Research ArticleCancer

Off-target toxicity is a common mechanism of action of cancer drugs undergoing clinical trials

See allHide authors and affiliations

Science Translational Medicine  11 Sep 2019:
Vol. 11, Issue 509, eaaw8412
DOI: 10.1126/scitranslmed.aaw8412

Drugs may not do what we think they do

The majority of proposed anticancer treatments do not succeed in advancing to clinical use because of problems with efficacy or toxicity, often for unclear reasons. Lin et al. discovered that a drug candidate in clinical development was effective at killing cancer cells even when its target protein was knocked out, suggesting that its proposed mechanism of action was incorrect. The researchers then identified multiple drugs with similar problems and also discovered the correct target for one of them, suggesting that more research and more stringent methods are needed to verify the targets of potential drugs before advancing them to the clinic.

Abstract

Ninety-seven percent of drug-indication pairs that are tested in clinical trials in oncology never advance to receive U.S. Food and Drug Administration approval. While lack of efficacy and dose-limiting toxicities are the most common causes of trial failure, the reason(s) why so many new drugs encounter these problems is not well understood. Using CRISPR-Cas9 mutagenesis, we investigated a set of cancer drugs and drug targets in various stages of clinical testing. We show that—contrary to previous reports obtained predominantly with RNA interference and small-molecule inhibitors—the proteins ostensibly targeted by these drugs are nonessential for cancer cell proliferation. Moreover, the efficacy of each drug that we tested was unaffected by the loss of its putative target, indicating that these compounds kill cells via off-target effects. By applying a genetic target-deconvolution strategy, we found that the mischaracterized anticancer agent OTS964 is actually a potent inhibitor of the cyclin-dependent kinase CDK11 and that multiple cancer types are addicted to CDK11 expression. We suggest that stringent genetic validation of the mechanism of action of cancer drugs in the preclinical setting may decrease the number of therapies tested in human patients that fail to provide any clinical benefit.

INTRODUCTION

Substantial progress has been made in the treatment of certain malignancies by targeting cancer “addictions” or genetic dependencies that encode proteins required for the survival and/or proliferation of cancer cells (1). Therapeutic agents that block the function of a cancer dependency—such as the kinase inhibitor lapatinib in HER2+ breast cancer—can trigger apoptosis and durable tumor regression (2). Discovering and characterizing druggable cancer dependencies are key goals of preclinical research.

While screening cancer drug targets, we found that maternal embryonic leucine zipper kinase (MELK), a protein previously reported to be essential in multiple cancer types, could be eliminated using CRISPR-mediated gene editing without any detectable loss in cancer cell fitness (3, 4). In addition, we demonstrated that OTS167, a small-molecule inhibitor of MELK undergoing phase 2 clinical trials, continued to kill MELK–knockout (KO) cancer cells with no decrease in potency. These findings suggested that a drug tested in human cancer patients had been designed to target a nonessential cellular protein and that its putative inhibitor killed cells by interacting with proteins other than its reported target. We hypothesized that problems in drug development and inhibitor validation, as exemplified by MELK and OTS167, could potentially contribute to the high failure rate of new cancer therapies. In particular, drugs that target superfluous proteins may display limited efficacy in human patients, and if these drugs are active only via off-target effects, then this could potentially contribute to patient toxicity. Moreover, clinical trials that use a biomarker to select patients for trial inclusion are about twice as likely to succeed as those without one (5). Misidentifying a drug’s mechanism of action (MOA) could hamper efforts to uncover a biomarker capable of predicting therapeutic responses, further decreasing the success rate of clinical trials. To test whether other cancer drugs had similarly been designed against nonessential targets or had been assigned an incorrect MOA, we set out to systematically analyze multiple cancer drugs and drug targets that have entered clinical trials or are in late-stage preclinical development.

RESULTS

CRISPR competition assays to investigate several putative cancer dependencies

Based on an analysis of the literature, we chose drug targets that met several criteria (described in detail in Materials and Methods). Notably, we selected drug targets that had been reported to play a cell-autonomous role in cancer growth, such that their loss or inhibition was reportedly sufficient to block cancer cell proliferation. In addition, we selected drug targets that lacked a known mutation capable of conferring resistance to their targeted inhibitors, which we hypothesized represents the gold standard for proving a drug’s MOA. We identified 10 cancer drugs targeting six proteins that met these criteria (Table 1). Five of these proteins are reported to represent cancer dependencies (HDAC6, MAPK14/p38α, PAK4, PBK, and PIM1) (615). One protein (CASP3/caspase-3) is reported to induce apoptosis when activated by a small molecule (16, 17) and is discussed separately. Among the putative dependencies, over 180 different publications indicate that they are required for cancer cell proliferation or fitness (listed in data file S1). For each of these genes, the majority of evidence supporting their designation as cancer dependencies comes from RNA interference (RNAi) studies, in which small interfering RNA (siRNA)– or short hairpin RNA (shRNA)–mediated knockdown was reported to impair cancer cell fitness. In addition, each protein is targeted by one or more small-molecule drugs, which have been described to exhibit potent cell killing in vitro and in vivo. On the basis of these preclinical results, the drugs listed in Table 1 have been used in at least 29 different clinical trials, with an estimated enrollment of more than 1000 patients.

Table 1 Anticancer drugs and drug targets.

View this table:

We first set out to validate the role of the putative dependencies targeted by these drugs in cancer cell fitness. To accomplish this, we applied a CRISPR-Cas9–based cell competition assay, in which cancer cells are infected at a low multiplicity of infection with green fluorescent protein (GFP)–expressing guide RNA (gRNA) vectors targeting a gene of interest (Fig. 1A) (18). If a CRISPR-induced mutation reduces cell fitness, then the untransduced cells within a population should outcompete the gRNA-expressing cells, and the fraction of GFP+ cells should decrease over time. To verify this approach, we designed gRNAs against pan-essential genes and against several confirmed cancer drug targets. In breast cancer, colorectal cancer, lung cancer, and melanoma cell lines, guides targeting the essential replication proteins RPA3 and proliferating cell nuclear antigen (PCNA) dropped out up to 100-fold, and guides targeting the validated pan-cancer dependencies Aurora A, Aurora B, and ERCC3 exhibited similar levels of depletion (Fig. 1B). Mutations in Aurora A (19), Aurora B (20), and ERCC3 (21) confer resistance to the cytotoxic agents MLN8054, ZM447439, and triptolide, respectively, thereby providing genetic evidence that they are required for cancer cell growth. In contrast, gRNAs targeting the nonessential Rosa26 and AAVS1 loci exhibited minimal dropout over five passages in culture. These GFP competition assays were also capable of identifying cell type–specific dependencies: Guides targeting the oncogenic kinase BRAF dropped out in a BRAF-mutant melanoma cell line but not a BRAF–wild-type (WT) colorectal cancer line, whereas guides targeting the gene encoding the estrogen receptor (ESR1) dropped out in an ER-positive breast cancer line but not in a triple-negative breast cancer line (Fig. 1C). We concluded that our CRISPR dropout assay can robustly identify both pan-essential and cancer type-specific genetic dependencies.

