The cohort comprised 34 NSCLC patients, 20 female and 14 male, ranging from 49 to 87 years old, whom underwent surgical resections of NSCLC as treatment with curative intent (details in Supplementary Table 1 ). Seventeen of the patients were diagnosed with adenocarcinomas, 12 with squamous cell carcinomas, three with large cell tumours and two with pleomorphic tumours. Using TNM classification (8^th edition), one patient had a T1, 17 had T2, 10 had T3 and six had T4 tumours (24). Lymph node (N) component included stage N0 in 24 patients, seven patients with N1 and three with N2 disease. As this was a curative intent cohort, no patients had distant metastasis. All patients except one were either previous or current cigarette smokers. The standardised uptake value of the PET imaging tracer 18F-FDG from the pre-operative CT scans of patients was high for all but two patients (one moderate, one not recorded). Non-cancerous lung tissue was retrieved from adjacent areas of resected lung tissue removed with the lung tumours ( Figure 1A ). Histopathological abnormalities within the non-malignant regions of the lung included single abnormalities or combinations of pneumonia, emphysema, inflammation, sarcoidosis and pleural fibrosis, and seven of the tissues were recorded as being histopathologically normal lung ( Supplementary Table 1 ).

To enrich extracellular proteins from patient-derived lung tissue for proteomic analysis, we used detergent and alkaline extractions and DNase treatment to deplete the lung tissue of cells and intracellular material, with minimal loss of relatively insoluble ECM proteins ( Figure 1A ; Supplementary Figure 1A ). We performed label-free liquid chromatography-coupled tandem mass spectrometry (LC-MS/MS) analysis of ECM isolated from NSCLC tumour and non-tumour tissue, identifying a total of 1,662,859 peptide–spectrum matches, quantifying 3,602 proteins with a false discovery rate (FDR) of 1% ( Supplementary Data 1 ). Extracellular proteins were significantly over-represented in the isolated ECM fractions ( Figure 1B ; Supplementary Data 2 ), and 1,420 identified proteins (39.4%) were annotated as extracellular in the Gene Ontology database (25) ( Supplementary Datas 1, 2 ). The median sequence coverage of extracellular proteins, determined by MS, was significantly higher than that of all other identified proteins ( Figure 1C ), and the spectral counts for extracellular proteins were significantly higher than for other proteins ( Figure 1D ).

Two hundred and fifty-six proteins in the dataset were classified as curated matrisome proteins (26, 27), which comprised 108 core matrisome proteins and 148 matrisome-associated proteins. The core matrisome group consisted of 70 glycoproteins, 26 collagens and 12 proteoglycans; the matrisome-associated group consisted of 37 ECM-affiliated proteins, 73 ECM regulators and 38 secreted proteins. This is comparable in number and matrisome subgroup distribution to recent proteomic analyses of lung tumour tissue (28–30) ( Supplementary Figure 1B ; Supplementary Table 2 ). Together, these data indicate strong detection and enrichment of extracellular proteins in the patient-derived lung ECM samples.

Of the 3,602 identified proteins, 1,805 significantly differed by at least 2-fold between tumour and non-tumour tissues (P < 0.05, paired two-sided Student’s t-test with Benjamini–Hochberg correction) ( Figure 1E ; Supplementary Data 3 ). Pairwise correlation analysis of all tumour and non-tumour samples revealed positive correlation between like sample types ( Supplementary Figure 1C ), and tumour samples generally clustered together ( Figure 1F ; Supplementary Figure 1C ). One hundred and sixty-one matrisome proteins were significantly differentially expressed ( Supplementary Data 4 ), 47 of which were increased in tumour samples compared to non-tumour samples, and 114 of which were decreased in tumour samples compared to non-tumour samples ( Figure 2 ; Supplementary Figure 2 ).

The majority of matrisome proteins up-regulated in tumour ECM were matrisome-associated proteins, predominantly ECM regulators and secreted proteins, whereas proteins down-regulated in tumour ECM included a substantial number of both matrisome-associated proteins and core matrisome proteins ( Figure 2 ; Supplementary Figure 2 ). Of those collagens that were differentially regulated, all were down-regulated in patient-derived tumour samples compared to matched non-tumour samples, ranging from 1.7-fold for type I collagen α1 chain (COL1A1) to 18.1-fold for type VI collagen α6 chain (COL6A6) ( Supplementary Figure 2 ; Supplementary Data 3 ). In addition, all detected laminins were down-regulated in tumour samples compared to matched non-tumour samples, from 1.8-fold for laminin α2 chain (LAMA2) to 18.6-fold for laminin α3 chain (LAMA3) ( Supplementary Figure 2 ; Supplementary Data 3 ). These findings agree with similar observations of decreased abundance of many collagen and laminin subunits in a murine model of lung adenocarcinoma (29).

