Report

MTOR signaling orchestrates stress-induced mutagenesis, facilitating adaptive evolution in cancer

See allHide authors and affiliations

Science  05 Jun 2020:
Vol. 368, Issue 6495, pp. 1127-1131
DOI: 10.1126/science.aau8768

How cancer cells adapt to stress

Bacteria adapt to harsh conditions such as antibiotic exposure by acquiring new mutations, a process called stress-induced mutagenesis. Cipponi et al. investigated whether similar programs of mutagenesis play a role in the response of cancer cells to targeted therapies. Using in vitro models of intense drug selection and genome-wide functional screens, the authors found evidence for an analogous process in cancer and showed that it is regulated by the mammalian target of rapamycin (mTOR) signaling pathway. This pathway appears to mediate a stress-related switch to error-prone DNA repair, resulting in the generation of mutations that facilitate the emergence of drug resistance.

Science, this issue p. 1127

Abstract

In microorganisms, evolutionarily conserved mechanisms facilitate adaptation to harsh conditions through stress-induced mutagenesis (SIM). Analogous processes may underpin progression and therapeutic failure in human cancer. We describe SIM in multiple in vitro and in vivo models of human cancers under nongenotoxic drug selection, paradoxically enhancing adaptation at a competing intrinsic fitness cost. A genome-wide approach identified the mechanistic target of rapamycin (MTOR) as a stress-sensing rheostat mediating SIM across multiple cancer types and conditions. These observations are consistent with a two-phase model for drug resistance, in which an initially rapid expansion of genetic diversity is counterbalanced by an intrinsic fitness penalty, subsequently normalizing to complete adaptation under the new conditions. This model suggests synthetic lethal strategies to minimize resistance to anticancer therapy.

Genetic diversity is central to adaptation and evolution in cancer. Mutagenesis in malignancies occurs incrementally (1, 2) or in transient bursts (37), providing increased cell-to-cell variation that facilitates adaptation to selective pressures, including anticancer therapies. Paradoxically, mutagenesis induced by some therapies, such as ionizing radiation, forms a core component of treatment for most cancers. The evolutionary effects of mutagenesis appear to be context dependent; they are potentially harmful under favorable conditions yet beneficial under stressed conditions. Stress-induced mutagenesis (SIM) programs that link environmental conditions to genetic diversity are a common atavistic theme across the phylogenetic tree (8) and are well described in prokaryotes (9) and lower eukaryotes (1013).

Here, we sought evidence for cognate programs in cancer. We first looked for accumulation of DNA damage (14) in human cancers exposed to therapies not considered directly genotoxic. Phosphorylated histone H2AX (γ-H2AX) was used as a marker of DNA double-strand breaks (DSBs) in pre- and posttreatment samples taken from diverse cohorts of chemo- and radiotherapy-naïve cancer patients. Increased DNA damage was consistently observed in prostate cancer patients treated with androgen deprivation therapy, in breast cancer patients treated with the aromatase inhibitor letrozole, in melanoma patients treated with the BRAF inhibitors dabrafenib or vemurafenib, and in patients with gastrointestinal stromal tumors treated with the KIT inhibitor imatinib (Fig. 1A and fig. S1). DNA DSBs were also increased in patient-derived xenograft (PDX) pancreatic cancer models treated with the CDK4/6 inhibitor palbociclib and the epidermal growth factor receptor (EGFR) inhibitor erlotinib, suggesting that increased DNA damage is a recurrent feature in human cancers exposed to nongenotoxic therapies.

Fig. 1 Selection with targeted anticancer therapies results in genetic diversity.

(A) Assessment of nuclear DNA DSBs in clinical samples and in PDX pancreatic cancer models. Shown is the fold change (mean ± SD) relative to pretreatment samples in patients and vehicle-treated controls in PDX models. Blue and black symbols represent negative and positive changes, respectively. Black diamonds indicate the overall fold change (FC) with 95% confidence intervals (95% CIs). (B) Schematic illustrating the time spans of the selection and evolution phases of in vitro model systems. Shaded area represents total population size, and lines represent the fate of individual clonal lineages after bottleneck. Dashed lines indicate extinction events; solid lines indicate evolution and clonal expansion. (C) Fold change values of DNA DSBs relative to untreated parental cells (mean ± SD). (D) Phylogenetic analysis of CNVs observed in single cell–derived clonal populations obtained from early-phase, tunicamycin-resistant 94T778 and vemurafenib-resistant SKMEL28 lines. The values below each set of parental (P) and drug-resistant (R) clones represent the average branch length. Values below the common trunks (orange) indicate their length. P values were obtained from Z tests. (E) Quantification of de novo SNVs generated during the expansion of the clonal populations. P values were obtained from quasibinomial logistic regression. (F) Colony formation assays performed in the absence of drug selection revealing the impact of the adaptive response on cellular fitness during the early phase of evolution. P values were obtained from two-tailed t test relative to drug-naïve parental cells.

