Integrative genome analyses identify key somatic driver mutations of small cell lung cancer

Abstract

Small-cell lung cancer (SCLC) is an aggressive lung tumor subtype with poor survival^1–3. We sequenced 29 SCLC exomes, two genomes and 15 transcriptomes and found an extremely high mutation rate of 7.4±1 protein-changing mutations per million basepairs. Therefore, we conducted integrated analyses of the various data sets to identify pathogenetically relevant mutated genes. In all cases we found evidence for inactivation of TP53 and RB1 and identified recurrent mutations in histone-modifying genes, CREBBP, EP300, and MLL. Furthermore, we observed mutations in PTEN, in SLIT2, and EPHA7, as well as focal amplifications of the FGFR1 tyrosine kinase gene. Finally, we detected many of the alterations found in humans in SCLC tumors from p53/Rb1-deficient mice⁴. Our study implicates histone modification as a major feature of SCLC, reveals potentially therapeutically tractable genome alterations, and provides a generalizable framework for identification of biologically relevant genes in the context of high mutational background.

Small-cell lung cancer (SCLC; ~15% of all lung cancer cases) typically occurs in heavy smokers and is characterized by aggressive growth, frequent metastases and early death^1,2,5. Unfortunately, no single molecularly targeted drug has yet shown any clinical activity in SCLC6. Genomic analyses have revealed genetically altered therapeutic targets in lung adenocarcinoma^7–16 and in squamous-cell lung carcinoma^17–19. By contrast, little is known about the molecular events causing SCLC beyond the high prevalence of mutations in TP53 and RB13. Systematic genomic analyses in SCLC are challenging because these tumors are rarely treated by surgery resulting in a lack of suitable fresh-frozen tumor specimens.

We have established a global lung cancer genome research consortium¹⁹, giving us access to approximately 6,600 surgically resected lung cancer specimens, out of which we retrieved 99 SCLC specimens. We conducted 6.0 SNP array analyses on 63 tumors, exome sequencing of 27 tumors and two cell lines, transcriptome sequencing of 15 tumors, and genome sequencing of two tumors (Supplementary Tab. 1).

We applied a novel algorithm in order to identify significant broad (Supplementary Fig. 1a) and focal copy number alterations (CNAs) (Fig. 1a, Supplementary Tab. 2) and observed almost universal deletions affecting 3p and 13q (containing RB1), frequent gains of 3q, 5p, and losses of 17p (containing TP53) (Supplementary Fig. 1a). Gains of 3q affected the region containing SOX2, recently shown to be focally amplified in squamous-cell lung cancer^19,20. However, 3q gains in SCLC were less focal than those in squamous-cell lung cancer (Supplementary Fig. 1b). Focal amplifications affected MYCL1 (5/63 cases) and MYCN (4/63 cases)^21,22 (Fig. 1a). A single case harbored a focal amplification of MYC. All MYC family member amplifications (16% of cases) were mutually exclusive suggesting genetic epistasis^21–23. Focal amplifications affected 8p12 including FGFR1 (6% with copy number ≥3.5; Fig. 1b) and 19q12 containing CCNE124. Fluorescent in-situ hybridization analyses in 51 independent specimens validated the occurrence of FGFR1 amplifications in SCLC (n=3, 6%, Fig. 1c). We and others have recently reported focal FGFR1 amplifications in squamous-cell lung cancer; FGFR inhibitors are currently being tested in such patients^17,19,25. Thus, FGFR1-amplified SCLC might benefit from FGFR inhibition. The only significant focal deletion involved FHIT²⁶ (Fig. 1a, Supplementary Tab. 2).

Figure 1.

a) Copy number analysis to detect significantly altered regions across 63 tumors. Statistical significance, expressed by q-values (x-axes), is computed for each genomic location (y-axis) (Supplementary Information). Deletions (blue lines, lower scale) and amplifications (red lines, upper scale) are analyzed independently and vertical dashed black lines indicate the significance threshold of 1%. Focally amplified and deleted regions were identified using narrow thresholds (upper quantile: 10%; lower quantile: 15%) to resolve CNAs down to candidate driver genes. b) CNAs of chromosome 8 containing FGFR1 (8p12). Samples are sorted according to the amplitude of FGFR1 amplification. c) FISH analysis to screen for FGFR1 amplifications in an independent set of 51 tumors. Quantification of green signals (FGFR1 specific probe) in comparison to red signals (centromere 8 probe) reveals three FGFR1 amplified samples. d) Copy number analysis based on array-CGH data of 20 SCLC tumors derived from p53/Rb1-deficient mice. Data was analyzed similar to the analysis presented in a). Due to the small sample size, we used a significance threshold of 5% (vertical dashed lines). e) Circos plot of all validated chimeric transcripts detected by transcriptome sequencing. f) Circos plot of validated genomic rearrangements obtained from whole genome sequencing. Both rearrangements affect only portions of the genome smaller than 500kbp. While the structural variant in sample S00841 affects non-coding DNA, the rearrangement in S00830 leads to a loss of exon 7 to 11 of the gene FOXP1.