Matrisome proteins down-regulated in tumours were enriched for ECM organisation functions ( Supplementary Data 5 ), and these proteins formed a functional subnetwork dominated by collagens and laminins that clustered based on protein interactions ( Figure 3A ; Supplementary Figure 3A ). Other down-regulated glycoproteins involved in ECM organisation clustered together, including elastin (3.5-fold), EMILIN-1 (2.0-fold), fibulin-5 (4.1-fold) and microfibril-associated glycoprotein 4 (2.8-fold) ( Figure 3A ; Supplementary Figure 3A ), all of which have been previously reported to be down-regulated in lung tumour ECM compared to normal ECM in mice (29). These data suggest that core matrisome proteins with key roles in ECM organisation and structural integrity are dysregulated in NSCLC tumours.

Matrisome proteins up-regulated in tumours were enriched for protein hydroxylation processes ( Supplementary Data 5 ), driven by procollagen-lysine,2-oxoglutarate 5-dioxygenases (also known as lysyl hydroxylases) and prolyl 3- and prolyl 4-hydroxylases ( Supplementary Figure 2 ; Supplementary Data 3 ), which formed a functional subnetwork of interacting proteins ( Figure 3A ; Supplementary Figure 3A ). These proteins catalyse the post-translational formation of hydroxylysine and hydroxyproline residues in collagen chains, which are critical for collagen helix stability (31), and have been reported to be up-regulated in fibrotic lung ECM in murine and human model systems (32, 33). LC-MS/MS analysis revealed that the majority of hydroxylated residues were found in type I and type III collagen chains (COL1A1, COL1A2, COL3A1) ( Figure 3B ). While there were both up- and down-regulated hydroxyproline-containing peptides quantified by LC-MS/MS ( Supplementary Figure 3B ), almost all hydroxylysine-containing peptides were up-regulated in tumour ECM, and all-but-one of these were derived from type I collagen chains ( Figure 3C ).

To identify clusters of residues that showed similar regulation of hydroxylation across the tissue samples, we mapped the changes in abundance of hydroxyproline- and hydroxylysine-containing peptides for matched tumour–non-tumour pairs. This analysis indicated variable modulation of proline hydroxylation in type I and type III collagen chains, including a substantial cluster of hydroxyproline-containing peptides derived from COL1A1 that were up-regulated in tumour ECM ( Figure 3D ; Supplementary Figures 3C, D ). Lysine hydroxylation, however, was predominantly up-regulated in type I collagen chains across most matched tumour samples ( Figures 3C, D ; Supplementary Figure 3D ), whereas modulation of lysine hydroxylation in type III collagen was more variable ( Figure 3C ; Supplementary Figure 3C ). These data imply that type I collagen lysine residues are extensively hydroxylated in tumour ECM, in concert with an up-regulation of enzymes that catalyse collagen hydroxylation.

In addition to enrichment of enzymes that regulate ECM hydroxylation in tumours tissue, there was an increase in matrisome proteins associated with ECM turnover in tumour samples. Several cathepsins, some of which have reported roles in the degradation of extracellular proteins, were up-regulated in tumour samples, including the secreted thiol proteases cathepsin B (2.1-fold) and cathepsin S (1.7-fold) ( Supplementary Figure 2 ; Supplementary Data 3 ). Several matrix metalloproteinases (MMPs) were up-regulated in tumour samples, namely MMP12 (5.1-fold), MMP14 (2.3-fold) and MMP2 (2.2-fold), which cleave elastin (and insulin), collagen (and other extracellular proteins, such as aggrecan) and certain collagens (and gelatin), respectively. In contrast MMP28, which cleaves casein, was down-regulated in tumour samples (7.3-fold) ( Supplementary Figure 2 ; Supplementary Data 3 ). A disintegrin and metalloproteinase with thrombospondin motifs 16 (ADAMTS16), another protease, was up-regulated in tumour samples (2.0-fold).

