Modelling asymmetric count ratios in CRISPR screens to decrease experiment size and improve phenotype detection

Abstract

Pooled CRISPR screens are a powerful tool to probe genotype-phenotype relationships at genome-wide scale. They are based on a library of target-specific guide RNAs (gRNAs) that is applied to a pool of cells, with the aim to induce a single genetic perturbation in each cell. Detection of viability phenotypes depends on statistical comparison of the frequencies of these gRNAs before and after cell proliferation. Here, we report empirical and theoretical evidence for asymmetries in gRNA count ratios, and we show how failure to consider asymmetric null distribution of the data leads to loss of detection power. We present a biology-based, generative probabilistic model to show that the asymmetry is generated during the proliferation phase of the screen and that it is mainly driven by the distribution width of the gRNA library. Based on these observations we develop a statistical test that improves hit detection at reduced experiment size. The method is implemented in the open source package gscreend (submission to Bioconductor pending).

A benchmark of algorithms for the analysis of pooled CRISPR screens

“…Because of this, we believe that the most promising area for future research is in identifying issues specific to the experimental design of CRISPR-pooled screens. The recent work in identifying biases such as copy number-associated effects [30][31][32], structural rearrangement effects [33], and bottleneck effects [57] is exemplary of promising directions. We believe that there are further questions to be answered: for example, Are such biases generalizable to CRISPR interference and CRISPR activation screens?…”

Combinatorial CRISPR screening reveals functional buffering in autophagy

“…Higher replicate correlation was computationally demonstrated to correlate with smaller library distribution skew and higher library representation 49 . Thus, we aimed at experimentally investigating to what extent a combinatorial gRNA library's distribution skew contributes to hit calling accuracy by generating combinatorial libraries with artificially distorted gRNA representations and screening them in different coverages.…”

“…Thus, our identified gene interactions likely resemble functional nodes within the human autophagy circuit in which redundancy exists and point towards gene paralogs as a critical factor in generating redundancy in autophagy. Passaging coverage and CRISPR library distribution skew have recently been computationally predicted to be critical factors for data quality 49 . Indeed, our experimental analysis of combinatorial libraries with varying distribution skews, applied with different passaging coverages, supports this prediction and identifies a distribution skew of below 2 as a threshold enabling passaging coverages below 100-fold.…”

ELP-dependent expression of MCL1 promotes resistance to EGFR inhibition in triple-negative breast cancer cells

“…A consensus "best" method for analyzing this type of screen has not yet emerged. The best performing method in any given scenario likely depends on experiment-specific parameters, which affect the distributions of targeting/nontargeting sgRNAs and the magnitude of changes within the data being analyzed (26). Thus, rather than selecting an analysis method a priori, we tested several analysis strategies in parallel (27)(28)(29)(30)(31).…”