Deciphering the genomic, epigenomic and transcriptomic landscapes of pre-invasive lung cancer lesions

Abstract

The molecular alterations that occur in cells before cancer is manifest are largely uncharted. Lung carcinoma-in-situ (CIS) lesions are the pre-invasive precursor to squamous cell carcinoma. While microscopically identical, their future is in equipoise with half progressing to invasive cancer and half regressing or remaining static. The cellular basis of this clinical observation is unknown. Here, we profile the genomic, transcriptomic and epigenomic landscape of CIS in a unique patient cohort with longitudinally monitored pre-invasive disease. Predictive modelling identifies which lesions will progress with remarkable accuracy. We identify progression-specific methylation changes on a background of widespread heterogeneity, alongside a strong chromosomal instability signature. We observe mutations and copy number changes characteristic of cancer and chart their emergence, offering a window into early carcinogenesis. We anticipate this new understanding of cancer precursor biology will improve early detection, reduce over-treatment and foster preventative therapies targeting early clonal events in lung cancer.

Introduction

Lung cancer is the commonest cause of cancer death worldwide with 1.5 million deaths per year¹. Lung squamous cell carcinoma (LUSC) is the most common subtype in parts of Europe and second in the U.S.A.² Before progression to invasive LUSC, there is step-wise evolution of ever more disordered pre-invasive lesions, ranging from mild and moderate dysplasia (low-grade lesions) to severe dysplasia and carcinoma-in-situ (CIS; high-grade lesions).³ The accessibility of the proximal airways allows detection and monitoring of these lesions using high-resolution diagnostic approaches such as autofluorescence bronchoscopy (AFB)⁴. This technique enables the acquisition of tissue throughout the natural history of LUSC, providing an excellent model to study early tumorigenesis in human patients.

Clinically, the optimal management of pre-invasive airway lesions remains unclear, despite the availability of surgery, radiotherapy and ablative techniques⁵. AFB with biopsy allows assessment of the size, gross morphology and histopathology of pre-invasive lesions (Fig. 1a, b) but cannot distinguish lesions that will ultimately progress to invasive tumours from those that will spontaneously regress. As such, indiscriminate surgical resection of pre-invasive lesions or external beam radiotherapy probably represent over-treatment: lesions will spontaneously regress in 30% of cases, patient co-morbidity and poor lung function impart considerable risk, and the presence of field cancerization means independent lung cancers frequently emerge at sites outside resection or therapy margins.⁶

Figure 1. Analysis of pre-invasive lung carcinoma-in-situ (CIS) lesions.

(a) Detection of bronchial pre-invasive CIS lesions by autofluorescence bronchoscopy. (b) Histological outcomes of bronchial pre-invasive lesions. (c) Overview of the study protocol. Patients with identified CIS lesions underwent repeat bronchoscopy and rebiopsy every 4 months. Definitive cancer treatment was only performed if pathological evidence of progression to invasive cancer was detected. The ‘index biopsy’ profiled in this study refers to the biopsy immediately preceding progression to invasive cancer or regression to low-grade dysplasia or normal epithelium. (d) Venn diagram of different -omics analyses performed on laser capture microdissection (LCM)-captured CIS lesions. Due to the small size of bronchial biopsies, not all analyses were performed on all samples

We reasoned that information on the future clinical trajectory of a pre-invasive lung lesion might be encoded in the genetic and epigenetic profile present at diagnosis. We therefore undertook a prospective cohort study of patients with pre-invasive squamous airway lesions. Patients were managed conservatively, undergoing surveillance AFB with biopsy and CT scanning every 4 and 12 months, respectively, with definitive cancer treatment only performed at the earliest pathological evidence of progression to invasive tumours (Fig. 1a, b).⁷ When a CIS lesion either progressed to invasive cancer or regressed to normal epithelium/low-grade disease, molecular profiling was performed on the preceding CIS biopsy from the same lesion – the ‘index biopsy’ (Fig. 1c). Index biopsies all demonstrated histologically and morphologically indistinguishable CIS and were classified as either ‘progressive’ or ‘regressive’. All such index CIS biopsies were subjected to a predetermined combination of transcriptomic, epigenetic and finally genomic profiling depending on DNA/RNA availability (Fig. 1d; Table 1; Extended Data Fig. 1; Supplementary Table 1).

Table 1. Demographic and clinical characteristics.

Table showing demographic and clinical characteristics of patients in the whole-genome sequencing, methylation discovery and validation, and gene expression discovery and validation datasets.

	Whole genome sequencing set (N=39)		Methylation discovery set (N=60)			Methylation validation set (N=27)			Gene expression discovery set (N=33)		Gene expression validation set (N=18) Clinical
Clinical Characteristics	Progression	Regression	Progression	Regression	Controls	Progression	Regression	Controls	Progression	Regression	Progression	Regression
Patients	21	8	13	7	16	9	7	8	16	14	9	8
Lesions Profiled	29	10	26	11	23	10	7	10	17	16	10	8
Gender
Male	18	8	11	7	15	7	7	7	14	10	7	4
Female	3	0	2	0	1	2	0	1	2	4	2	4
Age at bronchoscopy (years)
Mean	71.1	63.1	69.81	63.27	65.96	70.2	69.86	64.3	69.29	66.56	69.4	68.125
Median	72	65.5	70	67	68	73	68	63	70	67.5	71.5	68
Range	81	52-71	52-79	53-79	44-77	58-78	64-76	56-77	55-80	53-81	56-82	57-84
Smoking History (pack years)
Mean	54.4	54.9	58.08	31	41.95	57.3	62.14	37.71	57.07	47	49.125	59.2
Median	50	50	59.5	29	40	60	50	36	50	47.5	47.5	58
Range	100	9-141	32-141	5-88	20-65	40-75	30-141	20-60	22-141	5-141	30-75	30-96
COPD status
Yes	12	3	9	3	14	5	1	7	4	8	3	7
No	9	5	4	4	1	4	6	1	12	6	1	0
Previous History of Lung Cancer
Yes	12	2	6	2	9	7	4	3	5	4	3	4
No	9	6	7	5	7	2	3	5	11	10	6	4

Whilst molecular techniques have revolutionized our understanding of cancer biology, the key steps from normal cell to the point of cancer (uncontrolled growth and invasion) remain unclear. This is, to our knowledge, a unique collection of high-grade pre-invasive lung lesions for which prospective follow-up under conservative management enabled their natural history to declare.

Results

Patient Characteristics

Patients with pre-invasive lung cancer lesions were recruited through University College London Hospitals (UCLH) Early Lung Cancer Surveillance Programme (ELCSP). Full details of the surveillance protocol including eligibility criteria for patient inclusion have been previously described⁷. Briefly, the programme has recruited 140 patients to date with pre-invasive lung cancer lesions of varying histological grades. 129 index CIS biopsies were obtained from 85 patients and subjected to molecular analysis (Supplementary Table 1). Dependent on stored tissue quantity, in total, 51 samples from 42 patients underwent gene expression profiling; 87 samples from 47 patients underwent methylation profiling; and 39 samples from 29 patients underwent whole genome sequencing. Methylation and gene expression datasets were divided into independent discovery and validation groups.

Clinical characteristics within each analysis group are shown in Table 1. In comparing progressive and regressive samples, we found that progressive samples were associated with a higher pack-year smoking history in the methylation discovery group only (p < 0.01) and with increased age in the WGS group (p = 0.01). No clinical differences were consistently observed across the different analysis groups.

