Aurora A Kinase Inhibition Is Synthetic Lethal with Loss of the RB1 Tumor Suppressor Gene

An attractive strategy for cancer therapeutic discovery is to target enzyme functions that are dispensable in normal cells, but become essential for the survival of cells with mutated proto-oncogenes and tumor suppressor genes due to oncogene addiction (1), collateral vulnerability (2), and, more generally, synthetic lethality (3). Although numerous drugs that exploit oncogene addiction have proved successful for the treatment of cancer, mutated tumor suppressor genes (TSG) have thus far proved more challenging. A notable exception is the synthetic lethal interaction between BRCA-mutant cancer cells and PARP inhibitors, and it has been proposed that this example should encourage the search for synthetic lethal “gene–drug” interactions to target other TSGs in cancer (4).

One approach to the identification of new, synthetic lethal, gene–drug relationships is to perform genetic loss-of-function screens to identify dependencies, unique to cells with the mutated TSG, on genes that encode druggable enzymes (5). Drugs directed against the enzyme functions of hits from these screens must then be developed and tested to ensure that the gene–gene synthetic lethal relationship from the screen is preserved as a gene–drug interaction, not only in the screening models, but also, more broadly, among cells representing the clinical diversity of cancers with the TSG mutation. Indeed, depletion of a gene, or its mRNA (and therefore protein), as occurs in genetic screens, can have very different phenotypic consequences to inhibition of its encoded enzyme function with a drug, and examples of discrepancies between gene knockdown and enzyme inhibition are well known (6). Most importantly, genetic screens will miss genuine synthetic lethal relationships between a mutated cancer gene and enzyme blockade if the protein has additional functions, beyond the enzyme activity, which are essential to viability of all cells. Ultimately, of course, it is the “gene–drug” synthetic lethality that is required for therapeutic application, so an appealing alternative approach to “gene–gene” screens is to directly screen drug-like compounds across large panels of cancer cell lines to determine whether TSG-mutated cancers display enhanced vulnerability to particular drugs. Such gene–drug screens avoid many of the shortcomings of gene–gene screens but require that sufficiently specific, cell-active enzyme inhibitors against the relevant target already exist.

The prototypical TSG, retinoblastoma or RB1, is mutated in some of the most aggressive and hard-to-treat malignancies, including small-cell lung cancer (SCLC) and triple-negative breast cancer (TNBC; ref. 7). The function of the RB1 product, RB1, in controlling the G₁–S transition in the cell cycle is well understood (8). New cell cycles are triggered by mitogens and hormones that activate G₁ cyclins and these, in turn, promote the phosphorylation and neutralization of RB1. RB1 loss, therefore, subverts the normal requirement for external growth cues, and cancers with RB1 mutation are expected to be refractory to cancer therapeutics acting on upstream mitogen and hormone pathways. Indeed, RB1 loss can emerge as a mechanism of resistance to EGFR, CDK4, and ER antagonists in lung (9) and breast (10, 11) cancers. In addition to its well-characterized role in controlling entry to S phase, multiple groups have reported that loss of RB1 leads to a hyperactivated or “primed” spindle-assembly checkpoint (SAC; ref. 12). These results imply that RB1-mutant cancer cells must rely on a mechanism to overcome SAC priming to avoid the fitness cost of stalled mitoses (13), but the basis of that mechanism remains unknown.

The antiproliferative activities of various cell-cycle inhibitors, as well as inhibitors of pathways that regulate RB1, such as the RAS–RAF pathway, have been linked to RB1 status (14–17). However, none of these drugs have been developed specifically for RB1-mutant cancers. Cytotoxic chemotherapy regimens acting post G₁ in the cell cycle, e.g., tubulin-binding drugs, do have activity against RB1^mut tumors (16). However, these drugs are indiscriminate (18) and are widely used to treat malignancies, such as hormone receptor–positive breast cancer, that are predominantly RB1-positive. Nevertheless, these results hint that RB1 mutation may confer unique vulnerabilities at particular cell-cycle stages that could be exploited for the development of more effective treatments.

We previously reported a pharmacogenomic screening assay (19) that overcomes the confounding effects of the widely differing growth rates of human cancer cells (20). Importantly, as we show, this assay format permits uniform distinction between cytotoxic and cytostatic effects despite varying cell-cycle times and so is well suited to test drugs that act on the cell cycle for synthetic lethal interactions. Using this assay to find drugs toxic to RB1^mut cancer, we identified inhibitors of Aurora kinases as top-scoring hits and demonstrate that LY3295668, a highly specific AURKA inhibitor, can kill RB1-deficient cancer cells at doses that have minimal effects on normal cells.

In pharmacogenomic screens, the influence of cancer genomics on drug response is inferred from profiles of antiproliferative activity across large panels of cancer cell lines. This approach has been effective in uncovering associations between oncogenes that activate mitogenic signaling cascades and drugs that inhibit the same pathways (21, 22). However, the method has been less successful in identifying genetic associations for drugs that act directly on the cell cycle. This is surprising because many driver genes in cancer encode cell-cycle regulators, including frequently mutated TSGs such as CDKN2A and RB1. We were interested to tailor the conventional pharmacogenomics assay to cell-cycle inhibitors and to then test a collection of such compounds in parallel and determine which showed the most promising association to RB1.

We and others have previously shown that growth rate introduces substantial bias in the fixed duration (e.g., 72 hours) format that has been standard in pharmacogenomics assays (19, 20). Growth rate bias is expected to be particularly problematic for cell-cycle inhibitors and for distinguishing cytotoxic versus cytostatic effects. This is illustrated in the simulation in Supplementary Fig. S1, which shows how the dose–response curve of a cytostatic compound varies with the number of population doubling times (DT), converging after 3–4DT to resemble the curve of an otherwise equipotent cytotoxic compound. To overcome this artifact, we established the growth rates in vehicle control conditions for hundreds of genomically characterized cancer cell lines and developed a pharmacogenomic screening assay normalized for growth rate by running each assay for 2DT rather than for a fixed duration (19).

