Fig. 2.
Predictive models for dependency. a Multivariate regression of each CRISPR or RNAi gene effect profile (N = 15,221) using predictive features derived from omics datasets and cell line annotations. The accuracy of each predictive model is the correlation coefficient of measured and predicted values across cell lines. The number of accurate predictive models (r > 0.5) is split according to whether the gene target is also a CRISPR pan-dependency. b Mean predictive accuracy for high-confidence dependencies (N = 1703) as a function of the fraction of cell lines identified as dependencies using CRISPR (probability of dependency > 0.5). c CYCLOPS genes have a positive correlation (r > 0.5) between RNAi gene effects and copy number and account for ~ 35% of the accurate RNAi models for CRISPR pan-dependencies (N = 1719) after removing 148 RNAi models driven by confounding factors. d Genetic perturbation of PRMT5 using RNAi results in larger effect size between cell lines with MTAP copy number loss and MTAP wild-type compared to CRISPR knockout (Wilcoxon p-value: CRISPR = 4.6 × 10−8, RNAi = 2.5 × 10−34). e RNAi gene effect of RBBP4 is more correlated with expression of paralog gene RBBP7. Density (2D) contours represent 402 cell lines for each genetic dependency dataset (RNAi, CRISPR). f CRISPR knockout of RAB6A has a more negative viability effect on cells lacking expression of paralog gene RAB6B (Wilcoxon p-value: CRISPR = 6.2 × 10−24, RNAi = 6.1 × 10−7)