The systematic identification of essential genes in microorganisms has provided critical insights into the molecular basis of many biological processes (1). Similar studies in human cells have been hindered by the lack of suitable tools. Moreover, little is known about how the set of cell essential genes differs across cell types and genotypes. Differentially essential genes are likely to encode tissue-specific modulators of key cellular processes and important targets for cancer therapies. Here, we used two independent approaches for inactivating genes at the DNA level to define the cell-essential genes of the human genome.

The first approach employs the CRISPR/Cas9-based gene editing system, which has emerged as a powerful tool to engineer the genomes of cultured cells and whole organisms (2, 3). We and others have shown that lentiviral (single-guide RNA) sgRNA libraries can enable pooled loss-of-function screens and have used the technology to uncover mediators of drug resistance and pathogen toxicity (4-6). To systematically identify cell-essential genes, we constructed a novel library, which was optimized for high cleavage activity, and performed a proliferation-based screen in the near-haploid human KBM7 chronic myelogenous leukemia (CML) cell line (Fig. 1; Note S1).

The unusual karyotype of these cells also allows for an independent method of genetic screening. In this approach, null mutants are generated at random through retroviral gene-trap mutagenesis, selected for a phenotype, and monitored by sequencing the viral integration sites to pinpoint the causal genes (7). Positive selection-based screens using this method have identified genes underlying processes such as epigenetic silencing and viral infection (7-9). We extended this technique by developing a strategy for negative selection and conducted a screen for cell-essential genes (Fig. 1; Note S2).

For both methods, we computed a score for each gene that reflects the fitness cost imposed by inactivation of the gene. We defined the CRISPR score (CS) as the average log₂ fold-change in the abundance of all sgRNAs targeting a given gene, with replicate experiments showing a high degree of reproducibility (r=0.90) (Fig. 2A, S1A; Table S2). Of the 18,166 genes targeted by the library, 1,878 scored as essential for optimal proliferation in our screen, although this precise number depends on the cutoff chosen (Fig. 2A; Table S2-3). Overall, this fraction represents ∼10% of genes within our dataset or roughly 9.2% of the entire genome (many of the genes not targeted by our library encode olfactory receptors that are unlikely to be cell-essential). Gene products that act in a non-cell-autonomous manner are not expected to score as essential in this pooled setting (Fig. S1B).

We defined the gene-trap score (GTS) as the fraction of insertions in a given gene occurring in the inactivating orientation. Because the accuracy of this score depends on the depth of insertional coverage, we set a requirement on the minimum number (n=65) of anti-sense inserts in a gene needed for inclusion in our analysis by measuring the concordance between replicate experiments (Fig. 2B, S1C-D; Table S4; Note S2). For the 7,370 genes on the haploid chromosomes that exceeded this threshold, the GTS was well-correlated with the CS and with results from a co-published study employing a similar gene-trap approach (r=0.68) (Fig. 2C, S1E-F; Note S3). The strong correspondence between the overlapping sets of cell-essential genes defined by the two methods provides support for the accuracy of the CRISPR scores for the full set of 18,166 genes.

The two methods differed with respect to the diploid chromosome 8. Whereas the gene-trap screen failed to detect any cell-essential genes on this chromosome, the CRISPR screen uncovered a similar proportion of cell-essential genes on all the autosomes (Fig. 2A-C). These observations indicate that (i) the vast majority of cell-essential genes are haplosufficient and (ii) biallelic inactivation occurs at high frequency in our CRISPR screen (4).

To assess the accuracy of our scores with other measures of gene essentiality, we relied on functional profiling experiments conducted in yeast S. cerevisiae as a benchmark (1, 10). Specifically, we ranked genes common to all datasets by their scores in each dataset –CRISPR, gene trap, scores from similar loss-of-function RNA interference (RNAi) screens (11), and, as a naïve proxy for gene essentiality, gene expression levels determined by RNA-seq– and compared these rankings with the essentiality of yeast homologs. The CRISPR and gene-trap methods had significantly stronger correlations with the yeast results than did the RNAi screens or gene expression, which performed similarly to each other (both methods: p<10^-4, permutation test) (Fig. 2D). Based on additional comparisons with yeast gene essentiality, we also found that (i) our new optimized sgRNA library gave better results than those from screens using older unoptimized libraries (4, 5) and (ii) the coverage of this library (∼10 constructs per gene) approaches saturation, as evidenced by down-sampling (decreasing the coverage by randomly eliminating subsets of data) (Fig. S2A-B). Together, our results suggest that scores from the CRISPR and gene-trap screens both provide accurate measures of the cell-essentiality of human genes.

