2.1 Patients and specimens

The ethics committee of Hannover Medical School (MHH) approved the use of tissue resection materialof LC patients who received surgery at MHH during the period 2017–2022 (approval ID: 3381-2016). We obtained informed consent from all patients. Board-certified pathologists examined matched non-tumours adjacent and tumour tissue, and we summarize the clinicopathological data in Table 1 and Table S1.

The study cohort of 95 LC patients consisted of 54 adenocarcinomas (22f, 32m); 23 SQ (9f, 14m); 9 neuroendocrine (4f, 5m) and 9 metastatic tumours (4f, 5m). In regards to the NET, there are four and two cases, respectively, of large-cell and small-cell carcinoma and three carcinoids. The lung metastatic tumours originate from three renal cell carcinoma, two colorectal cancers, one melanoma, one endometrial, one cholangiocarcinoma and one neuroendocrine cancer. Further information is given in Table S2.

2.2 RNA extraction

We used the miRNeasy Mini Kit (Qiagen) to isolate total RNA from frozen lung tissues according to manufacturer's recommendation and determined its quantity and purity with the NanoDrop ND-1000 system. Additionally, we evaluated the purity of RNA with the nanoVette system (Beckman Coulter DU 730 UV/Vis Spectrophotometer) and considered RNA extracts with a ratio of ∼2.0 at 260/280 nm as acceptable for further analysis. Moreover, we assessed the integrity of the 18S and 28S ribosomal bands on denaturing agarose gels, and representative examples are shown in Figure S1.

2.3 Whole-genome gene expression profiling

We used the GeneChip 3’IVT Plus reagent kit (Thermo Fischer Scientific) according to the manufacturer's recommendation and initiated first- and second-strand cDNA synthesis with 100 ng of total RNA. We performed in vitro transcription with biotinylated ribonucleotide analogues and purified cRNA with magnetic beads. We quantified cRNA in a Beckman Coulter DU 730 UV/Vis Spectrophotometer (Beckman Coulter) and estimated the absorption ratio at 260/280 nm. We prepared cleaved cRNA with the fragmentation buffer and determined the size of the fragmented biotinylated cRNA by denaturing agarose electrophoresis. For this purpose, we mixed 9.4 µg labelled cRNA with 3′ fragmentation buffer and incubated the samples at 94°C for 35 min, followed by an incubation step at 4°C for ≥2 min. We assessed the quality of fragmented cRNA on denaturing agarose gels. Typically, we obtained fragments of the size of about 100 nt.

We hybridized 6 µg of the fragmented cRNA onto the GeneChip® HG-U133 array strips and placed them into a GeneAtlas Hybridization Station according to manufacturer's instruction. Following hybridization at 45°C for 16 h, we initiated the washing and staining step in a GeneAtlas® Fluidics Station. Subsequently, we scanned the microarrays on a GeneAtlas® Imaging Station. We processed microarrays, which passed the quality controls (polyA, oligoB2 and 20XEHC) and performed in-depth data analysis as described below.

2.4 Whole-genome miRNA profiling

We labelled 700 ng of total RNA with the FlashTag^TM Biotin HSR RNA Labelling Kit (Thermo Fisher Scientific) according to the manufacturer's instructions. The assay involves a two-step reaction, that is, poly(A) tailing and biotin ligation. The labelled samples were hybridized onto the Affymetrix GeneChip® miRNA array 4.1 (Affymetrix), at 48°C for 20 h. Subsequently, we placed the arrays into the GeneAtlas® Fluidics Station and initiated the wash and stain protocol according to the manufacturer's recommendation. We scanned the microarrays with the GeneAtlas® Imaging Station and processed the images with the GeneAtlas Instrument Control Software (Affymetrix) to generate CEL files. We processed microarrays, which passed the quality controls (polyA, oligoB2 and 20XEHC) and performed in-depth data analysis as described below.

2.5 Lung cancer kinome profiling

The assay is based on the PamGene technology (www.pamgene.com), which enables the real-time monitoring of kinase activity. Essentially, the assay determines the phosphorylation of canonical peptides of kinases. We performed kinome profiling for 54 patients by considering tumour and histologically proven adjacent non-tumour tissue, and we compile information regarding patient demographics and tumour pathology in Table S2.