We then used our 2DT drug-screening assay to identify drugs selectively toxic to RB1^mut cancer cells. Thirty-six inhibitors that act on the cell cycle, or cell-cycle regulating pathways, were selected for testing. These compounds inhibit a diversity of targets involved at all phases of the cell cycle with a bias toward drugs or drug-like compounds. The collection included compounds targeting G₁ to S phase regulators (e.g., CDK4/6), S phase processes (e.g., CDC7 and topoisomerase inhibitors), mitotic proteins (e.g., tubulin, PLK1, and Aurora kinases), mitogenic signaling pathways impacting G₁–S transition (e.g., RAS pathway, RSK1 and mTOR), and other proteins that have been shown to impact cell-cycle regulators such as MYC, p21^CIP1, and p53 (e.g., BRD4 and MDM2).

The antiproliferative activity of the 36 compounds against between 62 and over 500 cell lines from diverse epithelial, mesenchymal, and hematologic cancer lineages was determined (Supplementary Table S1). The test panel included, at minimum, 7 (range, 7–50) RB1-mutant cell lines to ensure sufficient power to detect RB1 synthetic lethal relationships. We ranked the 36 compounds for the strength of the association of sensitivity to RB1 mutation status using statistical methods described previously (19). As can be seen in Fig. 1A, RB1 mutation associates with resistance to RAF, MEK, and CDK4/6 inhibitors and, conversely, to sensitivity to all three tested Aurora kinase inhibitors. RB1^mut cells were, on average, also more sensitive than wild-type cells to inhibitors of other proteins active in mitosis, including the kinesin Eg5, and microtubules, but none of these compounds showed the same strength of association as the Aurora kinase inhibitors. This difference was not an artifact of the different number of cell lines tested for each drug because a similar relationship was derived from a common set of 443 cell lines that were tested with 12 of the compounds, including 7 different mitosis inhibitors (Supplementary Table S1). A weak preferential sensitivity of RB1-null cancers to inhibitors of p38, RSK, PLK1, Eg5, and microtubules has been described before (16). However, Aurora kinase inhibitors showed a much more pronounced effect than inhibitors of these targets in our profiling, suggesting that Aurora kinase inhibition might offer a unique therapeutic window that could be exploited for the treatment of RB1-mutant cancer. As explained below, we focused our subsequent investigation on further characterizing the relationship between RB1 status and response to Aurora A kinase inhibitors (AURKAi).

After 2DT, cell number will correspond to 25% of the DMSO control in response to the maximal effect of a purely cytostatic inhibitor. Therefore, we can infer that maximal inhibition values greater than 75% correspond to cytotoxic activity. As shown in Fig. 1B, the dose–response curves from the 2DT proliferation assay have a cytotoxic signature in RB1^mut cells, whereas a cytostatic signature typified RB1 wild-type (RB1^WT) cells treated with the top-scoring drug from our screen, the AURKAi MK5108. This difference in effect is presumably due to increased apoptosis because caspase 3/7 activation was enhanced in RB1-null versus RB1⁺ lung and breast cancer cells (Fig. 1C), and is not explained by weaker inhibition of AURKA kinase activity in RB1⁺ cells (Supplementary Fig. S2A and S2B). Artificial depletion of RB1 from RB1^WT lung and breast cancer cells (using either shRNA or siRNA reagents) led to enhanced cytotoxicity in response to the AURKAi (Fig. 1D and E; Supplementary Fig. S3A and S3B), whereas ectopic RB1 expression in RB1-null cells was protective (Fig. 1F and G). Altogether, these data indicate that our screening assay had, as intended, identified a synthetic lethal gene–drug interaction for RB1.

A similar, RB1-dependent, sensitivity to inhibitors of either AURKA or AURKB, as well as a synthetic lethal RB1–AURKB gene–gene interaction, is described in a companion paper in this issue by Oser and colleagues. Interestingly, AURKA gene knockout does not appear to associate strongly with RB1 (or MYC family gene) status, presumably because it is an essential gene in mammalian cells (23–25). Consistent with this, we found that AURKA has the profile of a pan-dependent gene from a recently published, genome-wide CRISPR screening data set across 342 cancer cell lines (26) scoring as similarly critical for RB1^mut and RB1^WT cells (data not shown). This contrasts with the ability of RB1^WT cells to survive high concentrations of AURKA inhibitors, and exemplifies the commonly observed phenomenon of protein depletion having a more severe phenotype than enzyme blockade (6).

In addition to alisertib, which inhibits both AURKA and AURKB (see below), both AURKA-dominant (MK5108 and MLN-8054) and AURKB-specific (barasertib) compounds showed a strong association with RB1, implying that inhibition of either Aurora kinase should be sufficient to achieve a synthetic lethality. Because myelosuppression has been encountered in clinical studies with AURKB-specific and dual AURKA/B inhibitors (27–29), but not AURKA-dominant drugs MK5108 and MLN-8054 (30, 31), we reasoned that specific inhibition of AURKA might be better tolerated than AURKB, resulting in a better therapeutic window and permitting a higher dose intensity to more effectively treat RB1-deficient cancers.

Crystal structures reveal several features that distinguish the active site of AURKA from AURKB, which we exploited to design a highly potent and selective inhibitor. The substituted isoquinoline (compound 1) derived from these efforts (Supplementary Fig. S4A and S4B). The carboxylic acid of compound 1 abuts Thr217 in AURKA but is sterically and electronically incompatible with the glutamate at the equivalent position in AURKB. The methylated piperidine ring likely further increases potency toward AURKA by establishing a novel water-mediated hydrogen bond to the main-chain carbonyl of Glu260. Compound 1 potently inhibited AURKA autophosphorylation (0.46 nmol/L IC₅₀) in RB1^mut NCI-H446 cells with over 1,000-fold selectivity against AURKB (measured by inhibition of phospho-histone H3). Capitalizing on these observations, further optimization of compound 1 led to the discovery of LY3295668 (Fig. 2A and B; Supplementary Information), an orally active compound with greater selectivity over AURKB than any Aurora inhibitor reported to date (Fig. 2C, and further described in a manuscript under review, J.D., R.C., J.H.). The fluorine atom on the pyridine ring of LY3295668 helps orient the carboxylate in LY3295668 closer to Thr217 of AURKA, likely contributing to its increased selectivity over AURKB. (AURKC is also more than 2 orders of magnitude less sensitive to LY3295668 than AURKA in enzyme assays; data not shown.)

