The Cancer Stem Cell Marker Aldehyde Dehydrogenase Is Required to Maintain a Drug-Tolerant Tumor Cell Subpopulation

Targeted cancer therapies that exploit the “oncogene addiction” phenomenon observed in many tumor cells can be highly effective in the clinic (1–3). However, responding tumors invariably recur due to acquired drug resistance, probably reflecting the outgrowth of a subpopulation of treatment-refractory cells. The maintenance of many cancers seems to require a subpopulation of tumor cells called cancer stem cells (CSC; ref. 4), which have been identified in hematopoietic and solid malignancies (5), and several studies have described a role for CSCs in drug resistance (6)

Cancer cells are continuously exposed to extrinsic and intrinsic stresses that promote increased reactive oxygen species (ROS) and DNA damage (7–9). The consequent accumulation of mutations in CSCs may render them more vulnerable to stress, especially as they enter S-phase. The activation of DNA damage response (DDR; refs. 10, 11), increased expression of antiapoptotic genes (12), activation of the β-catenin pathway observed in leukemic stem cells (13), and Notch and hedgehog pathways in prostate cancer cells (14) are some of the mechanisms reported to underlie drug resistance in CSCs.

CSCs express various markers, including CD44, CD133, and ALDH1A1, at levels substantially different from the bulk tumor cell population, and these markers are often used to isolate and functionally characterize CSCs. The association of CSC markers such as CD44 and CD133 with drug resistance has thus far been largely based on studies demonstrating reduced sensitivity of CSC marker–positive cells to chemotherapeutic drugs (15) or increased expression of these markers in treatment-resistant tumors (16, 17). There have also been functional roles in drug resistance reported for these markers, including activation of ABC transporters and antiapoptotic genes (18, 19). Elevated ALDH1A1 is observed in CSCs of multiple cancer types (20). ALDH proteins controls the oxidation of aldehydes to corresponding acids, and ALDH-mediated detoxification of toxic aldehyde intermediates produced in cancer cells treated with certain chemotherapy agents has been proposed to confer drug-resistant properties to ALDH1-positive tumor cells (21). Here, we report that cancer cell subpopulations that tolerate otherwise toxic drug exposures express elevated ALDH, which is critical for their survival. Inhibition of ALDH activity effectively eradicates this drug-tolerant subpopulation, revealing a potential beneficial effect of combination therapy that includes ALDH inhibition to delay cancer relapse.

Human cancer cell lines were grown in RPMI media supplemented with sodium pyruvate, 10% FBS, and the antibiotics penicillin and streptomycin at 37°C in 5% CO₂. Cell line identity is routinely monitored by SNP-based genotyping in the Genentech cell bank facility.

The Incucyte HD imaging system (Essen BioScience) was used to monitor cell growth in culture, as measured by percentage of confluence. This instrument captures high-definition phase contrast images of cells and uses a contrast based confluence algorithm to compute monolayer confluence of each image over multiple specified time points.

A bodipy-labeled ALDH substrate (Aldefluor Kit, Stem Cell Technology) was used to detect ALDH activity. The substrate was diluted in RPMI media (5 μL substrate/mL media) and added to adherent cells. After 30 minutes of incubation, cells were washed twice with RPMI media, and microfluorescence images were captured using an IncuCyte HD system (Essen BioScience) with a 10× objective. The quantification of ALDH^high cells (∼5% for MKN-45) was based on 16 images captured using the Incucyte System with a 10× objective.

The Aldefluor assay was used to detect ALDH activity in MKN-45 parental cells. ALDH^high and ALDH^low cells representing about 5% of parental cells with the highest and lowest ALDH activity, respectively, were sorted by flow cytometry. MKN-45 cells incubated with the bodipy-labeled substrate and DEAB, a cold competitive substrate, were compared as a negative control.

MKN-45 cells were plated on ten 15-cm dishes at 30 × 10⁶ cells per plate, and grown until 70% confluence and treated with 1 μmol/L crizotinib every 3 days for 4 weeks until drug-tolerant persisters (DTP) were established. Non–drug-treated cells in a 15-cm dish were harvested 3 days after plating. RNA was extracted using a Qiagen RNeasy kit, and RNA was amplified in both parental and DTP cells using the NuGen RNA Amplification Kit. Microarray analysis was carried out in triplicate as previously described (22). Total RNA was isolated from ALDH^high and ALDH^low cells using a RNeasy column (Qiagen), and microarray-based gene expression analysis was performed on triplicate samples.