We performed all reactions on ice and placed frozen tissue of the size of about 1–3 mm³ into a Eppendorf tube. We added 100 µL M-PER buffer (Thermo Fisher Scientific), which contained 2× protease and phosphatase inhibitor cocktails (Thermo Fischer Scientific) and gently mixed the samples. We incubated the lysates on ice for 60 min, followed by a centrifugation step at 4°C and 15 000 ×g for 15 min and collected 10 µL aliquots of the supernatant. These were immediately snap-frozen and stored at −80°C. We determined the protein concentration with the Bradford assay (Thermo Fischer Scientific) and performed protein tyrosine (PTK) and serine/threonine kinase (STK) assays according to the manufacturer's protocols using the PamStation 12 platform (PamGene International BV) as previously described.¹⁴ For the PTK assays, we added 1x PK buffer, 1% BSA (bovine serum albumin), 1 M DDT (dichloro-diphenyl-trichloroethane) solution, 10% of PTK additive and 4 mM ATP (PTK reagent kit; article 32112.5, PamGene) to 5 µg of protein lysate in a final volume of 20 µL. The PTK assay buffer contained a fluorescein isothiocyanate (FITC)-conjugated PY-20 antibody, which recognizes phosphorylation sites of 196 peptides. For the STK assay, we added 1 µg of protein lysate 1x PK buffer, 1% BSA and 4 mM ATP (provided by PamGene, STK kit) to 1 µg of protein lysate in a final volume of 20 µL. The STK chip allowed an assessment of 144 peptides simultaneously. Unlike the PTK assay, we added a FITC-conjugated antibody at the last cycle of the STK assay (Cycle 94) and measured the signal of the antibody, which recognizes phosphorylated peptides. We analysed the data with the Bio Navigator software version 6.2 (PamGene International BV).

2.6 Data processing

2.7 Identification of differentially expressed genes and miRNAs

2.8 Identification of differentially expressed kinases

We used the BioNavigator software version 6.2 (PamGene International BV) to normalize the data and to identify significantly regulated kinases. The BioNavigator software converts image files to signal intensities and corrects for local background noise. For the PTK assay, and for a given phosphopeptide, we fitted time-resolved measurement cycles to estimate an initial phosphorylation rate (Figure S3), whereas for STK assay, the signals of the last cycle were quantified. We used LIMMA to identify significantly regulated kinases by comparing their activity in tumours versus NATs. We used the following criteria to define significantly regulated kinases: FDR-adjusted p-value < 0.05 and a |FC| ≥ 1.5.

2.9 MiRNA–gene regulatory networks

We used the miRNet 2.0 database to search for target genes of DEMs. The database contains comprehensive information on experimentally validated miRNA targets (https://www.mirnet.ca/miRNet/home.xhtml).¹⁸ We considered DEMs of lung tumours to search for putative gene targets and used the information as input files to search for database entries in miRNet 2.0. We compared the list of putative target genes with DEGs of LC patients as determined in our genomic analysis, and this allowed us to identify validated DEG targets and to construct miRNA–gene regulatory networks. Subsequently, we visualized the networks with the software Cytoscape 3.9.1.¹⁹ Additionally, we computed correlations between DEMs and DEGs. Based on the normality test, we calculated Pearson and/or Spearman correlations. We compared the results to the miRNet proposed targets. Furthermore, we performed multiple query types in miRNet by searching for potential miRNA targets based on patient-specific DEGs (transcriptomic analysis), and vice versa, used patient-specific miRNAs as input file to search for potential DEG targets.

2.10 Survival analysis

For the survival analysis, we included 497 LUAD and 496 LUSC patients with OS information (https://xenabrowser.net/). The start of the survival analysis was defined as a definitive diagnosis of LC, the earliest years were 1991 (LUAD) and 1992 (LUSC), and the survival analysis ended in 2015 (LUAD) and 2013 (LUSC). The median follow-up time was 21.9 (LUAD) and 21.8 (LUSC) months, and the Interquartile range (IQR) was 23.5 (LUAD) and 31.4 (LUSC) months. We divided the patients into high- and low-expression individuals according to the median value of the gene/miRNA expression, and constructed Kaplan–Meier curves to determine overall survival (OS). We performed log-rank test and univariate Cox proportional hazards regression analysis to determine statistical significance and hazard ratio (HR) estimate with 95% confidence interval (CI). To evaluate the proportional hazard assumption, we computed the Schoenfeld residuals in R (Supporting Information 1) and assessed linearity of the Cox proportional hazard regression analysis by calculating fractional polynomials.

2.11 Search for master regulatory kinases in canonical pathways of LC patients

We uploaded DEGs with a |FC| ≥ 3 onto the geneXplain platform, and applied the function ‘find master regulators’. The underlying genetic algorithm has been published^{20, 21} and MRs were identified based on TRANSPATH® database entries. We only considered MRs with an FDR-adjusted p-value < 0.05 and used default settings of the algorithm, that is, a score of >0.2. The score defines the connectivity of the MR with other molecules of the same pathway, and we used a threshold >1 for the Z-score which measures specificity of the MR. We adjusted the algorithm to a maximum radius of 10 molecules upstream of input DEGs, and the software identified MRs in different canonical pathways. Next, we searched for significantly regulated kinases which qualify as MRs and identified pathways specific to a histological type of LC (AD, SQ, NET and MT). The kinase/kinome activity data of lung tumour samples informed on the possibilities to block signalling pathways by targeting hyperactive kinases. Finally, we constructed regulatory signalling networks, which consisted of DEGs, DEMs targeting the DEGs and kinases.

