Pharmacoproteomic characterisation of human colon and rectal cancer

We devised a multi‐omics data integration strategy to determine integrated proteomic subtypes of human colorectal cancer cell lines and patient samples in order to predict their sensitivity towards a variety of clinical and pre‐clinical drugs and combinations thereof (Figs 1A and EV1, Appendix Supplementary Methods). To accomplish this, we first used LC‐MS/MS‐based shotgun proteomics to measure the proteomes of 65 CRC lines and to quantify their expressed kinomes using Kinobeads (Bantscheff et al, 2007; Medard et al, 2015). This led to the identification of a total of 11,796 protein groups (median/cell line = 9,447) representing 10,951 genes (median/cell line = 9,068) and 235 human kinases (median/cell line = 155, from Kinobeads experiments; Fig 2A; Table EV1A and B). We next re‐analysed the proteomic profiles of 90 CRC patients published by the CPTAC (Zhang et al, 2014) using the same analysis pipeline and identified 7,005 protein groups (median/patient = 4,980) representing 6,727 gene groups (median/patient = 4,901; Fig 2C; Table EV1C, Appendix Supplementary Methods). The CRC65 data contained most proteins of the CPTAC CRC data (Fig 2B), but we noted that gene groups unique to the CPTAC dataset are enriched in extracellular matrix proteins and IgGs as one might expect for tissue samples containing blood vessels and connective tissue. For quantification of full proteomes, we used a modified version of the intensity‐based absolute quantitation (iBAQ) approach (Schwanhausser et al, 2011) termed gene‐centric iBAQ (giBAQ, Appendix Supplementary Methods), while kinomes were quantified using the label‐free quantification (LFQ) intensities provided by MaxQuant (Cox et al, 2014). For all subsequent data analyses, we used the proteomic data aggregated at the gene group level annotated with gene symbols as identifiers in order to be able to compare proteomics and transcriptomics data (referred to as protein expression/abundance throughout the manuscript). An overview of the data integration pipeline is depicted in Fig EV1.

A number of studies have reported CRC patient subtypes based on mRNA measurements, and these were recently consolidated into consensus molecular subtypes (CMS) and applied to patients from the CPTAC study (Guinney et al, 2015). On this basis, we sought to determine the CMS membership of CRC cell lines in order to identify cell lines representative of CRC tumour subtypes. We downloaded 10 public mRNA datasets (Appendix Supplementary Methods), eight for the cell lines and two for the patients (Fig 1B), analysed them from the data level closest to raw data available to us and aggregated the datasets using a scheme similar to the one used by Guinney et al. This involved the selection of a reference dataset, followed by the selection of a probe set for each gene based on a consistency criterion and subsequent cross‐dataset normalisation (Appendix Supplementary Methods). This resulted in a combined expression matrix of 9,737 transcripts (7,113 after filtering for low abundance transcripts) across 145 non‐redundant cell lines (including the CRC65 panel) and 89 tumours (Table EV1D). We adapted the single‐sample CMS classifier reported by Guinney et al to accept gene symbols as identifiers (rather than Entrez IDs; Appendix Supplementary Methods) and predicted the CMS for cell lines and patients based on 382 of the 692 classifier genes contained in the combined expression matrix. The correct classification of 65 out of 81 patients (80%, using the original CMS assignment as the ground truth) provided confidence that cell lines can be placed into CMSs with good accuracy and the resulting subtype labels for the CRC65 cell lines and the CPTAC patients are shown in Fig 1B. A subtype‐resolved evaluation of the prediction accuracy using a confusion matrix and a table containing a variety of commonly used metrics for evaluating classification performance can be found in Table EV2E.