Mice with conditional deletion of Rb1 and p53 develop SCLC^4,27–31 bearing amplifications of Mycl1, Mycn, and Nfib, which were subsequently also found in human SCLC²⁸. We analyzed CNAs in 20 SCLC tumors (15 primary tumors and 5 metastases) from p53/Rb1 conditional knockout mice⁴ in order to identify alterations shared by both human and mouse tumors. We found significant amplifications of Mycl1, Mycn, and of Nfib (Fig. 1d). In the 15 primary tumors (Supplementary Fig. 2), Nfib did not reach statistical significance, suggesting that Nfib amplifications occur later in tumor evolution. While NFIB was not significantly amplified in the human tumors, three samples exhibited copy number gain at this locus (data not shown). Furthermore, we identified significant amplifications affecting E2f2, a mediator of RB1 function³² and deletions of the histone acetyl transferase gene Crebbp in two mouse tumors (Fig. 1d).

By analyzing transcriptome sequencing data of 15 human tumors we next identified and validated three chimeric transcripts (Fig. 1e, Supplementary Tab. 3). Two contained a fusion partner that was also mutated, MPRIP-TP53 and CREBBP-RHBDF1 (Fig. 1e); both of which are predicted to create a loss-of-function of the genes involved (Supplementary Fig. 3a, b). Similarly, we also found a low genomic rearrangement frequency by reconstruction from paired-end whole genome sequencing data of two specimens (Fig. 1f). This low frequency is in accordance with the spectrum of CNAs in these samples exhibiting almost exclusively arm-level events (Supplementary Fig. 4a).

In order to identify possible differences in the overall genomic architecture between surgically resected (i.e., early stage) samples (n=17) and samples obtained by autopsy (i.e., late stage, n=10) we compared the spectrum of broad CNAs in these two sets. We computed absolute copy numbers from sequencing data in order to correct for admixture of nontumoral cells and for ploidy (Supplementary Note, Supplementary Fig. 4b), but found no significant difference between resected and autopsy cases (Fig. 2a). Furthermore, there was no difference in the total mutation frequency (Fig. 2b) and no segregation between resected and autopsy cases in an analysis of mutated “driver” genes (Fig. 2c, d). We further identified 5 triploid and 2 near-tetraploid cases (n=29) and found no statistical significant overrepresentation of samples with ploidy >2 between resected and autopsy cases (p=0.15). On average we observed a ploidy of 2.3, in line with previously reported studies based on DNA cytometry⁵. Thus, resected early-stage tumors and late-stage tumors are genomically similar, underscoring the representative nature of our analysis.

Figure 2.

a) Comparison of broad structural genome alterations between surgically resected and autopsy samples. The analysis is based on absolute copy numbers determined using a reconstruction of the allelic state (Supplementary Note). A broad alteration is assigned to be present if 1/4 of the chromosome arm is altered accordingly. Difference between resected and autopsy samples of broad CNAs in 3p, 3q, 5p, 13q, and 17p were statistically tested by a Fisher’s exact test. b) Distribution of the mutation frequency observed in SCLC (points: resected cases; squares: autopsy samples; diamonds: cell lines). The average of the mutation frequency in SCLC (red lines and label) is compared to various tumor types taken from recent large-scale sequencing studies of: melanoma (MEL)³⁷, SCLC38, breast cancer (BC)³⁵, ovarian cancer (OC)⁴⁰, multiple myeloma (MM)³⁴, ovarian clear cell carcinoma (OCC)³⁶, prostate cancer (PC)³³, renal cell cancer (RC)⁴¹, and chronic lymphocytic leukemia (CLL)³⁹. c) A schema showing the various steps of our integrated analysis and filtering procedures. All candidate driver genes extracted from sequencing are filtered against gene expression derived from transcription sequencing. CNAs are identified from SNP arrays and candidate CNA regions that are entirely driven by a single SCLC sample were subsequently removed. d) Candidate driver genes identified by significance analysis, presence in the COSMIC database, clustered mutations, and genes that are also involved in fusion events. The type of each mutation is shown for every sample including the gene specific total number of mutated samples.

