Lung neuroendocrine tumours: deep sequencing of the four World Health Organization histotypes reveals chromatin‐remodelling genes as major players and a prognostic role for TERT, RB1, MEN1 and KMT2D

Abstract Next‐generation sequencing (NGS) was applied to 148 lung neuroendocrine tumours (LNETs) comprising the four World Health Organization classification categories: 53 typical carcinoid (TCs), 35 atypical carcinoid (ACs), 27 large‐cell neuroendocrine carcinomas, and 33 small‐cell lung carcinomas. A discovery screen was conducted on 46 samples by the use of whole‐exome sequencing and high‐coverage targeted sequencing of 418 genes. Eighty‐eight recurrently mutated genes from both the discovery screen and current literature were verified in the 46 cases of the discovery screen, and validated on additional 102 LNETs by targeted NGS; their prevalence was then evaluated on the whole series. Thirteen of these 88 genes were also evaluated for copy number alterations (CNAs). Carcinoids and carcinomas shared most of the altered genes but with different prevalence rates. When mutations and copy number changes were combined, MEN1 alterations were almost exclusive to carcinoids, whereas alterations of TP53 and RB1 cell cycle regulation genes and PI3K/AKT/mTOR pathway genes were significantly enriched in carcinomas. Conversely, mutations in chromatin‐remodelling genes, including those encoding histone modifiers and members of SWI–SNF complexes, were found at similar rates in carcinoids (45.5%) and carcinomas (55.0%), suggesting a major role in LNET pathogenesis. One AC and one TC showed a hypermutated profile associated with a POLQ damaging mutation. There were fewer CNAs in carcinoids than in carcinomas; however ACs showed a hybrid pattern, whereby gains of TERT, SDHA, RICTOR, PIK3CA, MYCL and SRC were found at rates similar to those in carcinomas, whereas the MEN1 loss rate mirrored that of TCs. Multivariate survival analysis revealed RB1 mutation (p = 0.0005) and TERT copy gain (p = 0.016) as independent predictors of poorer prognosis. MEN1 mutation was associated with poor prognosis in AC (p = 0.0045), whereas KMT2D mutation correlated with longer survival in SCLC (p = 0.0022). In conclusion, molecular profiling may complement histology for better diagnostic definition and prognostic stratification of LNETs. © 2016 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of Pathological Society of Great Britain and Ireland.

This four-tier histological spectrum clinically corresponds to a prognostic scheme whereby TCs are low-grade tumours, with patients having a long lifeexpectancy, ACs are intermediate-grade tumours with varying clinical behaviour, and LCNECs and SCLCs are high-grade tumours with a dismal prognosis [2][3][4][5][6]. The clinical management of LNETs reflects this prognostic categorization: TC patients are treated with surgery, AC and LCNEC patients are treated with with multimodal approaches involving surgery and/or systemic therapy, and SCLC patients usually undergo systemic therapy [7]. The 10-year overall survival rate of TC patients is reported to be 87%, whereas the median overall survival of patients with untreated SCLC is only 2-3 months, reaching 1 year when they are treated [8].
A better understanding of the genomic changes occurring in LNET is essential to identify new therapeutic targets and markers for better prognostic stratification. Previous molecular studies on LNET [9][10][11][12] have highlighted frequent alterations in the chromatin-remodelling genes MEN1, PSIP1 and ARID1A in carcinoids (ACs and TCs) [9]. Conversely, inactivating mutations in TP53 and RB1, amplification of MYC family members and genetic alterations in the PI3K/AKT/mTOR pathway have frequently been observed in SCLC [9,10,12,13].
To complement these reports, we performed a direct comparison of the four histotypes by whole-exome sequencing (WES) and high-coverage targeted sequencing (HCTS) of 418 genes on a cohort of 46 LNETs, and then validated 88 genes in an independent set of 102 LNETs.