Despite the fairly deep proteomic measurements, the quantification of proteins across many cell lines (and patients) suffered from an increasing number of missing values for proteins of decreasing abundance (Fig EV2A). We addressed this frequently encountered issue by mRNA‐guided and minimum‐guided missing value imputation on the peptide level to generate one complete protein expression matrix consisting of 59 cell lines, 81 tumours and 6,254 proteins (Fig EV2, Table EV1E), of which 323 were contained in the CMS classifier by Guinney et al (CMSgene in Fig 3A; see Appendix Supplementary Methods for details). In order to estimate protein levels from mRNA levels, we removed systematic differences (Fig EV3A and B) between proteomics and transcriptomics data using MComBat (Stein et al, 2015; see Appendix Supplementary Methods). This increased the protein/mRNA correlation for both the CRC65 and CPTAC datasets (Figs EV2B and EV3C), enabling mRNA‐guided missing value imputation. After imputing missing values separately for both datasets (Fig EV2D and E, Appendix Supplementary Methods), we accounted for differences in their proteomic depth (Fig EV2C, Appendix Supplementary Methods) using ComBat (Johnson et al, 2007) before we merged the two protein expression matrices. Using the combined protein expression matrix and consensus clustering (Appendix Supplementary Methods), we identified three integrated Full Proteome Subtypes (short: FPSs, namely FPA, FPB and FPC; Fig 3A). Each subtype consisted of cell lines as well as patients in a ratio of 28/34 for FPA, 22/12 for FPB and 9/26 for FPC, indicating that indeed there are cell lines, which are molecularly more similar to tumours than they are to other cell lines. We measured the association of these FPSs with previously published subtypes, as well as genomic and epigenomic features using Fisher's exact test (Table EV2A), and found good overall concordance but with some differences in detail (see discussion). Interestingly, FPA was associated with TP53 mutations, while FPB was associated with ATM mutations, suggesting that p53 signalling in response to DNA damage is perturbed through distinct mechanisms in these two subtypes. FPC showed association with mutations in BCL9L, a transcriptional activator of b‐catenin activity, RNF43, an E3 ubiquitin‐protein ligase, which acts as a negative regulator of Wnt signalling, as well as with mutations in the histone acetyltransferase EP300, all pointing towards deregulated WNT signalling in conjunction with aberrant histone acetylation. Enrichment analysis of functional classifications using MetaCore (Fig EV4, Appendix Supplementary Methods) revealed that proteins significantly under‐ or overrepresented in the different FPSs fall into specific categories. Briefly, the analysis suggests that FPA is characterised by a “high metabolism, low cell cycle, microsatellite stable (MSI−)” signature, FPB harbours a “high immune response, low metabolism, microsatellite instable (MSI+)” signature and FPC shows a “low immune response, low inflammation, low adhesion (invasion)” signature.

Since the response to drug treatment cannot be determined as quickly and as broadly as would be desirable in clinical studies, we took advantage of the fact that CRC cell lines recapitulate the main molecular subtypes in CRC and were recently extensively characterised for drug sensitivity as part of major screening efforts. On the basis of the proteomic data, we used elastic net regression (Zou & Hastie, 2005) to predict the response (sensitivity/resistance) of CPTAC patients and CRC65 cell lines to 577 drugs or combinations thereof (Fig 3B, Appendix Supplementary Methods). Briefly, common cell lines (52 cell lines overlap in total; median overlap per drug is 27) between the CRC65 panel and three small‐molecule drug sensitivity screens (Barretina et al, 2012; Garnett et al, 2012; Rees et al, 2016) and one study investigating sensitivity of CRC cell lines towards cetuximab (Medico et al, 2015) served as the training set (Fig 1, Appendix Supplementary Methods). We used significance analysis of microarrays (SAM; median FDR of 0.001; Appendix Supplementary Methods) to identify drugs which showed specificity for certain FPSs and performed target space enrichment by measuring the association between these subtypes and recurrent annotated targets of these drugs using Fisher's exact test (Fig 3B, Table EV2B). Cetuximab, for example, was more effective in FPA than in the other FPSs, with EGFR target enrichment reaching statistical significance (P = 0.05). FPA was also significantly (multiple‐test adjusted P < 0.05) associated with drugs targeting the kinases MAP2K1 and MAP2K2 (MEK1 and 2; mainly selumetinib/AZD6244 and its combinations with other compounds), as well as with drugs and combinations targeting the bromodomain containing protein BRDT (mainly by the compound JQ‐1), histone deacetylases or HDACs (vorinostat among others) and the nicotinamide phosphoribosyltransferase NAMPT (daporinad among these). Even though more drugs were differentially effective in FPB than in FPA (205 versus 170), most of the former did not show significant enrichment in their target space, making it difficult to link the drug phenotype to the presumed mechanism of action. However, FPB did show positive association with drugs targeting inhibitors of apoptosis (IAPs; all these compounds are mimetics of second mitochondrial activator of caspases, SMAC) as well as negative association with drugs targeting EGFR, while FPC showed association with drugs targeting dihydroxyfolate reductase (DHFR, including methotrexate). The complete list of predicted drug sensitivities for the patients and cell lines can be found in Table EV3A, while mean effect sizes of all proteins from elastic net regression can be found in Table EV3B.