Fig. 1 Cell competition assays to test the essentiality of putative cancer dependencies.

(A) Schematic of the CRISPR-based cell competition assays used in this paper (18). (B) Cell competition assays comparing guides targeting AAVS1 and ROSA26 (nonessential, negative control genes), RPA3 and PCNA (pan-essential positive control proteins), and Aurora A, Aurora B, and ERCC3 (inhibitor-validated cancer dependencies). Full results from these competition experiments are included in data file S2. (C) Cell competition assays for the cell type–specific cancer dependencies BRAF and ESR1. (D) Western blot analysis of A375 populations transduced with the indicated gRNAs. (E) Cell competition assays with gRNAs targeting HDAC6, MAPK14, PAK4, PBK, or PIM1 in four different cancer cell lines.

We next designed gRNAs against the reported cancer dependencies HDAC6, MAPK14 (p38α), PAK4, PBK, and PIM1. To maximize the likelihood that a CRISPR-induced mutation results in a nonfunctional allele, guides were designed to target exons that encode key functional domains within a protein (fig. S1A) (18). We used Western blotting to verify that each guide resulted in strong protein depletion in four separate cell lines (Fig. 1D and fig. S1B), and we then further confirmed target ablation by performing a second set of Western blots with a different antibody that recognizes a distinct protein epitope (fig. S1C). Next, we conducted GFP competition assays in 32 cell lines from 12 different cancer types, which included multiple cell lines in which each gene had previously been reported to be essential (data file S1). In each experiment, four gRNAs targeting Rosa26 and AAVS1 were used as negative controls, whereas four gRNAs targeting PCNA and RPA3 were used as positive controls. These positive control guides dropped out between ~10- and ~200-fold over five passages in culture, whereas the negative control guides consistently exhibited <2.5-fold dropout. The variation in positive control dropout rates likely reflects cellular differences in Cas9 expression, proliferation, and the spectrum of indel mutations produced by the gRNA. Notably, all guides targeting HDAC6, MAPK14, PAK4, PBK, and PIM1 failed to drop out in every cell line that we tested (Fig. 1E, fig. S2, and data file S2). For instance, HDAC6 has been reported to be a genetic dependency in ARID1A-mutant ovarian cancer (6). However, in ARID1A-mutant ovarian cancer cell lines A2780, OVK18, OVTOKO, and TOV-21G, HDAC6-targeting guides failed to deplete above background levels. Similarly, PIM1 has been reported to be a genetic dependency in triple-negative breast cancer (14, 15), but PIM1-targeting guides were not depleted in any of the seven triple-negative breast cancer cell lines that we tested (data file S2). These results called into question whether these putative drug targets are required for cancer cell growth.

Generation and analysis of CRISPR-derived KO clones

To further test the essentiality of these genes in cancer, we derived clones harboring CRISPR-induced KOs in each gene in multiple cancer types. All five genes were knocked out in the triple-negative breast cancer cell line MDA-MB-231 and the melanoma cell line A375. HDAC6, MAPK14, PBK, and PIM1 were knocked out in the colorectal cancer cell line DLD1, whereas PAK4 was knocked out in the colorectal cancer cell line HCT116, because it has previously been reported that PAK4 is not a dependency in DLD1 (11). To minimize the possibility that downstream translational initiation or alternative splicing bypasses the effect of a single CRISPR-induced mutation, clones were made by cotransducing cancer cells with guides that targeted two different exons in a gene of interest (Fig. 2A and fig. S1A). Complete target ablation was then verified by Western blotting using two antibodies that recognized distinct protein epitopes (Fig. 2B and figs. S3 to S5A). We next compared these KO clones to control clones transduced with guides targeting Rosa26 or AAVS1. As a positive control, we confirmed that knocking out the verified drug target MEK1 decreased proliferative capacity in A375 clones (fig. S4). However, we found that clones lacking each putative genetic dependency listed in Table 1 proliferated at levels that were indistinguishable from control A375, DLD1, and HCT116 cancer cells (Fig. 2C). For instance, PAK4-KO melanoma cells underwent an average of 20.3 population doublings over the course of 15 days in culture, compared to 19.9 doublings for the Rosa26 gRNA-transduced clones. To test whether these genes were dispensable for cell division but required for growth in other environments, we also seeded the KO clones in soft agar and assessed their ability to grow in anchorage-independent conditions. Although MEK1-KO clones formed fewer colonies in soft agar (fig. S4E), every HDAC6, MAPK14, PAK4, PBK, and PIM1 KO exhibited WT rates of colony formation, further verifying that these genes are not required for cancer cell fitness (Fig. 2, D and E).

Fig. 2 Generating and analyzing single cell–derived KO clones of putative cancer dependencies.

(A) Schematic of the two-guide strategy used to generate clonal KO cell lines. (B) Western blot analysis of single cell–derived A375 KO clones. ab, antibody. (C) Proliferation assays for HDAC6, MAPK14, PAK4, PBK, and PIM1 KO clones. (D) Representative images of A375 and DLD1 Rosa26 or MAPK14-KO clones grown in soft agar. Scale bar, 2 mm. (E) Quantification of colony formation in control or KO A375, DLD1, and HCT116 clones. Boxes represent the 25th, 50th, and 75th percentiles of colonies per field, and the whiskers represent the 10th and 90th percentiles. For each assay, colonies were counted in at least 15 fields under a 10× objective.

Consistent with previously reported results, Rosa26 and AAVS1 control clones derived from MDA-MB-231 cell populations exhibited some variability in proliferative capacity (3, 4, 22). By analyzing a total of 12 single cell–derived control clones, we established a range of doubling times in which WT MDA-MB-231 cells can divide (fig. S5B). Every HDAC6, MAPK14, PAK4, PBK, and PIM1 KO clone proliferated at a comparable rate to these control clones (fig. S5C). All KO clones were also capable of forming colonies in soft agar at rates comparable to the control clones, further verifying that these putative dependencies are nonessential in breast cancer (fig. S5D).