Characterization of CIS genomic profiles

We believe that the 39 CIS lesions are the first pre-invasive LUSC lesions to be whole-genome sequenced, so we compared the burden and spectrum of mutations in CIS with publicly available LUSC exome sequencing data from The Cancer Genome Atlas (TCGA). Due to differences between whole-genome and exome sequencing, only broad comparisons can be made. We observe a similar mutation burden and copy number profile between CIS samples and TCGA LUSC tumours (Fig. 2). There is congruency of type and prevalence of potential driver mutations, broadly defined as any mutation in a gene previously implicated as a driver of lung cancer, between CIS and LUSC samples⁸. We observe frequent alterations in TP53, CDKN2A, SOX2 and AKT2, and less frequent alterations in FAT1, KMT2D, KEAP1, EGFR and NOTCH1 in CIS lesions (Fig. 2; Supplementary Table 2). CIS mutational signatures^9,10 showed a strong tobacco-associated signal and were similar to those found in LUSC (Extended Data Fig. 2).

Figure 2. Genomic aberrations in pre-invasive lung carcinoma-in-situ (CIS) lesions.

Circos diagram comparing CIS genomic profiles with TCGA LUSC data. The outer histogram (A), shows mutation frequencies of all genes in TCGA data. The inner histogram (D) shows mutation frequencies in our CIS data. Profiles appear similar and no statistically significant differences were identified between the two datasets. Genes previously identified as potential drivers of lung cancer are labelled. Between the two histograms, average copy number changes are shown for TCGA data (B) and CIS data (C). Copy number gains are shown in red, losses in blue. Although differences between whole-genome and whole-exome sequencing techniques makes these datasets difficult to compare, we observe many similar features between the two; for example, gains in 3q and 5p, which are well recognised features of squamous cell lung cancer. In the centre of the circos plot, 39 rings represent the copy number profiles of our 39 samples, illustrating the individual contribution of each sample to the average values presented (E).

Marked aneuploidy was observed in CIS lesions, with somatic copy number alterations (CNAs) present across the genome (Fig. 2; Extended Data Fig. 3). The most frequent changes were associated with gain and amplification of multiple locations on distal 3q: this is known to be the most common genomic aberration in LUSC¹¹. Other recognised copy number associations identified in our data include gain/amplification in 5p, 8q and 19q and regions of loss/deletion in 3p, 4q, 5q, 8p, 9p and 13q.^12–18

Whilst most CIS samples have the genomic appearance of neoplasms, we observe six lesions which show markedly lower mutational load and fewer copy number alterations than the others (Extended Data Fig. 3; PD21884c, PD21885a, PD21885c, PD21904d, PD38317a, PD38319a). These samples have very few genomic changes, despite being CIS histologically. All of these six samples regressed to normal epithelium or low-grade dysplasia on subsequent biopsy. Four further samples met this end-point for regression, despite widespread mutational and copy number changes. However, with longer follow up one of these cases developed CIS recurrence (Extended Data Fig. 4a; PD21893a), and two developed invasive cancer on further surveillance (Extended Data Fig. 4b,c; PD21884a, PD38326a). Only one sample, PD21908a, showed sustained clinical regression after 9 years of follow up despite widespread molecular changes.

All but one progressive sample and all highly mutated regressive samples showed amplification in a small region of distal 3q (chr3:172516434-178440382). This region contains the gene ECT2, a regulator of cytokinesis which is associated with chromosomal instability. Progressive sample PD38320a had little change outside this region and did not harbour a TP53 mutation, suggesting that this amplification may be a crucial early event in LUSC tumorigenesis.

We compared genomic features between the 29 progressive and 10 regressive lesions. The three samples which showed evidence of progression after meeting our end-point for regression were excluded from this analysis. Comparisons of mutation burden between progressive and regressive lesions were performed by mixed effects modelling, allowing us to account for samples that come from the same patient. Even after correcting for patient age, smoking history and sample purity, progressive lesions had more somatically acquired mutations than those from regressive lesions, across base substitutions (p<0.001), indels (p=0.018), structural variants (p<0.001) and copy number changes (p<0.001) (Extended Data Fig. 5a-d). When the analysis was restricted only to substitutions that were fully clonal in each lesion, there were still substantially more substitutions in progressive than regressive lesions (p<0.001) (Extended Data Fig. 5e), suggesting that the increase in mutation burden is not due to recent subclonal diversification in progressive lesions. All the mutational processes (or signatures^9,10) identified in the CIS lesions contribute to the excess of mutations in progressive compared to regressive samples; however, only tobacco-associated signature 4 showed proportionally more mutations (p=0.017) (Extended Data Fig. 2f-j). Progressive lesions contained more putative driver mutations than regressive lesions (p=0.001) (Extended Data Fig. 5h; Supplementary Table 2). Importantly, no single cancer mutation perfectly discriminated between progressive and regressive lesions.

Within the biopsied lesions, clonal architecture was similar between progressive and regressive lesions (Extended Data Fig. 5e-g). For four patients in whom we sequenced multiple progressive lesions, the lesions shared many somatic mutations despite their different locality in the bronchial tree, indicating their probable derivation from a common ancestral clone. By contrast, multiple regressive lesions from two further patients did not share common mutations and so are likely to have arisen independently (Extended Data Fig. 6). There were no differences in telomere lengths between progressive and regressive lesions (p=0.59) (Extended Data Fig. 5i).

CIS transcriptomic and epigenetic profiles

Gene expression microarrays were performed on a discovery set of 17 progressive and 16 regressive CIS lesions. We identified 1335 genes with significant expression changes (FDR < 0.01); 657 genes were up-regulated and 678 down-regulated in progressive CIS lesions (Fig. 3a and Supplementary Table 3).

Figure 3. Altered methylation and gene expression in lung carcinoma-in-situ (CIS) lesions.

(a) Hierarchical clustering of 1335 significantly differentially expressed genes in progressive (n=17) and regressive (n=16) CIS lesions, based on a discovery set. Biological and clinical factors including age at diagnosis, gender, smoking history (pack years) and COPD status had no effect on CIS lesion gene expression profile (high expression = purple, low expression = orange). (b) Hierarchical clustering of the top 1000 significantly differentially methylated positions (DMPs) between progressive (n=36) and regressive (n=18) CIS lesions and controls (n=33). Biological and clinical factors including age at diagnosis, gender and smoking history (pack years) status had no effect on the methylation profile (hypomethylated DMPs = blue, hypermethylated DMPs = orange). (c) Principle component analysis of all profiled genes in progressive (n=27) and regressive (n=24) CIS lesions showing a clear distinction between progressive and regressive groups (p=0.0017). (d) Principle component analysis of all methylation data in progressive (n=36), regressive (n=18) and control (n=33) CIS lesions showing a clear distinction between progressive and regressive groups (p=6.8x10^-25). P values were calculated using multivariate ANOVA.

Differential analysis of methylation profiles was performed on a discovery set of 26 progressive, 11 regressive and 23 control samples. Widespread methylation changes were observed with 12,064 differentially methylated positions (DMPs), associated with 2,695 genes, at which methylation was significantly different between progressive and regressive samples (FDR < 0.01; |Δβ| > 0.3). 6,314 DMPs were hypermethylated and 5,750 hypomethylated in progressive CIS (Fig. 3b and Supplementary Table 3). 260 differentially methylated regions (DMRs) were identified, of which 151 (58%) overlap with DMRs between TCGA cancer and control data (Extended Data Fig. 7). Finally, we identified 36,620 differentially variable positions (DVPs) for which probe variance was markedly different between progressive and regressive groups.