Validation of mutations and CNAs in 102 LNETs by targeted sequencing of 88 genes
Matched tumour/normal DNA from 102 FFPE LNETs was subjected to targeted next-generation sequencing (NGS) to validate the prevalence of somatic alterations in 88 genes, selected as follows: 42 from the discovery screen, and 46 from current LNET literature [9,10,12,13]. To this end, we used four multigene panels (supplementary material, Supplementary materials and methods and Table S1).  Table 1 and detailed in supplementary  material, Table S2, with indications of the molecular analyses performed. A discovery screen for mutations and CNAs was performed on DNA from 46 frozen samples, with WES on 20 cases (high-quantity DNA samples) and HCTS of 418 genes on these and an additional 26 (low-quantity DNA samples). An 88-gene panel was then selected as follows: 42 genes from this discovery screen, including 13 that showed CNAs, i.e .  BCL2, FGFR1, MEN1, MYC, MYCL, PIK3CA, RB1,  RICTOR, SDHA, SMAD4, SRC, TERT, and TP53 (panel  3 of supplementary material, Table S1), and 46 from a literature review [9,10,12,13]. These 88 genes were verified by targeted NGS in the discovery set, and then assessed on a validation set of an additional 102 FFPE LNETs. The prevalence of mutations and CNAs was computed on the whole 148 cases cohort. The results of the mutational analysis of 88 genes are detailed in supplementary material, Table S3, and summarized in Figure 2A for the 26 genes mutated in at least three cases. Eleven of these 26 genes met the criteria for statistical analysis of prevalence among LNET histotypes (genes altered in at least 10% of at least one histotype), with seven genes showing a statistically significant difference ( Table 2). The results of CNA analysis of 13 genes are shown in Figure 2B and reported in Table 3.

Mutational analysis
The 46-case discovery set was composed of two groups based on the amount of available DNA: 20 high-quantity DNA samples (10 TCs; four ACs; three LCNECs; and three SCLCs) and 26 low-quantity DNA samples (13 TCs; 10 ACs; two LCNECs; and one SCLC). Each group was analysed according to DNA availability: WES and HCTS for high-quantity DNA samples, and HCTS for low-quantity ones.
WES achieved an average coverage of × 94 (×75-161) in tumour samples and × 72 (×38-161) in normal samples. One thousand three hundred and twenty-two non-synonymous somatic mutations were found, including 1184 missense and 88 nonsense mutations, 34 frameshifts, six in-frame insertions, and 10 in-frame deletions (supplementary material, Figure  S1A and Table S4A). The mean numbers of mutations per sample were 14.6 for TCs, 15.7 for ACs, 64.7 for LCNECs, and 90.7 for SCLCs. One TC (no. 304) and one AC (no. 389) emerged among carcinoids as having a hypermutated profile, as they respectively contained 228 and 434 mutations, i.e. a gross excess of point mutations relative to the same tumour type (supplementary material, Figure S1A). These cases shared inactivating somatic mutations of POLQ, encoding polymerase θ (supplementary material, Table S4A), and hypermutated profiles have been associated with mutations of genes encoding polymerases [21,22].
HCTS of 418 genes on the 46 cases of the discovery set yielded an average coverage of × 864 (×298-1657) in tumour samples and × 686 (×670-1316) in normal samples. Mutations were identified in 36 genes, and were 145 missense mutations, 13 nonsense mutations, two frameshifts, five splice site alterations, and one in-frame deletion (supplementary material, Table S4B and Figure S1B). The mean numbers of mutations per sample were 0.7 for TCs, 1.8 for ACs, 4.6 for LCNECs, and 5.8 for SCLCs. One hundred and sixteen of the 1322 somatic mutations identified by WES were targetable by HCTS, and all of them were confirmed.
In conclusion, WES and HCTS sequencing on the discovery set of 46 cases identified 36 genes as recurrently mutated (i.e. in at least two cases of different histotypes on both WES and HCTS) (Figure 1; supplementary material, Table S4C).

Copy number alterations
CNA analysis of the discovery set was performed with data from both WES on the 20 cases with high DNA quantity and HCTS of 418 genes on the whole discovery set. Large structural CNA analysis, which was feasible only for WES cases, showed frequent gains of chromosomes 5p (7/20; 35.0%) and 8 (5/20; 25.0%), and loss of heterozygosity (LOH) of chromosome 3 (5/20; 25.0%) ( Figure 3A). Carcinoids showed rare CNAs. However, three cases showed an unusually unstable genomic profile: AC no. 450 and the hypermutated TC no. 304 showed alterations in several chromosomes ( Figure 3A), whereas AC no. 034 showed a copy number pattern compatible with chromothripsis of chromosomes 2, 11, and 20 (supplementary material, Figure S2). Six of the eight classification criteria used to define chromothripsis [23] are assessable by WES: five were fulfilled for chromosomes 2 and 11, and four for chromosome 20. Conversely, carcinomas had many large structural CNAs, including copy gains at chromosomes 5p and 8, and LOH at chromosome 13 and 17p, where RB1 and TP53, respectively, are located.