Compared to global sequencing studies of other tumor types^33–41, SCLC exhibits an extremely high mutation rate of 7.4 protein-changing mutations per million basepairs (Fig. 2b, Supplementary Fig. 5a). This high mutation rate is likely linked to tobacco carcinogens, reflected by an elevated rate of C:G>A:T transversions compared to the neutral mutation rate observed in evolution (Supplementary Fig. 5b)^38,42–44. In order to identify pathogenetically relevant driver genes in the context of frequent background mutations we applied several filters, including analyses of a signature of mutational selection and of gene expression (Fig. 2c, Supplementary Note). In particular, significantly mutated genes showing an expression level lower than 1 FPKM (fractions per kilobase of exon per million fragments mapped) in more than half of the 15 transcriptomes were removed. Using these adjustments only two genes had q-values of ≤ 0.1: TP53 and RB1 (Fig. 2d)^{22,29,30,45,46}. Remarkably, many of the significant genes were actually not expressed (Supplementary Tab. 4) and none of these mutations were called in the transcriptomes. By contrast, all known tumor suppressors exhibited expression in the upper part of the overall distribution (Supplementary Fig. 6) supporting our strategy for elimination of passenger mutations. Additional filters included an analysis of regional clustering of mutations in a given gene (defining a mutational hotspot) and integration with orthogonal datasets and databases (Fig. 2c)⁴⁷. Similar to the analysis of significantly mutated genes, we discarded genes that were enriched for silent mutations. Together, these filters yielded a list of likely driver genes in SCLC: TP53, RB1, PTEN, CREBBP, EP300, SLIT2, MLL, COBL, and EPHA7 (Fig. 2d).

SLIT2 showed a pronounced clustering of mutations (5 of 29 cases). The observed mutation spectrum (2 nonsense, 1 frame-shift deletion, 2 missense) (Fig. 3a) together with frequent genomic losses (Supplementary Fig. 7a) suggests that SLIT2 may be a novel tumor suppressor gene in SCLC. We sequenced SLIT2 in 26 additional tumors and 34 cell lines and found an overall mutation frequency of 10% (n=89). Slit proteins are secreted ligands for Robo receptors involved in axon guidance and cellular migration^48,49. Supporting a tumor suppressive function of SLIT2/ROBO1 in the lung, Robo1 knockout mice fail to develop normal lungs; surviving mice exhibit bronchial hyperplasia⁵⁰. Accordingly, a tumor suppressive role for SLIT2 has recently been implied in lung cancer cell lines⁵¹. Furthermore, ROBO1 was recently found to be a specific serum biomarker of SCLC⁵². EPHA7 was recently described as a tumor suppressor gene frequently lost in lymphomas⁵³. Given its role in embryonic development and neural tube closure⁵⁴, EPHA7 mutations may contribute to the invasive phenotype of SCLC.

Figure 3.

a) The spectrum of mutations affecting SLIT2. Red mutation labels indicate mutations detected by exome sequencing and black labels indicate the results of the extended screen using 454 sequencing. b) Mutations in CREBBP and EP300. Similar to a), red mutation labels indicate mutations discovered by whole exome sequencing, whereas black mutation labels show the results from the extended sequencing around the HAT domain. c) The structure of the chimeric transcript affecting RHBDF1 and CREBPP is shown. Note, that the genomic scale has been adapted to accommodate exons from both genes (axis break, dashes). Chimeric reads are shown below. d) Cell lines that show abnormal signals in the break-apart FISH assay of CREBBP/EP300. In case of CREBBP, both cell lines are showing a loss of the telomeric signal (red signal). For EP300 one cell line also showed a loss of the telomeric signal (here green signal). Break-apart FISH results for CREBBP in H209 are shown as a control38. e) Copy number status for CREBBP and EP300 of all samples that show a deletion in one of the two genes (copy numbers ≤ 1.6 are considered as being deleted). Copy numbers are sorted with respect to the minimal copy number between CREBBP and EP300.