Lack of homolog up-regulation in KO clones

Null mutations caused by CRISPR may trigger a different cellular response than RNAi-induced gene repression, potentially contributing to the discrepancies between our results and those that had previously been reported. In particular, a recent study suggested that CRISPR-induced nonsense mutations can trigger the up-regulation of the homologs of a targeted gene, potentially compensating for the effects of the lesion (23). We assessed the expression of the closest homologs of HDAC6, MAPK14, PAK4, PBK, and PIM1 in 33 different KO clones that we generated, but we observed no consistent up-regulation of any target homolog (fig. S6). In addition, we analyzed RNA sequencing (RNA-seq) data from 10 published experiments in gene-edited cancer cells from other laboratories and similarly failed to detect consistent evidence for the up-regulation of target gene homologs (fig. S7). In several experiments, we found that the homologs of the targeted gene were down-regulated. These results suggest that homolog up-regulation is not a common consequence of CRISPR mutagenesis in human cancer cells and that compensatory homolog overexpression is unlikely to explain the lack of a detectable growth defect in the CRISPR clones that we have analyzed.

Assessing putative cancer dependencies in whole-genome CRISPR and RNAi screens

Cell lines can exhibit interlaboratory variability that affects their response to different genetic and chemical perturbations (24). In addition, although we chose cancer types to study based on the dependency patterns reportedly exhibited by each gene (data file S1), it remained possible that these genes represent dependencies in a cancer lineage not included among the 32 cell lines that we studied. To test this possibility and to assess whether unique or nonrepresentative features of the cell lines used in our laboratory contributed to our discrepant results, we reanalyzed genetic dependency data from whole-genome CRISPR screens conducted in 485 cancer cell lines (fig. S8A). These screens consistently identified both pan-cancer and cell type–specific genetic dependencies (for example, Aurora B, BRAF, and PIK3CA; fig. S8, B and C). However, in accordance with our earlier results, these experiments also indicated that our chosen dependencies were fully dispensable for cancer cell fitness (fig. S8, A to C). For instance, MAPK14/p38α has previously been reported to be essential in breast cancer (9), but CRISPR screens conducted in 26 different breast cancer cell lines corroborate that its loss is tolerated without a substantial fitness defect (fig. S8D). Notably, we also reanalyzed 712 genome-wide shRNA screens, and these knockdown experiments similarly failed to identify HDAC6, MAPK14, PAK4, PBK, or PIM1 as cancer-essential genes (fig. S8, E to G). In total, these results indicate that our findings are unlikely to be explained by nonrepresentative features of the cell lines studied in our laboratory, by differences between partial and complete loss-of-function perturbations, or by these genes functioning as genetic dependencies only in certain cancer types. Instead, our data suggest that multiple genes targeted in cancer clinical trials are, in fact, fully dispensable for cancer cell growth.

Knocking down putative cancer dependencies with CRISPR interference

To further investigate whether differences between partial and complete loss-of-function perturbations could explain our discrepant results, we next performed competition experiments using the CRISPR interference (CRISPRi) system. In this approach, catalytically inactive Cas9 is fused to a transcriptional repressor and targeted to a gene’s promoter, resulting in down-regulation of gene expression without the generation of a complete loss of function–inducing frameshift mutation (25). We designed three gRNAs that recognized HDAC6, MAPK14, PAK4, PBK, and PIM1 and verified that these constructs blocked the expression of their targets (fig. S9A). We then conducted competition experiments in four different cell lines, and we found that gRNAs targeting the essential replication protein MCM2 exhibited ~10- to ~20-fold dropout, whereas gRNAs targeting HDAC6, MAPK14, PAK4, PBK, and PIM1 failed to deplete (fig. S9B). These assays further verify that our results cannot be explained by the existence of different cellular responses to partial and complete loss-of-function alterations.

Assessing the sensitivity of target-KO clones to chemotherapy agents undergoing combination clinical trials

Several of the proteins listed in Table 1 are currently undergoing combination clinical trials using their targeted inhibitors together with other chemotherapy agents. It is conceivable that a protein could be nonessential under normal conditions but that its loss sensitizes cells to specific chemotherapies. For instance, HDAC6 is capable of deacetylating microtubules (26), and HDAC6 inhibition has been reported to render cells vulnerable to drugs that interfere with microtubule dynamics (27). As a result of this preclinical work, two clinical trials are combining HDAC6 inhibitors with the microtubule stabilizer paclitaxel (NCT02632071 and NCT02661815). We therefore tested whether the KO clones that we had generated were sensitive to various chemotherapy agents (fig. S10, A to C). In contrast to previous results, loss of HDAC6 failed to sensitize cells to paclitaxel or to four other anticancer drugs (fig. S10A). Similarly, p38α inhibitors have been clinically applied in combination with bortezomib, gemcitabine, carboplatin, and temozolomide (NCT00087867, NCT00095680, NCT01663857, and NCT02364206), but MAPK14/p38α KO clones in multiple cell lines were as sensitive to these agents as Rosa26 control clones (fig. S10B). These results suggest that, in addition to being nonessential, these putative drug targets do not affect sensitivity to several chemotherapy agents that have been tested in combination trials.

Assessing RNAi promiscuity as a cause of the misidentification of cancer dependencies

If these genes do not drive cancer growth or chemotherapy resistance, then why have inhibitors targeting the proteins that they encode been tested in human patients with cancer? A review of the literature indicates that each of these genes has been described to be essential on the basis of RNAi-induced knockdown phenotypes (data file S1). Off-target toxicity has been reported to be a common problem in the design and interpretation of RNAi-based experiments (2830), although the impact of these issues on the therapeutic development pipeline is not known. We acquired four different RNAi constructs that were used in these prior studies and then tested their effects on the clones that we had generated. Although we were able to confirm that each construct decreased the expression of its putative target, we also found that these constructs impaired proliferation in both WT clones and clones in which the construct’s target had been knocked out (fig. S11, A to C). For example, a recent report found that PAK4-targeting siRNAs blocked cell division in HCT116 colon cancer cells and concluded that PAK4 was a genetic dependency in this cell line (31). However, we found that these same siRNAs induced an equivalent decrease in proliferation in both HCT116 PAK4-KO and HCT116 Rosa26 clones, suggesting that their effects on growth are a consequence of off-target toxicity (fig. S11A). Similarly, while knocking down PIM1 has been reported to block proliferation in the MDA-MB-231 breast cancer cell line (15), this construct had the same effect in MDA-MB-231 PIM1-KO cells (fig. S11B). Our results therefore suggest that these drug targets have advanced to clinical testing due, at least in part, to promiscuous RNAi constructs.