2.12 Statistics

We used R and the geneXplain (https://genexplain.com/) software to perform statistical analysis. If not otherwise specified, all tests were two-tailed, and an FDR-adjusted p-value < 0.05 was considered to be statistically significant.

3.1 Genomic profiling of different LC types

To identify disease-regulated genes (DEGs) and miRNAs, we compared the genomes of tumours to NATs. Shown in Figure S2 is the principal component analysis (Panel A), and the heatmaps (Panel B) clearly segregated tumours from NATs across the different histological LC subtypes. We identified 439, 1240, 383 and 320 significantly upregulated genes for AD, SQ, NET and MT cases, respectively, and there are 1092, 1477, 609 and 1267 downregulated DEGs (Table S3). Common to all tumours is the predominant repression of gene expression.

To identify DEGs specifically asscociated with a histological subtype of LC, we compared the genomes of AD, SQ, NET and MT cases to NATs. This defined 176, 895, 296 and 141 genes specifically upregulated and 255, 550, 172 and 433 repressed ones in the different LC subtypes as shown in Figure 2A.

Furthermore, we identified 442, 302, 85 and 88 significantly upregulated miRNAs for AD, SQ, NET and MT cases, and for the same patients, there are 17, 8, 10 and 15 downregulated miRNAs. In addition, we identified 150, 47, 1 and 1 specifically upregulated and 4, 2, 2 and 3 downregulated miRNAs in AD, SQ, NET and MT (Figure 2B). Collectively, miRNAs are primarily upregulated in tumours, whereas transcriptomes are predominantly repressed (Figure 2B and Table S4).

We used Metascape as an annotation tool for DEGs which were specifically regulated in the various histological subtypes of LC, and we compared significantly enriched terms between them. Note, in this comparison we used non-overlapping genes across histological subtypes. As shown in Figure S4, and with the exception of mTORC1 signalling, which is common to SQ and MT, none of the terms overlapped. Therefore, the gene ontologies and enriched terms are specific for a given histological subtype of LC.

For AD, and among unique terms associated with upregulated DEGs, we wish to emphasize KRAS and EGFR signalling and protein glycosylation. Significantly enriched terms for SQ are E2F and MYC targets, hypoxia and oestrogen response and for NET tumours, neuroepithelial cell differentiation and GABAergic signalling. Finally, enriched terms for MT are MTORC1 signalling, activation of GTPase activity, chromosome organization and DNA repair. We also considered enriched terms for downregulated DEGs. Here, TNFα signalling, response to cytokine stimulus, actin filament-based process and immune response, regulation of MAPK cascades and cell adhesion were specific for AD, SQ, NET and MT tumours (Figure 2C).

To independently validate the results of the present study (“Hannover cohort”), we retrieved data from the TCGA database (https://www.cancer.gov/ccg/research/genome-sequencing/tcga) and applied the following criteria: DEGs were defined with an FDR-adjusted p-value < 0.05, a |FC| ≥ 2 in ≥50% of cases. Note, the TCGA source does only compile LUAD and LUSC cases. Therefore, we compared 524 LUAD cases to 58 NATs and separated the data by clinical stages. We searched for common DEGs between the two cohorts and found > 70% of the up- and downregulated DEGs as commonly regulated. The results are similar when compared to the different histological subtypes of LC (Venn diagram, Figure 3A and Table S5). Therefore, the data agreed reasonably well.

Depicted in Figure 3B are the heatmaps for DEGs of the Hannover cohort, and the colour codes represent Z-scores which measures the standard deviations from the mean of a gene expression values. We obtained perfect segregation of AD and SQ tumours from NATs.

Subsequently, we searched for enriched gene ontologies with the Metascape software²² and selected the top 100 regulated genes common between the Hannover and TCGA cohorts. The FC of the selected genes ranged from 3- to 14-fold. Additionally, we used hallmark gene sets and KEGG pathways and show the results in Figure 3C. For upregulated genes, the hallmark gene sets underscored mTOR signalling, G2M checkpoint, E2F targets and glycolysis in LUADs (AD cases). Similarly, KEGG pathway analysis emphasized cell cycle, p53 signalling, cellular senescence and cancer pathway.

In regards to downregulated genes (range 5–20-fold), the hallmark gene sets highlighted apoptosis, inflammatory response and TNFα signalling via NFKB. We obtained similar results for KEGG pathways, and prominent examples of enriched terms were TNFα signalling, immune and chemokine signalling and leukocyte migration (Figure 3C).