Cells were fixed with 4% paraformaldehyde at the end of the assay period, and viability was determined using the nucleic acid stain Syto60 (Life Technologies) diluted 1:5,000 in water (23). Fluorescence was measured using a SpectraMax M5 instrument (excitation 635 nm and emission 695 nm; Molecular Device). Viability was expressed as fraction of a no-treatment control.

MKN-45- and GTL-16–derived DTPs were generated by treating parental cells with 1 μmol/L crizotinib for 15 to 30 days. PC-9 cells were treated with 1 μmol/L erlotinib for 6 to 9 days for DTP generation. Drug concentrations used for DTP generation for all cell lines except GTL-16, EBC-1, and A549 cells are based on previous studies (22). The appropriate concentration of each drug needed for DTP generation was established on the basis of a dose–response curve (Supplementary Table S3). The 200 nmol/L concentration of disulfiram used for most of the cell lines was based on an analysis of disulfiram effects on MKN-45 and GTL-16 DTPs, where an effective cytotoxic disulfiram concentration was empirically determined (Supplementary Fig. S3C). In all cases, media ± drug were changed every 3 days.

Proteins were extracted from cell pellets using NP-40 lysis buffer containing protease and phosphatase inhibitors. Protein concentration was measured using the Pierce BCA protein assay kit. Protein concentrations of samples were derived from a standard curve with BSA. Immunoblots were performed with equal amounts of protein for all samples. This assay generated erroneous concentration values for MKN-45 and GTL-16 DTPs. Therefore, immunoblots of DTP lysates were performed by loading gels with a maximum possible sample volume, and GAPDH levels were measured as a loading control. Proteins were separated using SDS-PAGE and immunodetection was performed using standard protocols. Antibodies were from various vendors: ALDH1A1 (R&D Systems); ALDH1A2 and ALDH8A1 (Santa Cruz Biotechnology); ALDH1A3, ALDH3A1, and phospho-EGFR (Abcam); GAPDH, cleaved PARP, and phospho-ATM/ATR substrates (Cell Signaling Technology); GAPDH (Pierce); and phospho-γH2A.x (Millipore).

shRNA constructs were obtained from Sigma-Aldrich. Lentivirus production was carried out in 293T cells as previously described (23). For knockdown studies, GTL-16 cells were plated in 10-cm dishes (2 × 10⁶ cells per plate). Following adherence overnight, cells were infected with lentivirus or control virus in the presence of polybrene (Sigma-Aldrich) overnight. Forty-eight hours after infection, cells were re-plated, and following adherence overnight, cells were treated as described above and cell viability was assayed using the Syto 60 nucleic acid stain (LifeTechnologies).

Notably, we were unable to develop suitable transfection and lentivirus infection conditions for MKN-45 cells. Attempts to infect MKN-45 cells with lentiviruses, including the control GFP shRNA virus, resulted in cell death. Attempts to transfect MKN-45 cells with siRNAs using various transfection reagents also failed because of the unusually poor transfection efficiency of this cell line.

RNA was extracted from parental cells and DTPs using the RNeasy kit and 1 μg RNA was reverse transcribed using the high capacity RNA-to-cDNA Kit (Applied Biosystems). The amount of amplicon was determined using the Applied Biosystems 7500 quantitative PCR system using SYBR green as the fluorescence reporter and normalized to GAPDH. All samples were analyzed in triplicate and C_t values were determined.

ROS measurement was performed by flow cytometry. The ROS detection reagent H2DCFH-DA (Molecular Probes) was added to cells in growth media (final concentration: 5 μmol/L), and after 30 minutes at 37°C, cells were harvested. FACS analysis was performed on 10,000 cells per sample. Parental cells not exposed to H2DCFH-DA served as negative control. ROS measurements from parental cells were used to normalize ROS levels for drug-treated cells.