Of the 1335 genes identified, TPM3, PTPRB, SLC34A2, KEAP1, NKX2-1, SMAD4 and SMARCA4 have previously been implicated as potential lung cancer drivers (Supplementary Table 4). Regarding methylation, the potential driver genes NKX2-1, TERT, DDR2, LRIG3, CUX1, EPHA3, CSMD3, MET, ZNF479, GRIN2A, PTPRD, NOTCH1, CD74, NSD1 and CDKN2A contain at least one significant DMP. Several genes which are significant in our gene expression analysis are also identified in our methylation data, including multiple genes in the homeobox family (HOXC8, HOXC9, HOXC10, HOXD10, HOXA11AS), previously implicated as an early epigenetic event in multiple cancers¹⁹. NKX2-1 (TTF-1) is the only putative driver gene to be identified in both gene expression and methylation analyses, and is also a member of the homeobox family. It is hypermethylated and underexpressed in progressive samples compared to regressive. This gene is widely used in diagnosis of lung adenocarcinoma and both underexpression and hypermethylation have been implicated in the development of this disease^20,21. NKX2-1 loss has been shown to drive squamous cancer formation in combination with SOX2 overexpression²²; focal gains in the 3q region containing SOX2 are commonly observed in progressive CIS (Extended Data Fig. 4).

Principal component analysis of all gene expression and methylation data showed a clear distinction between the progressive and regressive subgroups (p=0.0017 and p=6.8x10^-25, respectively) (Fig. 3c,d). In the methylation dataset, the regressive lesions closely clustered with the control normal epithelial cells. A history of chronic obstructive pulmonary disease (COPD) had an effect on case segregation (p=1.2x10^-5) but all other clinical and technical variables analysed, including smoking status and history of lung cancer, had no effect (Extended Data Fig. 8a-f). This was also the case for PCA analysis of the gene expression data (Extended Data Fig. 8g-k).

For methylation, one control and four regressive cases clustered with the progressive cases (Fig. 3d). Three of the four mis-classified regressive cases were subjected to whole-genome sequencing and were found to have more copy number alterations than other regressive samples (PD21884a, PD21893a, PD21908a). Two of these correspond to the samples discussed above, which showed signs of progression after meeting the clinical end-point of regression (Extended Data Fig. 4). For the control bronchial epithelium sample that was classified with the progressive lesions, CIS was detected in a biopsy specimen 12 months later from the same site. Thus, although we have formally treated these cases as mis-classifications, it is likely that the molecular data underpinning the apparent errors indicate a cellular phenotype that is not consistent with a straightforward regressive lesion.

Molecular signatures predict CIS outcome

The ability to predict if a pre-invasive lesion will progress to cancer has important clinical implications. For gene expression, we used the above pre-defined discovery set to define our classifier (n=33; 17 progressive, 16 regressive; 10-fold cross-validation applied). This was applied to a separate validation set (n=18; 10 progressive, 8 regressive). All samples in the validation set were classified correctly. When applied to external data from TCGA (n=551: 502 LUSC, 49 control), our 291-gene model was able to classify LUSC vs control samples with AUC=0.81 (Fig. 4a-c; Extended Data Fig. 9).

Figure 4. Carcinoma-in-situ (CIS) gene expression and methylation profiles are predictive of progression to cancer.

(a) Probability plot based on a 291-gene signature for correct class prediction (discovery set - red circles indicate progressive lesions, green circles indicate regressive lesions). (b) Challenging the 291-gene signature on a CIS validation set. Area under the curve (AUC) is 1 using Receiver Operating Characteristic (ROC) analysis. (c) Application of the 291-gene signature to TCGA LUSC data. Our signature classified TCGA LUSC vs TCGA controls samples with AUC of 0.81 (green circles indicate TCGA controls, orange circles indicate TCGA LUSC). (d) Distribution of methylation beta values across the genome in TCGA controls, CIS regressive and progressive and TCGA LUSC samples. Most probes are regulated at 0 or 1 in normal tissue but this regulation is reduced in both regressive and progressive CIS and TCGA LUSC samples. (e) Methylation Heterogeneity Index, defined as counts of methylation probes with 0.26 < ß < 0.88, for each sample. MHI is higher in regressive and progressive CIS and TCGA LUSC compared with TCGA controls and this can be used as an accurate predictor with AUC=0.96 for TCGA LUSC vs TCGA controls and AUC=0.74 for progressive vs regressive CIS. (f) Histogram of AUC values calculated by performing the same analysis used in (e) 10,000 times, with each run limited to a different random sample of 2,000 probes (AUC mean for TCGA LUSC vs TCGA controls is 0.95 (95% CI 0.92-0.98)). This demonstrates that a random sample of methylation probes can be an accurate predictor using this method.

An analogous analysis was performed for methylation using a discovery set of 60 samples and a validation set of 27 samples. This classified validation samples with AUC=0.99 and classified external TCGA samples (n=412: 370 LUSC, 42 controls) into LUSC vs controls with AUC=0.99, based on a 141-DMP classifier (Extended Data Fig. 10a-i).

We observed an increased number of methylation probes with intermediate methylation in TCGA LUSC cancer vs TCGA control samples (Fig. 4d), reflecting methylation heterogeneity in these samples. We therefore developed a methylation heterogeneity index (MHI), defined as the number of probes per sample with t_lo < ß < t_hi. Optimization based on our discovery set of 26 progressive and 11 regressive samples defined values of tlo = 0.26 and thi = 0.88. Control samples were not used in this analysis. This model classified progressive vs regressive CIS samples in our validation set with AUC=0.74 and TCGA LUSC vs TCGA control samples with AUC=0.96 (Fig. 4e; Extended Data Fig. 10j-n). Multivariate logistic regression in our CIS cohort demonstrated that this index was a predictor of progression status (p=0.017); previous history of lung cancer was also significantly associated (p=0.02), whereas smoking status, COPD status, age and gender were not.

Given the widespread nature of methylation changes, we hypothesised that this increase in heterogeneity may be a genome-wide process rather than specific to functional pathways. To test this theory, we assessed the predictive value of MHI calculated from a sample of 2,000 probes, randomly selected from across the genome. Running 10,000 simulations with each using a different random sample of 2,000 probes gave a mean AUC for TCGA LUSC vs TCGA control of 0.95 (95% CI 0.92-0.98) (Fig. 4f), and for progressive vs regressive CIS of 0.75 (95% CI 0.69-0.82) (Extended Data Fig. 10n). These results are similar to those obtained using the entire set of 450,000 probes, suggesting that methylation heterogeneity is a genome-wide process. However, these AUC values are lower than those obtained from our predictive model based on just 141 differentially methylated positions, suggesting that specific methylation changes are also important, on this background of generalised change.

To build a predictive classifier based on copy number, we used copy number derived from methylation data to increase sample size and classified 46 of 54 samples correctly (Extended Data Fig. 9g-i). The 154 predictive cytogenetic bands that we identified overlap with, but are not limited to, a model previously proposed by van Boerdonk et al.. Our model replicated their results, classifying 24/24 regressive samples and 9/12 progressive samples correctly²³ (Extended Data Fig. 9j-l). When applied to external data from TCGA (n=763: 524 LUSC, 239 control), our model was able to classify LUSC vs control samples with AUC=0.98 (Extended Data Fig. 9m-o).

We performed further analyses using only one sample per patient to demonstrate that our results are not dependent on multiple sampling. The first available sample for each patient was selected, with CIS samples prioritized over control samples for methylation data. Results are similar to our analysis above, validating our initial results (data not shown).

Although we cannot fully exclude that lesions meeting our end point for regression will progress in future, most patients in this cohort now have several years of follow up. Of 35 regressive lesions undergoing molecular profiling (Supplementary Table 1), mean follow up was 67 months (median 57 months, range 11-150 months).

CIN is an early marker of progression to cancer

To investigate possible drivers of tumorigenic progression, we performed a differential analysis of gene expression data between the progressive and regressive groups. 5 of the top 100 genes identified have been previously associated with chromosomal instability (CIN)²⁴, as defined by the previously published CIN70 signature²⁵ (ACTL6A, ELAVL1, MAD2L1, NEK2, OIP5). All five are up-regulated in progressive compared with regressive samples. CIN-related genes can predict progression (Fig. 5a); NEK2 expression alone predicts progression with AUC=0.93 (Fig. 5b).

Figure 5. Chromosomal instability is associated with progression to cancer.