Because the Aurora kinase inhibitors tested in our original screen are not exquisitely specific for AURKA (Fig. 2C; ref. 32), we wanted to determine whether LY3295668 maintained the strong association with RB1 status observed for those drugs across our cell line panel using the 2DT assay. (As shown in Supplementary Fig. S5, there was no correlation between cell line sensitivity and growth rate, suggesting that, as anticipated, the 2DT format counteracts growth rate bias.) Across 517 cancer cell lines, those most sensitive to LY3295668 included RB1^mut lung cancer, breast cancer, myeloma, retinoblastoma, and glioblastoma (Supplementary Table S2; Supplementary Fig. S6), and RB1 was the mutation most significantly associated with response to LY3295668 (Fig. 3A). The statistic for RB1 association is best powered in lung and breast cancer cells (Supplementary Fig. S6), so we focused the remainder of our investigation on these tumor types. Like MK5108, LY3295668 gave cytotoxic profiles in RB1-null, but not RB1⁺, SCLC cells, and in RB1⁺ NSCLC cells depleted of RB1 by shRNA (Fig. 3B and C), and activated apoptosis pathways much more strongly in RB1-null versus RB1⁺ lung and breast cancer cell lines (Fig. 3D and E; Supplementary Fig. S7). These data confirm that AURKAi-RB1 synthetic lethality does not require AURKB inhibition.

To complement studies in which RB1 is artificially depleted from cancer cells, we also examined the consequence of natural loss of RB1 as an adaptive response to selective pressure from drug treatment during acquired resistance (14). We took two CDK4/6i-sensitive ER⁺ breast cancer cell lines, MCF7 and MDA-MB-361, and from each derived palbociclib-resistant variants by prolonged drug selection (see Methods). Two independent, palbociclib-resistant variants of MDA-MB-361 cells were generated, one (MB-361-PR) from a conventional, rising dose selection method, the other (MB-361-PRENU) from a protocol using a pretreatment with the mutagen ENU and then selecting a resistant clone at a fixed, high drug concentration (33). The two palbociclib-resistant variants of MDA-MB-361 cells exhibited reduced RB1 expression and dramatically enhanced sensitivity to AURKAi, whereas the drug-resistant MCF7 variant had higher levels of RB1, and was not sensitized to LY3295668 (Fig. 3F). By CRISPR-mediated removal of RB1 from MDA-MB-361 cells, we confirmed that depletion of RB1 confers resistance to the CDK4/6i but sensitizes to LY3295668-induced apoptosis (Supplementary Fig. S8A–S8C).

Aberrations in cyclin-D and MYC family genes, which, like RB1 loss, release E2F-driven transcription at the G₁–S cell-cycle checkpoint (34), were also strongly associated with the LY3295668 profile. A relationship between MYC and AURKAi sensitivity has been described previously (35, 36), and the sensitivities of 87 cancer cell lines to a dual AURKA/B inhibitor were linked to MYC genes using an assay format partially adjusted for growth rate (37). MYC family gene amplification and RB1 mutation commonly co-occur, especially in SCLC. In the prior study across a small panel of cancer cell lines (37), all SCLC cell lines were RB1 mutant, precluding isolation of RB1 from MYC effects. In our profiling, although SCLC cells bearing both MYC family and RB1 aberrations are more sensitive than SCLC with either feature alone, neither amplification nor high expression of MYC family genes is required for sensitivity of RB1-mutant cells: Several highly sensitive RB1-mutant cell lines have normal copy number and expression of MYC genes (Supplementary Fig. S9A and S9B). Furthermore, among RB1-mutant cancer cells, expression levels of MYC family genes do not correlate with Abs IC₅₀ (Supplementary Fig. S10A). Interestingly, among cells with amplification of one of the MYC genes, most (10/11) of those with low RB1 expression are sensitive whereas those expressing RB1 have only about a 30% chance of being sensitive (Supplementary Fig. S10B). We also examined the AURKA-activating cofactors TACC3 and TPX2 to see whether their expression distinguished RB1^mut from RB^WT cells. As shown in Supplementary Fig. S11A and S11B, there is no significant difference in expression of either cofactor in RB1^mut cells.

The specificity of LY3295668 helped establish AURKA as the target most likely responsible for the RB1 association. In a kinome-wide survey, only 5 of 386 kinases were potently inhibited by LY3295668 (<10 nmol/L; manuscript under review, J.D., R.C., J.H.), and none of these kinases overlapped with targets of the other AURKAi (MK5108, alisertib) that showed an RB1 association. Notably, no inhibition of SYK, which has been reported essential in the context of RB1 loss (38), was observed up to 20 μmol/L. We verified that there is minimal SYK inhibition, relative to the IC₅₀ for AURKA, for the different AURKAi. The AURKA IC₅₀ is approximately 4,000-fold, 12,000-fold, or 50,000-fold lower than the SYK IC₅₀ for MK5018, alisertib, and LY3295668, respectively (Supplementary Fig. S12). Conversely, the relatively potent inhibition of AURKA observed with the SYK inhibitor R406 (87 nmol/L AURKA vs. 42 nmol/L SYK) may contribute to the activity of R406 in retinoblastoma (38).

We reasoned that if AURKA was indeed the critical target for the toxicity of LY3295668 toward RB1^mut cells, then a drug-binding defective form of AURKA would be able to rescue RB1^mut cells from LY3295668 treatment. Our crystallography experiments had demonstrated that Thr217 is critical for high-affinity binding of LY3295668, so we expressed AURKA^T217D in RB1^mut H446 cells to determine whether AURKA inhibition is required for the drug's activity. AURKA^T217D is, as predicted, resistant to LY3295668 because cells expressing AURKA^T217D are resistant to inhibition of AURKA autophosphorylation (Fig. 4A). RB1^mut H446 cells expressing AURKA^T217D are also protected from the cytotoxicity of LY3295668 (Fig. 4B and C), demonstrating that AURKA inhibition is indeed critical for RB1-LY3295668 synthetic lethality.

