹۱۹ genes, known cancer genes such as KRAS, NRAS, BRAF, PIK۳CA, CTNNB۱, TP۵۳,and CDK۴ were reassuringly identified by the proposed method in Pan-cancer analysis. We demonstratedthat the proposed approach outperformed ATARiS and DEMETER in terms of statistical power through subsamplingapproaches. In analyzing breast cancer data, we identified both putative oncogenes such as PAX۵and RASGRP۲ and known oncogenes such as KRAS.Conclusion: By using all information in RNAi screening data, the proposed method models on- and offtargeteffects and can identify oncogene addictions in cancer.