A method for benchmarking genetic screens reveals a predominant mitochondrial bias

Abstract

We present FLEX (Functional evaluation of experimental perturbations), a pipeline that leverages several functional annotation resources to establish reference standards for benchmarking human genome-wide CRISPR screen data and methods for analyzing them. FLEX provides a quantitative measurement of the functional information captured by a given gene-pair dataset and a means to explore the diversity of functions captured by the input dataset. We apply FLEX to analyze data from the diverse cell line screens generated by the DepMap project. We identify a predominant mitochondria-associated signal within co-essentiality networks derived from these data and explore the basis of this signal. Our analysis and time-resolved CRISPR screens in a single cell line suggest that the variable phenotypes associated with mitochondria genes across cells may reflect screen dynamics and protein stability effects rather than genetic dependencies. We characterize this functional bias and demonstrate its relevance for interpreting differential hits in any CRISPR screening context. More generally, we demonstrate the utility of the FLEX pipeline for performing robust comparative evaluations of CRISPR screens or methods for processing them.

Recovering false negatives in CRISPR fitness screens with JLOE

“…It should also be noted that experimental assay length is different between the Avana and Sanger screens (21 days versus 14 days, respectively) ( 46 ) and therefore different timepoints may result in differences in dependency profiles of certain genes. Consistent with this notion, dependency on oxphos genes, which is correlated with cell growth rate ( 47 ), represents a large source of bias in screen hits, and these genes are likely excluded from the Sanger core essentials for the same reason. We did not observe any strong functional enrichment among Sanger-specific core essentials, compared to the other classes.…”

Automated assembly of molecular mechanisms at scale from text mining and curated databases

“…In this assay, the effect of gene inactivation is assessed by determining the rate at which a specific knockout (or knockdown) disappears from a co-culture comprising cells transfected with a genome-scale RNAi or CRISPR-Cas9 library. It has previously been observed that that genes whose knockouts have similar effects on viability across a large number of cell lines —a phenomenon known as codependency—frequently participate in the same protein complex or pathway (Doherty et al, 2021; Meyers et al, 2017; Pan et al, 2018; Rahman et al, 2021; Shimada et al, 2021; Tsherniak et al, 2017). For example, CHEK2 and CDKN1A have a correlation coefficient of 0.359 and 0.375 in DepMap CRISPR and RNAi data, respectively ( Figure 7A ), and this codependency can be explained by the fact that the CHEK2 kinase is an activator of CDKN1A (also known as p21) and that the two genes jointly regulate cell cycle progression.…”

“…To obtain robust measures of gene co-dependencies, we combined the CRISPR and RNAi perturbation data by converting the Pearson correlation coefficients for each gene pair into signed z-scores and computing the combined z-score between the two datasets using Stouffer’s method ( Figure 7A ). In analyzing the data, we first accounted for a bias also observed by others (Dempster et al, 2019; Rahman et al, 2021), namely that many of the strongest correlations are between mitochondrial genes ( Figure 7B). These correlations have been described as an artifact of the screening method (such as the timepoint of the viability measurements relative to cell doubling time) rather than reflecting true co-dependencies (Rahman et al, 2021).…”

“…In analyzing the data, we first accounted for a bias also observed by others (Dempster et al, 2019; Rahman et al, 2021), namely that many of the strongest correlations are between mitochondrial genes ( Figure 7B). These correlations have been described as an artifact of the screening method (such as the timepoint of the viability measurements relative to cell doubling time) rather than reflecting true co-dependencies (Rahman et al, 2021). We considered the correlations among these genes to be “explained” a priori due to their shared mitochondrial function, and using the mitochondrial gene database MitoCarta as a reference (Rath et al, 2021), we excluded correlations among them from subsequent analysis.…”

Optimal construction of a functional interaction network from pooled library CRISPR fitness screens

“…S2). Since oxphos genes are known to be a source of bias in coessentiality networks and differential essentiality [ 23 ], we removed these six pathways from the Reactome reference set, hereafter referred to as CleanReactome.

Fig.

…”