Assessing the specificity of cancer drugs undergoing clinical trials

Off-target toxicity from small-molecule drugs can cause dangerous side effects and is a major cause of clinical trial failure (32, 33). Our results suggested that the drugs listed in Table 1 were designed to target nonessential cellular proteins, raising the possibility that the anticancer effects of these drugs could be due to off-target interactions. We therefore sought to apply CRISPR to differentiate between the on- and off-target effects of each clinical cancer drug. First, we confirmed that CRISPR could be used to verify the MOA for several genetically validated therapies. The natural product rapamycin is reported to bind to the prolyl isomerase FKBP12, and this complex inhibits the essential mTOR (mammalian target of rapamycin) kinase (fig. S12A) (34, 35). We knocked out FKBP12 using CRISPR, and we verified that these KO clones exhibited increased resistance to rapamycin treatment (fig. S12, B and C). Similarly, knocking out p53 conferred resistance to the experimental p53-activating drug nutlin-3a (fig. S12, D to F). Last, we sought to test whether CRISPR could be used to validate a published resistance-granting point mutation. We used CRISPR-mediated homology-directed repair (HDR) to introduce a missense mutation into the kinase domain of the essential mitotic kinase MPS1, and we verified that this substitution was capable of granting resistance to the small-molecule MPS1 inhibitor AZ3146 (fig. S12, G to I) (36). Thus, CRISPR-derived KO and knock-in cell lines can be used to validate on-target drug activity.

Next, we applied CRISPR to interrogate the MOA of two caspase-3 activating compounds: PAC-1 and 1541B. These drugs are reported to function by catalyzing the conversion of caspase-3 from its inactive, procaspase state to its active, cleaved form, thereby causing cellular apoptosis (fig. S13A) (16, 17). Currently, PAC-1 is undergoing three different clinical trials in patients with cancer (NCT02355535, NCT03332355, and NCT03927248). We knocked out the CASP3 gene in four different cell lines and then verified protein ablation using two different antibodies (Fig. 3A and fig. S13). However, these CASP3-KO lines exhibited identical sensitivity to PAC-1 and 1541B compared with Rosa26 controls (Fig. 3B and fig. S13, D and F). These results suggest that a putative caspase-3 activator undergoing clinical trials actually kills cancer cells in a caspase-3–independent manner.

Fig. 3 Target-independent cell killing by multiple anticancer drugs.

(A) Western blot analysis for caspase-3 in A375 and HCT116 cells. (B) Seven-point dose-response curves of Rosa26 and CASP3-KO A375 and HCT116 cells in the presence of two putative caspase-3 activators: 1541B and PAC-1. (C) Seven-point dose-response curves of Rosa26 and HDAC6-KO A375 and DLD1 cells in the presence of two putative HDAC6 inhibitors: ricolinostat and citarinostat. (D) Seven-point dose-response curves of Rosa26 and MAPK14-KO A375 and DLD1 cells in the presence of two putative MAPK14 inhibitors: ralimetinib and SCIO-469. (E) Seven-point dose-response curves of Rosa26 and PBK-KO A375 and DLD1 cells in the presence of two putative PBK inhibitors: OTS514 and OTS964. (F) Seven-point dose-response curves of Rosa26 and PIM1-KO A375 and DLD1 cells in the presence of a putative PIM1 inhibitor: SGI-1776. (G) Seven-point dose-response curves of Rosa26 and PAK4-KO A375 and HCT116 cells in the presence of a putative PAK4 inhibitor: PF-3758309.

We next tested each putative HDAC6, MAPK14, PAK4, PBK, and PIM1 inhibitor in control and KO clones. If these drugs act by specifically inhibiting their reported targets, then cancer cells that totally lack the expression of their targets would be expected to be resistant to these drugs’ effects. In contrast, if a drug kills cells in which its reported target has been knocked out, then this drug necessarily kills cells by affecting another protein or proteins. In every instance that we tested, cancer cells in which HDAC6, MAPK14, PAK4, PBK, or PIM1 had been knocked out exhibited WT sensitivity to their putative targeted inhibitors (Fig. 3 and fig. S14, A and B). For example, we found that the PAK4 inhibitor PF-3758309 blocked the growth of both Rosa26 and PAK4-KO melanoma cells with a GI50 value of ~9 nM (Fig. 3G). Given that this drug is fully capable of killing cells in which its putative target has been deleted, the ability of PF-3758309 to block cancer cell growth must be through an off-target effect. To further interrogate whether the drugs studied in this manuscript could exhibit an on-target MOA in an additional genetic background, we knocked out HDAC6 in TOV-21G cells, an ARID1A-mutant ovarian cancer cell line in which this gene has been reported to be a dependency (fig. S15A) (6). However, TOV-21G HDAC6-KO cells exhibited WT fitness in vitro and in soft agar (fig. S15, B and C), and these cells remained sensitive to citarinostat and ricolinostat, two putative HDAC6 inhibitors in clinical trials (fig. S15D). In total, all 10 different anticancer agents targeting CASP3, HDAC6, MAPK14, PAK4, PBK, or PIM1 exhibited clear evidence of target-independent cell killing in every KO cell line that we examined.

Last, we applied these putative inhibitors to investigate several combination chemotherapy trials. As described above, HDAC6 inhibitors are currently undergoing testing in cancer patients along with paclitaxel, and p38α inhibitors have also been combined with several different therapeutic agents. We verified that cotreatment with a targeted inhibitor and a second agent generally caused a greater decrease in cancer cell viability than either agent alone (fig. S14B). However, this synthetic enhancement was observed in both Rosa26 and target-KO clones, suggesting that these additive effects are also due to an off-target interaction.

Discovering the true target of OTS964

If these clinical anticancer therapies do not kill cells by inhibiting their reported targets, then how do they block cancer growth? We note that, although the MOA of each drug has previously been characterized using biochemical and biophysical approaches, there is little genetic evidence linking each drug to its reported target. We hypothesized that an alternative genetic methodology could shed light on the true target of a therapeutic agent whose MOA was in question.

For this work, we chose to focus on the putative PBK inhibitor OTS964, because it exhibited nanomolar potency in multiple cancer types and our CRISPR experiments had provided clear evidence that PBK was not required for cell proliferation. Moreover, OTS964 has been reported to affect mitotic progression (13), and antimitotic drugs have been historically proven to be highly successful anticancer agents (37). To identify mutations that conferred resistance to OTS964, we used HCT116 colorectal cancer cells, which harbor an increased mutation rate caused by a defect in mismatch repair (38). We cultured HCT116 cells in the presence of a lethal concentration of OTS964 and successfully isolated 12 clones that were capable of growing under these conditions (Fig. 4A). We found that these clones exhibited stable resistance to OTS964, as they failed to revert to OTS964 sensitivity after prolonged growth in normal medium (fig. S16A). Cancer cells commonly acquire chemotherapy resistance by amplifying the P-glycoprotein drug efflux pump (39). However, the OTS964-resistant clones remained sensitive to paclitaxel, a verified P-glycoprotein substrate, suggesting that they had not acquired a multidrug resistance phenotype (fig. S16B) (40). These experiments indicated that our drug-resistant clones could harbor a mutation or mutations that specifically altered OTS964 sensitivity.

Fig. 4 Discovery of CDK11 as the in cellulo target of the mischaracterized anticancer drug OTS964.