Mutations in CREBBP and EP300 were significantly clustered around the histone acetyltransferase (HAT) domain (Fig. 3b). Of these, mutations affecting the homologous Asp1399 (EP300) and Asp1435 (CREBBP) residues both affect acetylase activity in vitro^55–57. Furthermore, Gly1411Glu in CREBBP has been previously identified in lung cancer⁵⁸ and follicular lymphoma⁵⁹ and Gly1411Val as well as Asp1435Gly were found in relapsed acute lymphoblastic leukemia⁶⁰, suggesting a mutational hotspot. By contrast, the Arg386fs mutation and the CREBBP–RHBDF1 gene fusion truncate the open reading frame in the amino terminus (Fig. 3c, Supplementary Fig. 3a). Together with the observation of Crebbp deletions in mouse SCLC (Fig. 1d) and the recently described CREBBP–BTBD12 gene fusion in the NCI-H209 SCLC cell line³⁸, inactivation of CREBBP and EP300 likely plays a major role in SCLC. Focused sequencing of the HAT domain-encoding exons of CREBBP and EP300 in a validation set of 26 additional SCLC tumor specimens and 45 cell lines as well as break-apart FISH performed in 34 SCLC cell lines, confirmed an overall mutation frequency of 18% (point-mutations, indels, and gene rearrangements) (Fig. 3b, c, d). CREBBP/EP300 mutations have recently been described in relapsed acute lymphoblastic leukemia and B-cell lymphoma^57,61 but have not been observed at such high frequency in solid tumors so far. Furthermore, all mutations and most of the deletions of CREBBP and EP300 occurred in mutually exclusive fashion in the total set of 101 samples analyzed suggesting epistasis (Fig. 3e). The observed alterations are predominantly heterozygous supporting haploinsufficiency^57,62. Thus, even hemizygous deletions occurring in at least 10% of non-mutant samples (Fig. 3e; Supplementary Fig. 7b) may be considered inactivating.

Further supporting the relevance of CREBBP/EP300 mutations in SCLC, all but one (Asn1286Ser in EP300) of the missense mutations were classified as being damaging by computational analyses⁶³. Furthermore, all HAT domain mutations were located at the interface of substrate binding⁵⁶ (Fig. 4a), thus supporting the notion that they may impact catalytic activity. We assessed the functional impact on histone acetylation of the Gly1411Arg, Asp1435Tyr, and Ser1432Pro CREBBP mutations (homologous to Gly1375Arg, Asp1399Tyr, and Ser1396Pro in EP300) in reconstitution experiments in Crebbp^{Δflox/Δflox}, Ep300^{Δflox/Δflox} (Crebbp/Ep300 Cre-deleted double knockout, or dKO) murine embryonic fibroblasts (MEFs)^64–66. All three mutations significantly reduced acetylation of histone 3 lysine 18 (H3K18) (Fig. 4b, c). Specifically, Asp1435Tyr induced complete, Gly1411Arg pronounced, and Ser1432Tyr moderate loss of H3K18 acetylation. Furthermore, knockdown of CREBBP in the cell line DMS114 that lacks CREBBP HAT domain mutations resulted in a moderate but significant increase of cell proliferation (Fig. 4d, e). Tumors with mutations and hemizygous deletions in CREBBP/EP300 did not exhibit a significantly different pattern of gene expression as compared to wild-type tumors after correcting for multiple hypothesis testing (data not shown), suggesting that global changes in gene expression are not the predominant mechanism by which loss of HAT activity contributes to SCLC pathogenesis. Together, these results support a role for loss of CREBBP/EP300 function in the biology of SCLC.

Figure 4.

a) CREBBP/EP300 mutations mapped to the crystal structure of the EP300 HAT domain⁵⁶. All mutations are positioned at the molecular interface involved in Lys-CoA inhibitor binding. In particular, Asp1399 and Gln1455 (equivalent to CREBBP Asp1435 and Gln1491) are located on the substrate-binding loop L1 (red). b) Immunofluorescence was applied to measure levels of acetylated lysine 18 on histone H3 (H3K18Ac) in wild-type MEFs, Crebbp/Ep300 dKO MEFs and dKO MEFs transduced with retroviruses expressing wild-type or SCLC-derived mutants of mouse Crebbp. Human mutations were made at the equivalent murine amino acid, but human numbering is shown in labels. Crebbp-HA signal, red (CY3); H3K18Ac, green (Alexa 488); nuclei, blue (DAPI). The functionally defective Trp1502Ala/Tyr1503Ser81 was included as a control. c) Quantification of H3K18Ac mean signal intensity per nucleus relative to the HA-tagged Crebbp mean signal intensity. P-values shown are from Bonferroni post test of one way ANOVA. * P<0.05, **** P<0.0001. d) Whole cell lysates of DMS114 cells stably infected with lentiviruses expressing shRNAs targeting CREBBP were analyzed for CREBBP protein levels by immunoblotting. e) DMS114 cells stably infected with lentiviral shRNAs targeting CREBBP or the indicated control cells were seeded in 6-well plates and counted as triplicates at the indicated time points (x-axis). Absolute numbers are given on the y-axis and error bars are showing one standard deviation of the mean.