In the same way, we analysed SQ tumours (LUSC), and for upregulated DEGs (range 6–194-fold), the hallmark gene sets were similar to AD cases, that is, mitotic spindle, E2F targets, G2M checkpoints and mTOR signalling, while KEGG pathway analysis highlighted p53 signalling, cell cycle and cellular senescence. For repressed DEGs (range 9–72-fold), the hallmark gene sets emphasized inflammatory response, KRAS signalling and DNA repair, whereas KEGG highlighted phagosome, complement and coagulation cascades and ECM receptor interactions. We summarize the results in Table S5.

Unlike the cancer transcriptomic data described above, there was little overlap among DEMs of the Hannover and TCGA cohort (Figure 3D). In fact, for AD only 6% of up- and downregulated DEMs are common to both cohorts, whereas for SQ 5% and 25% DEMs are commonly regulated.

3.2 Gene networks in different histological subtypes of LC

To gain insight into the molecular wiring across different histological subtypes of LC, we constructed regulatory gene networks, and to ascertain the relationship between DEMs and DEGs, we analysed data deposited in the miRNet database. This revealed experimentally proven miRNA gene targets which we compared to DEGs specific for the different subtypes of LC (Figure 2A). Based on the paradigm that miRNAs repress the expression of gene targets, we identified 253, 417, 121 and 332 DEGs specific for AD, SQ, NET and MT (Figure 4A). Note, all the target genes were repressed in lung tumour samples (Figure 2A). Conversely, for repressed DEMs, we discovered 48, 291, 29 and 0 upregulated DEGs. Furthermore, there are 85 targets in common and therefore were irrespective of the histological subtype of LC. The gene ontologies emphasized UV response (DNA repair), hallmark gene set EMT and TNFα signalling for upregulated DEMs (Figure 4A). For repressed DEMs, only one target gene is common, that is, nucleoside diphosphate kinase 1 which we found upregulated in all lung tumour samples, and this kinase supports metastatic spread.

Additionally, we computed correlations between DEMs and DEGs. Based on the normality test, we calculated the Pearson and/or Spearman correlations and summarize the results in Table S6. In regards to AD, we observed 1231 significant correlations. We compared the results to the miRNet proposed targets and found 846 DEGs (64.2%) to be in common. For SQ, we computed 1425 significant correlations, and when compared to the miRNet targets there are 947 DEGs (50.9%) in common. Together, the correlation analysis identified more potential miRNA-DEG associations.

Furthermore, we computed multiple query types by searching for potential miRNAs which target patient specific DEGs (Hannover cohort data). Additionally, we used significantly regulated miRNAs of the Hannover cohort data as input file to search for potential DEG targets. We compared the results, and for ADs 96% of the targets are in common based on the multiple query approach, and similar results were obtained for SQ, that is, 98% (Figure S5A). We used Metascape gene annotations to construct regulatory miRNA–gene networks. We selected the top 10 biological processes based on FDR-adjusted p-values and required the term to cover > 80% of DEGs specific for a given tumour subtype (Figure 4A). We likewise selected the top 10 miRNAs based on the number of target genes for each biological process and compiled the individual data in Table S6. In this way, we constructed miRNA–gene networks specific for AD, SQ, NET and MT LCs. We compared the terms among the different types of LC, and the Venn diagram in Figure 4B shows the non-overlapped terms among the various histological subtypes of LC. We obtained similar results for repressed DEMs. Here, we considered terms associated with upregulated DEGs. In fact, in the comparison of AD versus SQ, there is only one term in common, that is, peptidyl amino acid modification. Together, this demonstrates specificity for the constructed miRNA–gene networks.

Shown in Figure 4C–F are the regulatory miRNA–gene networks for the different histological types of LC, and we used Cytoscape to create the network. Although the terms are generic in nature, the underlying DEGs are highly specific and encompass potential drug targets. For instance, for the SQ network (Figure 4D1), the Metascape analysis revealed cell cycle phase transition as a statistically significant enriched term and embedded in this term are highly regulated DEGs, notably CDK1, CHEK1 and GSK3B which were induced by eight-, four- and twofold, respectively, in expression. Strikingly, the activities of the coded kinases were also significantly increased in the same patient derived tumour biopsies (see below). A further example relates to peptidyl-amino acid modification, and among genes embedded in this term is the fivefold upregulated NTRK2. Importantly, the kinase activity of NTRK2 is also significantly increased, and its inhibitors entrectinib and larotrecitinib are already approved for the treatment of solid tumours in patients diagnosed with NTRK2 fusion proteins.

Together, the miRNA–gene networks bear great relevance for an identification of potential drug targets.