PC-9 cells were suspended in a 1:1 mixture of Hank's Balanced Salt Solution with Matrigel [growth factor–reduced; catalog #356231 (BD Biosciences)] to a final concentration of 5 × 10⁷ cells/mL. Nude (nu/nu) mice (Charles River Laboratories) were inoculated subcutaneously with 5 × 10⁶ PC-9 cells in 100 μL in the dorsal right flank. When tumor volumes reach about 100 to 200 mm³, mice were separated into groups of 10 or 15 animals with similarly sized tumors, and treatment was initiated the following day. Mice were dosed via daily (qd) oral gavage (per os) with erlotinib (50 mg/kg in 7.5% captisol) and/or disulfiram (Sigma; Tetraethylthiuram, Catalog # 86720, dosed at 200 mg/kg formulated in safflower oil 95%, benzyl alcohol 5%), or with vehicle only. Tumor volumes were determined using digital calipers (Fred V. Fowler Company, Inc.) using the formula (L × W × W)/2. Tumor growth inhibition (%TGI) was calculated as percentage of the area under the fitted curve (AUC) for the respective dose group per day in relation to the vehicle: %TGI = 100 × 1 − (AUC treatment/day)/(AUC vehicle/day). Curve fitting was applied to log₂-transformed individual tumor volume data using a linear mixed-effects model using the R package nlme, version 3.1–97 in R v2.12.0 (24, 25). Mean time to tumor progression (TTP) was measured as the time (days) to reach twice (2×) or 5 times (5×) initial tumor volume. Partial regressions (PR) are defined as greater than or equal to a 50% decrease from initial tumor volume, and complete regression (CR) is defined as a 100% decrease in tumor volume. Plots were generated and statistical analysis was run using GraphPad Prism 6 (GraphPad Software).

Approximately 5,000 parental cells and 15,000 DTPs were plated per well in XF 96-well microplates (SeahorseBioscience) and incubated for 24 hours at 37°C in 5% CO₂. Disulfiram and N-acetyl cysteine (NAC) treatments were done for 48 hours in the presence of kinase inhibitor. Oxygen consumption rate (OCR) and extracellular acidification rate (ECAR) measurements were performed in bicarbonate-free, serum-free, 37°C media. Cells were then fixed with 4% paraformaldehyde, stained with Hoechst, and four quadrants/well were imaged using a Molecular Devices ImageXpress HCS. Average nuclei number per quadrant was determined. Bar graphs represented the mean ± SEM of normalized (cell number) OCR and ECAR measurements from six wells.

To identify gene expression differences in the drug-tolerant subpopulation, we analyzed the MET-amplified gastric cancer cell line, MKN-45, before and after treatment with the MET kinase inhibitor crizotinib, to which these cells are highly sensitive (Fig. 1A). Following several days of treatment, about 5% of the cells remain viable and largely quiescent indefinitely in continuous crizotinib. Among the gene expression differences (Supplementary Table S1), ALDH1A1, one of several ALDH isoforms, was relatively highly expressed in the drug-tolerant subpopulation. Therefore, we further investigated ALDH activity in MKN-45 cells and in a second MET-amplified gastric carcinoma cell line, GTL-16.

The ALDH substrate Aldefluor was used to detect ALDH activity in individual cells. We observed clear heterogeneity in ALDH activity within untreated MKN-45 and GTL-16 cell populations, even though they were derived from single cell clones. About 5% of MKN-45 cells expressed detectable ALDH activity (ALDH^high)) before treatment. Following several days of crizotinib treatment, ALDH^high MKN-45 cells were clearly enriched among those that remained viable DTPs (Fig. 1B), and after 30 days, most DTPs were Aldefluor-positive (Supplementary Fig. S1A). Similar findings were made in GTL-16 cells (Supplementary Fig. S1B). These results indicate significant overlap between the Aldefluor-positive and the drug-tolerant subpopulations.

To confirm that the ALDH^high cells present before treatment correspond to the drug-tolerant population, we used two approaches. First, viability of MKN-45 ALDH^high and ALDH^low cells, separated by fluorescence-activated cell sorting (FACS; Supplementary Fig. S2A), was measured following 72-hour crizotinib treatment. The second assay involved treating ALDH^high and ALDH^low cells with crizotinib for 3 weeks and then allowing DTPs to expand as colonies in the absence of drug. Both assays revealed increased crizotinib tolerance by MKN-45 ALDH^high cells compared with ALDH^low cells (Supplementary Fig. S2B), whereas there was no difference in growth rates of ALDH^high and ALDH^low cells (Supplementary Fig. S2C). ALDH^high cells yielded more colonies of larger size (Supplementary Fig. S2D), and ALDH^high-derived colonies consist of both ALDH^high and ALDH^low cells, indicating that ALDH^high cells can give rise to ALDH^low cells (Supplementary Fig. S2E).