Another histone-modifying enzyme mutated in SCLC was the methyltransferase gene MLL, which was recurrently mutated at isoleucin 960 (Ile960Met)⁴⁷. MLL rearrangements occur in acute leukemia^67,68. Similarly, recurrent genetic alterations in histone modifying genes appear to be a novel hallmark feature of SCLC.

Confirming previous reports⁶⁹, we found mutations in PTEN (3 of 29 cases), all of which and are likely (Gly165Glu) or proven (His61Arg, Arg130Gly) to affect phosphatase activity⁷⁰, thereby activating the PI3-kinase pathway. We did not observe any mutations in PIK3CA⁷¹.

We developed a mathematical model that gives insight into the allelic state of each tumor and yields estimates of tumor heterogeneity (Supplementary Note). On average, we observed a rather low heterogeneity of about 6.5% (Supplemetary Tab. 5). Using the reconstructed allelic states of each tumor, we found that copy-neutral loss of heterozygosity (CNLOH) events (i.e., complete loss of one allele at a given locus combined with a match of the absolute copy number at that locus with the overall ploidy of the sample) were enriched at the TP53 and RB1 loci (Fig. 5a, b). Furthermore, all TP53 and RB1 mutations in CNLOH regions were early events (Fig. 5b) as their allelic fractions were compatible with the tumor purity. By integrating the different datasets we found that at least one allele of TP53 and RB1 was affected by any genomic event (i.e., mutation (including rearrangements), or hemizygous deletion, LOH) in all cases (Fig. 5c). Thus, similar to genetically manipulated mouse models of SCLC, inactivation of TP53 and RB1 are early and necessary events in the development of SCLC in humans as well^4,27–31. Finally, we identified one case, in which the patient had undergone surgery for lung adenocarcinoma three years prior diagnosis of SCLC. While both tumors contained the identical TP53 mutation (Val73fs), the RB1 mutation (Arg251X) was restricted to the SCLC tumor (Supplementary Fig. 8), compatible with trans-differentiation of adenocarcinoma cells to SCLC cells, mediated in part through loss of RB1. Acquired resistance of EGFR-mutant lung adenocarcinomas to EGFR inhibition has been linked with trans-differentiation to SCLC^72,73. It is tempting to speculate that loss of RB1 may be mechanistically involved in such cases of acquired resistance as well.

Figure 5.

a) Analysis of copy-neutral LOH events (CNLOH) in SCLC. The allelic state of each exome-sequenced sample was reconstructed by applying a detailed mathematical model (Supplementary Information). Genomic portions that showed a loss of heterozygosity (LOH) and an absolute copy number equal to the overall samples’ ploidy are classified CNLOH events. Only samples showing at least one CNLOH event are shown. An analysis of the allelic fraction of somatic mutations in CNLOH regions yields information about the timing of these mutational events. b) TP53 and RB1 mutations in CNLOH regions. c) Distribution of mutations (including rearrangements), hemizygous deletions, and LOH affecting TP53 and RB1 across all exome-sequenced samples. d) SCLC driver genes and their mutation frequency mapped to signaling pathways. We classified the occurring mutations into 5 major groups: receptor tyrosine kinase (RTK) alterations, PI3-kinase and p53 pathway, cell cycle control, histone modifiers, and regulation of actin cytoskeleton.

Despite methodological challenges (limited sample set, high mutation frequency), integrative genome analyses of human and mouse SCLC afforded sketching a molecular map of this tumor type, condensed in 5 categories (Fig. 5d). The tumor suppressive functions of p53 rely on its acetylation by CREBBP or EP30074-79. However, given the universal loss of p53 function in SCLC, the tumor suppressive functions of CREBBP that we observed are likely independent of p53. One of the best-studied functions of SLIT2 is its involvement in actin polymerization mediated by Cdc4280. We speculate that this property might enhance invasive capabilities and thus contribute to the aggressiveness of SCLC. The reported functions of EPHA753,54 may also contribute to this phenotype. Beyond universal losses of TP53 and RB1 and amplifications of MYCL1, MYCN, MYC, we present PTEN mutations and FGFR1 amplifications as potentially therapeutically tractable genome alterations. Finally, we define genomic alterations affecting histone modifying enzymes CREBBP, EP300, and MLL as the second most frequently mutated class of genes in SCLC. In summary, our study represents a significant extension of the current molecular concept of SCLC and, more broadly, provides an example of how integrative computational genome analyses can provide functionally tractable information in the context of a highly mutated cancer genome.