(a) Mean expression of CIN-associated genes in CIS samples. Progressive (n=27) and regressive (n=24) CIS samples are well differentiated with AUC=0.96. Green circles indicate regressive CIS lesions; red circles indicate progressive CIS. (b) Plot of NEK2 expression across CIS samples demonstrates increasing expression with progression to cancer. Expression of this gene alone classifies progressive vs regressive CIS with AUC=0.93. (c) Pathway analysis of gene expression data between progressive (n=17) and regressive (n=16) CIS shows a strong chromosomal instability (CIN) signal, based on a discovery set. This signal remains strong when cell cycle genes are removed from the CIN70 signature. (d) Pathway analysis of methylation data demonstrating several cancer-related pathways up-regulated in progressive CIS compared with regressive CIS. Quoted significance values in (c) and (d) are calculated using 2-sided t-tests adjusted for multiple testing using a False Discovery Rate method, as implemented in the GAGE Bioconductor package.

Pathway analysis was performed using the gage Bioconductor package²⁶ to compare our differentially expressed genes to KEGG gene sets. The CIN70 gene set was the most significant gene set identified (adjusted p value 8.9x10^-32; up-regulated in progressive group), suggesting a role in early tumorigenesis. Cell cycle and DNA repair pathways were also implicated (Fig. 5c; Supplementary Table 5). Results were similar when cell-cycle associated genes were removed from the CIN70 signature, suggesting that this is a genuine CIN signal rather than a marker of proliferation.

Performing similar differential analysis of differentially methylated probes found widespread changes. The top probes identified were associated with cancer-associated cell signalling pathways, including TGF-beta, WNT and Hedgehog, as well as cell cycle and CIN-associated genes (Fig 5d).

This CIN signal is consistent with the observed pattern of widespread copy number change (Fig. 2). Overall copy number variation for a sample, as measured by Weighted Genome Integrity Index (wGII)²⁷, correlates with mean CIN-associated gene expression of that sample (Pearson r²=0.473) (Extended Data Fig. 5j). We also observe a correlation between local copy number of a gene and expression of that gene, consistent with previous results^28,29.

Discussion

In summary, we have delineated changes in the genomic architecture, genome-wide gene expression and DNA methylation of pre-invasive cancers with known histological evidence of subsequent disease progression or regression. The CIS genome shares many of the hallmarks of advanced, invasive LUSC but marked genomic, transcriptomic and epigenetic differences exist between lesions that are benign and those that will progress to cancer. Our data demonstrate the potential use of these differences in predicting outcome over current clinical practice.

Among the strongest pathways associated with progression is chromosomal instability, defined as a high rate of gain or loss of whole (or parts of) chromosomes. CIN is implicated in many human cancers, including lung, and has been suggested both as a prognostic marker and therapeutic target^30,31. Regressive lesions do not have the wholesale genomic instability of those that will progress and their epigenetic and transcriptional profiles more closely resemble normal bronchial epithelium than invasive cancers. Despite this, CIS lesions that spontaneously regress are genuine neoplasms; they harbour many somatic mutations, which can include known potential driver mutations. The mechanism of regression remains mysterious: it is unclear whether clones become exhausted and die out, potentially abetted by immune surveillance, or whether clones persist but phenotypically revert to an architecturally normal, physiological epithelium. Likewise the mechanisms of CIN are not well understood; our study paves the way for investigation of these CIN-associated genes in model systems to elucidate their role.

We present here the first major whole genome sequencing data of pre-invasive lung lesions. We acknowledge that, despite using the world’s largest cohort of such lesions, the study remains underpowered to detect less common genomic alterations. Expanding our knowledge in this area will require a major international collaboration. Likewise we acknowledge that whilst our predictive signatures demonstrate the power of molecular data in guiding management decisions, a prospective clinical trial using predictors derived from our data will be required before clinical use. Again, international collaboration will be required to develop an appropriately powered trial.

Despite these limitations, our data offer the first insight into the molecular map of early lung squamous cancer pathogenesis, foretelling an era in which molecular profiling will enable personally tailored therapeutic decisions for patients with pre-invasive lung disease.

Methods

Ethical approval

All tissue and bronchial brushing samples were obtained under written informed patient consent and were fully anonymised. Study approval was provided by the UCL/UCLH Local Ethics Committee (REC references 06/Q0505/12 and 01/0148). All relevant ethical regulations were followed.

Biological samples

All patients with pre-invasive lung cancer lesions were recruited through University College London Hospitals (UCLH) Early Lung Cancer Surveillance Programme (ELCSP). Full details of the surveillance protocol including eligibility criteria for patient inclusion have been previously described.² Briefly, the programme has recruited 140 patients to date with pre-invasive lung cancer lesions of varying histological grades. Patients undergo autofluorescence bronchoscopy (AFB) and CT/PET scans every four to six months during which multiple biopsy specimens are collected. This longitudinal sequential AFB procedure provides biopsies of the same lesion sampled repeatedly over time, allowing us to monitor whether the individual lesions have progressed, regressed or remained static².

For a given CIS lesion under surveillance, when a biopsy from the same site showed evidence of progression to invasive cancer or regression to normal epithelium or low-grade dysplasia, we define the preceding CIS biopsy as the ‘index’ lesion. An index lesion was defined as progressive if the subsequent biopsy at the same site showed invasive cancer, or as regressive if the subsequent biopsy showed normal epithelium or low-grade disease (metaplasia, mild or moderate dysplasia). Lesions which do not satisfy one of these end-points were excluded from this study. Patients with multiple fresh-frozen (FF) and formalin-fixed, paraffin-embedded (FFPE) tissue biopsies were identified for DNA methylation and gene expression analysis, respectively. Laser-capture micro-dissection (LCM) was used to selectively isolate CIS cells for molecular analysis, reducing the extent of contamination by stromal cells.

The following protocol was used to determine which profiling methods were applied to a given CIS lesion during our initial data collection phase:

• If FFPE samples were available, gene expression profiling was performed. For the first 33 samples (17 progressive and 16 regressive), gene expression profiles were generated using Illumina microarrays. Our predictive models are trained on this discovery set. Subsequently, a further set of 10 progressive and 8 regressive samples from 18 patients were profiled using a different microarray platform (Affymetrix) to validate our findings on an independent platform.

• If FF samples were available, DNA from these samples was first used for methylation profiling. Samples with sufficient DNA after DNA profiling were additionally subjected to whole-genome sequencing. After acquisition of sufficient samples for our methylation dataset (54 samples; 36 progressive, 18 regressive), only 29 samples had sufficient DNA for WGS, therefore we prioritised WGS over methylation for the subsequent 10 samples.

Tissue processing and laser-capture micro-dissection

FF or FFPE tissue sections (7-10μM thickness) were mounted on a MembraneSlide 1.0 PEN. Prior to cryosectioning, the slides were heat-treated for 4 h at 180°C in a drying cabinet to inactivate nucleases. To overcome the membrane’s hydrophobic nature and to allow better section adherence, the slides were then UV-treated for 30 min at 254nm. Prior to laser-capture micro-dissection (LCM), the slides containing the FF tissue sections for DNA extraction were washed in serial ethanol dilutions (50, 75, 100%) to remove the freezing medium (OCT) and to avoid any interference with the laser’s efficiency. For RNA extraction, FFPE sections were dewaxed using the Arcturus® Paradise® PLUS Reagent System (Applied Biosystems, Foster City, CA, USA). For each case, epithelial areas of pre-invasive disease were identified by haematoxylin and eosin staining of the corresponding cryosection (~7 μM thick). The presence of epithelial areas of interest was confirmed by histological assessment of each case by two histopathologists. LCM to isolate the tissue area/cells of interest was performed with the PALM Microbeam™ system (Carl Zeiss MicroImaging, Munich, Germany) on unstained sections. The micro-dissected material was catapulted into a 500μl AdhesiveCap that allows capture of the isolated tissue without applying any liquid into the cap prior to LCM, thus minimizing the risk of nuclease activity. The captured cells were stored at -80°C until DNA extraction or processed immediately for RNA.

