Genomic Amplifications Cause False Positives in CRISPR Screens

Reverse-genetic screens are powerful tools for decoding how genotype translates to phenotype (1). RNAi and CRISPR are two major tools for loss-of-function, reverse-genetic studies in mammalian cells (2). High-throughput RNAi- and CRISPR-based screens are typically performed by transducing cells with a lentivirus library of shRNAs or single-guide RNAs (sgRNA), and the shRNAs or sgRNAs that produce a desired phenotype are identified in a population as enriched (e.g., inactivation promotes cell growth) or depleted (e.g., inactivation reduces viability). Recent studies reveal profound differences between genes identified by RNAi- and CRISPR-based screens even in the same genetic context (3). The disparate cellular responses to RNAi-based knockdown and CRISPR-based genome editing indicate the need for more scrutiny when interpreting the results of genetic screens using these technologies.

RNAi-based genetic screens have been widely used in mammalian cells, and false-positive and false-negative results of RNAi screens have been well characterized. Partial silencing of a target gene by RNAi may be insufficient to produce a phenotype, leading to false-negative results for some genes. Unintended or “off-target” silencing of mRNAs homologous to a target mRNA or quenching of the miRNA pathway by overproduction of an RNAi construct can lead to false-positive results.

The commonly used S. pyogenes CRISPR system targets the DNA rather than the transcript of a gene and introduces a double-strand DNA break. Imprecise repair of the break by nonhomologous end-joining results in a small insertion or deletion that disrupts the reading frame. Thus, CRISPR inactivates target genes more readily than RNAi does. Nevertheless, potential false-positive and false-negative results of CRISPR screens should be expected. Indeed, complete gene knockout can trigger compensatory transcriptional activation of functionally redundant genes (4), causing false-negative results. As with RNAi, false positives may result if CRISPR modifies off-target genomic sites with sequence similarity to the intended CRISPR target site (5). But the caveats of CRISPR-based screens have not been systematically documented. In this issue of Cancer Discovery, Munoz and colleagues (6) and Aguirre and colleagues (7) identify target site copy number as an unexpected trigger of false positives in CRISPR screens for essential genes.

The technical features of the two CRISPR screens are summarized in Fig. 1A. Aguirre and colleagues used a genome-wide library at six sgRNAs per gene. This study was performed in a large collection of 33 cell lines representing a variety of cancer types and genetic contexts. Munoz and colleagues used a focused CRISPR library targeting approximately 2,700 genes in five cell lines. Though not at genome-wide scale, a focused CRISPR library can achieve more sgRNAs per gene (20 vs. 6) and higher representation (1,000 cells per sgRNA vs. 500) compared with a genome-wide library.

Aguirre and colleagues show that their CRISPR screening approach identifies essential genes in cancer cells, including both oncogene drivers and nononcogene dependencies. Their study parallels previous studies in which they performed genome-wide RNAi screens using the same cell lines, allowing them to make some comparisons between the CRISPR and RNAi screens (8). Because copy-number amplification frequently leads to oncogene overexpression, they sought to identify cancer drivers associated with genomic amplification. They unexpectedly found that CRISPR guides targeting genes within genomic amplifications reduce proliferation or survival as compared with guides that target genes outside of the amplifications (Fig. 1B).

Seeking to compare RNAi and CRISPR screening technologies, Munoz and colleagues constructed complementary shRNA and sgRNA libraries (Fig. 1A). Screening in five cancer cell lines—three diploid and two aneuploid—they found that CRISPR screens identify 2 to 5 times more essential genes than RNAi screens, possibly due to more complete inactivation by CRISPR. To rule out the possibility that the CRISPR screens had lower false-negative rates or higher false-positive rates, the authors examined the lethality scores of nonexpressed genes. Notably, the CRISPR screens in the three diploid cell lines produced virtually no false positives. But the screens in the two aneuploid cell lines did produce false positives, and the false positives mapped to genomic amplifications.

Both studies proceed to more carefully examine the effects of CRISPR target site copy number on cell viability. Consistent with the finding that CRISPR-mediated lethality is independent of transcriptional status, CRISPR guides targeting intergenic sequences in genomic amplifications are as lethal as those targeting essential genes. Thus, reduced proliferation or survival by sgRNAs targeting amplified genes is not due to gene inactivation. Comparing aggregate analysis of apparent essentiality due to amplified genes, Aguirre and colleagues found that increasingly essential genes (based on level of CRISPR guide depletion) were more likely to reside in genomic amplifications in CRISPR screens than they were in RNAi screens—also consistent with the finding that lethality does not result from gene inactivation.

Both studies found that the antiproliferative effect of CRISPR targeting genomic amplifications positively correlates with target site copy number. They show that CRISPR guides that target multiple sites throughout the genome induce lethality, and that the best predictor of off-target lethality is the number of genomic sites with perfect complementarity to an intended CRISPR target site. The authors of both studies therefore propose that excessive DNA damage due to CRISPR cutting underlies the lethality, and they show that CRISPR guides targeting genomic amplifications or multiple sites throughout the genome activates the DNA damage response (i.e., γH2AX phosphorylation and foci) and increases G₂–M cell-cycle arrest. These findings help to explain observations of CRISPR-induced off-target lethality as well as previous work showing that sgRNAs targeting a nongenic region of the BCR–ABL amplification decrease cell viability (9).

Aguirre and colleagues suggest two cellular responses to CRISPR genome editing in cancer cells: an early antiproliferative DNA damage response and a later target gene inactivation. Overall, the antiproliferative effect is independent of the target gene or chromosome structure, and increases with the number of cuts conferred by individual sgRNAs. The authors propose that the early antiproliferative cell response induced by CRISPR represents a critical vulnerability of cancer cells with genome amplifications that might enable cancer-specific therapy. In the second response, sgRNAs targeting essential genes are depleted following loss of protein expression, representing a true positive in a CRISPR screen. They note the caveat, however, that copy-number amplifications may protect some essential genes from complete knockout by CRISPR.

The study by Munoz and colleagues also provides insight into the design and functionality of sgRNAs used in CRISPR screens. Screening three different cell lines with a CRISPR tiling array against 139 essential genes and approximately 364 sgRNAs per gene, they determined that the best predictor of sgRNA performance is targeting a conserved PFAM protein domain—consistent with a recent study by Shi and colleagues (10)—followed by sequence conservation across vertebrate species. In addition to these features, the authors advise the following criteria to avoid excessive double-strand DNA breaks: Design sgRNAs with minimal matches across the genome, transduce lentiviral sgRNA libraries at low multiplicity of infection, and use sgRNA targeting nonexpressing or known nonessential regions as controls instead of scrambled sgRNAs.

In summary, these papers report comprehensive loss-of-function CRISPR screens across a panel of human cancer cell lines. Both papers show that cancer driver genes can be readily identified, demonstrating the feasibility of CRISPR-based functional genomic screens. Both studies also highlight potential false-positive results in CRISPR screens in cancer cell lines harboring copy-number variations, adding a layer of complexity to interpreting CRISPR screens. The findings call for improved CRISPR libraries, use of diverse cell lines, and scrutiny of CRISPR screen data. Alternative approaches—including loss-of-function RNAi knockdown, gain-of-function cDNA rescue, and CRISPR-based transcriptional inactivation—are needed to screen for cancer drivers in genomic amplifications and should be used to complement or validate CRISPR screens.

No potential conflicts of interest were disclosed.

This work was supported by the NCI (5R00CA169512), the American Cancer Society (129056-RSG-16-093), and the Lung Cancer Research Foundation (to W. Xue).