All buffers, gels and SELDI chips were purchased from Bio-Rad Laboratories Ltd. (Hemel Hempstead, UK), and all other chemicals were obtained from Sigma-Aldrich; Merck KGaA (Gillingham, UK), unless stated otherwise in the text.

Urine samples were obtained from 83 patients with upper GI cancers undergoing potentially curative resection. The participant demographics are summarised in Table I and provided in detail in Table SI. The participants' age ranged between 43 and 83 years. Fasting urine samples were obtained at induction of anaesthesia. One-third of the patients had pancreatic tumours, one-third had oesophageal cancer, approximately one-sixth had malignancies of the OGJ and one-sixth suffered from gastric cancer. All procedures were approved by the local research ethics committee and written informed consent was obtained from the patients. The study conformed to the standards set by the Declaration of Helsinki. All urine samples were stored at -40˚C. Long-term storage of samples (>1 month) was at -80˚C.

SELDI chips (CM10 and IMAC30) were prepared for sample application according to the manufacturer's recommendations. Briefly, IMAC30 chips were loaded with 0.1 M CuSO₄, washed with water, neutralised with 0.1 M NaHAc (pH 4.0) and again washed with water, followed by two washes with 0.1 M NaHPO₄ and 0.5 M NaCl; CM10 chips were washed twice with 0.1 M NaHPO₄ (pH 4.0). All chips were processed in a bioprocessor assembly by incubating 0.1 ml urine and 0.1 ml binding buffer [IMAC30: 0.1 M NaHPO₄, 0.5 M NaCl; CM10: 0.1 M NaHPO₄ (pH4.0)] for 1 h at room temperature with vigorous shaking, followed by three washes with 0.2 ml binding buffer for 5 min each at room temperature with vigorous shaking and two washes with 0.2 ml water at room temperature with vigorous shaking. All chips were removed from the bioprocessor assembly, air-dried and 1 µl energy-absorbing matrix [a saturated solution of sinapinic acid in 50% acetonitrile (ACN) and 0.5% trifluoroacetic acid] was added twice. Air-dried chips were analysed in a PCS4000 SELDI-TOF instrument (Bio-Rad Laboratories, Ltd.) by measuring the 1,000-25,000 Da range with a low laser setting of 2.5 µJ, and spectra were exported as ‘.xml’ files. The SELDI instrument was calibrated using the ProteinChip All-In-one peptide standard (Bio-Rad Laboratories, Ltd.). The source voltage was 25,000 V and the detector voltage was 2,946 V. Quality control and consistency were ensured by using one random pool of urine samples on one spot per chip each. Spectra of the full analysis were recorded in two large batches to minimize instrument variability and drift. Spectral alignments of all quality controls ensured consistency of all spectra.

ProteinChip Data Manager Software version 4.1 with integrated Biomarker Wizard cluster analysis (Bio-Rad Laboratories, Ltd.) was used for analysis. SELDI-TOF-MS traces were split into the four cancer type groups. The baseline was subtracted from individual m/z traces and the profiles were normalised using total ion current, followed by identification of peak clusters using the cluster analysis tool. Peaks were selected in the first pass with a signal-to-noise (S/N) ratio of >5 and a valley depth of at least 3, and in the second pass with a S/N of 2 and a valley depth of 2. The cluster mass window was set to 0.2% of the mass. Clustered peaks were only included if they occurred in at least 10% of all spectra. The resulting P-values, mean and median m/z values, and the intensities of the clustered peaks were exported and saved as ‘.csv’ files. A two-sample t-test was used to compare mean normalized intensities between the groups. The P-value was set at 0.05 to indicate statistically significant differences. Clustered peak lists were analysed with the Biomarker Pattern Software (Bio-Rad Laboratories, Ltd.) and m/z vs. intensity matrices were analysed using decision tree analysis, selecting the standard error rule of minimum-cost tree regardless of size, and using the Gini method. V-fold testing was set to 1,000. All samples from one cancer type were used to build exploratory networks by using all other cancer samples as ‘negative’ controls. Tree model building was performed using a selected m/z peak as the only deciding factor to obtain sensitivity and specificity values. Sensitivity was defined as the probability of predicting specific cancer cases, and specificity was defined as the probability of predicting other cancers.