DNA extraction

DNA from the micro-dissected tissue and bronchial brushing samples was extracted using QIAGEN’s QIAmp DNA Mini and Micro kits, respectively (Crawley, UK). Soluble carrier RNA was used to increase tissue DNA yield. Concentration was measured using the Qubit® dsDNA High-Sensitivity assay and Qubit® 2.0 Fluorometer (Life Technologies, Paisley, UK). Nucleic acid quality and purity was estimated based on the A_260/280 absorbance ratio readings using the NanoDrop-8000 UV-spectrophotometer (Thermo Scientific, Hertfordshire, UK). Only samples with an A_260/280 ratio of 1.7-1.9 were included in the study.

RNA extraction

RNA was extracted using the High Pure FFPE RNA Kit (Roche Applied Science, West Sussex, UK) according to manufacturer’s protocol. Quantification was carried out using the Quant-iT RNA assay kit and the Qubit® 2.0 fluorometer (Life Technologies, Paisley, UK). RNA integrity was analyzed using a BioAnalyzer 2100 (Agilent, Stockport, UK).

Bisulfite conversion

For each sample undergoing methylation profiling, 200 ng of DNA were bisulfite converted using the EZ DNA methylation kit (Zymo Research Corp., Orange, CA, USA) according to the manufacturer’s modified protocol for Illumina’s Infinium 450K assay. This protocol incorporates a cyclic denaturation step to improve the conversion efficiency³. The 10 μl final conversion reaction was concentrated down to 4 μl with a vacufuge plus vacuum concentrator (Eppendorf AG, Hamburg, Germany) and sent to UCL’s Genomics Core Facility for hybridization on the 450K BeadArray according to Illumina’s Infinium HD protocol (Illumina Inc., San Diego, CA, USA) as previously described.⁴

Infinium HumanMethylation450K raw data extraction and pre-processing

Illumina’s iScan fluorescent system was used to scan and image the arrays. DNA methylation data were extracted as raw intensity signals without any prior background subtraction or data normalization and were stored as IDAT files.

CpG-specific methylation levels (β-values; continuous value ranging from 0 to 1) for each sample were calculated as the ratio of the fluorescent signal intensity of the methylated (M) and unmethylated (U) alleles according to the following formula:

$$\begin{array}{cc}\beta =& \frac{\text{intensity}\text{of}\text{methylated}\text{allele}\text{(M)}}{\text{intensity}\text{of}\text{[unmethylated}\text{(U)}+\text{methylated}(\text{M})\text{allele}]+100}\end{array}$$

All subsequent raw β-value pre-processing, normalisation and down-stream analysis was performed using the Chip Analysis Methylation Pipeline (ChAMP) Bioconductor package with default settings.⁵

Analysis of differentially variable positions (DVP) was performed using iEVORA⁶. Beta values from ChAMP were used as input to iEVORA following normalization and batch correction.

Genome-wide gene expression array

The extracted FFPE RNA used to generate the gene expression profiles on the discovery set was sent to UCL’s Genomics Core Facility for hybridization on the Human Whole-Genome DASL (cDNA-mediated Annealing, Selection, extension and Ligation) beadarrays according to Illumina’s protocol (Illumina Inc., San Diego, CA, USA).

The extracted FFPE RNA used to generate the gene expression profiles on the validation set was sent to UK Bioinformatics Limited for hybridization on the Clariom™ D Transcriptome Human Pico Assay 2.0 according to Affymetrix’s protocol (Thermo Fisher Scientific Waltham, MA, USA).

Principal Component Analysis (PCA)

In order to identify any potential factors of variability affecting sample/group segregation, we applied principal component analysis on all probes passing filters defined above (implemented in the prcomp method of the R stats package). Technical and biological variation was investigated for batch arrays, smoking (pack-years), age at initial diagnosis, gender and previous lung cancer history. The ability of these features to predict the first principal component was quantified using ANOVA analysis, implemented in the R aov method. p-values quoted are derived from this method.

Gene expression analysis

Raw gene expression data were expressed as log₂ ratios of fluorescence intensities of the experimental samples. Quantile normalization was applied to Illumina data, using Illumina GenomeStudio Gene Expression Module v1.0 software. For Affymetrix data, RMA normalization was applied as defined in the affy Bioconductor package. For analyses utilizing both data sets, only genes represented on both arrays were included and ComBat⁷ was used to adjust for batch effects.

Differential expression analysis was performed using the limma⁸ Bioconductor package. Raw p-values were adjusted by the Benjamini-Hochberg procedure to give a FDR.⁹ A significance threshold of FDR < 0.01 was used to select differentially expressed genes. Cluster analysis and visualization was performed using the pheatmap¹⁰ Bioconductor package.

Real Time PCR Validation

For microarray validation, total RNA from the 33 pre-invasive LUSC lesions undergoing Illumina gene expression profiling was reverse transcribed using qScriptTM cDNA Super-Mix (Quanta Biosciences, Lutterworth, UK) according to the manufacturer’s protocol. Real-time quantitative PCR was carried out in eight genes using the SYBR-green master mix (Applied BioSystems, Bleiswijk, Netherlands) in an Eppendorf real-time PCR Machine (Eppendorf, Stevenage, UK). Findings were validated using quantitative PCR (qPCR) for four up-regulated (GAGE5, GPNMB, MMP12 and STC2) and four down-regulated (SPDEF, LMO7, OBSCN and MT1E) genes. Gene-specific primers were designed inside or nearby the microarray sequence targeted, using Primer Express Software v2.0 (Thermo Fisher Scientific). Relative gene expression was quantified using the threshold cycle (Ct) method and normalized to the amount of CTBL and CEP250, which met the criteria of less variation between samples and compatible expression level with the studied genes. Each sample was tested in triplicate and a sample without template was included in each run as a negative control. Correlations between microarrays and real time PCR data were measured using the Pearson coefficient. From microarray and real time PCR data, we calculated the progressive/regressive ratio for each gene expression. All eight genes tested were significant in our differential microarray analysis with FDR < 0.05. A high degree of correlation (r=0.982) was observed between qPCR and array data.

Predictive modelling

For methylation, gene expression and copy number data we applied Prediction Analysis of Microarrays (PAM)¹¹ to predict whether a sample was progressive or regressive based on its molecular profile. The Bioconductor pamr package was used. In all presented analyses we select a threshold which minimizes the number of data inputs required whilst maintaining the minimum possible number of classification errors.

PAM calculates the probability of each sample being progressive. We describe this value as a ‘Progression Score’. ROC analytics were performed on these progression scores to determine their value as a diagnostic test, using the pROC¹² and PRROC¹³ Bioconductor packages.

For methylation and gene expression data a predictive model was trained on the training set and subsequently applied to an independent validation set. Regressive and control samples were grouped together for the methylation data analysis. ROC analytics were performed only on the validation set. Internal cross-validation was used for methylation-derived copy number data due to smaller sample size (control samples are used as a baseline to calculate copy number, therefore are excluded from predictive analysis).

When multiple lesions from one patient were included in an analysis, these were treated as independent events as they were always taken from different sites in the lung. The outcome of a lesion (whether it progressed or regressed) was determined on a perlesion basis; the lesion was assigned to the progressive group only if cancer developed at the same site in the lung, and to the regressive group only if normal or low-grade dysplasia was obtained from the same site in the lung.

In some cases different technologies were used, for example our gene expression discovery set used Illumina microarrays whereas our validation set used Affymetrix. In such instances, both data sets were reduced to the subset of genes covered by probes in both platforms prior to creating a predictive model. The ComBat method from the sva Bioconductor package was used to correct for batch effects between the different platforms. In the case of RNAseq data, we used the voom transformation defined in the limma Bioconductor package to derive data comparable to expression data prior to batch correction with ComBat.