In addition to many matrisome-associated proteins, several core matrisome glycoproteins were up-regulated in tumour tissue. The most enriched glycoproteins in tumour ECM compared to non-tumour ECM included thrombospondin-2 (16.9-fold), MXRA5 (14.4-fold), peroxidasin (2.5-fold), thrombospondin-1 (2.5-fold), SPARC (2.3-fold), FRAS1 (2.3-fold) and tenascin-C (2.1-fold) ( Supplementary Figure 2 ; Supplementary Data 3 ). Four of the top five most up-regulated glycoproteins in tumour ECM – thrombospondin-1 and -2, SPARC and peroxidasin – formed a connected subnetwork ( Supplementary Figure 3A ), pointing to their common functional roles in organising the collagen ECM and regulating cell–ECM adhesion.

Of the extracellular proteins we quantified as up-regulated in tumours, we selected peroxidasin and ADAMTS16 for further analysis ( Figure 4A ). The activity of peroxidasin – an enzyme that mediates collagen cross-linking – has recently been linked to the promotion of lung cancer through inhaled air-pollution particles, which adsorb peroxidasin and induce aberrant ECM thickening (34), providing a rationale to examine peroxidasin expression in early-stage NSCLC ECM. ADAMTS16 DNA methylation has been shown to be dysregulated in several epithelial cancers, including lung cancer (35), suggesting that regulation of ADAMTS16 protein expression in tumour-associated ECM, and its potential as a protein biomarker in NSCLC, warrant investigation. We constructed a tumour microarray (TMA) containing control and tumour samples from a separate cohort of lung cancer patients undergoing curative intent surgery ( Supplementary Table 3 ). Selected proteins were analysed by immunohistochemistry (IHC) and scored on a four-point scale (scored from 0–3) based on increasing immunostaining by IHC from no staining to strong staining. For lung cores in the TMA derived from tumour samples, the distribution and scoring of immunostained proteins was assessed separately for tumour stroma and tumour cells. In tumours, peroxidasin and ADAMTS16 were found variably distributed between tumour stroma and tumour cells ( Figure 4B ). Importantly, both proteins had significantly increased immunostaining intensities in tumour cells compared to non-tumour cells ( Figures 4B, C ). These data confirm the results from the MS-based proteomics analysis and imply that peroxidasin and ADAMTS16 are up-regulated in NSCLC tumours.

To examine the expression of the genes encoding the selected extracellular proteins, we analysed an integrated dataset of RNA-seq data derived from paired tumour and adjacent normal tissue from lung adenocarcinoma or lung squamous cell carcinoma patients (36). Expression of both PXDN (which encodes peroxidasin) and ADAMTS16 were significantly up-regulated in adenocarcinoma tumours compared to matched adjacent normal lung tissue (1.96-fold and 10.7-fold change in median expression, respectively) ( Figure 4D ). Both genes were also significantly up-regulated in squamous cell carcinoma tumours compared to matched adjacent normal lung tissue (1.78-fold and 5.62-fold change in median expression, respectively) ( Figure 4E ). These data indicate that PXDN and ADAMTS16 are transcriptionally up-regulated in NSCLC tumours.

We next assessed whether the protein expression levels of peroxidasin and ADAMTS16 were predictors of survival in the TMA cohort ( Supplementary Table 3 ). Although up-regulated in tumour tissue ( Figures 4A–C ), the degree of tumour stroma immunostaining of these two markers, as determined by IHC scoring of all interpretable lung cores, did not significantly correlate with survival outcome within the limited TMA cohort ( Supplementary Figure 4 ). To examine patient survival in larger NSCLC cohorts, we used the Cancer Genome Atlas (TCGA) data derived from primary tissue samples from lung adenocarcinoma (513 patients) or lung squamous cell carcinoma (501 patients). We found that higher expression of PXDN was significantly associated with shorter survival of adenocarcinoma patients (hazard ratio (HR) = 1.57; 95% confidence interval (CI), 1.05–2.37; P = 0.028, log-rank test), whereas that of ADAMTS16 was not ( Figure 4F ). Median survival of adenocarcinoma patients was 54.4 months for the low PXDN expression subgroup and 39.0 months for the high PXDN expression subgroup. In contrast, there was no statistical association of expression of PXDN with survival of squamous cell carcinoma patients, whereas higher expression of ADAMTS16 was significantly associated with shorter survival of squamous cell carcinoma patients (HR = 1.72; 95% CI, 1.18–2.50; P = 0.0043) ( Figure 4G ). Median survival of squamous cell carcinoma patients was 74.1 months for the low ADAMTS16 expression subgroup and 33.4 months for the high ADAMTS16 expression subgroup. Together, these results indicate that these proteins, identified in the patient lung tumour matrisome, are increased in NSCLC and their corresponding gene expression provides prognostic information for lung adenocarcinoma (PXDN) or squamous cell carcinoma (ADAMTS16) patients undergoing curative intent surgery.