Assays to measure the cytotoxicity of drugs to human bone marrow (BM) cells growing in vitro are widely used for understanding the risk for clinical myelosuppression because they have a good track record of predictive power (39). Therefore, to test our hypothesis that truly specific AURKA inhibitors might avoid the myelosuppression seen with AURKB inhibitors, we tested the antiproliferative activity of LY3295668, and other Aurora kinase and mitosis inhibitors, toward human blood mononuclear cells to compare the relative cytotoxicity to BM versus RB1-mutant cancer cells. As shown in Fig. 5A, BM cells were 10-fold less sensitive to LY3295668 than RB1^mut H446 SCLC cells. In contrast, other mitotic inhibitors were toxic to BM cells at similar or even lower concentrations than those required to inhibit RB1-null cancer cells (Fig. 5A).

These data encouraged us to examine the efficacy of LY3295668 against SCLC tumors growing in vivo. Mice bearing SCLC xenografts were administered twice-daily doses of 50 mg/kg LY3295668 by oral gavage continuously for at least 3 weeks as shown in Fig. 5B. In accord with the in vitro findings, clear evidence of regression was observed for the three RB1-null SCLC xenograft models, but not for the RB1^WT PDX model LXFS-1129. The oncolytic activity of LY3295668 toward RB1-null tumors was rapid and persistent, provided the drug was continuously administered (Fig. 5B). The dosing regimen of LY3295668 used was chosen because it is well tolerated in mice and corresponds to plasma concentrations that exceed the IC₉₀ for inhibition of AURKA for the entire dosing interval yet does not reach the IC₉₀ for inhibition of AURKB, even at C_max (Fig. 5C). Similar exposures of LY3295668 in rats had insignificant toxicologic effects on BM cells. In a 1-month toxicology study, continuous dosing that resulted in comparable steady-state plasma concentrations to the mouse xenograft experiments caused no histologic changes in BM of the femur and sternum, and caused minimal changes in absolute numbers of circulating red blood cells, total white blood cells, lymphocytes, and eosinophils, with no effect on reticulocytes, platelets, neutrophils, or monocytes (data not shown).

Various functions have been ascribed to AURKA, the best-characterized being roles in orchestrating normal centrosome maturation and mitotic spindle assembly. In an effort to understand why RB1^mut cells are so highly dependent on AURKA, we examined whether LY3295668 has different effects on the morphology of the mitotic apparatus in RB1^mut compared with RB1^WT cells. These experiments are somewhat confounded by the rapid apoptosis apparent in RB1-null cells, but we saw no obvious differences in the effects of LY3295668 on centrosome biogenesis and mitotic spindle assembly in RB1-positive DMS-53 and MDA-MB-231 cells compared with RB1-null H2228 and HeLa cells (see below; data not shown).

These results hinted that a different function of AURKA may be responsible for the RB1 interaction. We reasoned that depletion of gene products critical for the sensitivity of RB1^mut cells to AURKAi would rescue cell growth in the presence of drug, and hence an unbiased genetic suppressor screen might give clues to the mechanistic basis of the RB1-AURKAi synthetic lethality. To this end, we used a genome-scale shRNA drug suppressor screen (40) in two RB1-null cells, NCI-H446 and MDA-MB-468, to identify genes critical for LY3295668 cytotoxicity (Supplementary Table S3). Several hits were identified, many common to both cell lines and therefore more likely true positives. These common suppressors might include both universal mechanisms that would explain the RB1-AURKAi pattern observed across our cell panel, as well as less general mechanisms restricted to these two cell lines. If depletion of a gene reduces sensitivity to AURKAi as a general mechanism, then cell lines with naturally low levels of the gene's transcript might be expected to be less sensitive to AURKAi across our cancer cell panel. Therefore, to enrich screening hits for genuine and broadly relevant suppressors, we triaged for genes where low mRNA associates with less sensitivity to LY3295668 across our entire cell line panel. We defined “universal suppressors” as genes whose expression (i) scores as required for sensitivity to LY3295668 in both NCI-H446 and MDA-MB-468 suppressor screens [false discovery rate (FDR) < 0.5] and (ii) associates with LY3295668 sensitivity in the pharmacogenomic cancer cell panel screen (P < 0.05; Fig. 6A; Supplementary Table S3). Seven genes meet these criteria (Fig. 6B) and include BUB1B and BUB3, two of the three mitotic checkpoint complex (MCC) genes responsible for enforcing the SAC by inhibiting the activity of the anaphase promoting complex/cyclosome (APC/C). APC/C is the ubiquitin ligase responsible for degradation of securin and cyclin B to permit anaphase progression and mitotic exit (41). The third MCC gene, MAD2L1, was not represented in our shRNA library, but expression of MAD2L1 also correlates very well with sensitivity to LY3295668 across the panel (Fig. 6C).

The link with MCC genes is notable because RB1 loss can prime expression of BUB1B/MAD2L1 (42, 43), and it has been argued that RB1-null cancer cells must, therefore, acquire a mechanism to overcome the primed SAC and its negative impact on fitness (13). To explain our combined results, we wondered whether AURKA kinase activity, which, uniquely among the mitotic drug targets, is capable of overriding the SAC (44, 45), could contribute to this mechanism. If so, RB1-negative cells would be strictly dependent on AURKA activity to exit mitosis. This model predicts a profound mitotic arrest in RB1-null cells exposed to AURKAi concentrations that have minimal effect on mitotic duration in RB1-positive cells. AURKAi have been shown to cause a transient mitotic arrest in RB1-proficient cells, but this is typically readily resolved (46). In contrast, and consistent with this model, we find that in RB1^mut cells, LY3295668 prolongs the stability of the APC/C substrate cyclin B1 (Supplementary Fig. S13A–S13E) and causes a substantial mitotic arrest (Fig. 7A), without any evidence of AURKB inhibition and without obviously different consequences to the mitotic spindle morphology in RB1^mut versus RB^WT cells (Supplementary Figs. S13 and S14). The timing of apoptosis coincides with the mitotic arrest (Supplementary Fig. S13E), suggesting that AURKAi-treated RB1-null cells predominantly die as a consequence of failure to escape mitosis. Depletion of BUB1B (Fig. 7B) or MAD2L1 (data not shown) reverses these effects, diminishing cyclin B1 accumulation and other markers of mitotic block, and reducing apoptosis. The enhanced LY3295668 sensitivity of breast and lung cancer cells depleted of RB1 is also associated with elevated markers of apoptosis and mitotic arrest (Fig. 7C). Similarly, natural RB1 loss in MDA-MB-361PR cells is associated with elevated BUBR1 and MAD2, but not MYC, levels and, in response to LY3295668, stabilized cyclin B1 and increased apoptosis (Fig. 7D and E; Supplementary Fig. S15). Together, these data suggest a unifying model (Fig. 7F) in which AURKA can promote mitotic exit despite an activated SAC. This function, which presumably requires higher levels of catalytic activity than the various functions of AURKA in mitotic entry and is therefore more sensitive to AURKAi, is only vital to cells with mitotic stress caused by loss of RB1, MYC amplification, or spindle poisons such as tubulin inhibitors.