A second predictive model based on methylation probe variation was also developed. For a given sample we defined Methylation Heterogeneity Index (MHI) by counting all probes with beta values between 0.26 and 0.88. These thresholds were optimized by calculating MHI for a range of different threshold values, and choosing those with the highest AUC for progressive vs regressive in our discovery cohort. We used ROC analytics to assess this model as a predictor of TCGA cancer vs control samples, and of progressive vs regressive samples in our validation cohort. We demonstrate in the main text that applying this method to a random sample of 2,000 probes performs similarly to using the entire array. We ran simulations using different sample sizes and found that performance with n=2000 was similar to that of the entire array. To investigate potential confounding variables we use binomial logistic regression, implemented in the R glm method, to assess whether outcome (progression/regression) could be predicted by MHI, smoking status, COPD, previous history of lung cancer, age or gender. Control samples derived from brushings were excluded from these analyses.

Copy number variation analysis

For samples with whole-genome sequencing available we used ASCAT¹⁴ to derive local copy number estimates as described below. To increase our sample size for comparative analyses, Copy number variation (CNV) data were obtained from non-normalised methylated and unmethylated signal intensities of probes in the 450K array as previously described¹⁵ using the ChAMP Bioconductor package with default settings. Copy number (CN) profiles for progressive and regressive cases were obtained using the control cases for baseline normalisation. A previously defined threshold of ±0.3 was used for the identification of single CNV. Probes associated with highly polymorphic regions (e.g. major histocompatibility complex) were removed from the analysis. The analysis generated group CN frequency plots and CN profiles for each sample. For samples with both methylation and sequencing data available we observed good correlation between copy numbers derived from the two different methods (data not shown).

For comparison with previous results, the ChAMP pipeline was then modified to return CNV values per-probe. Probe locations were matched to cytogenetic bands using the Ensembl GRCh37 assembly, obtained from http://grch37.rest.ensembl.org/info/assembly/homo_sapiens?content-type=application/json&bands=1, such that copy number variation could be assessed by cytogenetic band. The mean CNV value for each of 778 cytogenetic bands was calculated for each of our 54 samples. Limma analysis was used to identify bands that differed significantly between progressive and regressive samples with BH-adjusted p-value < 0.05. Predictive modelling was performed using PAM to find bands predictive of progression, using the same method as for gene expression data. Due to the low number of regressive samples, an internal cross-validation method was used rather than separate discovery and validation sets.

Following identification of predictive cytogenetic bands, PAM modelling was repeated with the dataset limited to only those bands identified by van Boerdonk et al: 3q26.2–29, 3p26.3–p11.1 and 6p25.3–p24.3.^16,17 This model was also accurate.

Finally, we applied our model to the validation data set of 24 regressive and 12 progressive samples used by van Boerdonk et al (GEO accession number GSE45287). These data were measured using a different microarray platform (arrayCGH). We assigned each probe to a cytogenetic band, and took the mean values to create a matrix of expression values by band. Our model was applied to the subset of chromosomal bands present in both data sets (760 of 778 bands). ComBat was used for batch correction between the two platforms. Our model correctly predicted 24/24 regressive samples and 9/12 progressive samples, replicating the results of van Boerdonk et al.

External validation using TCGA

Lung cancer methylation datasets publically available through The Cancer Genome Atlas (TCGA) were downloaded using GenomicDataCommons download tools¹⁸. We obtained the normalized β-values of 370 LUSC samples and 42 normal controls. ComBat was used to correct for batch effects between our data and TCGA data. These data were used as an external validation set to test our predictive models, and as input for our differential analysis of progression drivers from control through CIS to cancer.

Gene-expression microarray data sets comparable to our data were not publically available. RNAseq data was available from TCGA for 502 LUSC samples and 49 control samples. We applied a voom transformation¹⁹ to these data, which uses normalized log-counts-per-million as an approximation for expression values, and hence allows comparison of RNAseq data with our gene expression pipeline. ComBat was used to correct for batch effects. The predictive model generated using PAM on our gene expression microarray data was applied to voom-transformed RNAseq data from TCGA and shown to be predictive (Fig. 4C). We therefore demonstrate the applicability of our model to this fully independent data set. These data were again used as input to our differential analysis of progression drivers.

Pathway analysis

For gene expression data, the GAGE Bioconductor package²⁰ was used with KEGG gene sets^21–23 to identify pathways associated with genes differentially expressed in our analysis of progression to cancer (BH-adjusted p-value <0.01). In addition to these pathways we use the CIN70 signature defined by Carter et al.²⁴ to assess for a chromosomal instability signal. We also use a subset of the CIN70 genes with cell-cycle associated genes²⁵ removed to ensure that our signal is genuinely CIN-related, rather than a measure of proliferation.

Methylation data was analysed in the same way, using beta values as input to GAGE. In cases where there are multiple methylation probes for a single gene we use the mean beta value over that gene as input to pathway analysis. We acknowledge that using mean signal may be insensitive to single-probe methylation changes, however given the scale of changes observed we believe it will identify areas of large methylation change.

Genomic sequencing

We created genome-wide shotgun libraries (insert size 331-367 bp) from native DNA using the Agilent Technologies Custom SureSelect Library Prep Kit library (cat no. 930075). 150 bp paired-end sequence data were generated using the Illumina HiSeq X Ten system. Sequenced data were realigned to the human genome (NCBI build 37) using BWA-MEM. Unmapped reads and PCR duplicates were removed. A minimum sequencing depth of 40x was required.

Somatic mutation calling and annotation

Single base somatic substitutions were identified by our in-house algorithm Cancer Variants through Expectation Maximisation (CaVEMan: https://github.com/cancerit/CaVEMan)²⁶. This algorithm compares the sequence data from each tumour sample to its matched normal and calculates a mutation probability at each locus. This calculation incorporates information from aberrant cell fraction and copy number estimates from the Allele-Specific Copy number Analysis of Tumours (ASCAT) algorithm (https://www.crick.ac.uk/peter-van-loo/software/ASCAT).^14,27 Additional post-processing as described previously²⁸ was implemented. Any putative driver mutations were visually inspected with Jbrowse.²⁹ For every substitution that passed all filters in at least one sample, we counted the number of wild-type and mutant reads at the same position in all other samples from the same patient to see if that mutation was also present in related samples but had not been called.

Somatic small insertions and deletions

These were identified using our in-house algorithm Pindel.^30,31 As with substitutions, all putative driver mutations were visualised with Jbrowse.

Somatic structural variant detection”

Abnormally paired read pairs were grouped using an in-house tool, “Brass”.³² Read groups overlapping genomic repeats, reads from the matched normal, or from a panel of unmatched normals were ignored. Read pair clusters were then filtered by read remapping. Read pair clusters with >50% of the reads mapping to microbial sequences were removed. Finally, candidate SV breakpoints were matched to copy number breakpoints as defined by ASCAT within 10 kb. Candidate SVs that were not associated with copy number segmentation breakpoints and with a copy number change of at least 0.3 were removed. All putative driver rearrangements were visually inspected using IGV.^33,34

Somatic copy number events, ploidy, and stromal contamination

Copy number changes were derived from whole-genome sequencing data using the ASCAT algorithm. This algorithm compares the relative representation of heterozygous SNPs and the total read depth at these positions to estimate the aberrant cell fraction and ploidy for each sample, and then to determine allele-specific copy number.

Weighted Genome Integrity Index

To estimate the overall chromosomal instability of a sample, we use the Weighted Genome Integrity Index (wGII) score³⁵. This is calculated by measuring the percentage of the genome which is abnormal, corrected such that each chromosome is equally weighted.