Patient samples were collected with ethical approval and written patient consent. Ethical approval was granted by Lothian NRS Bioresource, REC number 15/ES/0094 (reference SR419). All samples were assigned an anonymised code, and researchers were blinded to patient details for experiments. The TMA dataset was approved by Lothian NRS Bioresource, REC number 15/ES/0094 (reference SR1208), and approved by the NHS Lothian Caldicott Guardian (reference CRD19031).

Tissues samples used for mass spectrometry (MS) were collected from patients with NSCLC undergoing curative-intent surgical procedures. Following resection, samples were handled by an experienced thoracic pathologist, and samples of tumour and non-cancerous lung (from the most distal portion of the resection specimen) were dissected and provided for MS analysis. Samples were snap frozen and stored at −70°C until required. Anonymised patient details were recorded, including age, gender, smoking history, histopathological diagnosis of tumour and non-tumour tissues, degree of differentiation of the tumour tissue, tumour stage and lymph node stage (determined by TNM status), survival and PET tracer uptake.

Tissue samples were enriched for predominately insoluble matrisome proteins by depleting soluble intracellular proteins ( Figure 1A ). Methods were adapted from previously published work (54). Tissue samples were finely minced with scalpels, and the presence of any necrosis and tissue pigmentation was noted. Samples were homogenised twice in 1 ml of chilled phosphate-buffered saline (PBS) (without Ca²⁺ or Mg²⁺), containing 1% (v/v) protease inhibitor cocktail (Sigma-Aldrich), using a Precellys 24 tissue homogeniser (Bertin Instruments) at 6,500 rpm for 50 s. Samples were incubated for 5 min on ice between homogenisation cycles. The homogenate was centrifuged at 14,000 × g for 10 min at 4°C. The supernatant was removed (fraction 1), and the remaining pellet was resuspended and incubated in 10 mM Tris-HCl, pH 8, 150 mM NaCl, 25 mM EDTA, 1% (v/v) Triton X-100, 1% (v/v) protease inhibitor cocktail for 30 min on ice. The lysate was centrifuged at 14,000 × g for 10 min at 4°C. The supernatant was removed (fraction 2), and the remaining pellet was resuspended and incubated in 20 mM NH₄OH containing 0.5% (v/v) Triton X-100 in PBS (without Ca²⁺ or Mg²⁺). The lysate was centrifuged at 14,000 × g for 10 min at 4°C. The supernatant was removed (fraction 3), and the remaining pellet was incubated with 10 µg/ml DNase I in PBS (with Ca²⁺ and Mg²⁺) for 30 min on ice. Samples were centrifuged at 14,000 × g for 10 min at 4°C, and the supernatant was removed (fraction 4). The remaining pellet was washed in ice-cold PBS and centrifuged at 14,000 × g for 10 min at 4°C three times. The final insoluble pellets enriched for ECM proteins were then stored at −70°C until further use.