We interrogated this phenomenon further in vitro using multiple human cancer cell lines exposed to nongenotoxic drug selection. Drugs specific to genetic targets in each cell line were titrated to near-extinction conditions, consistently generating one to 20 resistant colonies per 100,000 cells (Fig. 1B and table S1), which were analyzed for genetic and molecular features, as well as drug sensitivity (fig. S2). All model systems displayed increased DSBs early during evolution (Fig. 1C), which decreased to baseline during subsequent culture (fig. S3). The levels of DSBs were stable in the absence of selection.

To further characterize this phenomenon, we undertook whole-genome sequencing (average read depth 116×) on single cell–derived clonal populations obtained from untreated and early-phase, drug-resistant 94T778 human liposarcoma and SKMEL28 human melanoma lines. Phylogenetic trees were computed on the basis of copy number changes, revealing in both lines distinct clusters of untreated and drug-resistant clones sharing a common trunk. The observed lengths of the trunks (0.414 and 0.539 arbitrary units for 94T778 and SKMEL28 clones, respectively) were significantly longer than the average branch lengths of the respective parental clones (0.118 in 94T778, P = 1.8e-27, and 0.209 in SKMEL28, P = 1.4e-6), consistent with accelerated genomic evolution induced by the therapeutic bottleneck (Fig. 1D and figs. S4 and S5). To formally estimate mutation rates, we expanded clones using an equal number of generations (20 for 94T778; 21 for SKMEL28). Taking into account differential senescence and cell death (figs. S6 and S7, tables S2 and S3, and materials and methods), we quantified the percentage of de novo single nucleotide variants (SNVs) as the number of subclonal SNVs divided by the number of total unique SNVs. Intraclonal diversity, as measured by the number of de novo SNVs, was significantly higher in the resistant populations in both cell lines (Fig. 1E), consistent with higher mutation rates. We sought evidence for resistance-specific mutational signatures but observed no consistent patterns (fig. S8). We also performed targeted sequencing (figs. S9 to S11) on bulk populations, which revealed transiently increased structural and single nucleotide variation early during drug selection.

To further characterize the dynamics of genomic instability, the 94T778 cell line, containing two amplified targetable oncogenes (CDK4 and MDM2), was exposed to palbociclib or to nutlin-3a, an inhibitor of the p53–MDM2 interaction. Cells were transduced at both early and late time points of evolution with a single-copy fluorescent reporter gene (mCherry). In this system, genomic alterations lead to loss of gene expression. In agreement with the targeted sequencing data, drug-resistant cell populations assayed early during evolution demonstrated a significant loss of fluorescence compared with baseline or late time points (fig. S12). To approximate the relative contribution of de novo versus preexisting genetic variation to drug resistance, clonal population dynamics were tracked by engineering 94T778 and the human melanoma A375 cell line with a high-complexity DNA barcode library (15). Although we observed selection of preexisting genetic variation, the high frequency of individual barcodes (94T778 line average 69.6%, range 42.9 to 93.3%; A375 line average 17.9%, range 3 to 64.7%) is consistent with a contribution of de novo mutagenesis to resistance (fig. S13). This is also suggested by the late (37 weeks, T3) emergence of subclonal TP53 mutations (C277F, R273L) induced by nutlin-3a, neither of which is visible at weeks 11 (T1) or 21 (T2) (fig. S11).

As noted, mutagenesis may directly entail a fitness penalty driven by the accumulation of deleterious mutations. To study the counterselective effects of SIM, we examined the clonogenic potential of both parental and resistant lines in the absence of drug selection. We observed universally decreased clonogenicity relative to drug-naïve controls (Fig. 1F). Consistent with these effects, colonies obtained from nutlin-3a–resistant 94T778 contained up to 45% senescent cells at 6 weeks, falling to 6% at 25 weeks, compared with 1.5% in untreated controls (fig. S14). Extinction events may explain the loss of some initially resistant clones; for example, the C238F resistance mutation in TP53 was detected at 11 weeks under nutlin-3a selection but was lost thereafter (fig. S11). These data support a SIM-induced intrinsic fitness penalty early during adaptation.