Although Aldefluor is the preferred substrate for the ALDH1A1 isoform (26), Aldefluor can be converted by other ALDH isoforms (27). To determine which ALDH family member(s) contribute to increased ALDH activity in ALDH^high cells, we FACS-sorted MKN-45 cells into ALDH^high and ALDH^low cells and observed about 8-fold upregulation of ALDH1A1 levels in ALDH^high cells but no significant expression difference for 18 other ALDH isoforms (Supplementary Fig. S1C). Similarly, ALDH1A1 protein expression was higher in ALDH^high cells (Fig. 1C). Increased ALDH1A1 was also observed in crizotinib-treated MKN-45 within 24 hours (Fig. 1D) before drug-induced apoptosis was detected (Fig. 1E). ALDH1A1 protein was induced in ALDH^high and ALDH^low MKN-45 cells, albeit at much lower levels in ALDH^low cells (Fig. 1C). Consistent with the enrichment of ALDH^high cells in the drug-tolerant subpopulation, increased ALDH1A1 was observed in MKN-45 and GTL-16 DTPs (Fig. 1F). These results indicate that ALDH^high cells, present before treatment, and crizotinib-tolerant DTPs are largely overlapping populations, both expressing elevated ALDH1A1.

To examine a functional role for ALDH in drug tolerance, we used the drug disulfiram, which irreversibly inhibits ALDH enzymatic activity (28, 29). Disulfiram-treated MKN-45 and GTL-16 cells exhibited no significant viability effects (Supplementary Fig. S3), but when combined with crizotinib, disulfiram potently eliminated crizotinib-tolerant MKN-45 and GTL-16 cells (Fig. 2A and B). Although a brief preexposure to disulfiram modestly reduced DTPs for some cell lines, it did not eliminate all DTPs, indicating that continuous disulfiram exposure is required to substantially reduce the drug-tolerant subpopulation (Supplementary Fig. S4A). When MKN-45- and GTL-16-derived DTPs, established by treatment with crizotinib, were subsequently exposed to disulfiram, they were effectively eliminated (Supplementary Fig. S4B), confirming a critical requirement for ALDH for DTP survival. Multiple ALDH isoforms are increased in GTL-16 DTPs (Supplementary Fig. S5A and S5B), and knockdown of ALDH1A1 had no effect on drug sensitivity or DTP formation (Supplementary Fig. S6A and S6B), implicating additional ALDH enzymes. Although we were unable to knockdown ALDH1A1 in MKN-45 cells for technical reasons, expression of 12/19 ALDH isoforms was detected in MKN-45 parental cells and DTPs (Supplementary Fig. S5D), implicating potential redundancy.

We also tested disulfiram with the ALDH1A1-negative EGFR-mutant lung carcinoma cell line PC-9 (Supplementary Fig. S5C), which expresses other ALDH isoforms (Supplementary Fig. S5A and S5B). PC-9 cells are EGFR-addicted and very sensitive to treatment with the EGFR inhibitor erlotinib (22). Disulfiram plus erlotinib effectively killed erlotinib-tolerant PC-9 cells (Fig. 2C). PC-9–derived DTPs begin expanding as colonies when treated with erlotinib beyond 10 days (Supplementary Fig. S3A and S3B) while maintaining drug tolerance. We previously described these expanded clones as drug-tolerant expanded persisters (22), which, unlike DTPs, are relatively refractory to disulfiram (Fig. 2C). Consistent with the observed effects of disulfiram on PC-9–derived DTPs, expression of several ALDH family members was increased (Supplementary Fig. S5A and S5B). These results implicate multiple ALDH family members in drug tolerance and suggest that the role for ALDH in maintaining the viability of this subpopulation is most critical before their expansion as drug-tolerant clones.

We extended these findings to additional kinase-dependent cancer cells. In eight cancer cell lines derived from breast, colon, and lung tumors, with previously established kinase dependency, combining disulfiram with kinase inhibition substantially reduced drug-tolerant clones (Fig. 3A and B). Disulfiram alone did not cause significant cell death in any tested cell lines.