To date, no drugs specifically tailored to the prototypical TSG, RB1, have been discovered. Drug screens across large panels of cancer lines are, in principle, well suited for the discovery of novel “gene–drug” synthetic lethal relationships but have so far failed to identify strong candidates for RB1. Two large screens using conventional 72-hour assays (21, 22) did not report drugs with significantly enhanced activity toward RB1^mut cancer cells, although the Garnett et al. study (22) found a significant association between RB1 mutation and resistance to both CDK4/6 and MEK inhibitors. Using an assay that corrects for the growth rate bias of conventional fixed duration assays, we report that, among a collection of 36 cell-cycle inhibitors, inhibitors of AURKA or AURKB show the strongest differential cytotoxicity toward RB1^mut versus wild-type cancers. A companion paper in this issue by Oser and colleagues describes a dependency of RB1-mutant cells on the AURKB gene and, in accord with our data, a corresponding sensitivity of RB1-mutant cells to inhibitors of either AURKA or AURKB kinase activities.

To establish that RB1-AURKAi synthetic lethality is indeed a consequence of AURKA inhibition, we introduce and use LY3295668, a new, truly specific AURKAi, and show that LY3295668 cytotoxicity toward RB1^mut cells is counteracted by a drug-binding defective AURKA variant. We have followed modern usage (47, 48) in using the term “synthetic lethality” to describe a gene–drug interaction: the effect of concomitant inhibitory perturbations to the RB1 gene and the AURKA function. Interestingly, despite surviving high concentrations of AURKA inhibitors, RB1–wild-type cells do not appear to tolerate knockout of the AURKA gene, which is essential in mammalian cells (23–25). Therefore, this appears to represent an example where protein depletion has a more severe phenotype than enzyme inhibition (6), and hence there is no evidence of a corresponding RB1–AURKA gene–gene synthetic lethality. Such a discrepancy could be caused if AURKA has important, kinase-independent functions (49).

The model we propose (Fig. 7F) to explain the striking sensitivity of RB1-mutant cells to AURKAi is premised on the unique role of high-level AURKA kinase activity in overriding an activated SAC (44, 45), a function that is likely distinct from the essential functions of AURKA in normal mitosis. On the basis of prior reports showing that RB1 loss is associated with increased priming of the SAC (12, 13, 43), we hypothesize that RB1-mutant, but not RB1 wild-type, cells constitutively require this function of AURKA to exit mitosis. The model invokes an evolutionary process. It is well established that RB1 loss, by itself, can be detrimental to mammalian cells, and an unknown mitotic escape mechanism has been postulated to necessarily occur during the evolution of RB1 loss in cancer cells to compensate for the fitness disadvantage otherwise incurred (13). Our hypothesis is that loss of RB1 function imparts a fitness advantage on cancer cells only if AURKA kinase activity is sufficiently high to overcome a mitotic delay. By examining a model of natural evolution of RB1 loss in the context of resistance to CDK4/6i, we have been able to begin to test this hypothesis. We show that cells that lose RB1 as a mechanism of resistance to the CDK4/6i palbociclib show increased BUBR1 and MAD2 and are sensitized to AURKAi cytotoxicity.

Future work will be necessary to rigorously test and refine our model. It will also be important to further tease apart the relative contributions of MYC proteins and RB1 loss to AURKAi sensitivity. RB1 loss and MYC amplification commonly co-occur, especially in SCLC, and we have not been able to identify RB1-mutant cells that do not coexpress a MYC protein to completely rule out a role for MYC in the synthetic lethal effect. Indeed, MYC proteins may play an essential role in cells with defective RB1, necessitating expression of at least one MYC family member in RB1-null cells (50). However, we have been able to show (i) that RB1^mut SCLC is more sensitive than RB1^WT SCLC, despite similar expression levels of MYC proteins, both in vitro and in vivo; (ii) that artificial RB1 depletion by three different methods (siRNA, shRNA, and CRISPR/Cas9), as well as (iii) natural selection of RB1 loss in response to CDK4/6i, in various RB1-positive cancer cells, can clearly sensitize to AURKAi without obvious increase in MYC protein levels. In fact, our model may help explain the association of MYC family genes with AURKAi because MYC has been reported to activate MAD2L1 expression (43, 51). Other perturbations that prime the SAC would also be predicted to sensitize cells to AURKA. In this regard, we note that the potentiation of taxane cytotoxicity by dual AURKA/B has been attributed to AURKA (52). AURKA amplification, which is a frequent event in various epithelial tumors (53), was not associated with AURKAi hypersensitivity in our profiling, perhaps because it is a vestige of adaptation to an SAC primed by a perturbation no longer present, such as from prior cytotoxic chemotherapy.

In summary, the data presented here predict that truly specific AURKAi will provide a better therapeutic window than classic cytotoxic agents for the treatment of RB1^mut malignancies, permitting more aggressive dosing regimens. The poor prognosis of patients whose cancers lack functional RB1, coupled with the recent descriptions of RB1 mutations in cancers that become resistant to inhibitors of EGFR (9) and CDK4/6 (11), highlights the pressing need for more effective treatments directed toward this tumor suppressor. To this end, a clinical trial is under way to test continuous dosing of LY3295668 for the treatment of patients with RB1-deficient cancers (NCT03092934).