Since many targeted cancer drugs act on kinases, we performed independent experiments focussing on the expressed kinome of the CRC65 cell line panel to identify kinases associated with drug sensitivity or resistance. Pulling down protein kinases using immobilised kinase inhibitors (Kinobeads; Bantscheff et al, 2007; Medard et al, 2015) led to the reproducible (average R = 0.91 between replicates; Fig EV5A) quantification of 138 kinases by mass spectrometry (Appendix Supplementary Methods), which correlated well (R = 0.94; Fig EV5B) with data obtained by Western blotting and densitometry for the kinases EPHA4 (Fig EV5C; variable expression across the panel) and ABL1 (Fig EV5D; relatively stable expression across the CRC65 panel). While kinase expression as measured using Kinobeads correlated reasonably well with measurements in full proteomes, an enrichment of kinases was clearly visible in the Kinobeads data (Fig EV5E) and led to the identification and quantification of more than 50 kinases not detected in the deep proteome analysis (Fig EV5F). Consensus clustering identified three Kinobeads Subtypes (KSs) KA, KB and KC (Fig EV5G) that fully recapitulated the Full Proteome Subtypes and showed significant association with other subtype classifications (Table EV2C).

By searching for differentially expressed kinases between the different Kinobeads Subtypes using SAM (median FDR of 0.01), we identified proteins frequently mutated in microsatellite instable (MSI+) tumours with concomitant decrease in expression of proteins such as ACVR2A and TGFBR2 (Kim et al, 2013), as well as the receptor tyrosine kinase EPHA2, which was overexpressed in KC (Table EV4). Overexpression of EPHA2 is known to predict resistance towards cetuximab in CRC (Strimpakos et al, 2013). Elastic net regression was used to identify kinases associated with drug sensitivity or resistance and confirmed the strong association of EPHA2 with resistance towards cetuximab (Fig EV5H, Table EV3C and D). We also noted that overexpression of MAP2K1 (MEK1) was associated with resistance to two‐thirds (12/18) of all inhibitors targeting EGFR (Fig 4 for example). Activating mutations in K57 of MAP2K1 were previously shown to be a potential mechanism of primary resistance towards EGFR‐targeted therapy (Bertotti et al, 2015), and our data suggest that high expression can have a similar effect. SAM also identified MERTK—a receptor tyrosine kinase correlating with disease progression in melanoma (Schlegel et al, 2013) and not detected in the full proteome measurements—to be differentially expressed between the different Kinobeads Subtypes (q‐value < 0.01). MERTK is down‐regulated in KRAS‐mutant CRC (Watanabe et al, 2011) and up‐regulated in pancreatic cancer cell lines resistant towards selumetinib (Beech & Kelly, 2014). Interestingly, we also recurrently found high MERTK expression associated with resistance towards MEK1/2 inhibitors as well as other drugs from multiple drug sensitivity datasets (Figs 5A and B, and EV6A, Appendix Supplementary Methods). To validate hypotheses arising from the above predictions, we performed a series of in vitro drug treatment experiments.

