MAGeCK enables robust identification of essential genes from genome-scale CRISPR/Cas9 knockout screens

Abstract

We propose the Model-based Analysis of Genome-wide CRISPR/Cas9 Knockout (MAGeCK) method for prioritizing single-guide RNAs, genes and pathways in genome-scale CRISPR/Cas9 knockout screens. MAGeCK demonstrates better performance compared with existing methods, identifies both positively and negatively selected genes simultaneously, and reports robust results across different experimental conditions. Using public datasets, MAGeCK identified novel essential genes and pathways, including EGFR in vemurafenib-treated A375 cells harboring a BRAF mutation. MAGeCK also detected cell type-specific essential genes, including BCR and ABL1, in KBM7 cells bearing a BCR-ABL fusion, and IGF1R in HL-60 cells, which depends on the insulin signaling pathway for proliferation.

Figure 1

Overview of the MAGeCK algorithm. Raw read counts corresponding to single-guided RNAs (sgRNAs) from different experiments are first normalized using median normalization and mean-variance modeling is used to capture the relationship of mean and variance in replicates. The statistical significance of each sgRNA is calculated using the learned mean-variance model. Essential genes (both positively and negatively selected) are then identified by looking for genes whose sgRNAs are ranked consistently higher (by significance) using robust rank aggregation (RRA). Finally, enriched pathways are identified by applying the RRA algorithm to the ranked list of genes.

Figure 2

A comparison of MAGeCK with two other RNA interference screen algorithms, RIGER and RSA. (a) The numbers of significantly selected genes identified by MAGeCK, RIGER and RSA in different comparisons. For comparisons between control samples or between replicates of the same condition (highlighted in yellow), ideally no significantly selected genes should be detected. Comparisons between treatments and controls are highlighted in green. See Table S1 in Additional file 2 for a complete comparison. (b) The overlap of top-ranked genes between CRISPR/Cas9 knockout screening and RNAi screening on the melanoma dataset. The positive screening experiment was performed in the same way as for the melanoma dataset [17], except that pooled shRNA screening was used instead of CRISPR/Cas9 knockout screening.

Figure 3

MAGeCK is robust against sequencing depth and the number of targeting sgRNAs per gene. (a) The number of significantly selected sgRNAs and genes in the leukemia dataset using various sequencing depths. The maximum sequencing depth for all samples is 30 million. See Materials and methods for sampling details. (b) The number of significantly selected sgRNAs and genes in the melanoma dataset using various sequencing depths. The maximum sequencing depth for all samples is 17.5 million. See Materials and methods for sampling details. Error bars represent the standard deviation from three independent sampling experiments.

Figure 4

The number of identified positively and negatively selected sgRNAs at different sequencing depths. (a,b). The numbers of positively and negatively selected sgRNAs in the leukemia dataset (a) and melanoma dataset (b) under different sequencing depths are shown. The numbers are normalized by the number of identified sgRNAs at the maximum sequencing depths (30 million for the leukemia dataset, 17.5 million for the melanoma dataset).

Figure 5

MAGeCK is robust to the number of targeting sgRNAs per gene. This figure shows the effect of different numbers of targeting sgRNAs per gene. Each bar indicates the percentage of top-ranked, 'reference' genes that are identified by MAGeCK, RIGER and RSA using different numbers of sgRNAs per gene. 'Reference' genes are those that are in the top 5% of ranked genes in all three methods when using 10 sgRNAs per gene. See Materials and methods for sampling details. Error bars represent the standard deviation from three independent sampling experiments.