Peaks observed in the CM10 and IMAC30 chip types (Table SIII) that exhibited marked expression differences between tissue-specific cancer samples, and statistical significance in the cluster analyses with P<0.05, were further investigated. Urine (0.5 ml) from positive or negative samples in relation to specific peaks was added to 30 µl CM10 or IMAC30 (Cu²⁺-complexed) spin column resin (Bio-Rad Laboratories, Ltd.) and 0.75 ml binding buffer [0.1 M NaHPO₄ (pH 4.0) for CM10 resins, and 0.1 M NaHPO₄ (pH 7.0) including 0.5 M NaCl for IMAC30 resins] and incubated for 1 h at room temperature under constant agitation. Unbound material was removed and the resin was washed four times with 0.3 ml binding buffer. Bound material was separated by electrophoresis on a 16.5% Tris-Tricine gel (Bio-Rad Laboratories, Ltd.), and gel bands in the region of 2-10 kDa were excised following Coomassie staining (BioSafe Coomassie; Bio-Rad Laboratories, Ltd.). Positive and negative samples were both selected on the presence and absence of a specific m/z peak to be identified based on SELDI-TOF-MS analysis. Proteins and peptides from gel bands were digested in situ with trypsin, the resulting peptides eluted with ACN, and analysed by LC-MS/MS as described previously (17). Data-dependent acquisition was controlled by Xcalibur software and fragmentation spectra were then processed by Xcalibur and BioWorks software (Thermo Fisher Scientific, Inc., Loughborough, UK) and submitted to the Mascot search engine (Matrix Science, London, UK) using UniProt/SwissProt (release July 2010, Homo sapiens, 18055 sequences) as the reference database. The Mascot search parameters were as follows: Enzyme specificity, trypsin; maximum missed cleavage, 1; fixed modifications, cysteine carbamidomethylation; variable modification, methionine oxidation; precursor mass tolerance, ±3 Da; and fragment ion mass tolerance, ±0.4 Da. Only Mascot hits with a false discovery rate of #x003C;0.05 were taken into consideration.

Observed proteins with at least two peptide matches from the LC-MS/MS analysis were then further analysed by pattern matching based on SELDI-TOF-MS measured expression levels of peaks of interest (expected abundance in selected samples). This was performed using software written in-house, which compares observed protein expression patterns in a pre-defined set of samples (LC-MS/MS results) against a matrix of peak patterns (SELDI-TOF clustered peak intensities, where estimated peaks are set to null) in the same set of samples. The scoring is based on sensitivity (percent observed over expected) and specificity (percent not observed over not expected), and the results are presented in descending order of cumulative scores.

Cross-validation of identified peaks was performed by western blotting of raw urine samples for anti-cathepsin B (CTSB) and anti-cystatin B (CSTB) blots (20 µl per sample) using standard protocols (37). The antibodies used were rabbit anti-human CTSB (G-60; 1:1,000; cat. no. 3373; Cell Signaling Technology, Danvers, MA, USA), rabbit anti-human serum albumin (1:1,000; cat. no. A3293; Sigma-Aldrich; Merck KGaA), mouse anti-human CSTB (1:400; cat. no. sc-101510; Santa Cruz Biotechnology, Santa Cruz, CA, USA), and the peroxidase-coupled secondary antibodies were from Upstate (Lake Placid, NY, USA), used at a dilution of 1:5,000. Detection of signals was performed by chemiluminescence using ECL western blotting reagents (Thermo Fisher Scientific, Inc., Cramlington, UK).

Urine samples from 83 patients diagnosed with various types of upper GI cancer were analysed in the course of this study (Table I). A full demographic with additional clinical data is provided in Table SI. The mean age of the participants was 65 years, 34% of the cohort suffered from pancreatic cancer, 33% from oesophageal cancer, 18% from gastric tumours and 15% from cancer of the OGJ. SELDI-TOF-MS analysis was selected to screen the samples mentioned above based on our previous study, where we described global specific markers for upper GI cancer (20). We found that both the metal chelator resin IMAC30 (Cu²⁺-chelated) and the weak cation exchanger CM10 yielded the best and most reproducible results. All samples were measured by SELDI-TOF-MS to obtain a peak pattern of identified molecular constituents, allowing us to stratify tissue type-specific cancers. We selected to compare only samples from cancer patients against each other to circumvent the issue of accidentally tracking lead markers that may be associated with additional underlying conditions, such as inflammatory responses, which are commonly observed in cancer patients.