(A) A schematic of the strategy to use the highly mutagenic HCT116 cell line to isolate mutations that confer OTS964 resistance. (B) Sanger sequencing validation of two heterozygous mutations in the CDK11B kinase domain. (C) Constructs used to introduce the G579S mutation into CDK11B via CRISPR-mediated HDR. Yellow arrowhead indicates the site of Cas9 cleavage. Red bar indicates the G579S substitution, and blue bars indicate silent mutations introduced to prevent recutting after HDR. (D) Crystal violet staining of cancer cells transfected with the indicated constructs and then cultured in a lethal concentration of OTS964. (E) Seven-point dose-response curves of Rosa26, PBK-KO, and CDK11BG579S clones grown in varying concentration of OTS964. (F) Titration experiments reveal that OTS964 binds to CDK11B with a KD of 40 nM. (G) Pancreatic cancer cell line MiaPaca-2 was transduced with guides specific for CDK11A, guides specific for CDK11B, or guides that harbored cut sites in both genes. (H) A375 H2B-mCherry cells (left) or A375 H2B-mCherry cells that express CDK11BG579S (right) were arrested at G1/S with a double-thymidine block and then were released into normal medium or medium containing OTS964. The percentage of mitotic cells in each population was scored every hour. (I) Representative images of the experiments in (H), 9 hours after release from thymidine. Scale bar, 25 μm.

To identify genetic alterations capable of conferring OTS964 resistance, we subjected 10 OTS964-resistant clones, 1 Rosa26 control clone, and the parental cell population to whole-exome sequencing (WES). Notably, all 10 resistant clones were found to harbor heterozygous missense mutations in the poorly characterized cyclin-dependent kinase (CDK) CDK11B (fig. S16C). Eight clones harbored two mutations in this gene, H572Y and G579S, in trans, whereas two clones harbored only the G579S substitution. No CDK11B mutations were observed in the parental population or in the Rosa26 control clone. Sanger sequencing verified the presence of the CDK11B mutations in two independent drug-resistant clones that were not subjected to WES (Fig. 4B and fig. S16C). These mutations were also absent from additional control clones that were analyzed (fig. S16C) and have not been previously observed in the Catalog of Somatic Mutations in Cancer database (41).

The human genome encodes two CDK11 proteins, CDK11A and CDK11B, that are 97% identical and that arose from an evolutionarily recent gene duplication event (42). The CDK11 family has been reported to support various cellular processes, including transcription, splicing, and chromosome segregation (43), but its role in cancer is unknown. No drugs have previously been reported to target CDK11, and inhibitors that are specific for single CDKs are difficult to discover due to the sequence similarity among these kinases (44). We aligned the sequences of the human CDKs, and we noted that 19 of 20 of these proteins harbored an alanine residue immediately upstream of the magnesium-coordinating DFG motif (fig. S16D). Only CDK11 contained a glycine at this location, and this glycine was mutated to serine in every OTS964-resistant clone that we sequenced (fig. S16, C and D). This amino acid position (called “xDFG”) has previously been identified as a key residue that affects kinase inhibitor binding (45), suggesting a potential basis for CDK11-selective inhibition. To test whether the xDFG Gly→Ser mutation was sufficient to confer resistance to OTS964, we designed a strategy to use CRISPR-mediated HDR to introduce this substitution into the endogenous CDK11B gene (Fig. 4C). These experiments revealed that this point mutation was sufficient to restore viability in A375, A2780, DLD1, and MDA-MB-231 cancer cells grown in a lethal concentration of OTS964 (Fig. 4, D and E). To verify that our results were not an off-target effect of CRISPR, we generated a retrovirus to stably express CDK11BG579S cDNA, and we confirmed that this construct was also sufficient to confer OTS964 resistance (fig. S16E). In an HCT116 clone that had spontaneously evolved resistance to OTS964, eliminating mutant CDK11B with CRISPR restored OTS964 sensitivity, demonstrating that this alteration is both necessary and sufficient for drug resistance (fig. S16F). Introducing an alanine substitution into residue 579, so that the CDK11B xDFG motif was identical to the other human CDKs, was also sufficient to decrease the efficacy of OTS964 (fig. S16, G and H). Last, to confirm a direct interaction between OTS964 and CDK11B, we assessed the binding of OTS964 to different CDKs. OTS964 bound to CDK11B with a KD (dissociation constant) of 40 nM, and it displayed greater than 10-fold selectivity for this kinase compared with several other CDKs (Fig. 4F and fig. S16I). In total, these results indicate that the putative PBK inhibitor OTS964 actually functions by targeting CDK11, and its specificity for this kinase is conferred by CDK11’s distinct xDFG motif.

Discovering the consequences of CDK11 inhibition in cancer

We next determined the cellular effects of OTS964 treatment and CDK11 ablation with CRISPR. In cell competition assays, cancer cells transduced with gRNAs specific for either CDK11A or CDK11B exhibited minimal dropout. However, guides designed to recognize both isoforms exhibited substantial dropout in every cell line that we tested, including pancreatic cancer and triple-negative breast cancer (Fig. 4G and fig. S17A). Flow cytometry revealed that cells transduced with pan-CDK11 guides accumulated in G2/M with 4C DNA content, suggesting that CDK11 function is required for mitotic progression (fig. S17B). To test whether OTS964 phenocopied the CDK11 gRNAs, we arrested A375 cells expressing the chromosomal marker H2B-mCherry at G1/S with a double-thymidine block and then released them into normal medium or medium containing OTS964. Cells treated with a low concentration of OTS964 exhibited delayed nuclear envelope breakdown and progressed slowly through mitosis (Fig. 4H, fig. S17C, and movies S1 and S2). Cells treated with a lethal concentration of OTS964 arrested in G2, before mitotic entry (movie S3). OTS964 treatment did not perturb DNA replication, because the arrested cells displayed 4C DNA content and did not accumulate 53BP1 foci, a marker of DNA damage (fig. S17, D and E). Introducing the G579S substitution into A375 cells rescued normal mitotic entry and progression in the presence of a lethal concentration of OTS964 (Fig. 4, H and I, fig. S17C, and movie S4). These results establish CDK11 activity as necessary for mitosis in human cancer and suggest that CDK11 is the key in cellulo target of OTS964.

DISCUSSION

It is generally known that small molecules can exhibit off-target effects that may confound the design of specific chemical inhibitors (46). Our data suggest that, rather than simply being the side effect of a drug, these off-target interactions are frequently the mechanism by which small molecules block cancer growth. Every inhibitor tested in this manuscript that lacked a previously described resistance mutation was found to kill through an off-target effect; these results therefore identify this phenomenon as a common problem that affects cancer clinical trials. As 97% of drug-indication pairs tested in clinical trials fail to receive U.S. Food and Drug Administration (FDA) approval (5), the misidentification of essential genes in cancer and the mischaracterization of reportedly target-specific inhibitors likely contribute to their exceedingly high failure rate. The adoption of more stringent genetic target and activity validation studies may alleviate this problem and decrease the failure rate of new cancer drugs.