Essential genes should be under strong purifying selection and should thus show greater evolutionary constraint than non-essential genes (12). Consistent with this expectation, the essential genes found in our screens were more broadly retained across species, showed higher levels of conservation between closely related species, and contain fewer inactivating polymorphisms within the human species, as compared to their dispensable counterparts (Fig. 2E-G). Essential genes also tend to have higher expression and encode proteins that engage in more protein-protein interactions (13-15). These patterns were also observed in our CRISPR dataset (Fig. 2H-I).

In S. cerevisiae, genes with paralogous copies in the genome show a lower degree of essentiality –due presumably to at least partial functional overlap (16). Surprisingly, meta-analysis of knockout mouse collections has suggested that there is no such correlation in mammals (17, 18). However, others have challenged this interpretation because the genes analyzed were far from a random sample (19). Using the results from our genome-wide screens, we revisited this question and observed that genes with paralogs are indeed less likely to be essential, consistent with the idea that paralogs can provide functional redundancy at the cellular level (Fig. 2J).

To examine the functions of the cell-essential genes we used gene set enrichment analysis (GSEA) and found strong enrichment for many fundamental biological processes, such as DNA replication, RNA transcription and mRNA translation (Fig. S3A) (20). Whereas most of the genes could be assigned to such well-defined pathways, no function has been ascribed to about 330 of the cell-essential genes (18%) (Fig. 3A). For this set of uncharacterized genes, an analysis of the domains within their encoded gene products and comparisons to proteomic datasets from organellar purifications revealed substantial enrichment in proteins found in the nucleolus and those containing domains associated with RNA processing (Fig. S3B-C) (21).

We characterized three such genes, C16orf80, C3orf17, and C9orf114, whose mRNA expression patterns across the cancer cell line encyclopedia (CCLE) were correlated with that of genes involved in RNA processing (Fig. 3B). We validated the essentiality of these genes in short-term proliferation assays and detected localization of their products to the nucleus (C16orf80) or nucleolus (C3orf17 and C9orf114) (Fig. 3C-D) (22). Additionally, mass spectrometric analyses of anti-FLAG-immunoprecipitates prepared from KBM7 cells expressing FLAG-tagged C16orf80, C3orf17, and C9orf114 revealed interactions with multiple subunits of the spliceosome, ribonuclease (RNAse) P/MRP, and H/ACA small nucleolar ribonucleoprotein (snoRNP) complexes, respectively (Fig. 3E). These results implicate C16orf80 in splicing, consistent with its association with mRNAs; C3orf17 in rRNA/tRNA processing; and C9orf114 in RNA modification (23). More broadly, our results indicate that the molecular components of many critical cellular processes, especially RNA processing, have yet to be fully defined in mammalian cells.

To determine how the set of essential genes differs among cell lines, we screened another CML cell line (K562) and two Burkitt's lymphoma cell lines (Raji and Jiyoye) using the CRISPR system (Table S2-3). Overall, the sets of essential genes in the four cell lines showed a high degree of overlap (Fig. 4A). Importantly, out of these four cell lines, the KBM7 CRISPR results showed the highest correlation with the KBM7 gene-trap dataset, suggesting that the few differences observed are likely to be biologically meaningful (Fig. S4A).

We focused first on genes found to be essential in only one of the four cell lines. The Raji, Jiyoye, and KBM7 cell lines had 6, 7, and 19 such genes, respectively (Fig. S4B; Table S5). One example was DDX3Y, which resides in the non-pseudoautosomal region of the Y chromosome and was required only in Raji cells (Fig. 4B). Its X-linked paralog, DDX3X, was essential in KBM7 and K562 cells (Fig 4E). Both genes encode DEAD-box helicases that likely have similar cellular functions (24). Thus, the dependence on one paralog might reflect functional absence of the other paralog. Indeed, DNA sequencing of DDX3X in Raji cells revealed hemizygous mutations in the 5′-splice site of intron 8 that resulted in the production of a truncated mRNA transcript (Fig. 4C-D). Conversely, DDX3Y was not expressed in KBM7 cells and was not present in K562 cells, which are of female origin (Fig. 4E). Introduction of wild-type DDX3X cDNA into Raji cells fully rescued the proliferation defect resulting from DDX3Y loss, indicating that the paralogous genes are essential and functionally overlapping (Fig. 4F). Essential paralogous gene pairs, involved in glucose metabolism (HK1/2 and SLC2A1/3) and cell-cycle regulation (CDK4/6), were also observed in the Jiyoye line (Fig S4C; Note S4). Vulnerabilities due to the loss of a paralogous partner may serve as targets for highly personalized anti-tumor therapies (25).