SELDI analysis of the CM10 chip type-based screen of the 83 cancer urine samples resulted in 9,379 peaks, and the IMAC30 chip analysis provided a cumulative peak list of 3,346 features (Table SII). Clustering of observed peaks using the thresholds described in Materials and methods resulted in 328 cluster peaks for the CM10 chips, and the IMAC30 chip type gave yield to 92 common peak clusters above the set thresholds (Table SIII). Both analyses were performed by omitting estimated peaks in order to restrict and raise the specificity of potential marker peaks and, therefore, only included well-defined and separated individual peaks from all spectra. Statistical analysis revealed that 8 peak clusters are potentially associated with the various cancer types, namely m/z 2444 and 2557 for pancreatic cancer in the IMAC30 set, and m/z 2447 and 9618 in the CM10 chip-based set for the same cancer type, m/z 5511 and 4908 for OGJ cancer and m/z 4639 for gastric cancer in the IMAC30 chips, and m/z 4141 for oesophageal cancer in the CM10 chip set (Fig. 1; Table II). All potential markers have P<0.05 and exhibit upregulation in the associated disease. It is noteworthy to mention that any given peak cluster can comprise more than one molecular entities. Therefore, it is possible that both the 2444 and 2447 m/z peaks share an overlap in their molecular constituents; however, they include other peptides or proteins which are unique in this specific peak cluster based on the sensitivity and specificity distribution. The frequency distribution analysis of those 8 peaks shows that the m/z 9618 cluster from the CM10 chip screen displays the best stratifier for pancreatic cancer (Fig. 1D), with a sensitivity of 61% and a specificity of 82% in upper GI cancer cases. Comparison of all 8 peaks in non-cancer control cases measured in our previous study (17) also demonstrated that the m/z 9618 peak cluster exhibits the same frequency distribution in both non-cancer and non-pancreatic cancer cases of ~30%, whereas this frequency is doubled in pancreatic cancer for this peak cluster. The best predictor, based on a low frequency in non-cancer cases, was the m/z 2577 cluster in pancreatic cancer (Fig. 1B), with a 79% frequency in this cancer type, a 62% frequency in other cancers, and a 10% occurrence in non-cancer patient urine, with a sensitivity of 86% and a specificity of 44%. The elevated amount of the observed m/z 4141 peak cluster associated with oesophageal cancer (Fig. 1H) demonstrates the best specificity of 100%, but exhibits a relatively poor sensitivity of only 19% due to the relatively low frequency of one in four samples where this peak can be measured. Low frequency values were also observed for the OGJ cancer-associated elevated peak cluster at m/z 4908 (Fig. 1F), with good sensitivity and specificity values of 92% and 66%, respectively. A more pronounced potential OGJ marker in terms of quantitative changes was measured at m/z 5511 (Fig. 1E) with a high specificity (94%), but a relatively low sensitivity (38%).

We then endeavoured to identify the molecular identity of the m/z 2447 peak, since it showed the highest sensitivity of 100% in the detection of pancreatic cancer. This was performed by selecting 8 samples, of which 4 either contained the cluster peak at m/z 2447, or did not show the peak according to the SELDI-TOF scans. The samples were individually batch-absorbed on CM resin; bands in the 2-4 kDa range were excised after peptide gel electrophoresis and processed as described previously (17). The molecular cluster of interest was shown to consist of fragments from CTSB, α-1-antichymotrypsin precursor and immunoglobulin γ and κ chains (Table III). The same result was obtained using an independent approach, as described below.