Mutation annotation

Lung cancer driver genes were selected from the COSMIC Cancer Gene Census (CGC) v85 (cancer.sanger.ac.uk)³⁶. CGC data was downloaded on 20^th June 2018. Genes annotated in the CGC as potential drivers in lung cancer or NSCLC were included. Those specific to adenocarcinoma were excluded as our samples are precursors to squamous cancers. Genes identified in two large studies of squamous cell cancer, and some additional genes based on expert curation of the literature (ARID1A, AKT2, FAT1, PTPRB) were included if they were present in the CGC – even if they were not annotated explicitly as implicated in lung cancer. Both Tier 1 and Tier 2 genes were included. A total of 96 genes were selected as putative lung squamous cell carcinoma drivers (Supplementary Table 4).

Mutations affecting these putative driver genes were annotated as driver mutations if they passed the following filters:

• The mutation type (e.g. missense, frameshift, amplification) must have been validated in the CGC for the affected gene.

• For genes annotated as tumour suppressors, mutations determined to have High or Moderate impact using Ensembl’s Variant Effect Predictor³⁷ were classed as driver mutations.

• For genes annotated as oncogenes, we checked the specific mutation against COSMIC mutation data for lung carcinomas. If the specific mutation occurred 3 or more times in this dataset it was classed as a driver mutation.

• For genes annotated as fusion proteins, translocations with a translocation partner gene matching validated tranlocation partner genes in the CGC were classed as driver events.

• Copy number amplifications and deletions were all classed as driver events if amplifications/deletions in the affected gene have been previously validated in the CGC. We included homozygous deletions of tumour suppressor genes and amplifications to more than double the sample ploidy for oncogenes.

Driver mutation discovery was also attempted using dndscv³⁸. This was underpowered, however, and only yielded TP53 and CDKN2A as genes under positive selection. This package was also used to estimate the global dNdS for both progressive and regressive lesions.

Subclonality analysis

The number of subclones contributing to a sample and their relative contribution was estimated by using a modified version of the sciClone Bioconductor package³⁹. sciClone uses a Bayesian method to allocate mutations to clusters based on their variant allele frequency (VAF). By default, sciClone only considers regions that are copy number neutral and LOH-free. Given the significant aneuploidy in our data set we overcame this limitation by clustering on cancer cell fraction (CCF) rather than VAF. Briefly, cancer cell fraction represents the fraction of cancer cells in which a given mutation is present, therefore clonal mutations will have CCF=1. Following the method of McGranahan et al.⁴⁰, we estimated the CCF for each mutation with a 95% confidence interval. Mutations for which 1 lay within this confidence interval were labelled as ‘clonal’, other mutations as ‘subclonal’.

CCF values for each mutation were then used as input to sciClone in place of VAF values to quantify clusters present (divided by 2 such that clonal mutations have a value of 0.5). As CCF corrects for local copy number, all regions were assumed to have copy number of 2, allowing sciClone to group mutations based only on their CCF estimates. A minimum tumour sequencing depth of 10 was required for each mutation.

Where more than one sample from a given patient was available, both one dimensional and multi-dimensional clustering were performed. Results from one dimensional clustering were used in the comparison of numbers of clones and proportion of clonal mutations between progressive and regressive lesions, in order to provide as fair a comparison as possible.

Extraction of mutational signatures

To obtain an approximate estimate of the contribution of different known mutational signatures to each sample, we used the MutationalPatterns Bioconductor package⁴¹. As a reference set of mutational signatures, we used a table with the relative frequency of each of the 96 trinucleotide substitutions across 30 known mutation signatures,^42,43 available through the COSMIC website (http://cancer.sanger.ac.uk/cosmic/signatures).

After a first run which indicated the most likely contribution of each signature, it seemed that the majority of substitutions were contributed by signatures 1, 2, 4, 5, and 13, which have been described to be the strongest signatures in lung squamous cell cancer.⁴⁴ Some contribution was identified from signatures 16, 8, 18 and 3 in our initial analysis; however, in this context it is likely that these represent overfitting given that signature 16 is similar to signature 5, and signatures 8, 18 and 3 are similar to signature 4. We therefore ran the algorithm a second time, this time only using a 5x96 matrix of mutational signatures 1, 2, 4, 5 and 13. All mutations were thus forced to belong to one of these five mutational signatures.

For a comparison of the clonal vs subclonal mutational processes in each sample, substitutions were annotated as clonal or subclonal based on CCF as described above. These were then run through the MutationalPatterns package.

Comparison of mutational burden and signatures with other cancer types

Signatures of mutations in our CIS dataset were compared with mutational signatures found in lung squamous cell cancer. Raw whole-exome sequencing data for this cancer type was downloaded from TCGA, and run through our substitution-calling algorithm CAVEMaN as described above. We then looked at the total number of subsitutions called, and estimated the contribution of each mutational signature using the methods described above. Only coding regions of the CIS whole-genome sequencing data were compared to these exomes.

Estimation of telomere lengths

Telomere lengths were estimated using telomerecat⁴⁵, and were compared in progressive and regressive groups. Telomerecat is a de novo method for the estimation of telomere length (TL) from whole-genome sequencing samples. The algorithm works by comparing the ratio of full telomere reads to reads on the boundary between telomere and subtelomere. This ratio is transformed to a measure of length by taking into account the fragment length distribution. Telomerecat also corrects for error in sequencing reads by modeling the observed distribution of phred scores associated with mismatches in the telomere sequence. Samples were analysed in two groups corresponding to two separate sequencing batches, as per the telomerecat documentation.

Extended Data

Extended Data Fig. 1. Experimental workflow.

Flow diagram illustrating which profiling techniques were applied to which samples. Biopsies taken fromindex CIS lesions were stored as fresh frozen (FF) and formalin-fixed paraffin embedded (FFPE). DNA was extracted from FF biopsies. The first 54 samples studied that had sufficient extracted DNA passing quality control (QC) underwent first methylation profiling, then whole-genome sequencing (WGS) when sufficient remaining DNA was available. Due to the low DNA quantity extracted from some biopsies, the methylation dataset (n=54) was larger than the WGS dataset (n=29), therefore the subsequent 10 samples underwent WGS directly without methylation profiling. RNA was extracted from FFPE samples and underwent gene expression profiling when RNA passed QC. To ensure validity of our conclusions across orthogonal platforms we usedIllumina microarrays to profile a discovery set of 33 samples, then subsequently used Affymetrix micro arrays to profile an independent validation set of 18 further samples.

Extended Data Fig. 2. Mutational signatures of CiSlesions.

a–d, The contribution of each of five pre-selected mutational signatures to each lesion is shown. These five mutational signatures, associated with CpG deamination (1), APOBEC (2 and 13), tobacco (4) and unknown aetiology (5), were selected based on an initial run using all 30 mutational signatures, which showed that these were present in the data and in signature extractions from lung squamous cell cancer (LUSC) datasets. The number of substitutions attributed to each signature is shown (a-b) as well as the proportion of mutations attributed to each mutational signature (c-d). Samples from the same patient share the same identifier except for the final letter; for example, PD21883a and PD21883d are two samples from the same patient. e, Comparison of the mutational signatures of CIS lesions to those found in lung squamous cell cancer (LUSC). LUSC data were downloaded from TCGA and mutations called with our algorithms. All mutations from all samples from each cancer type were pooled for this analysis. The colour scale indicates the proportion of substitutions in each sample that are attributed to each signature. f-j, Comparison of the relative proportion of mutations attributed to each signature between progressive (red; n = 29) and regressive (green; n = 10) CIS samples. P values were calculated using likelihood ratio tests of a mixed effects model with outcome (progressive or regressive) included as a fixed effect versus a model that was identical but for the fact that outcome was not included as a fixed effect. Only signature 4 (smoking-associated) was significantly different between the two groups. Boxplots are generated using the R boxplot function, which displays the first and third quartile as hinges and places whiskers at the most extreme data point that is no more than 1.5 times the length of the box away from the box.