Each gene that we studied has been reported to be required in a cell-autonomous manner for cancer proliferation by more than 180 publications, and it is this specific claim that our work sought to test (data file S1). Toward this end, we generated KO clones in at least three cancer cell lines, we performed CRISPR competition assays in 32 cell lines, and we performed CRISPRi-knockdown competition assays in an additional four cell lines, which all consistently demonstrated that HDAC6, MAPK14, PAK4, PBK, and PIM1 are dispensable for cancer cell fitness. Although cancer cells harbor the ability to evolve in response to various perturbations, we do not believe that this adaptability is sufficient to explain the robust growth of the CRISPR-modified cells that we have generated. First, in the competition assays that we conducted, cells are analyzed immediately after gRNA transduction, allowing them no time to adapt to the loss of the targeted gene. For instance, although BRAF-addicted melanoma cells are capable of evolving BRAF independence over time by acquiring secondary mutations in MEK or NRAS (47), we still observed a strong depletion of BRAF-targeting guides in these dropout assays. Second, we knocked out the verified drug target MEK1 and confirmed that MEK1-KO clones grow substantially more slowly than Rosa26 control clones, demonstrating our ability to validate genetic dependencies in CRISPR-modified clones. Third, although each of the genes that we studied has previously been reported to be essential, our experiments provide a mechanism to explain these discrepant results. In particular, we demonstrate that several RNAi constructs previously used to inhibit these genes exhibit identical antiproliferative effects in target-WT and target-KO cancer cells, suggesting that RNAi promiscuity contributed to the misidentification of these genes as drug targets. Fourth, although it has been proposed that cells can compensate for CRISPR-induced mutations by up-regulating homologs of the targeted gene (23), we failed to detect any evidence of this in our KO clones or in genetically modified cell lines from several independent laboratories. Last, high-throughput screens conducted in hundreds of cancer cell lines using both RNAi and CRISPR technologies have also failed to identify these genes as cancer dependencies. We therefore believe that cellular evolution after CRISPR mutagenesis is unlikely to explain the robust growth of the cancer cells lacking the drug targets that we have studied.

The cell lines studied in this paper were chosen based on the literature on each target, but we have not attempted to recapitulate every individual published result with each drug or drug target. Thus, it remains possible that these drug targets exhibit a cell type–specific dependency pattern not uncovered in this work. To partially address this concern, we analyzed published whole-genome CRISPR and RNAi screening data from >700 cell lines, which consistently revealed that the genes studied in this work could be eliminated without substantially affecting cell fitness. Because of the breadth of cell lines tested both within our laboratory and through high-throughput screening, it is unlikely that these genes are genetic dependencies in a common cancer lineage. Nonetheless, we do not rule out the possibility that these genes are essential in a rare cancer type not included among those studied here. In addition, it remains possible that these genes play a non–cell-autonomous role during tumorigenesis. For instance, whereas MAPK14/p38α has been reported to be essential for proliferation in breast cancer (9, 48), colon cancer (8), ovarian cancer (49), and several other cancer types, it has also been proposed to mediate inflammatory signaling (50). Thus, although our work provides strong evidence that these proteins are dispensable for cancer proliferation, we do not rule out the possibility that these proteins have some function in other non–cell-autonomous processes related to tumor development in vivo.

Our results indicate that many cancer drugs in clinical trials kill cells independently of their reported targets. As the application of a predictive biomarker doubles the likelihood that a clinical trial will succeed (5), the inability to decipher a drug’s true target may prevent successful biomarker identification and contribute to trial failure. Moreover, our findings may provide evidence that cancer drug polypharmacology is a common MOA for reportedly target-specific compounds. For example, whereas ricolinostat has been reported to be a selective HDAC6 inhibitor (51), our work shows that HDAC6 expression is fully dispensable for ricolinostat sensitivity. These results are similar to those reported in (52), which found that ricolinostat continued to kill HDAC6-KO HAP1 cells. The human genome harbors 18 different histone deacetylases (53), and it is possible that this drug kills cells by inhibiting HDAC6 and several synthetically redundant HDAC family members. In addition, the invalidation of a drug’s putative target does not necessarily mean that a drug will be ineffective in the clinic, as some broadly nonspecific inhibitors have been proven efficacious in certain circumstances. In many cases, these successes derive from a thorough understanding of a drug’s MOA. For instance, the multitargeted kinase inhibitor midostaurin has received FDA approval for use in FLT3+ leukemias due to its demonstrated activity against FLT3 (54). Thus, strong validation of on-target drug activity remains essential.

Alternately, these mischaracterized drugs may kill cells by inhibiting single, specific proteins that are not closely related to their reported targets. For instance, our results demonstrate that the putative PBK inhibitor OTS964 functions by blocking CDK11 activity. Several CDK inhibitors have received FDA approval or are in late-stage trials for various malignancies, underscoring the clinical potential for targeting members of the CDK family (55). However, no CDK11-specific inhibitors have been previously described (56). OTS964 and its derivatives could therefore allow us to block cancer growth by inhibiting a previously undruggable mitotic CDK. Furthermore, our work identifies CDK11’s distinct xDFG motif as a key determinant of drug sensitivity, suggesting a potential structural basis for CDK11-specific inhibition. Although CDK11 has previously been reported to function in chromosome segregation (57), our results demonstrate that its activity is required for entry into mitosis. It will therefore be crucial to investigate whether CDK11 inhibitors are capable of synergizing with PLK1 inhibitors, Aurora A inhibitors, or any other drugs that similarly target mitotic entry (58). Last, the CDK11 locus on chromosome 1p has been reported to be deleted or translocated in several cancer types, including melanoma and neuroblastoma (43), raising the exciting possibility that alterations in this gene family could serve as predictive biomarkers for CDK11 inhibitor sensitivity.

More broadly, our results underscore the power of genetic approaches to improve the preclinical characterization of cancer drugs and drug targets. In particular, CRISPR-mediated gene editing is a powerful methodology for interrogating the effects of loss-of-function alterations in disease-relevant genes, and head-to-head comparisons have verified that CRISPR is less susceptible to off-target effects than RNAi (59, 60). Although biochemical and biophysical approaches can demonstrate target engagement by a potential therapeutic molecule, these assays alone are insufficient to demonstrate the relevance of this interaction in cellulo. Mutagenesis experiments, using either spontaneous or CRISPR-directed approaches, can complement these assays to verify or discover a drug’s true MOA and indicate potential biomarkers for sensitivity and resistance. We suggest that the adoption of stringent genetic characterization assays in the preclinical setting will decrease the number of drugs used in human cancer patients that fail to provide any clinical benefit.