We then analysed the molecular constituents of human cancer patient urine in the 2-10 kDa molecular weight range using chromatographic protein and peptide enrichment on CM10 and IMAC30 resins, followed by gel separation, trypsin digestion and LC-MS/MS fragmentation, as before. Samples were selected based on the SELDI-TOF analysis results to either contain the 8 m/z peaks of interest (50% of samples per subset) or not in subsets of groups of 16 samples each. This resulted in the analysis of 145 urine samples, including repeats, from 62 upper GI cancer patients, of which 42 unique patient urines were enriched on CM10, and 40 unique patient samples on IMAC30 chromatography resins. 950 non-redundant proteins were identified in the CM10 resin-based approach by Mascot searching, and 600 unique proteins could be observed in the IMAC30-based enrichment, totalling 1,228 unique and non-redundant proteins in the combined datasets (Table SIV). Protein expression pattern matching (i.e., molecules identified by LC-MS/MS and Mascot searching, and matching to peak expression patterns observed by SELDI-MS) was performed by using an automated computer program written in-house based on the filtered Mascot dataset, where each individual identification was based on Mascot scores >16 and consisting of at least 2 peptides each. All proteins were identified in at least three independent samples. Table III lists all found by this approach, including a list of publications describing the relevance of these molecules in tumour growth. A fully detailed list of these molecules, including peptide sequences, is supplied in Table SV.

The m/z 2444 peak observed in pancreatic cancer using the IMAC30 chromatography resin, as well as the pancreatic cancer-associated m/z 2447 peak in the CM10 dataset, matched the expression pattern of α-1-antichymotrypsin (SERPINA3 or AACT). The same molecule, although a different proteolytic fragment of m/z 4639, could also be a potential biomarker for gastric cancer. A fragment of CSTB matched the observed pancreatic marker of m/z 2577 in the IMAC30 resin type-derived dataset, and other potential pancreatic markers of m/z 2477 and 9618 matched the presence of peptides from immunoglobulins. The m/z 2477 cluster distribution pattern also coincides with the measured expression pattern of CTSB.

Potential biomarkers for OGJ tumours include programmed cell death 6-interacting protein (PDCD6IP), matching the observed pattern of m/z 5511 in the IMAC30 resin dataset, vitelline membrane outer layer protein 1 homolog (VMO1) and triosephosphate isomerase (TPI1), which matched the m/z 4908 peak pattern in SELDI-TOF-MS measurements. The m/z 4141 clustered peak in the CM10 enriched dataset, which may be indicative of oesophageal cancer, matched the LC-MS/MS observed expression pattern of several molecules, namely complement C4-A (C4A), prostatic acid phosphatase (ACPP), azurocidin (AZU1) and histone H1.2, H1.3 or H1.4.

Both CTSB and CSTB were also cross-validated by western blot analysis (Fig. 2), and both molecules exhibited a very good correlation to the expected cluster-peak pattern observed by SELDI analysis, thereby validating both the predictor model and the Mascot identification, as well as the SELDI peak clustering of these molecules, and may represent viable prognostic biomarkers for pancreatic cancer.

Over the last decade, the emergence of high-throughput screening platforms opened the possibility to mechanistically characterize at the molecular level disease phenotypes, disease subtypes, assess disease progression and monitor response to therapy. However, despite a crescendo in the availability of these large-scale multi-omics data for numerous conditions, no apparent or adequate effort has been made to curate and integrate those data resources.

Therefore, large-scale multi-omics data handling differential expression of several molecular entities, for example, microRNA, gene, protein and metabolite across multiple tissues and biological fluids in several cancer types (with special emphasis on GI neoplasms) were collected from the literature and general scope databases, and subjected to extensive manual curation and annotation. This effort yielded a cancer profiling database: the Multi-Omics Cancer database MoCadb (www.padb.org/mocadb), a database module of the Pan-Omics Analysis Database (PADB) (www.PADB.org) which ensures a long-term lifecycle of the created repository, along with providing the appropriate framework that encloses curated resources, in order to assist in the development of integrative -omics disease models, straightforward translation of findings from experimental animal models using established ortholog maps, creation of cellular regulatory networks based on the association of miRNAs and transcription factors to targets, protein-protein and gene-gene interactions, enzymatic reactions, delineation of pathways and drug interactions.

MoCadb aims to incorporate existing molecular information and clinical metadata, ultimately holding the potential to unravel and allow an in-depth understanding of the regulation of key molecules modulating pathophysiology and progression in several cancer types.