All cell lines were obtained from commercial vendors and were cultured in conditions recommended by vendors. Cell line histology and site of origin annotation was derived from the source vendor or the Catalogue of Somatic Mutations in Cancer (COSMIC) cancer cell line database (www.cancer.sanger.ac.uk). Prior to use, cell lines were tested for Mycoplasma using a PCR-based method, and cell line authenticity was confirmed by STR-based DNA finger printing and multiplex PCR (IDEXX-Radil). For Mycoplasma-free cultures with authentic STR fingerprints, growth curves were determined to establish average population doubling time in the absence of drug treatment for each cell line. Cell density was optimized to ensure robust, logarithmic cell growth for the duration of compound exposure. All cell lines were used within 10 passages after recovery.

The 2DT cell panel screening assays using CellTiter-Glo (CTG; Promega Corporation) and IC₅₀ determinations were performed as previously described (19). For BM cell assays, human bone marrow mononuclear cells (AllCells LLC) were prepared according to the manufacturer's recommendations. BM cells were carefully thawed and suspended in warm growth media. The cells were resuspended the next day in media with human growth factors (erythropoietin, GM-CSF, IL3, and SCF/c-KIT ligand) and cultured for 72 hours. Then, the cells were seeded into 96-well plates and treated with DMSO control or various compounds for 48 hours before cell viability (CTG assay) was determined. In parallel, NCI-H446 cells (SCLC) were cultured and treated with DMSO control and various compounds for 48 hours before cell proliferation was measured.

To measure apoptosis by high content imaging, cells were fixed in 3.7% formaldehyde (Sigma, cat #F-1268) or Prefer (Anatech), permeabilized with 0.1% Triton X-100 (Sigma) in PBS (Gibco) for 10 minutes, washed several times with PBS and blocked with 1% bovine serum albumin (BSA; Invitrogen #15260-037) in PBS for 1 hour at room temperature. All subsequent dilutions and washes were performed in PBS. Cells were incubated with primary antibodies to cleaved caspase-3 (Cell Signaling Technology; #9661, RRID:AB_2341188) and cyclin B1 (BD Pharmingen; #554177, RRID:AB_395288) diluted in 1% BSA overnight at 4°C. Cells were then washed 3 times and incubated with secondary antibody, 5 μg/mL goat α-rabbit-Alexa-647 (Molecular Probes; #A-21244, RRID:AB_141663), and 200 ng/mL Hoechst 33342 (Molecular Probes; #21492) or DAPI (Sigma) to detect nuclear material for 1 hour at room temperature. Cells were washed again 3 times and imaged using a CellInsight NXT (Thermo Scientific) or Acumen eX3 (TTP Labtech Ltd). For CellInsight NXT, a minimum of 1,500 individual cellular images or 20 fields were captured for each well. Analysis was performed using the TargetActivation V.4 Bioapplication (Thermo Scientific). Arbitrary responder levels (percent positive for desired marker) were set based on the control groups for each cell line.

TUNEL staining was assayed with the In Situ Cell Death Detection, Fluorescein kit (Roche Applied Science) following the manufacturer's protocol. Fluorescence was captured using Acumen eX3.

To measure apoptosis by IncuCyte Zoom instrument (Essen BioScience), NCI-H446 and Calu-6 cells were plated on Costar 3596 plates and treated with LY3295668 at different concentrations for 24, 48, and 72 hours. Caspase 3/7 activation was measured with CellPlayer 96-Well Kinetic Caspase 3/7 reagent (Essen BioScience 4440). Green fluorescent images were acquired every 2 hours. Green objects counted (y-axis) were plotted against either real time (x-axis) or raw numbers as percent control.

Cell lines were grown overnight in the appropriate growth media recommended by ATCC. For RB1 shRNA experiments, MDA-MB-231 and Calu-6 cells were transduced with RB1 shRNA (TRCN194866 and TRCN196261) for 48 hours and treated with LY3295668 at the different concentrations for a further 48 hours. In addition, Calu-6 cells were also treated with MK5108 at different concentrations for 48 hours after RB1 shRNA transduction. Cells were then stained with propidium iodide (PI) for 1 hour at room temperature. Fluorescence was read using Acumen eX3. For siRNA-mediated RB1 knockdown in Calu-6 cells, cells were transfected with either Dharmacon RB1 Smart Pool (cat #L-003296-00-0005) or nontargeting control (cat #D-001810-01-05) siRNA according to the manufacturer's protocol. Cells were then treated with DMSO or MK5108 at different concentrations. Cell growth was continuously measured using an ACEA instrument for 160 hours. For siRNA-mediated MAD2L1 and BUB1B knockdown, NCI-H446 and NCI-H2228 cells were transfected with control siRNA2 (Ambion AM4637), MAD2L1 (Ambion S20468, Dharmacon; cat #L-003271-00-0005), or BUB1B (Ambion S261, Dharmacon; cat #L-004101-00-0005) siRNA for 72 hours. Cells were then treated with DMSO, 100 nmol/L LY3295668, or 200 nmol/L LY3295668 for 24 hours.

MDA-MB-468 and NCI-H446 cells were infected with the Module 1 Decipher library targeting the signaling pathways (Cellecta; cat #DHPAC-M1-P) at a multiplicity of infection of <1. The lentiviral-based library is composed of 5,043 genes with 5 to 6 plasmid pools per gene. Lentiviral particles were generated as described by the manufacturer. Infected cells were selected with puromycin for 72 hours. Following selection, cells were pooled, plated, and treated with DMSO or LY3295668 at IC₉₀ concentration (200 nmol/L for MDA-MB-468 and 400 nmol/L for H446) for 6 (MDA-MB-468) or 4 (H446) days, refreshing media once during the experiment. Genomic DNA was extracted using the Qiagen kit (cat #13362) as described in the manual. The barcodes tagged to each shRNA were amplified by PCR and sequenced on Illumina NextSeq 500 according to Cellecta's manual. “Common suppressors” of AURKA were identified as genes whose shRNA abundance was enriched in LY-treated cells relative to DMSO control–treated cells in both H446 and MDA-MB-468 cells (FDR < 0.5).