In some cases, cell line-specific essentiality of paralogous genes did not reflect differential expression. For example, the transcription factors GATA1 and GATA2 are expressed in both K562 and KBM7 cells, but the first is specifically essential in K562 cells and the second in KBM7 cells (Fig. S4D). These master regulators are known to promote proliferation and survival during distinct developmental stages in the hematopoietic lineage – GATA1 is required for the survival of erythroid progenitors and GATA2 for the maintenance and proliferation of immature hematopoietic progenitors (26). These two cell types likely correspond to the cells-of-origin of the two CML lines (27, 28). We also identified similar instances of genes required for cell line-specific functions such as nf-κb pathway regulation and homology-directed DNA repair in Raji and KBM7 cells, respectively (Fig. S4E-F; Note S5).

Whereas the other three cell lines showed only a few cell line-specific essential genes, the K562 cell line was an outlier, with 63 such genes (Fig. S4B). Oddly, these genes showed no discernible common functions and many were not even expressed in the K562 cell line. (Additionally, a few encoded secreted factors whose loss would not be expected to be lethal in a pooled screen.) This mystery was resolved when we examined the chromosomal location of the genes: strikingly, the majority (39 of the 63 genes) reside near 9q34 and 22q11. These two regions are translocated to produce the BCR-ABL oncogene and are present in a high-copy tandem amplification in K562 cells (Fig. S5A-B) (29). Notably, all 61 contiguous genes within the amplicon on 22q11 scored as essential, suggesting that sgRNA-mediated cleavage in this repeated region induces cytotoxicity in a manner unrelated to the function of the target genes themselves (Fig. 4G, S5F). Indeed, two sgRNAs targeting non-genic sites within the amplicon were toxic to K562, but not KBM7 cells. They increased the abundance of phosphorylated histone H2AX (γH2AX), a marker of DNA damage, and induced erythroid differentiation, which occurs upon DNA damage in this cell line (Fig. 4H-J, S5C) (30). We obtained similar results in another cell line, HEL, which contains a highly amplified region surrounding the JAK2 tyrosine kinase (Fig. S5D-E). Together, these findings indicate that lethality upon Cas9-mediated cutting may also reflect chromosome structure and therefore should be evaluated in light of copy-number information.

Finally, we looked for consistent differences in essential genes between the two CML and two Burkitt's lymphoma lines. Such genes might represent attractive targets for anti-neoplastic therapies as their inhibition is less likely to be broadly cytotoxic. Overall, we identified 33 genes that were specifically essential in the CML lines and 15 genes in the Burkitt's lymphoma lines (Fig. 4K; Table S6). As a control, permuted comparisons – that is, a set containing of one CML and one Burkitt's line vs. the complementary sets – showed roughly half as many “set-specific” essential genes (Fig. S6A-C).

In the CML lines, the top two genes were BCR and ABL1, consistent with the known essentiality of the BCR-ABL translocation product and the therapeutic effect of BCR-ABL inhibitors such as imatinib (31). Additional members of the BCR-ABL signaling pathway, son of sevenless homolog 1 (SOS1), growth factor receptor-bound protein (GRB2), and GRB2-associated binding protein 2 (GAB2) scored strongly as well (ranked 3, 4, and 7, respectively). Network analysis of the other top hits also uncovered several genes encoding assembly factors for the electron transport chain, as well as enzymes involved in folate-mediated one-carbon metabolism. These results suggest additional potential targets for CML therapy (Table S6).

In the B cell-derived Burkitt's lymphoma cell lines, the top genes included three B cell lineage transcription factors early B-cell factor 1 (EBF1), POU class 2 associating factor 1 (POU2AF1), and paired box 5 (PAX5) (ranked 3, 6, and 8, respectively). Each of these genes is the target of recurrent translocations in lymphoma (32-34). Enhancers of the corresponding three gene loci all show a high level of bromodomain containing 4 (BRD4) occupancy in Ly1 cells, a related diffuse large B cell lymphoma cell line, suggesting bromodomain inhibitors such as JQ1 as potential treatments (35). Other selectively essential genes included MEF2B, a transcriptional activator of BCL6, B-cell lymphoma 6, and CCND3, cyclin D3, both of which are frequently mutated and implicated in the pathogenesis of various lymphomas (36). Intriguingly, the top two hits, CHM and RPP25L, do not appear to have specific roles in B cells; rather, their differential essentiality is likely explained by the lack of expression of their paralogs, CHML and RPP25, in both of the Burkitt's lymphoma cell lines studied (Fig. S6D).

In conclusion, we used two complementary and concordant approaches, CRISPR and gene-trap, to define the cell-essential genes in the human genome. Although the gene-trap method is suitable only for loss-of-function screening in rare haploid cell lines, the CRISPR method is broadly applicable. Extending our analysis across different cell lines and tumor types, we developed a framework to assess differential gene essentiality and identify potential drivers of the malignant state. The method can be readily applied to more cell lines per cancer type to eliminate idiosyncrasies particular to a given cell line and to more cancer types to systematically uncover tumor-specific liabilities that might be exploited for targeted therapies.