To identify common genetic pathways mediating SIM in human cancer, we conducted a genome-wide functional screen using the 94T778 cell line. The presence of multiple critical drug targets in these cells allowed us to identify shared genetic determinants of resistance. Cells were engineered with a whole-genome short hairpin RNA (shRNA) library and subjected to five different selective conditions (Fig. 2A): (i) culture for 2 weeks without selection to estimate selective drift; (ii) equivalently titrated near-extinction selective conditions using nutlin-3a; (iii) palbociclib; (iv) tunicamycin targeting the unfolded protein response; and (v) incrementally increasing concentrations of nutlin-3a (to determine the relevance of selective stringency).

Fig. 2 Whole-genome RNAi screen identifies MTOR as a common evolutionary capacitor in cancer.

(A) Schematic diagram of the genome-wide RNAi screen. (B) Enrichment (left panel) and depletion (right panel) of hairpins targeting components of MTOR signaling. Genes are ranked by q values, with genes in red and light red being highly significant and approaching significance, respectively. The dashed line indicates the 0.05 statistical significance threshold. (C) Silencing of MTOR signaling imposes a fitness penalty in normal culture conditions (white bars) but fosters adaptability in response to pharmacologic pressures (gray bars). P values were obtained from two-tailed t test versus nonsilencing (NS) controls. (D) Silencing of MTOR affects therapeutic responsiveness to palbociclib in PDX pancreatic cancer models. P values were obtained from two-tailed t test versus first time point after the end of the second treatment cycle.

Initially, the screen identified condition-specific enrichment of shRNAs for genes known to confer resistance, as well as previously unknown genes, in these important pathways (“solution” genes). For examples of known resistance genes, shRNAs targeting TP53 and RB1 were among the top-ranked genes for the nutlin-3a and palbociclib conditions, respectively (table S4 and fig. S15). This provided evidence that the assay was performing as expected. We next looked for genes with shRNAs that were enriched across all near-extinction conditions (“facilitator” genes). Facilitator genes were associated with lower levels of enrichment compared with solution genes (table S5), which was perhaps due to an indirect role in mediating resistance through subsequent stochastic events, or possibly to counterselection. Top-ranked among facilitators were the genes encoding the Chaperonin Containing TCP1 Subunit 3 (CCT3) and the RNA-processing factor Heterogeneous Nuclear Ribonucleoprotein L (HNRNPL) (16). CCT3 has an important role in telomere maintenance (17, 18) and genome stability (19), and hnRNP protein–mediated alternative splicing has a role in cancer progression (20). The next most enriched gene was that coding for the mechanistic target of rapamycin (MTOR) (table S6 and fig. S16). In addition, the analysis revealed the enrichment for shRNAs directed to positive regulators of MTOR and the depletion of hairpins targeting the negative regulator PTEN (Fig. 2B).

The MTOR signaling pathway is widely expressed in human tissues and functions as an evolutionarily conserved sensor of environmental and endogenous stress (2123). The 94T778, SKMEL28, and human breast cancer SKBR3 cell lines were engineered with MTOR-directed shRNAs and were exposed to tunicamycin, vemurafenib, and the EGFR/HER2 inhibitor lapatinib, respectively. All cell lines showed a reduction in MTOR protein levels and a reduced phosphorylation of the downstream kinase p70-S6K (fig. S17A). In addition, we observed a transcriptional repression of genes positively regulated by MTOR and an increased expression of genes repressed by MTOR (24) (fig. S17B). As expected, MTOR silencing impaired clonogenicity in the absence of selection, but paradoxically increased clonogenicity under drug selection (Fig. 2C and supplementary materials), recapitulating the models described earlier. This effect was not caused by altered drug sensitivity in short-term assays (fig. S18). To exclude off-target effects from gene knock-down, these experiments were validated using shRNAs targeting the 3′ untranslated region of MTOR. The stress-induced adaptive potential was abrogated using an RNA interference (RNAi)–resistant open reading frame of the gene (fig. S19). In vivo, repression of MTOR accelerated emergence of palbociclib resistance in a PDX pancreatic cancer model (Fig. 2D) despite enhancing the effect of palbociclib in short-term ex vivo assays (fig. S20). Consistent with previously described kinetics of mutagenesis and impaired intrinsic fitness, a time course analysis of MTOR and p70-S6K phosphorylation levels revealed a significant reduction early during evolution and a restoration to pretreatment levels late during adaptation (fig. S21).