On the basis of MERTK expression, we selected three cell lines (CC07, HDC‐143 and SK‐CO‐1) predicted to be sensitive to two MEK1/2 inhibitors (RDEA119 and PD‐0325901) as well as three cell lines (C10, CaCo‐2 and T84) predicted to be resistant to these drugs (Fig EV6B). Cell viability assays and nuclei counting of cells (Fig 6C and D) showed that the predictions could be confirmed, as cell lines expected to be resistant did not respond as well as cell lines expected to be sensitive (P ≤ 0.05, one‐sided Mann–Whitney test). Similar results were obtained for the predicted association of ACVR2A expression with drug sensitivity to AUY922 (an HSP90 inhibitor) and BAY 61‐3606 (a designated SYK inhibitor) in LS‐180, OXCO‐1 and RKO cells (sensitive) or C125‐PM, HDC‐111 and HT55 cells (resistant; Fig 5C and D). Since ACVR2A expression is reduced in microsatellite instable CRC, these tumours might benefit from treatment with these drugs.

We next asked whether knockout of MERTK in MEK1/2‐inhibitor‐resistant C10 cells could re‐sensitise them to treatment with RDEA119 (MEK1/2 inhibitor). A Western blot‐based validation of the CRISPR/Cas9‐mediated knockout is depicted in Fig 5E. We also confirmed the knockout by sequencing of the targeted genomic region, which showed two insertions: one in exon 7 (FN3 domain) and one in exon 14 (kinase domain), inducing frameshifts. As depicted in Fig 5F, $\mathrm{C}{10}^{{\mathrm{MERTKN}}^{\mathrm{KO}}}$ was more sensitive to RDEA119 than the parental $\mathrm{C}{10}^{{\mathrm{MERTKN}}^{\mathrm{WT}}}$ cell line. We next tested whether the combination of MEK (by RDEA119 or PD‐0325901) and MERTK inhibition (UNC569) could re‐sensitise MEK1/2‐inhibitor‐resistant cell lines. As shown in Fig EV6C–F, co‐treatment indeed significantly (P ≤ 0.05, one‐sided Mann–Whitney test) reduced the viability of MEK1/2 inhibitor‐resistant cell lines to levels comparable to sensitive cell lines. However, this reduction appeared to be mainly due to UNC569 on its own, since both MEK1/2 inhibitor‐sensitive and resistant cell lines show comparable response to the co‐treatment when treated with UNC569 alone.

In order to evaluate the general expression status as well as the clinical impact of MERTK expression in CRC, we quantified its abundance in 1,074 patients enrolled in the QUASAR 2 trial by immunohistochemistry (IHC). After establishing cell‐based positive (CaCo‐2) and negative (RKO) controls with high and low expression of MERTK, respectively (Fig 6A and B), tissue microarrays (TMAs) were stained for MERTK and the fraction of positive cells, as well as the staining intensity, were quantified by pathologists. While MERTK expression was found primarily at the membrane in CaCo‐2 cells, the TMA data showed combined cytoplasmic and membrane staining. Therefore, TMAs were categorised to have either high or low MERTK expression (more or < 5% cytoplasm/membrane‐positive tumour cells, Appendix Supplementary Methods) without distinguishing between these compartments. Figure 6C shows three representative TMA samples alongside their quantification of MERTK, also highlighting the strong variability of MERTK expression in primary tumours. Modelling the outcome variables of the QUASAR 2 trial as a function of MERTK expression showed that high MERTK expression was prognostic of worse 5‐year overall survival (OS), disease‐free survival (DFS) and recurrence‐free survival (RFS) in both univariate and multivariate Cox proportional hazards regression (Table EV5, Fig 6D). Notably, MERTK expression is more informative in predicting these outcome variables than the patients’ treatment status, since its coefficient was significant in the final multivariate Cox proportional hazards model, while the coefficient of the treatment variable was not (see also Appendix Supplementary Methods). Comparing tumours with high versus low MERTK expression resulted in a multivariate hazard ratio (HR) of 1.61 for OS (CI_0.95 = 1.1–2.36; significant coefficient for MERTK with P = 0.015; Wald test), 1.70 for DFS (CI_0.95 = 1.24–2.33; significant coefficient for MERTK with P = 0.00094; Wald test) and 1.77 for RFS (CI_0.95 = 1.26–2.48; significant coefficient for MERTK with P = 0.00087; Wald test), respectively. High expression of MERTK was also significantly associated with BRAF^V600E mutations (12% of the CRC65 cell line panel and 13% of the QUASAR 2 cohort carry the mutation, Fig EV5D) and stage T4 tumours, while low cytoplasmic/membranous expression was associated with wild‐type BRAF and stage T2/T3 tumours (two‐sided Fisher's exact test P < 0.05, Table EV2D). We note that more work is needed in order to determine the direction of causality for these observations and whether MERTK might qualify as a drug target in CRC in addition to being a prognostic survival marker in CRC patients.