A similar effect of disulfiram was observed on DTPs derived from treating GTL-16 cells with the DNA-damaging agent, etoposide (Supplementary Fig. S4C). Gossypol, another ALDH inhibitor (30), also showed effectiveness, although with less potency than disulfiram, in killing drug-tolerant cells (Supplementary Fig. S7A). To exclude the possibility that the cytotoxic effect of disulfiram reflects activity on low-density cultures (DTPs), we plated GTL-16 cells and DTPs at different densities and treated with disulfiram. Irrespective of plating density, disulfiram effectively killed most DTPs, whereas no cytotoxic effect was observed on parental cells plated at the same densities (Supplementary Fig. S6C). A similar analysis of PC-9 parental cells revealed disulfiram sensitivity only in DTPs (Supplementary Fig. S6D). These results collectively support the potentially broad ALDH requirement in a drug-tolerant subpopulation of cancer cells.

The ALDH requirement in drug-tolerant cells might reflect increased levels of toxic aldehydes, resulting from peroxidation of membrane lipids (31), due to oxidative stress associated with increased ROS. The drug-resistant properties of CSCs may reflect, in part, the activation of adaptive responses to continuous exposure to low level stress (32), often associated with increased ROS and a DDR (33). Because mitochondria are the major source of ROS, we examined the bioenergetics of ALDH^high cancer cells by measuring their OCR, an indicator of mitochondrial electron transport chain (ETC) activity, and their ECAR, an indicator of glycolytic activity. Although the extent of energy production through the glycolytic pathway is similar between GTL-16 ALDH^high and ALDH^low cells, the basal OCR in ALDH^high cells is significantly higher (Fig. 4A and Supplementary Fig. S8A). These results reveal increased mitochondrial respiration in ALDH^high cells, which might lead to increased ROS. Indeed, we detected relatively high ROS levels (ROS^high) in ALDH^high cells (Supplementary Fig. S8C). We then sorted cells based on ROS levels and measured ALDH1A1 (Fig. 4B and Supplementary Fig. S8B). Consistent with flow cytometric results, ROS^high cells expressed higher ALDH1A1 than ROS^low cells. Increased DNA damage, a consequence of increased ROS, was also observed in ROS^high cells, measured as increased phospho-γH2A.x and phospho-substrates of ATM/ATR, indicators of double-stranded DNA breaks and DDR, respectively (Fig. 4B and Supplementary Fig. S8B). Both indicators were higher in ROS^high cells, further supporting the CSC-like properties of ALDH^high cells. Furthermore, as in ALDH^high cells, DTPs exhibit increased dependence on mitochondrial respiration (Fig. 4C), suggesting elevated oxidative stress in DTPs. Increased DNA damage sensor and transducer activity was also observed in DTPs (Fig. 4D). These results suggest that ALDH^high cells and DTPs use specific mechanisms to maintain viability under conditions of increased oxidative stress.

ALDH enzymes function as detoxifying agents by reducing toxic aldehydes resulting from lipid peroxidation in cells undergoing oxidative stress (34, 31). To determine whether selectively killing of drug-tolerant cells by disulfiram reflects the role of ALDH in suppressing ROS-induced DNA damage, we used the ROS scavenger, NAC (N-acetylcysteine). In PC-9 cells, erlotinib alone caused a 6-fold increase in ROS in DTPs, and an additional 3-fold ROS increase was observed in cells cotreated with erlotinib and disulfiram. Significantly, the disulfiram effect on ROS was almost completely abrogated when NAC was added (Fig. 5A). Similar results were obtained with GTL-16-derived DTPs. A small but significant increase in ROS was observed in crizotinib/disulfiram cotreated cells, and the addition of NAC caused about 75% decrease in ROS. NAC also reduced ROS in GTL-16 DTPs by about 60% but failed to reduce erlotinib-induced ROS in PC-9–derived DTPs (Fig. 5A). These results suggest that while kinase inhibitor treatment elevates ROS in kinase-addicted cancer cells, the drug-tolerant subpopulation uses ALDH to protect against ROS-mediated cell death.

We then examined DNA damage in DTPs in the presence or absence of disulfiram and NAC. We observed significantly increased DNA damage in DTPs exposed to kinase inhibitors plus disulfiram compared with kinase inhibitor alone, and the effects of disulfiram on DNA damage were abrogated by NAC (Fig. 5B). Notably, ALDH inhibition increased the turnover of some of the ALDH isoforms (Fig. 5B and Supplementary Fig. S7B). The ROS increase in cells cotreated with kinase inhibitor and disulfiram also triggered apoptosis of DTPs as revealed by increased cleaved PARP in GTL-16–derived DTPs and BimS in PC-9–derived DTPs. Consistent with the ability of NAC to suppress disulfiram-induced DNA damage, NAC cotreatment also decreased cleaved PARP and BimS (Fig. 5B). NAC is also sufficient to rescue DTPs from the lethal effects of disulfiram (Fig. 5C). Collectively, these results suggest that increased ALDH in the drug-tolerant subpopulation maintains sufficiently low levels of ROS to prevent apoptosis.