3.2.1 Kinome profiling in LC

We used the PamGene® platform to identify aberrant kinase activities in tumours and NATs of 54 patients. The assay measures the phosphorylation of peptides, and each peptide contains canonical amino acid sequences which are recognized by distinct serine/threonine and tyrosine kinases. We obtained two sets of data. First, direct measurement of the phosphorylation of kinases, and second, an identification of regulated kinase based on the phosphorylation of canonical amino acid sequences. Here, the upstream kinase phosphorylates a specific substrate, and therefore, the phosphopeptide is a read-out of the kinase activity. Depicted in Figure 5A is a Venn diagram of phosphorylated peptides across the different histological types of LC. We measured an increased phosphorylation of 156 peptides with 95 being common between the different types of LC (Figure 5A). Conversely, there are 28, 1, 2 and 1 phosphorylated peptides uniquely associated with AD, SQ, NET and MT tumours. Of the 156 phosphorylated peptides, there are 72 which allow direct measurement of regulated kinases. Note, for some kinases, such as the EGFR, we measured the phosphorylation of several peptides which contain specific tyrosine residues, and in Figure 5B we show significantly regulated kinases between the different types of LC. Together, we identified 53 significantly regulated kinases of which 46 are common (Figure 5B and Table S7). We did not observe repressed kinase activities but upregulation of primarily PTKs, that is, 67% (Figure 5C), and identified 35 serine/threonine kinases specifically regulated in AD (Figure 5D), and these kinases phosphorylate 28 peptides (Figure 5A). Moreover, upstream kinase analysis revealed 71 kinases regulated in common (Figures 5D and S6a–d) of which 67.6% are tyrosine kinases, and in Figure S4 we show examples of kinetic plots for tumour and NAT biopsies. Converging the results of Figure 5B (direct measurement of 53 phosphorylated kinases) and upstream kinase analysis of phosphorylated peptides (see Figure 5D) yielded a total of 137 distinct kinases which were significantly upregulated in LC (Figure S6e).

The heatmap shown in Figure 5E informs on tumour regulated kinases, and the data are log2 fold increases across different histological subtypes of LC. In Table S7, we specify the amino acid sequence of the target protein which becomes phosphorylated, and a more detailed information is given in Supporting Information 2.

Next, we addressed the questions whether the genes coding for the kinases were also regulated at the gene expression level and show the results in Figure 5F. There are 4 (PTK6, CDK1, AURKA and EPHB2) and 10 kinases (PAK1, PKMYT1, AURKB, CHEK1, CDK1, JAK3, EPHB2, PRKCG, PTK6 and AURKA) of the Hannover-AD and TCGA cohort whose gene expression and kinase activities were upregulated in the same way. Similarly, there are eight (CDK1, CHEK1, EPHA4, GSK3B, MAP2K6, NTRK2, PAK1 and PTK2) and five kinases (CDK1, CHEK1, EPHB2, PAK1 and PKMYT1) in common between the two cohorts (Figure 5F). For neuroendocrine LC tumours, RET is the only kinase where the gene expression (11-fold) and kinase activity (5-fold) were increased in the same way. Together, the data imply that only a few kinases are simultaneously regulated at the gene expression and kinase activity level. Therefore, kinase profiling is of critical importance.

Additionally, we determined the prognostic value and computed univariate Cox regression proportional hazard models for genes that code for 110 significantly regulated kinases based on upstream analysis (Figure 5D and Table S9). We identified 37 kinase coding genes in AD with a statistically significant HR. However, only 18 kinase coding genes had a HR > 1 (range 1.1–1.79). For SQ, there are three kinase coding genes with a statistically increased HR (range 1.2–1.37) (Table S9). Together, the prognostic value of the univariate Cox regression proportional hazard models based on the expression of the kinase coding genes is less obvious.

Notwithstanding, we were able to compute meaningful Kaplan–Meier survival curves for kinases where the change in gene expression and kinase activity agreed, and depicted in Figure 5G are the plots for the kinases AURKA, AURKB, CDK1, CHEK1, EPHB2 and PKMYT1. Here, the prognostic value for these kinases could be established, and in the case of CDK1, clinical Phase II trials in NSCLC and SCLC are ongoing (NCT05651269 and NCT02161419).

Furthermore, for the different histological subtypes of LC, we constructed networks that describe the relationship between miRNAs, their gene targets and the activity of the kinase protein (Figure 5H). As an example, the expression of the genes coding for PDGFRA, KDR and NTRK2 kinase are downregulated, and the miRNAs targeting these kinases are upregulated. Although this fits the paradigm, that is, miRNAs repress gene translation, we assayed increased kinase activity, and this implies independent mechanisms for their regulation. We also identified upregulated miRNAs linked to an increased expression of kinase coding genes, as denoted for CDK1, ARUKA and PTK6, and their kinase activity was likewise induced. The results reinforce the notion that expression of the kinase coding gene and the activity of the coded protein cannot be correlated easily. However, the networks are of critical importance to define therapeutic targets as described below (see master regulatory networks).