All cell lines in the CRC65 panel apart from SW480 (kind gift from Ulrike Stein from the MDC, Berlin) were collected by the laboratory of Prof Sir Walter Bodmer FRS at the University of Oxford and were previously HLA‐typed and characterised for other genetic changes to determine whether they are derived from cancers of different donors (Emaduddin et al, 2008). Information on the CRC65 cell lines and the CPTAC tumour samples like MSI status, original sources (cell lines) or subtype membership is compiled in Table EV6A.

We analysed the expression of MERTK in 1,074 patients from the QUASAR 2 clinical trial cohort, with approval from the West Midlands Research Ethics Committee (Edgbaston, Birmingham, UK; REC reference: 04/MRE/11/18). All participants provided written informed consent for treatment, and separate consent was obtained regarding the use of tumour tissue. QUASAR 2 is a phase III international randomised controlled trial, which collected data on toxicity, overall survival (OS), disease‐free survival (DFS) and recurrence‐free survival (RFS) for 1,941 stage II/III CRC patients, with the aim to determine the efficacy of adjuvant capecitabine±bevacizumab after resection of the primary tumour. A biobank comprising 1,350 FFPE blocks was established, and tissue microarrays (TMAs) were generated from 1.2‐mm cores.

Slightly altering the culture conditions described by Emaduddin et al (2008), cells were grown in high glucose Dulbecco's modified Eagle medium (DMEM, including GlutaMAX and pyruvate; PAA) containing 1% Pen‐Strep (penicillin at 100 units/ml and streptomycin at 100 μg/ml final concentration, PAA) and 10% foetal bovine serum (FBS, SLI EU‐000F Batch 503005) in a humidified incubator at 37°C and 10% CO₂ using T‐175 flasks (Corning). Hereafter, this medium composition is referred to as “culture medium” and the culture conditions apart from the culture vessel are referred to as “standard culture conditions”. Adherent cells were harvested at ~80–90% confluency. Suspension cells (e.g. HDC‐135) growing in 250 ml of culture medium were harvested by centrifugation in 250‐ml centrifuge tubes (Corning) at 300 ×g and 4°C for 5 min. We used RIPA100 buffer (20 mM Tris–HCl pH 7.5, 1 mM EDTA, 100 mM NaCl, 1% Triton X‐100, 0.5% sodium deoxycholate and 0.1% SDS) containing protease (Complete™ mini with EDTA; Roche) and phosphatase inhibitors (Phosphatase Inhibitor cocktail 1 and 2; Sigma Aldrich) at 2× and 5× the final concentration recommended by the manufacturer, respectively, and lysed the cells for 30 min at 4°C. After clearing the total cell lysates (TCLs) at 22,000 ×g for 30 min at 4°C, the protein concentration was determined using a Coomassie‐based protein assay kit (Thermo Fisher) and the TCLs were stored at −80°C until further use.