To investigate the ability of disulfiram to delay tumor relapse in vivo, we used a PC-9 xenograft model. The treatment regimen was first tested in vitro; thus, cells were treated with either erlotinib or disulfiram alone or in combination, and after 6 days, cells were maintained with or without disulfiram for four more days. We observed a significant growth delay in cells treated with erlotinib + disulfiram compared with erlotinib alone, and a further delay in erlotinib + disulfiram–treated cells that continued to receive disulfiram (Fig. 6A). A pharmacodynamic study of treated tumors indicated decreased p-EGFR in both erlotinib and erlotinib + disulfiram–treated tumors (Fig. 6C).

In the xenograft tumor growth study, mice were divided into four groups, namely, vehicle control, disulfiram control, erlotinib alone, and erlotinib + disulfiram groups. As expected, erlotinib was highly effective, whereas single-agent disulfiram caused minimal efficacy (Fig. 6B and Supplementary Table S2). After 11 days of erlotinib alone or in combination with disulfiram, most tumors were substantially reduced. Therefore, to assess the added effect of disulfiram, all treatments were stopped, except for disulfiram, which was continued until the end of the study. No significant loss in body weight was observed in mice as a result of any of the treatments (Supplementary Table S2).

We observed near-complete regression of tumors with erlotinib with 11 of 15 mice responding [10 PRs (defined by >50% decrease in tumor volume) and 1 CR (no measurable tumor)]. Tumors quickly rebounded following treatment cessation, with a mean TTP of 41 days to reach twice the starting tumor volume (Supplementary Table S2). Treatment with erlotinib + disulfiram was more effective, with 15 of 15 animals responding (9 PRs and 6 CRs). Significantly, the combination treatment resulted in a prolonged response, with tumors not progressing beyond the initial tumor volume (Supplementary Table S2). These data indicate a highly significant (P < 0.0007) delay in tumor relapse in the erlotinib + disulfiram–treated mice compared with those treated with erlotinib alone, consistent with a role for ALDH in drug tolerance in tumors.

We have described a drug tolerance mechanism associated with a cancer cell subpopulation derived from various tissues of origin, which involves the ALDH enzyme family. Our observations highlight similar properties of ALDH^high CSCs and drug-tolerant cancer cells, consistent with previously reported findings supporting a role for CSCs in drug resistance (4–6). These observations suggest potential benefit of combining ALDH inhibition with targeted cancer therapeutics as a strategy to prevent or delay cancer relapse.

Our studies initially focused on the CSC marker, ALDH1A1, which is enriched in a subpopulation of MET-addicted gastric cancer cells that display crizotinib tolerance. This difference in crizotinib sensitivity between ALDH^high and ALDH^low gastric carcinoma cells perhaps provides a mechanism by which CSCs can maintain viability upon an initial onslaught of a cytotoxic drug. The observed crizotinib-induced ALDH expression in ALDH^high cells may provide a survival advantage over ALDH^low cells during prolonged drug exposure. Significantly, we observed disulfiram sensitivity among DTPs derived from many cell line models, regardless of whether they specifically expressed elevated levels of the ALDH1A1 isoform. Indeed, we observed consistently increased expression of multiple ALDH family members in drug-tolerant cells, pointing to the effectiveness of disulfiram in inhibiting multiple ALDH isoforms. Our results are consistent with the previous observation that simultaneous knockdown of two ALDH members results in increased cyclophosphamide sensitivity of lung cancer cells compared with knockdown of individual members (35). These findings may explain why ALDH1A1 alone, as a prognostic biomarker for worse clinical outcome and metastasis in cancer, has at best, yielded mixed results (20). In metastatic breast cancer, such a correlation improved significantly when expression of ALDH1A3, ALDH2, and ALDH6A1 isoforms, as well as ALDH1A1, was considered (36). Together, these results point to a likely role for multiple ALDH family members in drug resistance.