Apart from miRNet searches, and given that the miRNA–DEG correlation analysis identified more potential targets (see above and Table S6), we searched for new kinases based on the correlation analysis. By applying a threshold of R² ≥ 0.7, we used the results of the correlation analysis to identify DEGs coding for kinases and compared the results with their actual enzyme activity, as determined by kinome profiling of tumour samples from LC patients. As shown in Figure S5B, none of the correlation analysis revealed additional DEGs coding for kinases whose enzyme activity was likewise upregulated in tumour samples of AD patients. However, for SQ patients, we identified six kinases of which five were also flagged as potential targets based on the miRNet query (Figure 5H). Together, the DEM–DEG correlation analysis identified the FGR kinase as an additional potential drug target for SQ (Figure S5b).

3.2.2 Commonly regulated kinases in LC

Depicted in Figure 6A are the complex signalling networks in LC, and we separate the signalling networks in tumour and immune cells. Based on the paradigm that ≥70% of patients show the same change, that is, induced kinase activity, we identified 45 and 47 significantly regulated kinases in AD and squamous LC cases of which 34 are regulated in common (Figure 6B, Venn diagram). We did not observe repressed kinase activities. In the same way, we analysed the neuroendocrine and metastatic tumours and identified 74 and 27 kinases as significantly upregulated. Additionally and irrespective of the histologic subtype of LC, we identified 17 tyrosine and 4 serine-threonine kinases as commonly upregulated across 54 LC patients (Figure 6C).

As shown in Figure 6A, RAS is not regulated in the present study (grey coloured), while the rectangular marked proteins are based on upstream kinase analysis. Proteins marked by the symbol ℗ are direct measurements of phosphorylated kinases. Additionally, we identified regulated kinases by upstream analysis. Here, the phosphorylation of canonical peptides was assayed. An important finding of our study is an upregulation of a wide range of receptor tyrosine kinases in tumours such as the vascular endothelial growth factor receptor 2 (VEGF2/KDR), EGFR and ERBB2, ephrin type-A&B receptor (EPHA & EPHB2), platelet-derived growth factor receptor, the RET and the MER, MET and RON proto-oncogenes. Additionally, we assayed peptides typically phosphorylated by the AXL (Tyro3-Axl-Mer (TAM) receptor) tyrosine kinase, the insulin (InSR) and insulin-like growth factor 1 receptor (IGF1R), the ALK, neurotrophic TK, the kit proto-oncogene and the fibroblast growth factor, and we obtained clear evidence for their increased activity in tumours of LC patients (range in AD 2–4-fold, in SQ 2–24-fold). Moreover, for some of the receptor tyrosine kinases, we determined the phosphorylation sites and show the downstream signalling of kinases which were regulated as well. For instance, we assayed the phosphorylation of three peptides of the EGFR: the amino acid sequence (AS) 1103–1115 (GSVQNPYHNQPL), the AS 1165–1177 (ISLDNPDYQQDFF) and AS 1190–1202 (STAENAEYLRVAP) and observed differences between the various histological subtypes of LC. The differences in the tyrosine phosphorylation can be explained due to variant SRC and ABL1 kinase activity with implications for downstream signalling events. Similarly, we assayed the tyrosine phosphorylation of the MAPK1 peptide HTGFLTEYVATRW (AS 180–192) which was significantly increased in SQ cases only (range 1.2–3.7-fold). This peptide can be phosphorylated by the kinases RAF1, RET, JAK2, ERK1 and LYN, and all of them are upregulated in the Hannover cohort of LC patients. Likewise, we observed increased phosphorylation of the tyrosine residue 185 of the p38gamma/MAPK12 peptide ADSEMTGYVVTRW, and one study reported the precisely ordered phosphorylation reactions of the p38 mitogen-activated protein (MAP) kinase which involves the kinases MKK3/6, MEKK6, SEK1 and ASK.²³

3.3 Different kinomes in lung AD and SQ

Unlike adenocarcinomas, targeted therapy in SQ is still in its infancy.³³ We compared the kinase activities between different histological subtypes of LC and shown in Figure 6D is the fold change difference of significantly regulated kinases between AD and SQ. We measured the activity of kinases in tumour tissue relative to NATs and compared the results between AD and SQ patients. Remarkably, of the 46 commonly regulated kinases (Figure 5B), the activities of 17 kinases are higher in SQ, and for some patients, we observed unprecedented high activities as denoted for the FGFR1, PGFRB; ANXA2, CD3ζ, GAB2, LYN and MAPK12. It was not unexpected to see higher FGFR1 activities given its common alterations in SQ. However, we also observed up to 10-fold induced activities of the phosphoinositide-3-kinase regulatory subunit 1 and therefore demonstrate high activity of the phosphatidylinositol 3-kinase subunit. Together, the data underscore the great opportunities for targeted therapies in LC, and we obtained strong evidence for a broad range of kinases which are also potential targets for the treatment of SQ.