To confirm matrisome proteins were enriched in the final protein pellet prior to proteomics experiments, the supernatants from serially extracted fractions from a subset of samples were probed by SDS-PAGE and western blotting. Three percent of the supernatant from each fraction was used for western blotting to allow comparison between fractions. Fractions 1–4 were incubated in 1× Laemmli buffer containing 50 mM dithiothereitol for 10 min at 95°C. The final ECM pellet was precipitated using TCA–acetone (see below), homogenised and incubated in 8 M urea, 100 mM NH₄HCO₃, pH 8, 10 mM dithiothereitol for 30 min at 37°C. Samples were resolved by SDS-PAGE using 4–20% Tris-glycine gels. Proteins were transferred to PVDF membranes using an iBlot 2 dry blotting system (Thermo Fisher Scientific) according to manufacturer’s instructions. Membranes were blocked using milk blocking buffer (5% (w/v) non-fat skimmed milk powder (Marvel) in 1× Tris-buffered saline containing 0.1% (v/v) Tween 20 (TBS-T)) for 1 h at room temperature with shaking. Following blocking, membranes were incubated with primary antibodies, diluted 1:1,000 in milk blocking buffer, overnight at 4°C with rolling. Antibodies used were rabbit polyclonal anti-fibronectin (#ab2413, Abcam), mouse monoclonal anti-lamin A/C (clone 4C11; #4777, Cell Signaling Technology) and rabbit monoclonal anti-GAPDH (clone 14C10; #2118, Cell Signaling Technology). Membranes were washed three times in TBS-T and then incubated with anti-rabbit or anti-mouse horseradish peroxidase-conjugated secondary antibodies, diluted 1:10,000 or 1:5,000, respectively, in milk blocking buffer, for 45 min at room temperature with rolling. Membranes were washed three times in TBS-T. Membranes was developed using Pierce ECL western blotting substrate (Thermo Fisher Scientific) according to the manufacturer’s instructions. Membranes were imaged using an Odyssey Fc imaging system (LI-COR Biosciences).

Ice-cold TCA (10% (v/v) final concentration) was added to final ECM-enriched fractions for 20 min at 4°C. The sample was centrifuged at 16,000 × g for 30 min at 4°C. The supernatant was discarded, and the ECM-enriched pellet was resuspended in ice-cold acetone, using vortexing and sonication, and incubated for 20 min at −20°C. The sample was centrifuged at 16,000 × g for 30 min at 4°C, and the supernatant was discarded. The acetone wash step was then repeated. Samples were air dried at room temperature until residual solvent had evaporated. Precipitated protein pellets were stored at −70°C until further use.

Precipitated protein pellets were resuspended in 300 μl solubilisation buffer (8 M urea, 100 mM NH₄HCO₃, pH 8, 10 mM dithiothreitol). Samples were sonicated on ice using a probe sonicator and then incubated for 30 min at 37°C. After cooling to room temperature, sample pH in the range pH 8–9 was verified using pH indicator paper. To 50 µl sample, 8.3 μl of 175 mM iodoacetamide in 100 mM NH₄HCO₃, pH 8, was added (final concentration 25 mM) and incubated for 30 min at room temperature in the dark to block thiol groups of cysteine residues. Excess iodoacetamide was quenched by the addition of 3 µl of 100 mM dithiothreitol in 100 mM NH₄HCO₃, pH 8. For protein digestion, urea was diluted from 8 M to 2 M using 100 mM NH₄HCO₃, pH 8. Samples were incubated with 833 units of PNGase F for 2 h at 37°C with shaking (900 rpm), then with 800 ng of Lys-C for 2 h at 37°C, then with 1.6 µg of MS-grade trypsin for 16 h at 37°C with shaking (1,200 rpm). Samples were then incubated with an additional 0.8 µg of trypsin for 2 h at 37°C with shaking (1,200 rpm). Samples were acidified using trifluoroacetic acid (TFA) to obtain sample pH in the range pH 3–4, and samples were clarified by centrifugation at 18,000 × g for 15 min.

Peptide concentrations in the digested samples were estimated using a Nanodrop spectrophotometer measuring absorbance at 280 nm. Stop-and-go extraction (Stage) tips were made in-house, using a method adapted from Rappsilber et al. (55). Briefly, using an 18-gauge blunt-ended needle, two disks were cut from C18 solid-phase extraction material and placed on top of each other inside a 200-µl EasyLoad pipette tip (Greiner Bio-One). Stage tips were loaded into a custom-built tip holder over a deep 96-well plate. Methanol was added to each Stage tip, and tips were centrifuged at 300 × g for 2 min. Stage tips were then equilibrated with 0.1% (v/v) TFA and centrifuged at 500 × g for 5 min. Samples estimated to contain 10 µg of acidified peptide (based on Nanodrop readings) were then added to Stage tips, and tips were centrifuged at 500 × g for 5 min. Stage tips with bound protein were stored at −20°C until elution.