Gene-level CNAs were investigated on the discovery set with both WES data ( Figure 3B; supplementary material, Table S5A) and HCTS data (supplementary material, Figure S1B and Table S5B). The most recurrent copy loss events involved one or both copies of RB1 (TCs, 13 Although a specific CNA hallmark of each histological subtype was not apparent, the average number of genes affected by CNA progressively increased from TCs to SCLCs (TCs, 1.3; ACs, 2.0; LCNECs, 10.8; SCLCs, 14.7; p < 0.0001), with ACs showing a hybrid profile between TCs and carcinomas.
To identify prognostic molecular markers in all LNETs, we screened the following common alterations among the three subtypes showing disease-specific mortality (ACs, LCNECs, and SCLCs): KMT2D mutation, MYC copy gain, PIK3CA copy gain, RB1 mutation, RB1 homozygous deletion, RICTOR copy gain, TERT copy gain, TP53 mutation, and TP53 LOH. TC cases were excluded from this analysis, given that their features (i.e. no death of disease after long follow-up) would have represented a confounding effect on analysis of the impact of molecular markers on the rest of the series. Each marker yielding a p-value of <0.2 on univariate survival analysis (supplementary material, Table S6 and Figure S3) was tested in a multivariate model against the significant clinicopathological predictors of tumour stage and histotype. The final analysis confirmed RB1 mutation, TERT copy gain, histotype and tumour stage as independent factors for poor prognosis (Table 4).
We also performed a stratified univariate analysis for mutations clustering to specific tumour histotypes, which showed a significant association of MEN1 mutation with poor prognosis in ACs (p = 0.0045), in which MEN1 mutations were mainly confined to early-stage tumours. Among SCLCs, KMT2D mutation was almost exclusive to stage III tumours [five mutated of 13 stage III (38.5%) versus one mutated of 11 stage IV (9%)], and correlated with longer survival (p = 0.0022).

Discussion
Although the WHO classification defines clinically relevant subgroups of LNET, there is still a need for better diagnostic definition and prognostic stratification within subtypes. Therefore, we performed a direct comparison of molecular features of the four LNET subtypes, which to date has been lacking, on a cohort of 148 cases classified by five independent pathologists according to WHO criteria [1], including morphology, appropriate immunohistochemistry, and proliferation rate.
The results of our study can be summarized as follows: (1) carcinoids and carcinomas shared most of the altered genes, but with different prevalence rates among subtypes; (2) when mutations and copy number changes were combined, MEN1 alterations were almost exclusive to carcinoids, whereas alterations of the TP53 and RB1 cell cycle regulation genes and PI3K/AKT/mTOR pathway genes were significantly enriched in carcinomas; (3) chromatin-remodelling genes showed comparably high mutation rates in carcinoids and carcinomas; (4) rare carcinoids showed a hypermutated phenotype, which was associated with damaging POLQ mutations; (5) the prevalence of CNA was lower in carcinoids than in carcinomas -however, ACs showed an intermediate profile, with gains of several genes at rates similar to those in carcinomas, and loss of MEN1 at a frequency comparable to that in TCs; and (6) RB1 mutation and TERT gain were independent unfavourable prognostic markers in all LNET series, whereas MEN1 mutation was associated with poor prognosis in ACs, and KMT2D mutation was associated with longer survival in SCLCs.
No genetic alteration was exclusive to a specific LNET subtype, but the differences between carcinoids and carcinomas reside in the prevalence rates of the most frequently mutated genes, with the exception of SMARCA2, which was altered in LCNECs only (p = 0.011). Overall, this is in line with previous reports, and further demonstrates that the phenotypic classification of LNETs into carcinoids (ACs and TCs) and carcinomas (LCNECs and SCLCs) is supported by the different molecular landscapes of these tumours [9,12,25].
When mutations and CNAs were combined, inactivating alterations of TP53 and RB1 were enriched in carcinomas (p < 0.0001), whereas MEN1 alterations were almost exclusive to carcinoids (p = 0.0004). However, the fact that the same gene alterations found in carcinomas are identified in low-grade tumours at lower prevalence rates may suggest the existence of a progression of malignancy, with the development of secondary high-grade neuroendocrine tumours from pre-existing carcinoids. This has been suggested on molecular grounds in pancreatic and thymus neuroendocrine tumours [26,27], reinforcing the biological relevance of intratumour heterogeneity.