We performed in vitro cell viability assays in order to test our in silico predictions. Assays were performed as described previously (Garnett et al, 2012), with minor modifications. Cell viability assays were carried out in technical triplicates in order to generate 10‐point dose–response curves for drugs with the resistance of which either ACVR2A or MERTK was associated. For each drug, we selected three cell lines predicted to be resistant and three cell lines predicted to be sensitive towards the respective drug and seeded them in 100 μl culture medium at their optimal seeding density on day zero (C10 = 4,000, CaCo‐2 = 2,000, CC07 = 4,000, HDC‐143 = 12,000, SK‐CO‐1 = 8,000, T84 = 11,000, RKO = 4,000, LS 180 = 2,000, HDC‐111 = 4,000, HT55 = 12,000, OXCO‐1 = 8,000, C125‐PM = 11,000). Following an overnight incubation under standard culture conditions, 50 μl of fresh medium containing either 1% DMSO or the respective drug in a 9‐point twofold dilution series in 1% DMSO was added to the corresponding wells. This resulted in a final DMSO concentration of 0.33%, while the highest final drug concentration was 10 μM for UNC569 (Merck), as well as 0.5 μM for PD‐0325901, 5 μM for RDEA119, 0.5 μM for AUY922 (all from Cambridge Bioscience) and 16 μM for BAY 61‐3606 (Insight Biotechnology), respectively. For drug co‐treatments, the respective compounds were combined at constant ratios over the entire concentration range used for the single‐agent treatments, keeping the final DMSO concentration in the assay at 0.33%. After 72 h of incubation time under standard culture conditions, the medium was either replaced with 150 μl of fresh culture medium containing 1 μM of Hoechst 33342 (Thermo Fisher #H3570) or 10 μl AlamarBlue (Thermo Fisher #88952) added to each well (only $\mathrm{C}{10}^{{\mathrm{MERTKN}}^{\mathrm{KO}}}$ and $\mathrm{C}{10}^{{\mathrm{MERTKN}}^{\mathrm{WT}}}$). Cells were incubated under standard culture conditions for 1 h or 4 h, respectively, before they were either imaged using an In Cell Analyzer 6000 automated confocal microscope (GE Healthcare) with four fields of view (FOVs) per well or before AlamarBlue fluorescence was quantified using a FLUOstar Omega plate reader (BMG Labtech). The Hoechst 33342 channel of all images was subsequently analysed with Columbus v2.6.0 (PerkinElmer) using two of the built‐in algorithms “B” and “C” to automatically segment nuclei. For each well, we then counted the number of nuclei satisfying standard quality control criteria in all four FOVs. Subsequently, the nuclei count or AlamarBlue fluorescence was normalised to the mean of the corresponding DMSO controls, followed by dose–response modelling and parameter extraction (Appendix Supplementary Methods).

Knockout of MERTK in C10 cells was installed by CRISPR/Cas9 gene targeting as described previously (Ran et al, 2013). Guide RNA sequences were selected by using the CRISPR Design Tool (http://www.genome-engineering.org/crispr/?page_id=41). Three guide RNAs targeting exons 7, 8 (encoding the FNIII domain) and 14 (encoding the kinase domain) of MERTK were obtained, which were designed to induce double‐strand breaks at position 982, 1,273 and 1,882 bp. The guide RNA sequences were cloned into pSpCas9(BB)‐2A‐GFP (the vector was a gift from Feng Zhang; Addgene plasmid #48138) by standard Golden Gate Assembly using the BbsI site, followed by transformation of chemically competent DH5α cells. Three colonies of each sgRNA transformation were picked and the correct insertion confirmed by sequencing. C10 cells were transfected with a mixture of all three sgRNAs using Lipofectamine 3000 (Invitrogen) according to the manufacturer's instructions. Transfected cells were microscopically identified by expression of GFP and sub‐cloned 3 days post‐transfection using cloning rings. After 2–4 weeks, individual cell clones were tested for successful knockout of MERTK by Western blot and sequencing.