MATERIALS AND METHODS

Study design

In this work, we sought to determine whether several drug targets were truly essential for cancer cell fitness. After discovering that many of these drug targets were nonessential, we investigated whether the drugs used to target them killed cells through an off-target effect. The cell lines used in this study were selected based on an analysis of the literature on these drug targets (summarized in data file S1). No predetermined sample sizes were used for this analysis. No randomization or blinding was performed. Cell competition experiments represent single biological replicates. Proliferation assays represent two to three biological replicates. Drug sensitivity curves were generated with three to six technical replicates. Soft agar experiments represent three technical replicates with at least 15 independent fields counted for each experiment. Raw data for the cell competition experiments are included in data file S2.

Selecting drug targets to study

Our laboratory previously investigated the role of MELK in cancer (3, 4). We found that, contrary to previous results obtained with RNAi, cancer cells tolerated CRISPR-induced ablation of MELK with no loss in cell fitness. In addition, we found that OTS167, a small-molecule inhibitor of MELK in clinical trials, killed cells in a MELK-independent manner. These findings led us to investigate whether MELK and OTS167 were aberrations or whether other drugs and drug targets had been similarly mischaracterized.

To begin this project, we generated a list of drugs and drug targets to study. We constructed this list using a few criteria, informed in part by our experience studying MELK. First, we sought to identify cancer genes that reportedly played a cell-autonomous role in cancer growth, so that we could study the most relevant phenotypes that resulted from their ablation in cell culture. Thus, we did not consider drugs that primarily target angiogenesis, the immune checkpoints, or related in vivo processes. (Importantly, our work does not rule out in vivo roles for the genes studied in this paper.) Second, we only considered drugs that were reported to act by targeting single, specific proteins. If a drug was believed to act by inhibiting multiple proteins, then this would confound our CRISPR experiments, because the genetic ablation of a single gene would not be expected to phenocopy the effects of the inhibitor. Third, we focused on genes that were reported to have broad dependency patterns, allowing us to study the consequences of their inhibition in a wide range of cell lines. Fourth, we focused on genes that had been largely characterized using RNAi, although we did not exclude genes that had been previously studied using CRISPR, transgenic mice, dominant-negative alleles, or other approaches. Fifth, we chose drugs that were in advanced preclinical or clinical testing. Sixth, we posited that the gold standard for showing on-target drug activity was the identification of a mutation that confers resistance to a targeted inhibitor, and we sought to study drugs that lacked known resistance-granting mutations.

Using these criteria, we searched PubMed, the database of American clinical trials (https://clinicaltrials.gov/), and other related resources (61) for drugs and drug targets that fit these criteria. We did not aim to comprehensively identify every drug that met the above criteria but instead chose to limit ourselves to a small number of targets such that we could deeply characterize each one. Using these approaches, we chose to study CASP3, HDAC6, MAPK14, PAK4, PBK, and PIM1. We first became aware of the putative PBK inhibitor OTS964 because it was developed by the same company that created the MELK inhibitor that we previously studied. A press release on this drug reported that clinical trials would soon be initiated, although, to our knowledge, these clinical trials have not yet begun (62). In addition, our list of drugs initially included the putative PAK4 inhibitor KPT-9274. However, while performing the research described in this paper, a report was published that identified mutations in NAMPT that granted resistance to KPT-9274; therefore, we did not further pursue this compound (63). After the initial submission of this manuscript, a second group independently demonstrated that the putative HDAC6 inhibitor ricolinostat kills HDAC6-KO HAP1 cancer cells, in accordance with our results (52).

We also chose to study two drugs, PAC-1 and 1541B, that reportedly function by activating the apoptosis protein caspase-3. Although caspase-3 is not considered to be a “cancer dependency,” its activation by small molecules has been reported to trigger cancer cell apoptosis (16, 17). The first activator, PAC-1, was introduced in a publication that used in vitro methods to demonstrate the ability of PAC-1 to induce procaspase-3 cleavage and activation. Furthermore, the group implied specificity of PAC-1 for caspase-3 by showing that it exhibited a higher IC50 (median inhibitory concentration) value in MCF7 cells, a caspase-3–deficient breast cancer cell line, compared with caspase-3–expressing cell lines (16). After this study, however, concerns were raised over whether caspase-3 activation was the true MOA of PAC-1. In a letter to Nature Chemical Biology, another laboratory reportedly failed to see substantial activation of procaspase-3 in in vitro studies of PAC-1 (64). In addition, the group stated that the caspase-3–deficient cell line, MCF7, displayed similar sensitivities to the drug as two caspase-3–expressing cell lines. Although this letter raised questions as to the true MOA of PAC-1, the original developers of this compound disputed these concerns, arguing that the different in vitro results were a consequence of different procaspase concentrations and buffer conditions used in the assays (65). In support of this claim, in later publications, the group detailed the in vitro MOA of PAC-1 as a zinc chelator, describing a mechanism where zinc prevents procaspase-3 activation; therefore, the in vitro efficacy of PAC-1 is highly dependent on the concentration of zinc in the buffer (66). This group also argued that the activity of PAC-1 against the MCF7 cell line only occurs under conditions of low cell density and high drug concentration and that the mechanism of death seems to resemble necrosis more than caspase-mediated apoptosis (65). A number of other publications using PAC-1 and second-generation caspase-3 activators reported a similar resistance of MCF7 cells to caspase-3–activating compounds (17, 67). On the basis of this evidence, many in the field have continued to use PAC-1, not only in biological investigations but also in a number of clinical trials, under its listing as a caspase-3 activator (Table 1 and data file S1). In contrast to PAC-1, later caspase-3 activators, namely 1541, were reported to have direct interactions with caspase-3. 1541, the parental compound to the 1541B inhibitor used in our study, was not only able to induce caspase-3 activation in in vitro conditions where PAC-1 exhibited no effect, but 1541 resistance mutations in the CASP3 gene were also described (17). Because caspase-3 deficiencies have been linked with decreased sensitivity to a wide range of chemotherapeutic agents, we considered it possible that this putative resistance could be caused by an indirect effect on apoptosis (68, 69). Thus, because of the controversy and conflicting data concerning the MOA of different caspase-3 activators, we decided to include these compounds in our study.