Cells were washed with PBS and were then lysed in lysis buffer containing protease and phosphatase inhibitors (Thermo Scientific, 1861281). Protein concentrations were determined by Bio-Rad Protein Assay Reagent (50000002). Cell lysates were cleared of debris by centrifugation at 14,000 rpm for 10 minutes. Western blots were performed with antibodies directed to RB1 (Cell Signaling Technology; 9309, RRID:AB_823629), AURKA (R&D Systems; AF3295), AURKA pT288 (Cell Signaling Technology; 3079, RRID:AB_2061481), cl-caspase 3 (Cell Signaling Technology; 9661, RRID:AB_2341188), cl-PARP (Cell Signaling Technology; 9541, RRID:AB_331426), cyclin B1 (Cell Signaling Technology; 4138, RRID:AB_2072132 or 12231), MAD2 (BD Transduction Laboratories; 610679, RRID:AB_398006), BUBR1 (Bethyl Labs; A300-386A), phospho-histone H3 ser10 (Millipore; 06-570, RRID:AB_310177), C-MYC (Abcam; ab32072, RRID:AB_731658), N-MYC (Thermo Fisher; MA1-16638, RRID:AB_2235735), MYCL1 (R&D Systems; AF4050, RRID:AB_2282440), GAPDH (Cell Signaling Technology; 32233), and actin (Sigma; A5441, RRID:AB_476744). The images were captured with Odyssey (LI-COR) or Amersham Imager 600 using the manufacturer's protocol.

Wild-type and mutant T217D AURKA and wild-type full-length RB1 constructs were cloned into pcDNA3.1 and, for AURKA, tagged on the N-terminus with a 3× FLAG tag. Cells were transfected with lipofectamine 2000 (Invitrogen; 11668-109), and selected with Geneticin for AURKA (Gibco, 10131-035) at 400 μg/mL. Cells were plated onto poly-d-lysine plates (Corning; 354640) and treated in duplicate with a dose curve of either DMSO or LY3295668 for 2 hours (AURKA) or MK5108 72 hours (RB1). For RB1 expression in MDA-MB-436 cells, expression levels were robust for at least 96 hours. AURKA activity was measured with phospho-AURKA (Thr288) Assay kit (MSD, K150JCD) as described by the manufacturer. Proliferation was assessed by CellTiter 96 (Promega, 63580) 4 days after cells were grown in RPMI-1640 (Gibco, 11875-093) and 10% FBS (Hyclone, SH30071.03).

Mitotic timing was determined by live-cell phase contrast imaging by IncuCyte using a 20× objective. Briefly, cells were plated in multiwell plates, treated with LY3295668, and imaged every 30 minutes for 72 hours. Time in mitosis was determined by first following cell round up and chromosome condensation as mitotic entry and then exit from mitosis by cell flattening and chromosome decondensation. Quantification was done by manually tracking 25 cells per cell line and treatment in ImageJ and time in mitosis was graphed as individual mitotic events with average hours in mitosis ± SD.

Cells were treated for 24 hours in glass bottom chamber slides. Following treatment, cells were fixed in 4% paraformaldehyde in PBS for 20 minutes, permeabilized with 0.2% Triton X-100 in PBS for 5 minutes, and blocked with 1% BSA in PBS for 20 minutes. Cells were incubated with primary antibodies to pericentrin (ab4448; Abcam) and alpha-tubulin (T5168; Sigma) diluted in 1% BSA/PBS for 1 hour, washed 3 times with PBS, incubated with secondary antibodies goat anti-rabbit Alexa Fluor 568 (A11011; Thermo Fisher) and goat anti-mouse Alexa Fluor 633 (A21053; Thermo Fisher) and sytox green (S7020; Thermo Fisher) to detect DNA for 1 hour. Cells were washed 3 times with PBS and imaged on a Zeiss LSM880 confocal microscope using the 63 × 1.4 NA oil objective. Z stacks were taken of cells representative of cellular mitotic phenotypes. Maximum intensity projections are shown.

All in vivo studies were performed according to the Institutional Animal Care and Use Protocols of the party or provider conducting the experiments. NCI-H69 cells were harvested, washed, and resuspended in a 1:1 mixture of serum-free media and Matrigel (BD Biosciences, 354234) prior to subcutaneous implantation (6 × 10⁶ cells/mouse) in the rear flank of athymic nude female mice (Harlan; 7–8 weeks). The SCLC patient-derived tumor models (LXFS 615, LXFS 650, and LXFS 1129) were derived from surgical specimens from patients at Oncotest (Oncotest GmbH, Charles River Laboratories). Following their primary implantation into nude mice (passage 1, P1), the tumor xenografts were passaged until stable growth patterns established. Stocks of early-passage xenografts were frozen in liquid nitrogen according to the relevant SOP for subsequent compound testing. Tumors were implanted subcutaneously in the left flank and randomized when the volume reached 80 to 200 mm³ to start compound treatment. Tumor volume was estimated by using the formula: V = L*W²*0.536, where L = larger of measured diameter and W = smaller of perpendicular diameter. Ten animals were used per treatment group and 8 animals in the vehicle group. Standard-of-care (SOC): etoposide 30 mg/kg subcutaneously (s.c.) Q7D×4 plus cisplatin 3.2 mg/kg s.c. Q7D×4.