Additionally, we compared kinase activities of patients diagnosed with driver mutations in EGFR, KRAS, p53, BRAF and MYC (Table S2) to patients without driver mutations and compile the results in Table S10. Essentially, we did not observe significant differences in kinase activities between carriers of KRAS mutations and KRAS wild type. For instance, we compared the kinase profile of > 3-fold significantly regulated kinases of five AD patients with the KRAS mutations G13D, G12A, G12C, G12V and KRAS amplifications to nine AD patients without driver mutations and did not observe significant differences between the two cohorts (Figure 6E). The data broaden the perspective of KI-based therapies in patients without oncogenic driver mutations.

Finally, we identified kinases which were specifically regulated in AD, and shown in Figure S7 are the signalling networks for 28 peptides. The peptides are phosphorylated by 35 different serine/threonine kinases (Table S7) and with the exception of PI3K, GSK3B, RAF and ERK, the kinases were defined by upstream analysis, that is, kinases known to phosphorylate these peptides (Figure 5D). Our findings underscore the therapeutic opportunities in blocking tumour associated signalling networks.

3.4 Identification of molecular hub proteins to block tumour-specific signalling networks

We define hub proteins as master regulatory molecules which are at the apex of signalling cascades, and their inhibition results in disintegration, fragmentation and abrogation of signalling networks. By combining the kinome and genomic data, and based on computational analysis, we searched for MRs in LC-associated signalling networks. Based on the knowledge of TRANSPATH, we obtained information on signalling molecules in different pathways and downstream reactions. We identified several kinases as MRs, and in Figure 7 we show the results for the different histological types of LC. Although the number of MRs differed between AD and SQ, that is, 7 vs 3, we did not identify specific MRs for these tumour entities. Conversely, for NET and MT, there are four (NTRK2, RET, SYK and MAPK14) and three MRs (EGFR, MAPK3 and LYN), respectively, which were specific for these histological types of LC (Figure S8). The networks shown in Figure 7 depict significantly regulated MRs, which function as kinases. Moreover, depicted are significantly downstream regulated kinases, in addition to DEGs coding for signalling molecules in the networks. Finally, miRNAs control the expression of DEGs, and for a given network, we visualize the hierarchy of signalling molecules (Figures S9–S12). The MRs function primarily as receptor tyrosine kinases, MAPKs and cyclin-dependent kinases to influence cell proliferation, cytoskeletal dynamics and TCR/immune cell responses.

3.5 Validation of tumour-regulated kinases in human LC cell lines by siRNA and CRISPER knock out and cell viability assays

We used two different publicly available data sets to validate kinases as potential therapeutic targets. First, we considered the findings of Campbell and colleagues who performed large-scale profiling of kinase dependencies in cancer cell lines.³⁴ We defined a Z-score of ≥ −1.5 as threshold for significance, and only considered kinases that were significantly upregulated in > 70% of tumour samples. Among the 21 commonly upregulated kinases shown in Figure 6C, and following gene knock-down of CDK4, FER, FES and JAK1&2, we were able to confirm loss of cell viability in 9 out of the 12 human LC cell lines. In regards to 11 kinases specifically regulated in AD (Figure 6B), siRNA-mediated gene-silencing of CDK1, EPHA1, JAK3, RET and ZAP70 caused loss of cell viability in a panel of six human adenocarcinoma cell lines (Table S11). Moreover, siRNA-mediated gene silencing of the kinase ephrin receptor EPHA1 caused cell death in the large-cell lung carcinoma cell line BEN. Concerning SQ, gene silencing of FGFR2 and PDPK1 led to significant reductions in cell viability of the human LC cell lines H460, H23 and A427. Finally, we confirmed significantly reduced cell viability in the carcinoid cell line H727 following siRNA of tyrosine kinase 2 which is specifically upregulated in carcinoid NET (Table S11).

To independently validate the relevance of regulated kinases as therapeutic targets, we interrogated the DepMap database (https://depmap.org/portal/) and retrieved kinase inhibition data for 50 human lung AD and 19 squamous carcinoma cell lines. We evaluated the impact of gene knock down on cell viability, and compared kinases significantly regulated in patient tumour samples with kinase inhibition data deposited in DepMap.³⁵ Given that we compared results from different studies and experimental settings, we used −0.5 as threshold for effect size. The results are summarized in Table S11, and for each cell line, we specify the Z-scores and/or effect size. Together, we confirmed 62 and 27 kinases, respectively as potential therapeutic targets in 33 human lung AD and 15 squamous carcinoma cell lines (Table S11). Therefore, by utilizing the DepMap data, we were able to validate nearly half of the kinases as potential therapeutic targets.