Chromatin-remodelling genes, which include those encodig covalent histone modifiers and subunits of the SWI-SNF complex, emerged from the present study as the group of genes with the highest mutation rate in both carcinoids (45%) and carcinomas (55%). Fernandez-Cuesta et al reported a similar mutation frequency of chromatin-remodelling genes (52%) in their study of 44 carcinoids, including 35 TCs and nine ACs [9]. A more recent study by Rekhtman et al reported a 78% mutation frequency of chromatin modifiers in 45 LCNECs [25]; in the same publication, the authors also re-analysed data of 42 SCLCs from Hyman et al [28] and Paik et al [29], and showed that 64% of SCLCs have chromatin-remodelling gene alterations. Our direct comparison of all subtypes complements these previous observations, and suggests that chromatin modifiers are major players in the pathogenesis of all LNETs. However, their role in carcinoids may be more relevant, as these tumours have a low mutational background, whereas carcinomas have a high mutational background and feature inactivation of genes involved in cell cycle control, such as TP53 and RB1 (carcinomas versus carcinoids, p < 0.0001). Carcinomas are also enriched in mutations of the oncogenic PI3K/AKT/mTOR pathway (carcinomas versus carcinoids, p = 0.0313), a potential therapeutic target to be investigated in prospective trials [10].
In our mutational analysis, carcinoids showed a lower mean number of mutations (TCs, 0.7; ACs, 1.8) than carcinomas (LCNECs, 4.6; SCLCs, 5.8), in line with current literature. However, two carcinoids (one TC and one AC) showed a hypermutated phenotype, which is defined as a gross excess of point mutations relative to the same tumour type [30]. The total numbers of mutations of these two cases were nine in the TC and 14 in the AC. Moreover, the total numbers of mutations determined on WES were 228 for the TC and 434 for the AC, well beyond the average mutational burden found on WES in our carcinoids (14.9 mutations/case) and those reported in the literature (14.0 mutations/case, as calculated from Fernandez-Cuesta et al) [9]. Common to these cases was the presence of inactivating mutations of POLQ that could be linked to this peculiar genomic profile. POLQ encodes DNA polymerase θ, an enzyme implicated in DNA repair [31]: POLQ-defective mammalian cells are susceptible to DNA double-strand breaks, and show increased rates of homologous recombination [21,32]. Interestingly, one of these cases also showed CNAs in several chromosomes, suggestive of defective DNA repair. The hypermutated carcinoids had proliferation rates within the range of their respective WHO categories. Mitotic counts were TC = 1 and AC = 6 per 2 mm 2 , with the WHO criteria setting being 0-1 for TC and 2-10 for AC [1]. The Ki67 proliferation index was TC = 7% and AC = 15%, which fits in the range of their respective categories [33]. Moreover, these cases had no clinicopathological evidence of aggressiveness, and both patients were disease-free at 21 months (TC) and 17 months (AC). As this is a short follow-up for carcinoids, we have no definitive indication as to the clinical significance of this phenotype. A third carcinoid showed CNAs consistent with chromothripsis of chromosomes 2, 11, and 20. Chromothripsis has been recently described in a stage III atypical carcinoid of a smoker patient [9], whereas our case was a stage IV atypical carcinoid of a non-smoker patient.
There were fewer CNAs in carcinoids than in carcinomas. Within individual subtypes, losses of MEN1 in ACs and losses of RB1 in SCLCs were the most frequent events, in agreement with previous literature [12,34]. However, an increase in the number of CNA events was observed from low-grade to high-grade tumours, especially for concomitant gains of TERT, SDHA and RICTOR located on the p-arm of chromosome 5. On WES analysis, we observed gain of chromosome 5p, with SDHA and RICTOR being located at the edges of and TERT within the altered chromosomal region. Therefore, as the limited material available did not allow us to perform fluorescence in situ hybridization or WES analysis on the whole LNET series, these three genes were included in the validation NGS panels, and their simultaneous CNAs were assessed as a surrogate of 5p gains. Gain of 5p has already been reported as a common event in all lung cancer types, suggesting that genes located on this chromosomal arm might be involved in development of all lung tumour types [35]. Of those genes, RICTOR encodes for a protein of the mTORC2 complex, whose amplification has recently been demonstrated to define the sensitivity of lung cancer cell lines to mTORC1/2 inhibitors [36]. When mutational and CNA analysis were combined, ACs were shown to be hybrid tumours sharing molecular features of both low-grade (TC) and high-grade (LCNEC and SCLC) tumours. In particular, ACs had frequencies of MEN1 losses and mutations comparable to those in TCs, and their overall mutational burden was 2.6 times higher than that in TCs and 2.9 times lower than that in carcinomas. However, they shared with carcinomas a similar prevalence of copy gains in several genes, including the chromosome 5p genes TERT, SDHA, and RICTOR, but also SRC and MYCL.