Because inhibition of MTOR impairs the DNA-damage response (25, 26), we examined DNA DSBs in 94T778 cells engineered with MTOR-specific shRNAs. The analysis of confocal microscopy image data revealed elevated levels of DSBs comparable to those seen in cells spontaneously resistant to nutlin-3a (fig. S22A). Furthermore, the selective MTOR inhibitor torkinib (pp242) delayed the repair of ionizing radiation–induced DNA DSBs, assessed as the number of γ-H2AX (fig. S22B) and 53BP1 (fig. S22C) foci. To quantitate the effects of MTOR silencing on mutagenesis, we undertook whole-genome sequencing on single cell–derived clonal populations obtained from 94T778 cells expressing nonsilencing or MTOR-specific shRNAs. We found a significant increase in intraclonal diversity in MTOR-silenced cells (figs. S23 to S25, tables S2 and S3, and materials and methods). No MTOR-specific mutational signatures were evident in 94T778 clonal populations (fig. S8). In our lines, pharmacologic inhibition of MTOR repressed expression of the homologous recombination (HR) and Fanconi anemia pathways and selectively repressed high-fidelity, but not error-prone DNA polymerases (Fig. 3A and table S7).

Fig. 3 Inhibition of MTOR fosters adaptive mutagenesis by repressing accurate DNA repair.

(A) Transcriptional changes induced by torkinib (pp242). P values were obtained from one-sided Wilcoxon rank-sum test and are shown in table S7. (B) Transcriptional changes observed during the early phase of the adaptive evolution in 94T778 cells exposed to palbociclib (P), nutlin-3a (N), and tunicamycin (T) sampled at 10 weeks for P and N and at 9 weeks after bottleneck for the T condition. P values were obtained from one-sided Wilcoxon rank-sum test and are shown in table S8. (C) Effect of silencing of HR genes in the absence (white bars) or presence of pharmacologic pressures (gray bars). P values were obtained from two-tailed t test versus NS control shRNA. (D) Effect of palbociclib and rucaparib combination therapy on tumor growth inhibition (bar plot, 30 days) and overall survival (Kaplan–Meier plot) in a PDX pancreatic cancer model. P values for tumor growth were obtained from two-tailed t test; a log-rank test on palbociclib + rucaparib versus palbociclib monotherapy was used in the Kaplan–Meier plot.

To ascertain whether similar transcriptional changes are elicited across a broader range of cancer types, we interrogated the LINCS L1000 database, a transcriptome dataset of 100 lines exposed to ~400 drugs across multiple doses and time points (27). Almost identical patterns were seen in multiple cell lines exposed to three different MTOR inhibitors (fig. S26), as well as in 94T778 cells developing spontaneous drug resistance, as described earlier (Fig. 3B and table S8). In the LINCS L1000 database, multiple cytostatic drugs similarly repress accurate DNA repair (fig. S27). This suggests that inhibiting cell proliferation may affect genetic diversity through coordinated regulation of multiple components of DNA repair. To test whether altered DNA repair would enhance adaptation, we reinterrogated the genome-wide screen and found specific enrichment for shRNAs targeting HR genes (table S9). To determine whether the repression of HR directly facilitates drug resistance, we engineered 94T778, SKBR3, and SKMEL28 cells with shRNAs targeting BRCA1, BRCA2, and PALB2 and subjected the cells to selection. Recapitulating both spontaneous drug resistance models and silencing of MTOR, impairment of HR reduced clonogenicity in the absence of selection and increased clonogenicity under pharmacologic pressure (Fig. 3C and supplementary materials). The induction of a defective DNA-damage response by nongenotoxic cytostatic therapies suggests synthetic lethality-based strategies. In a preliminary proof-of-concept experiment, we assessed the impact of palbociclib combined with a poly(ADP-ribose) polymerase inhibitor (rucaparib) in a PDX pancreatic model. Combination therapy enhanced the antitumor effects compared with either agent alone. (Fig. 3D).