3.6 Translating multi-omics data into individualized therapeutic concepts in distinct LC subtypes

Based on the seminal works of Hanahan and Weinberg, sustained proliferative signalling is one of the hallmark traits in cancers. In order to define kinases at the apex of signalling networks, we used kinome and genomic data and searched for master regulatory molecules as detailed in the Methods section. Depicted in Figure 7 are the MR regulatory networks. We show significantly regulated MR and downstream kinases and their associated gene networks all of which were highly regulated. We hypothesized that inhibition of MR kinases will lead to disintegration of an entire signalling trait and propose their combined inhibition with downstream kinases to be even more effective. Consequently, the gene silencing and/or drug-dependent inhibition of the MR kinases AKT1, CDK1&2, ERBB2 and the downstream kinase EGFR markedly reduced cell viability in a large panel of human adenocarcinoma cell lines (Table S11). Similarly, gene silencing and/or drug-dependent inhibition of CDK1&2 and the downstream kinases EGFR and MAPK1 in SQ led to marked reductions in cell viability. Noteworthy, and with the exception of AKT1 and ERBB2, estimates for the mean effect size differences ranged between −3 and −1.8. Clearly, this underscores the large effect of MR kinase inhibition on cell viability in LUAD cell lines. In regards to SQ and with the exception of MAPK1, the mean effect size differences ranged between −3.3 and −1.9. Altogether, this demonstrates the potency of MR kinase inhibition on cell viability.

By harnessing the kinase profiling data, we were able to propose new targets and/or combinations of kinase inhibitors for the treatment of LC. Specifically, of the 62 kinases upregulated in AD, and based on cell viability data following drug inhibition or gene silencing, we were able to validate 48 kinases as potential therapeutic targets across a panel of human LUAD cell lines (Table S11). We also validated 21 kinases across a panel of human lung SQ cell lines whose inhibition or gene knockdown halted or diminished cell proliferation. While some of the kinase inhibitors are already exploited for LC therapy, that is, four in AD (MET, EGFR, NTRK1 and RET) and four in SQ (EGFR, MAP2K2, RET and BRAF), there are another seven kinase targets (SRC, JAK1, TYK2, FGFR1/2, CDK6 and PDGFRA) which are used for other indications such as clear cell renal cell carcinoma, osteosarcoma, acute myeloid leukaemia, urothelial bladder cancer, bile duct and breast cancer (Table S8). We propose repurposing of these drugs and consider them as worthwhile candidates for clinical trials in LC.

Of the 62 kinases regulated in AD, there are 29 for which no approved drug is available. Similarly, for SQ, we identified three new kinases as potential therapeutic targets (Figure 8A, and Table S11). Given their significant upregulation in LC, there is good evidence for these kinases to be bona fide therapeutic targets. Some of the kinases function in cell cycle (AURKA, AURKB) or Ca2+/calmodulin dependent, mitogen-activated protein kinases/ribosomal proteins, AKT/mTOR and various protein kinase signalling pathways as well as ephrin. Additionally, we propose PDFGRB, MAPK3 and the serine-protein kinase Sgk1 for clinical evaluation in SQ. We also identified 10 kinases regulated in common in AD and SQ (Figure 8A).

Based on the MR networks described above (Figure 7), we hypothesized that the simultaneously use of MR and downstream kinase inhibitors will improve potency and therefore searched for LC patients with common upregulation of these therapeutic targets. The results are shown in Figure 8B, and in the case of AD, we propose the combined use of CDK1 and EGFR, CDK2 and EGFR and AKT1 and EGFR. In the case of SQ, we propose CDK1 and EGFR as well as CDK1 and MAPK1. Recently, the importance of CDK1 inhibition in cancer therapy was the subject of a review,³⁶ and overexpression of EGFR is common in LC.

Apart from targeting kinases which function as MR at the apex of signalling networks, we searched for commonly regulated kinases in LC. Here, we followed the idea of dual kinase inhibition with the kinases functioning in different pathways. We set a threshold of ≥50% of patients showing upregulation of two kinases in the same tumour sample. This defined 10 kinase doublets as potential therapeutic targets (Figure 8C). Once again, the new kinase targets have been validated based on cell viability assays (Table S11) and are suitable candidates for in-depth evaluation. In the same way, we considered kinase doublets in SQ. Outstandingly, there are 15 kinase doublets regulated in common in AD and SQ (Figure 8C).