Survival analysis highlighted the prognostic relevance of RB1 and TERT regardless of subtype, whereas KMT2D was prognostically relevant for SCLCs, and MEN1 for ACs. Mutation of RB1 and copy gain of TERT are reported herein for the first time as independent predictors of poor prognosis in patients with LNETs. A correlation between alterations of these genes and poor prognosis has been observed for RB1 in non-small-cell lung cancers [37] and anaplastic astrocytoma [38], and increased mRNA expression of TERT was previously reported as a poor prognostic marker in breast [39] and lung [40] cancers. We report for the first time that mutations affecting KMT2D in SCLC patients correlate with longer survival, as previously observed for pancreatic adenocarcinoma [41]. The value of MEN1 mutation as an unfavourable prognostic predictor in patients with AC is in line with our previous finding of a significant association between inactivation of MEN1 and poor prognosis [34].
The fact that our discovery set included a higher number of carcinoids than of carcinomas was attributable to the following: (1) the availability of high-quality DNA from adequate amounts of frozen material -this is obtainable from surgically treated tumours, and SCLCs and LCNECs are rarely surgically resected; and (2) permission to use tissue for research, which is limited by the need to preserve material for clinical use. This could be considered a weakness of the study, as it might have led to a low-grade-driven validation gene list. However, to overcome this potential bias, we increased the gene list derived from the discovery screen with genes identified as altered in the current LNET literature [9,10,12,13].
In conclusion, our comprehensive molecular study encompassing all four histological variants of LNET indicated that a featured molecular portrait stands behind the commonly adopted histological classification, with a mixture of genetic similarities and differences among the four WHO LNET histotypes. This was mirrored by survival analysis, in which alterations in RB1 and TERT were common to all subtypes, and showed an independent poor prognostic role in all LNETs, whereas mutations of MEN1 and KMT2D were specifically enriched and showed a prognostic impact in ACs and SCLCs, respectively. Other genes that we found to be altered in LNETs are, instead, amenable to therapeutic targeting, including alterations leading to PI3K/AKT/mTOR pathway activation, namely PIK3CA mutations and copy gains of PIK3CA and RICTOR [10]. Our data suggest that molecular profiling may complement histology to provide better diagnostic definition and prognostic stratification of LNETs that would be helpful for their clinical management.     Table S1. Ampliseq custom panels used for validation of mutations in 88 genes and copy number alterations in 13 genes. Table S2. Clinicopathological features of 148 lung neuroendocrine tumours and analyses executed. Table S3. Prevalence of mutations in 88 genes for 148 lung neuroendocrine tumours. Related to Figure S2A and Table 2. Table S4A. Discovery screen, whole exome sequencing: list of mutations found in 20 lung neuroendocrine tumours. Related to Supplementary Figure S1A. Table S4B. Discovery screen, high coverage targeted sequencing of 418 genes: list of mutations found in 46 lung neuroendocrine tumours. Related to Supplementary Figure S1B. Table S4C. Discovery screen, integration of whole exome sequencing and high coverage targeted sequencing of 418 genes: list of 36 genes mutated in at least two cases of 46 lung neuroendocrine tumours. Related to Supplementary Figures S1A and S1B .   Table S5A. Discovery screen, copy number alterations by whole exome sequencing: histotype-specific distribution in 20 lung neuroendocrine tumours. Related to Figure 3. Table S5B. Discovery screen, copy number alterations by high coverage targeted sequencing of 418 genes: histotype-specific distribution in 46 lung neuroendocrine tumours. Related to Supplementary Figure S1B. Table S6. Univariate analysis on 78 lung neuroendocrine tumours for selection of candidate molecular prognostic predictors to be included in multivariate survival analysis. Related to Supplementary Figure S3.

The Journal of Pathology
Understanding Disease