MDA-MB-361 and MCF-7 ER⁺ breast cancer cell lines were used to derive variants with acquired resistance to palbociclib. For MDA-MB-361 cells, palbociclib selection was performed with or without a prior mutagenesis step to increase the diversity of resistant mechanisms available for selection. For mutagenesis we adapted a published method (33). The mutagen N-Ethyl-N-nitrosourea (ENU; Sigma, N3385) was dissolved in DMSO at 50 mg/mL and stored in aliquots at −80°C. MDA-MB-361 cells were cultured in complete medium at an exponential growth rate when ENU was added at a concentration of 50 μg/mL for 16 hours. The cells were then washed 3 times with fresh medium, replated in complete medium, and allowed to expand for about 2 weeks under optimal conditions. After ENU exposure and recovery, cells were cultured in 96-well plates at 5,000 cells/well in complete media with graded concentrations of the respective inhibitors. Wells were observed for cell growth by visual inspection under an inverted microscope. Fresh medium was supplemented when medium color changed. When growth in a well occurred, cells were transferred to 24-well plates and expanded in the presence of the corresponding inhibitor concentration used in the screen. Palbociclib-resistant derivatives of MDA-MB-361 and MCF-7 cell lines were also developed without ENU mutagenesis. To generate these variants, cells at 50% to 60% confluence were treated with inhibitors at a concentration approximating the IC₅₀ for cell growth for about 1 to 2 weeks. Cells were passaged when they grew to 80% confluence. Upon every passage, cells were left untreated overnight for attachment and then retreated with incrementally higher doses. This process was repeated several times until the cells were able to grow in the presence of drugs at a high concentration with no apparent off-target effects.

Variants that showed at least a 10-fold decreased sensitivity to palbociclib (i.e., >10-fold increase in palbociclib IC₅₀) were identified from these experiments leading to the palbociclib-resistant MCF7 cells (MCF7-PR) and two variants of MDA-MB-361 cells using either ENU mutagenesis (MDA-MB-361-PRENU) or the rising dose method (MDA-MB-361-PR).

Gene mutation, copy-number, and expression data were compiled from public domain data sets from COSMIC (www.cancer.sanger.ac.uk) and the Cancer Cell Line Encyclopedia (CCLE; www.broadinstitute.org/ccle/home).

Statistical analysis of the cell panel screening results was conducted as previously described (19). Linear regression and one-way ANOVA models were applied to test the significance of the association between abemaciclib potency across the cell panel and gene expression and mutation, respectively. Abemaciclib IC₅₀ was modeled on a log scale, and a generalized Tobit model was applied to account for censored IC₅₀ data. FDR was computed using the Benjamini–Hochberg method.

J. Du has ownership interest (including stock, patents, etc.) in Eli Lilly and Company. R.D. Van Horn has ownership interest (including stock, patents, etc.) in Eli Lilly and Company. A.M. Mc Nulty is a researcher at Eli Lilly and Company and has ownership interest (including stock, patents, etc.) in the same. M. Dowless has ownership interest in Eli Lilly stock. S.W. Eastman has ownership interest in Eli Lilly stock. M.Z. Dieter has ownership interest in Lilly Restricted Stock Units. H.-R. Qian has ownership interest (including stock, patents, etc.) in Eli Lilly and Company. G.D. Plowman was a Director on the Board of AURKA Pharma, Inc., is VP of Oncology Research at Eli Lilly, and has ownership interest (including stock, patents, etc.) in Eli Lilly. C. Reinhard has ownership interest (including stock, patents, etc.) in Eli Lilly. R.M. Campbell is a Research Fellow at Eli Lilly and Company and has ownership interest (including stock, patents, etc.) in the same. J.R. Henry has ownership interest (including stock, patents, etc.) in Eli Lilly. S.G. Buchanan has ownership interest (including stock, patents, etc.) in Eli Lilly and Company. No potential conflicts of interest were disclosed by the other authors.

Conception and design: X. Gong, J. Du, S.H. Parsons, F.F. Merzoug, P.W. Iversen, B. Han, M. Dowless, M.J. Lallena, Y.-H. Hui, X.S. Ye, D.A. Barda, G.D. Plowman, C. Reinhard, R.M. Campbell, J.R. Henry, S.G. Buchanan

Development of methodology: X. Gong, J. Du, S.H. Parsons, F.F. Merzoug, L.-C. Chio, B. Han, S. Yao, H. Bian, C. Ficklin, L. Fan, D. Manglicmot, A. Pustilnik, M. Dowless, S. Chu, X.S. Ye, D.A. Barda, S.G. Buchanan

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Du, S.H. Parsons, F.F. Merzoug, L.-C. Chio, R.D. Van Horn, X. Lin, W. Blosser, B. Han, S. Jin, H. Bian, C. Ficklin, L. Fan, A. Kapoor, A.M. Mc Nulty, K. Weichert, S.R. Wasserman, M. Dowless, C. Marugán, C. Baquero, S.W. Eastman, M.Z. Dieter, S. Chu, G.D. Plowman

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): X. Gong, J. Du, S.H. Parsons, F.F. Merzoug, Y. Webster, P.W. Iversen, R.D. Van Horn, B. Han, S. Yao, C. Ficklin, S. Antonysamy, K. Froning, K. Weichert, S.R. Wasserman, C. Marugán, C. Baquero, Y.-H. Hui, M.Z. Dieter, T. Doman, S. Chu, H.-R. Qian, X.S. Ye, S.G. Buchanan

Writing, review, and/or revision of the manuscript: X. Gong, J. Du, S.H. Parsons, F.F. Merzoug, Y. Webster, P.W. Iversen, R.D. Van Horn,S. Antonysamy, Y.-H. Hui, M.Z. Dieter, H.-R. Qian, X.S. Ye, D.A. Barda, G.D. Plowman, C. Reinhard, R.M. Campbell, J.R. Henry, S.G. Buchanan

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S.H. Parsons, X. Lin, W. Blosser, D. Manglicmot, M.Z. Dieter, H.-R. Qian, S.G. Buchanan

Study supervision: X. Gong, J. Du, F.F. Merzoug, X.S. Ye, G.D. Plowman, C. Reinhard, S.G. Buchanan

Other (served as a nonsalaried board member on AURKA Pharma and was involved as an advisor in the external clinical development of the compound described in the manuscript): G.D. Plowman

This research used resources of the Advanced Photon Source, a U.S. Department of Energy (DOE) Office of Science User Facility operated for the DOE Office of Science by Argonne National Laboratory under Contract No. DE-AC02-06CH11357. Use of the Lilly Research Laboratories Collaborative Access Team (LRL-CAT) beamline at Sector 31 of the Advanced Photon Source was provided by Eli Lilly Company, which operates the facility.