Cell culture

The sources of each cell line are listed in data file S3. The identities of all human cell lines used in this study were confirmed using STR (short tandem repeat) profiling (University of Arizona Genetics Core). A375, A549, A673, Cal51, Cama1, DLD1, HCT116, HEK293T, MDA-MB-231, PC3, SK-MEL-28, and U87 cell lines were cultured in Dulbecco’s modified Eagle’s medium (DMEM) (Thermo Fisher Scientific, cat. no. 11995-073) supplemented with 10% fetal bovine serum (FBS) (Sigma-Aldrich, cat. no. F2442), 2 mM glutamine (Lonza, cat. no. 17-605F), and penicillin and streptomycin (100 U/ml; Life Technologies, cat. no. 15140122). A2780, DU-145, HCC1143, HCC38, HT29, K562, LNCaP, MDA-MB-453, MDA-MB-468, NCI-H82, OVK18, OVTOKO, SUIT2, SW480, and TOV-21G cell lines were cultured in RPMI 1640 (Lonza, cat. no. 12-115F/12) supplemented with 10% FBS, 2 mM glutamine, and penicillin and streptomycin (100 U/ml). MiaPaCa-2 cells were cultured in DMEM supplemented with 10% FBS, 2.5% horse serum (Thermo Fisher Scientific, cat. no. 26050088), 2 mM glutamine, and penicillin and streptomycin (100 U/ml). MCF7 cells were cultured in DMEM supplemented with 10% FBS, insulin (0.01 mg/ml; Thermo Fisher Scientific, cat. no. 12585-014), 2 mM glutamine, and penicillin and streptomycin (100 U/ml). T24 cells were cultured in McCoy’s 5A medium (Life Technologies, cat. no. 16600-108) supplemented with 10% FBS, 2 mM glutamine, and penicillin and streptomycin (100 U/ml). HepG2 cells were cultured in Eagle’s minimum essential medium (American Type Culture Collection; cat. no. 30-2003) supplemented with 10% FBS, 2 mM glutamine, and penicillin and streptomycin (100 U/ml). RPE1 cells were cultured in DMEM/F12 (Thermo Fisher Scientific, cat. no. 11320-033) supplemented with 10% FBS, 2 mM glutamine, and penicillin and streptomycin (100 U/ml). Sum149 cells were cultured in Ham’s F12 medium (Lonza, cat. no. 12-615F) supplemented with 5% FBS, insulin (0.01 mg/ml), hydrocortisone (500 ng/ml; STEMCELL Technologies, cat. no. 07926), 2 mM glutamine, and penicillin and streptomycin (100 U/ml). Additional details on the conduct of these studies are included in Supplementary Materials and Methods.

Statistical analysis

For box plots, the boxes represent the 25th, 50th, and 75th percentiles of the colonies per field, whereas the whiskers represent the 10th and 90th percentiles. In fig. S4, a Student’s t test (two sided) was used to compare control and MEK1-KO colony formation efficiency.

SUPPLEMENTARY MATERIALS

stm.sciencemag.org/cgi/content/full/11/509/eaaw8412/DC1

Materials and Methods

Fig. S1. Drug target ablation with CRISPR-Cas9.

Fig. S2. Cell competition assays in multiple cancer types.

Fig. S3. Verification of drug target-KOs.

Fig. S4. Knocking out the verified genetic dependency MEK1 in A375 melanoma cells.

Fig. S5. MDA-MB-231 clonal analysis.

Fig. S6. Analysis of homolog gene expression in CRISPR-KO clones.

Fig. S7. Analysis of homolog gene expression in published RNA-seq experiments.

Fig. S8. Assessing putative cancer dependencies in whole-genome CRISPR and RNAi screens.

Fig. S9. Targeting several putative cancer dependencies with CRISPRi.

Fig. S10. Lack of sensitivity to several clinical chemotherapy agents in putative cancer dependency KOs.

Fig. S11. Target-independent toxicity of RNAi reagents previously used to investigate several putative cancer dependencies.

Fig. S12. Using CRISPR to validate the MOA of several anticancer drugs.

Fig. S13. Off-target toxicity of two caspase-3–activating compounds in CASP3-KO clones.

Fig. S14. Target-independent cancer cell killing in single-agent and combination therapy experiments.

Fig. S15. Off-target toxicity of two putative HDAC6-inhibiting compounds in HDAC6-KO ovarian cancer clones.

Fig. S16. A mutation in the xDFG residue of CDK11B in OTS964-resistant clones.

Fig. S17. Requirement for CDK11 activity for progression through mitosis.

Data file S1. Literature supporting the designation of HDAC6, MAPK14, PAK4, PBK, and PIM1 as cancer genetic dependencies and CASP3 as a drug target.

Data file S2. Cell competition assay results.

Data file S3. Sources of the cell lines used in this manuscript.

Data file S4. CRISPR gRNA sequences.

Data file S5. CRISPRi gRNA sequences.

Data file S6. Quantitative polymerase chain reaction primers.

Data file S7. Antibody sources and concentrations.

Data file S8. Drugs and drug sources.

Movie S1. A375 cells expressing H2B-mCherry released from a double-thymidine block into normal medium.

Movie S2. A375 cells expressing H2B-mCherry released from a double-thymidine block into medium with 25 nM OTS964.

Movie S3. A375 cells expressing H2B-mCherry released from a double-thymidine block into medium with 100 nM OTS964.

Movie S4. A375CDK11B-G579S cells expressing H2B-mCherry released from a double-thymidine block into medium with 100 nM OTS964.

References (70257)

References and notes

Acknowledgments: We thank the following individuals for sharing plasmids and cell lines used in this work: C. Vakoc (CSHL), C. dos Santos (CSHL), K. Chang (CSHL), L. Trotman (CSHL), M. Jost and J. Weissman (UCSF), J. van den Berg and R. Medema (Netherlands Cancer Institute), H. Goodarzi (UCSF), D. Solomon (UCSF), D. Livingston (Harvard), J. Lum (University of Victoria), H. Tsao and K. Flaherty (MGH), S. Karuppaiyah (Ohio State University), and S. Murphy (Duke University). Funding: Research in the Sheltzer Lab is supported by an NIH Early Independence Award (1DP5OD021385), a Breast Cancer Alliance Young Investigator Award, a Damon Runyon-Rachleff Innovation Award, a Gates Foundation Innovative Technology Solutions grant, and a CSHL-Northwell Translational Cancer Research grant. This work was performed with assistance from the CSHL Flow Cytometry Shared Resource, which is supported by the CSHL Cancer Center Support grant 5P30CA045508. A.L. and C.J.G. are supported by NSF Graduate Research Fellowships. Author contributions: A.L., C.J.G., and J.M.S. designed the experiments, performed the experiments, interpreted the results, and wrote the paper. A.P., K.M.J., C.A., M.L.Y., E.L.S., D.A.L., L.L., A.R.C., Z.C.G., and C.T. performed the experiments and interpreted the results. Competing interests: The authors declare that they have no competing interests. Data and materials availability: All data associated with this study are present in the paper or the Supplementary Materials.
View Abstract

Navigate This Article