Total cell lysates were diluted to 1 mg/ml protein concentration and 1× final sample buffer concentration with 4× sample buffer (70 mM Tris pH 6.8, 5% v/v 2‐mercaptoethanol, 40% v/v glycerol, 3% w/v SDS, 0.05% w/v bromophenol blue) and stored at −80°C until further use. Samples were separated using 4‐12% NuPAGE Bis‐Tris mini/midi gels (70 μg/14 μg per sample) and subsequently blotted to Hybond‐P PVDF (Amersham) or nitrocellulose (iBlot Transfer Stack, Thermo Fisher) membranes according to the manufacturer's instructions. We used primary antibodies against EPHA4 (ab157588, 1:500, Abcam), ABL1 (OP20, 1:1,000, Merck), BRAF^V600E (E19290, 1:500, Spring Bioscience) and ERK1/2 (#4695, 1:1,000, Cell Signaling Technology) for PVDF and primary antibodies against MERTK (ab52968, 1:2,000, Abcam) and ACTB (sc‐47778, 1:1,000, Santa Cruz) for nitrocellulose membranes. Membranes were incubated in the dark with appropriate fluorophore‐coupled secondary antibodies (IRDye 800CW goat anti‐mouse IgG #926‐32210, IRDye 680RD donkey anti‐rabbit IgG #926‐68023, IRDye 800CW goat anti‐rabbit IgG #926‐32211 and IRDye 680RD goat anti‐mouse IgG #925‐68070). Membranes were measured using an Odyssey near‐infrared scanner (LI‐COR). Densitometry of Western blots was carried out using ImageStudioLite v5.2.5 (LI‐COR), expressing EPHA4 and ABL1 expression relative to the respective ERK1/2 signal, followed by dividing all expression values by the expression value of OXCO‐1 (present on each gel) and log2‐transformation.

In order to evaluate MERTK expression in tissue microarrays, a staining protocol was developed using positive and negative control cell lines for MERTK expression, selected based on the expression level as measured by LC‐MS/MS. For each of the four control cell lines (negative: RKO, CC07; positive: HT55, CaCo‐2), we grew one T‐175 flask under standard culture conditions in culture medium to ~70% confluency, washed the cells once with 10 ml ice‐cold PBS and subsequently scraped them into 2 ml 4% PFA in PBS, followed by centrifugation at 400 ×g for 10 min and fixation overnight. Afterwards, the cell pellets were embedded in paraffin using standard procedures. Immunohistochemistry (IHC) of QUASAR 2 TMAs and FFPE control cell pellets was carried out as already described (Schlegel et al, 2013), with minor modifications. Stainings quantifying MERTK expression were performed using the Bond‐MAX automated IHC system (Leica Biosystems) at room temperature if not specified otherwise. After deparaffinising and re‐hydrating the slides, heat‐induced epitope retrieval (HIER) was performed for 10 min at 100°C with epitope retrieval solution 1 (AR9961, pH 6). MERTK was detected using the Bond Polymer Refine Detection Kit (DS9800) together with a primary antibody against MERTK (ab52968, 1:1,000, Abcam) according to the manufacturer's instructions, followed by counterstaining nuclei with haematoxylin solution (< 0.1% haematoxylin). Slides were washed, dehydrated with an increasing alcohol series (30 s in 50, 70, 100 and 100% EtOH), cleared with two 30‐s washes in 100% xylene and subsequently mounted using DPX mountant (Sigma Aldrich). Finally, slides were scanned at 40× magnification using a ScanScope (Aperio).

Proteins detected in proteomics experiments were quantified based on MaxQuant output data, which was subsequently integrated with transcriptomics data from various CRC samples (see Appendix Supplementary Methods for details). All statistical analyses were carried out using R v3.2.4 (R Core Team, 2016).

Modified MComBat and computeAUC functions (Appendix Supplementary Methods) can be downloaded from https://github.com/mfrejno/pharmacoproteomics_crc.

The proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository (Vizcaino et al, 2016) under ID codes PXD005353–PXD005355.