Peptides were eluted from C18-containing Stage tips with 40 µl of 80% (v/v) acetonitrile, 0.1% (v/v) TFA, and tips were centrifuged at 200 × g for 5 min. Acetonitrile was evaporated using a vacuum centrifuge. Samples were adjusted to 15 μl volume with 0.1% (v/v) TFA, and peptide concentration was re-measured using a Nanodrop spectrophotometer measuring absorbance at 280 nm.

‘Bottom-up’ liquid chromatography-coupled tandem MS (LC-MS/MS) was used to elucidate the structure of isolated peptides and to detect post-translational modifications. LC-MS/MS analysis was carried out using an Orbitrap Fusion Lumos Tribrid mass spectrometer (Thermo Fisher Scientific) coupled to an UltiMate 3000 UHPLC Nano (Thermo Fisher Scientific), Aurora C18 column (IonOpticks), column oven (maintained at 50°C; Sonation) and Proxeon Nanospray ion source (Thermo Fisher Scientific). Peptides (1 µg) were injected onto an Aurora C18 column in buffer A (2% (v/v) acetonitrile, 0.5% (v/v) acetic acid) and eluted with a linear 120-min gradient of 2%–45% (v/v) buffer B (80% (v/v) acetonitrile, 0.5% (v/v) acetic acid). Eluting peptides were ionised in positive-ion mode before data-dependent analysis. A dynamic exclusion window of 30 s was enabled and lockmass was not used.

The MS data were normalised and quantified using MaxQuant software (version 1.6.10.43) (56). MaxQuant quantifies proteins using a label-free technique, which calculates a normalised peptide abundance from ion signal intensities (57). Peptide lists were searched against the human UniProt knowledgebase database (version 2019_09) (58) and a common contaminants database using the Andromeda search engine implemented in MaxQuant. Cysteine carbamidomethylation was set as a fixed modification, and methionine oxidation, lysine oxidation, proline oxidation, N-terminal deamidation and protein N-terminal acetylation were set as variable modifications, with up to five modifications per peptide. Peptide identifications were matched between runs if they eluted within a time window of 0.7 min. Enzyme specificity was C-terminal to arginine and lysine, except when followed by proline. A maximum of two missed cleavages were permitted in the database search; minimum peptide length was seven amino acids. At least two peptide ratios were required for label-free quantification, and large label-free quantification ratios were stabilised. Peptide and protein false-discovery rates (FDRs) were set to 1%, determined by applying the target-decoy search strategy implemented in MaxQuant. Proteins matching to the common contaminants or decoy databases, and matches only identified by site, were omitted. Label-free quantification intensities were log₂ transformed, and proteins quantified in less than one-third of samples were removed. Missing values were imputed from a width-compressed, down-shifted Gaussian distribution using Perseus (version 1.6.2.3) (59).

Gene Ontology (GO) over-representation analysis was performed using DAVID 2021 (DAVID Knowledgebase, version 2022q2) (60). The functional annotation tool category of GO term enrichment was used to filter out very broad GO terms. GO over-representation analysis of matrisome proteins used the entire matrisome database as background. Significant over-representation of terms was determined using a Fisher’s exact test with Benjamini–Hochberg correction.

Identified proteins were defined as belonging to the matrisome if they exist in searchable databases of matrisome proteins based on data from 17 studies of the ECM, MatrisomeDB (27), and in-silico and in-vivo data from the Matrisome Project (26). Label-free quantification intensities for proteins derived from tumour and non-tumour samples were compared using a paired two-sided Student’s t-test, with FDR set to 5% and artificial within-groups variance (s0) set to 1 using Perseus. Statistical data were visualised using Prism (version 9.2.0; GraphPad). For differentially expressed matrisome proteins, potential confounding variables were also considered. These variables included necrotic tumour samples versus non-necrotic tumour samples, adenocarcinoma versus squamous cell carcinoma, well and moderately differentiated tumours versus poorly differentiated tumours, tumours with local lymph node spread versus no known lymph node metastasis and normal non-tumour samples versus non-tumour samples with underlying pathology. Necrosis was crudely assessed based on the composition of the tumour samples when dissected and minced. PET standardised uptake value was scored as mild, moderate or marked. In some instances, absolute values were provided in reports rather than categorical values. These numbers were reclassified as no uptake for 0.6–0.8, low uptake for 1.0–2.0, moderate uptake for 1.5–2.0 and high uptake for >2.5. These variables were compared using multiple Student’s t-tests (P < 0.05, FDR 5%).