The ALDH requirement in drug-tolerant cells for survival under prolonged drug exposure implicates a drug-induced increase in aldehyde levels. We speculate that increased oxidative stress in drug-tolerant cells, leading to membrane lipid peroxidation, is a likely cause of increased aldehyde levels. The observed increased dependence on mitochondrial respiration, measured as increased OCR/ECAR ratio, is probably a major source of increased ROS in drug-tolerant cells. Interestingly, CSCs identified from gastric cancer cells, defined by high ALDH1A1 expression, also exhibit elevated ROS and increased OCR/ECAR, similar to what is seen in the drug-tolerant cells, and further supporting the presence of a subpopulation of preexisting (before treatment) cancer cells that exhibit characteristics of drug-tolerant cells.

One of the major consequences of increased oxidative stress levels is the accumulation of DNA lesions, which can lead to apoptosis. The activation of DDR in glioma-derived CSCs at a level higher than that seen in non-CSCs may contribute to the observed resistance to ionizing radiotherapy (10). Similarly, we detected increased double-stranded DNA breaks in our model systems, suggesting that activation of DDR may be a widely used mechanism of survival of a subpopulation of cancer cells during cancer therapy.

Other reported mechanisms of action of disulfiram, including inhibition of P-glycoproteins involved in drug efflux and proteasome inhibition, are seen at relatively higher disulfiram concentrations (37, 38). However, our findings indicating complete inhibition of phospho-MET in MKN-45 and GTL-16 DTPs (Supplementary Fig. S9A) by crizotinib (even in the presence of disulfiram), together with no apparent effect of a proteasome inhibitor MG132 on DTP formation (Supplementary Fig. S9B), would seem to exclude these alternative potential mechanisms of disulfiram action on DTPs. A recent study of paclitaxel-resistant MDA-MB-231 cells (39) demonstrated cytotoxic effect of disulfiram/Cu⁺⁺ on the resistant cells; however, in contrast to our observation of a lethal effect of disulfiram specific to drug-tolerant cells, disulfiram/Cu⁺⁺ was equally toxic to parental MDA-MB-231 cells.

We propose a model (Fig. 7) implicating ALDH family members in maintaining intracellular aldehyde levels (and consequently, ROS levels) below the threshold beyond which an apoptotic response is engaged. Our findings with NAC cotreatment emphasize the critical importance of managing ROS levels for the survival of drug-tolerant cells and raise the possibility of discovering additional therapeutics that can specifically elevate ROS levels within the drug-tolerant cells to delay tumor relapse. Notably, NAC does not protect drug-sensitive cells from the lethal effects of targeted therapies, consistent with a unique role in the CSC-like subpopulation.

Our in vivo findings demonstrate a beneficial effect of the combination of disulfiram and erlotinib in significantly delaying treatment relapse. Significantly, a phase II clinical trial performed almost two decades ago demonstrated an increased survival rate of patients with breast cancer who received chemotherapy plus an active disulfiram metabolite compared with chemotherapy alone (40). Altogether, these observations suggest a potential benefit of ALDH inhibition for the treatment of breast and other cancers. If the cell culture findings with disulfiram cotreatment can be similarly translated to the clinic, there may be broad opportunities to use this therapeutic strategy to delay tumor relapse in various cancer treatment contexts.

M. Merchant is employed with Genentech and Roche as Sr. Scientist and has ownership interest (including patents) with Roche. J. Settleman has other commercial research support from Genentech. No potential conflicts of interest were disclosed by the other authors.

Conception and design: D. Raha, T.R. Wilson, J. Settleman

Development of methodology: D. Raha, J. Peng, D. Peterson

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): D. Raha, J. Peng, D. Peterson, M. Merchant

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): D. Raha, T.R. Wilson, J. Peng, D. Peterson, P. Yue, M. Evangelista, M. Merchant, J. Settleman

Writing, review, and or revision of the manuscript: D. Raha, T.R. Wilson, M. Evangelista, M. Merchant, J. Settleman

Study supervision: M. Evangelista, M. Merchant, J. Settleman

Performed laboratory experiments: C. Wilson

The authors thank members of Settleman Laboratory, especially Ho-June Lee, for helpful discussions, R. Soriano and G. Fedorowicz for microarray services, the Genentech Laboratory Animal Resource, and the gCELL and FACS core laboratories for support.

This work was supported by Genentech, Inc., a member of the Roche group.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.