Collectively, these data reveal SIM as a common mechanism facilitating the evolution of human cancer and are broadly in agreement with recent observations in colorectal cancer (28). As with error-prone and faithful DNA polymerases, the advantage conferred by SIM in a metazoan context is not clear and may simply be an evolutionary remnant. MTOR-mediated SIM slows replication and fosters genomic instability by impairing accurate DNA repair, thereby enhancing genetic diversity and facilitating resistance to therapy (29) (Fig. 4A). Both suppression of MTOR signaling and DNA repair also directly confer an intrinsic fitness penalty by inhibition of growth signaling and by increasing the accumulation of deleterious mutations, respectively. Rather than being a simple process, our data suggest that drug resistance may be the compound result of two independent but related mechanisms. Initially, net fitness balances the intrinsic fitness penalty of MTOR-mediated SIM with the generation of resistant genotypes undergoing extreme selection. Subsequently, normalization of SIM and fixation of stably resistant genomic configurations establish a new adaptive equilibrium (Fig. 4B). From a clinical perspective, our findings may explain the observation that agents targeting MTOR or the upstream phosphoinositide 3-kinase pathway have greater effects on objective responses or progression-free survival than on overall survival (30). From a drug development perspective, enhanced adaptation leading to drug resistance is undetectable by most preclinical assays of anticancer activity or by short-term measures of objective responses in trials. Our findings also provide a rational framework for synthetic lethal combinations of cytostatic agents with genotoxic therapies. Such combinations could potentially generate a lethal mutational load during the initial phase of adaptive evolution, thereby reducing therapeutic failure.

Fig. 4 Conserved mechanisms underpinning SIM in human cancer.

(A) Selection by cytostatic anticancer therapies leads to increased levels of DNA damage and impaired mTOR signaling, providing genetic diversity and accelerating adaptation to anticancer treatments. (B) During the initial phase of adaptation, the fitness landscape is a dynamic balance between the deleterious effect of intrinsic selection and the beneficial effect conferred by resistant genomic configurations. Fit genotypes are progressively stabilized during the second equilibrium phase.

Supplementary Materials

science.sciencemag.org/content/368/6495/1127/suppl/DC1

Materials and Methods

Figs. S1 to S27

Tables S1 to S9

References (3172)

References and Notes

Acknowledgments: We thank K. Simpson, P. Madhamshettiwar, and J. Luu for providing the shRNA library and for QC analysis; G. Mir Arnau for assistance with RNA and DNA sequencing; C.L. Chan for assistance with the preparation of DNA libraries; W. Hughes for assistance with confocal microscopy; O. Martin and H. Reza Bigdeli for assistance with quantification of DNA damage; L. Lara-Gonzalez, N. Thio, and M. Zethoven for assistance with bioinformatics; and L. Caldon, G. Rancati, and D. Bowtell for their critical reading of the manuscript. Funding: This research was supported by the Australian National Health and Medical Research Council (NHMRC project grant no. 1088353 to D.M.T., A.C., and D.L.G.) and by the Girgensohn Foundation (A.C.). In vivo studies were supported by NHMRC 1162860, NHMRC 1162556, and Cancer Australia 1143699 project grants (M.P.). J.B. was supported by the Stafford Fox Centenary Fellowship in Bioinformatics and Computational Biology of Rare Cancers. Author contributions: A.C. and D.M.T. conceived and supervised the study, wrote the manuscript, and acquired research funding; A.C., M.J.M., A.T.P., S.R.J., U.N., and A.G.R. performed the experiments and analyzed the data; D.L.G., S.K., and J.B. performed bioinformatics analyses. D.G.G, N.M.C., G.V.L., M.P, and J.-Y.B. provided the clinical samples; A.C.V. and M.R.Q. analyzed the clinical samples; P.L. developed the JCountPro software. Competing interests: D.M.T. is a paid consultant and received research support from Amgen, Eisai, Pfizer, Roche, Astra Zeneca, Novartis, and Bayer. G.V.L. is a paid consultant for Aduro Biotech Inc., Pierre-Fabre Medicament, Bristol-Myers Squibb, Amgen Inc., Merck Sharp & Dohme, Novartis, Array Biopharma Pty Ltd., and Sandos. The other authors declare no competing interests associated with this work. Data and materials availability: The plasmids pLV-mCherry, psPAX2, pCAG-VSVG, and pPB-DEST-53BP1trunc, are available from AddGene under a material transfer agreement. The pCMV-hyPBase plasmid is available upon request to the Sanger Institute Archives (http://www.sanger.ac.uk/technology/clonerequests/). DNA-sequencing data have been deposited in the NCBI’s BioProject database (ID: PRJNA623123). RNA-sequencing data have been deposited in NCBI’s Gene Expression Omnibus (accession nos. GSE148342 and GSE148344). Computer codes are available at GitHub (https://github.com/dlgoode/Cipponi_Science_2020).

Stay Connected to Science

Navigate This Article