To enable relative comparison of protein enrichment, log₂-transformed label-free quantification intensities were standardised by row-wise (protein-wise) Z-scoring. Differentially expressed matrisome proteins were hierarchically clustered on the basis of Euclidean distance, computed using average linkage, using Cluster 3.0 (C Clustering Library, version 1.54) (61). The following variables were included in the protein enrichment analysis: sex, age, smoking history, if patients had died at the time of analysis, histopathology results, degree of tumour differentiation, tumour stage, lymph node stage and PET 18F-FDG tracer uptake. Modified peptide fold changes between matched non-tumour–tumour pairs were hierarchically clustered on the basis of Euclidean distance, computed using average linkage. For sample correlation analysis, Spearman rank correlation coefficient-based distance matrices were computed using average linkage. Clustering results were visualised using Java TreeView (version 1.1.5r2) (62).

Composite functional association networks were constructed using GeneMANIA (version 3.5.2; human interactions) (63) in Cytoscape (version 3.8.0) (64). Networks were based on reported physical and predicted protein–protein interactions; edges, representing protein–protein interactions, were weighted according to evidence of co-functionality using GeneMANIA. Networks were clustered using Markov clustering (granularity parameter 2.5), and graph layouts were determined using the force-directed algorithm in the Prefuse toolkit (65).

The TMA was constructed for sequential patients undergoing surgical resection for lung cancer over a 2-year period at a regional thoracic surgery centre. Formalin-fixed paraffin pathological blocks were annotated by an experienced pathologist, and 1-mm cores taken from tumours and non-cancerous lung were embedded into new blocks and, subsequently, 4-µm sections were cut onto glass slides. Slides were deparaffinised and rehydrated, and antigen retrieval was undertaken with citrate buffer (#ab64214, Abcam) twice for 5 min in a microwave. Slides were processed with a commercial DAB cell and tissue staining kit (#CTS019, R&D Systems). Primary antibody immunostaining was optimised for polyclonal rabbit anti-human peroxidasin (#abx101905, Abbexa) and polyclonal rabbit anti-human ADAMTS16 (#TA322059, AMS Biotechnology) (diluted 1:100 and 1:200, respectively, and incubated at room temperature for 1 h). Secondary antibodies were incubated at room temperature for 30 min and DAB was developed. Slides were counterstained with haematoxylin and mounted, and images were acquired on an Axioscan microscope slide scanner (Zeiss).

For peroxidasin immunostaining, there were 151 non-cancerous lung cores that were available for evaluation (149 interpretable following staining) and 138 tumour cores, of which 119 were paired samples ( Supplementary Table 3 ). For ADAMTS16 immunostaining, there were 150 available non-cancerous lung cores (148 interpretable for inflammatory cells and 149 for non-tumour cells) and 155 available tumour cores (154 interpretable for tumour stroma and 152 for tumour cells), of which 133 were paired samples ( Supplementary Table 3 ). Slides were scored for intensity of staining of tumour cells or stromal areas on a four-point scale: 0, no staining; 1, low staining; 2, moderate staining; 3, high staining. Data are presented as the proportion of positive staining within each category of staining for tumour and normal samples. Significant difference in the distribution of staining scores was determined by a chi-square test. Outcome data were recorded for each patient who had a tumour resection, with the median time to follow-up of 1,432 days (range 1,054–1,905 days). Outcome included a record of death as defined by the clinical care team. Immunohistochemistry (IHC) scoring was assessed against survival for tumour staining.

Gene expression data were derived from RNA-seq datasets of primary tissue samples from cohorts of lung adenocarcinoma or lung squamous cell carcinoma patients extracted from TCGA, the Genotype-Tissue Expression repository and the Therapeutically Applicable Research to Generate Effective Treatments database using TNMplot (36). Data from tumour samples were compared to paired data from matched adjacent normal samples.

The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier PXD041066.

Kaplan–Meier curves were computed from RNA-seq datasets of primary tissue samples from cohorts of lung adenocarcinoma or lung squamous cell carcinoma patients extracted from TCGA using KMplot (66). The lower and upper quartiles of gene expression were compared. Follow-up time was truncated to retain at least 10 patients at risk.