Extended Data Fig. 3. Genome-wide copy number changes of CIS lesions.

Visualization of copy number changes for 39 whole-genome-sequenced CIS samples. Rows represent samples, genomic position is represented on the x-axis. Local copy number gains are illustrated in red, losses in blue. We observe widespread changes in progressive CIS samples and a subset of regressive samples.

Extended Data Fig. 4. Documentation of biopsy history and chronology of lesion appearance in three misclassified regressive cases.

a, Case 1 (PD21893a) appeared to regress from a CIS lesion (07/2012) to squamous metaplasia (SqM; 11/2012). However, again, CIS was subsequently reconfirmed by biopsy (05/2013). b, Case 2 (PD21884a) had a lobectomy for T1N0 lung squamous cell cancer (LUSC) in the left upper lobe (LUL) and was under surveillance for carcinoma-in-situ (CIS) at the resection margins. A subsequent, high-grade CIS lesion (08/2009) profiled for genome-wide DNA methylation changes was considered regressive since a follow-up biopsy on the same anatomical site demonstrated the presence of a low-grade, moderately dysplastic (MoD) lesion (11/2009). A subsequent biopsy, however, was classified as CIS (02/2011) and the lesion then remained static for 26 months but eventually progressed into invasive cancer (04/2014). c, Case 3 (PD38326a) had an initial diagnosis of CIS (11/2015) followed by regression to normal epithelium (03/2016). CIS was subsequently identified at the same site (03/2017), with invasive cancer diagnosed on subsequent biopsy (07/2017).

Extended Data Fig. 5. Genomic aberrations in pre-invasive lung CIS lesions.

Comparisons of the number of substitutions (a), small insertions and deletions (b), genome rearrangements (c) and copy number changes (d), showing significantly more genomic changes in progressive (n = 29) than regressive (n = 10) lesions. Although there were more clonal substitutions in progressive than regressive lesions (e), the proportion of substitutions that were clonal and the number of clones were similar (f-g). Progressive lesions had more putative driver mutations (h). Telomere lengths (base pairs) were similar between the two groups (i). To confirm an association between CIN gene expression and copy number change we correlated Weighted Genome Integrity Index (wGII) with mean CIN gene expression for the CIS samples in which we have both gene expression and whole-genome sequencing data (n = 11). Pearson correlation coefficient r² = 0.473 (j). All P values were calculated using likelihood ratio tests of a mixed effects model with outcome (progressive or regressive) included as a fixed effect versus a model that was identical but for the fact that outcome was not included as a fixed effect. Boxplots are generated using the R boxplot function, which displays the first and third quartile as hinges and places whiskers at the most extreme data point that is no more than 1.5 times the length of the box away from the box.

Extended Data Fig. 6. Subclonal mutational structure in progressive and regressive CIS lesions.

Heatmap showing the proportion of overlapping mutations between samples taken from the same patient. For four patients with lesions that would ultimately progress to cancer (denoted ‘P’), over half the mutations were shared between any two given samples, suggesting that the lesions were derived from a common ancestral clone. By contrast, for two patients with lesions that would ultimately regress (denoted ‘R’), almost no mutations were shared, suggesting that the lesions arose independently. Samples from the same patient are shown in the same color; PD38321a and PD38322a do belong to the same patient and were mislabelled during processing.

Extended Data Fig. 7. Differential molecular changes between progressive and regressive lesions.

Visualization of differential changes across the genome. A, shows all identified differentially methylated regions (DMRs) (hypermethylated regions in yellow, hypomethylated in blue) alongside a similar analysis comparing cancer and control samples from The Cancer Genome Atlas. We observe that 58% of DMRs identified in our progressive vs regressive analysis are also identified in cancer vs control. B, shows copy number changes across the genome in regressive CIS, progressive CIS and TCGA cancer samples. We observe congruency of copy number change, suggesting similar processes in the two cohorts.

Extended Data Fig. 8. Principal component analysis investigating effect of various biological, clinical and technical factors affecting correct case segregation for all DMPs and gene expression data.

a-f, Principal component analysis based on all methylation probes (n = 87; 36 progressive, 18 regressive, 33 control). (a) Smoking history (pack years). (b) Chronic obstructive pulmonary disease (COPD) status. (c) Previous lung cancer history referring to the presence of lung squamous cell cancer (LUSC) prior to identification of pre-invasive lesions. (d) Age at bronchoscopy (years); age of individual when pre-invasive lesion was first biopsied. (e) Gender. (f) Sentix ID. g-k, Principal component analysis for all gene expression data. (g) Smoking history (pack years). (h) COPD status. (i) Previous lung cancer history referring to the presence of LUSC prior to identification of pre-invasive lesions. (j) Age at bronchoscopy (years); age of individual when pre-invasive lesion was first biopsied. (k) Gender. P-values were calculated using multivariate ANOVA.

Extended Data Fig. 9. Predictive modeling and ROC analytics of gene expression and CNA data.

ROC and precision-recall curves for the predictive model based on gene expression data shown in Fig. 4A-C. Curves are shown for the CIS discovery set (a-b), CIS validation set (c-d) and application to TCGA LUSC data (e-f). Using an analogous method to gene expression and methylation we used copy number data derived from methylation arrays to predict lesion outcome. Probe-level copy number changes were aggregated over cytogenetic bands; these data were used as input to Prediction Analysis of Microarrays (PAM). g-i, Probability plot based on a 154 cytogenetic band signature for correct class prediction (red circles indicate progressive lesions, green circles indicate regressive lesions). The area under the curve for the 154-cytogenetic band signature is 0.86. j-l, Application of our predictive model to previously published data (van Boerdonk et al.) replicates their result, classifying all regressive and 9/12 progressive samples correctly. This dataset included pre-invasive samples of various histological grades, rather than only CIS. m-o, Application of our predictive model to TCGA copy number data. Samples were correctly classified into TCGA LUSC and TCGA control samples with an AUC of 0.98.

Extended Data Fig. 10. Predictive modeling of methylation data.

In addition to the predictive modeling based on probe variation shown in Fig. 5, we used differentially expressed methylation probes to create a predictor using a Prediction Analysis for Microarrays (PAM) method. The model was trained on a training set (a-c) consisting of 26 progressive samples, 11 regressive samples and 23 control samples, shown in red, green and blue, respectively. A predictor based on 141 DMPs was created. This was applied to a validation set of 10 progressive, 7 regressive and 10 control samples (d-f), predicting outcome with AUC = 0.99. g-i, Application of our predictive model to TCGA methylation data. Samples were correctly classified into TCGA LUSC and TCGA control samples with AUC = 0.99. j-m, ROC analytics and precision-recall curves for Methylation Heterogeneity Index (MHI) model presented in Fig. 4. Curves apply to cancer vs control (j-k) and progressive vs regressive (l-m), respectively. n, Histogram of AUC values using MHI model with random samples of 2000 probes, applied to progressive vs regressive data. This demonstrates that a similar AUC is achieved with a random sample of probes as when using the entire array.

Data Availability Statement

Whole-genome sequencing data have been deposited at the European Genome Phenome Archive (https://www.ebi.ac.uk/ega/ at the EBI) with accession number EGAD00001003883. All gene expression and methylation microarray data reported in this study have been deposited in the National Center for Biotechnology Information Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) public repository, and they are accessible through GEO accession number GSE108124.

Code availability

All code used in our analysis will be made available at http://github.com/ucl-respiratory/preinvasive on publication. All software dependencies, full version information, and parameters used in our analysis can be found here.

Unless otherwise specified, all analyses were performed in an R statistical environment (v3.5.0; www.r-project.org/) using Bioconductor¹ version 3.7.