Mass spectrometry protein expression profiles in colorectal cancer tissue associated with clinico-pathological features of disease

Tissue samples were collected from a total of 36 patients with confirmed CRC at the time of surgical resection at Colchester General Hospital, Essex UK. All specimens were obtained following informed consent in accordance with local UK NHS Ethics Committee approval (protocol reference: MH 528). Surgically excised specimens were washed extensively in ice-cold 150 mM NaCl and samples of normal colonic mucosa (>10 cm from tumour margin) and tumour tissues were excised using a scalpel and then snap frozen and transferred to a - 80°C freezer. The total time from surgical resection to snap freezing of specimens was <30 mins.

Frozen tissue samples (approximately 250 mg) were ground using a mortar and pestle and then lysed for 30 mins at 4°C in 1.0 ml of 10 mM Tris-HCl pH 7.5, 200 mM NaCl containing Protease inhibitor cocktail (Roche Pharmaceuticals) and 1% N-octyl-β-D-glucopyranoside (Sigma Aldrich). The cell lysate was then centrifuged at 12,000 × g for 30 mins and the supernatant representing the solubilised fraction was removed. Protein was further purified by reversed phase hydrophobic interaction chromatography using a commercially available super-paramagnetic microparticle kit (MB-HIC-C8, Bruker Daltonics). Briefly, 10 μl of 30-35 mg/ml protein solution was adsorbed to 10 μl of beads after addition of 20 μl kit binding buffer. After three washes with 200 μl 0.1% trifluoroacetic acid, protein was eluted in 20 μl of 50% (v/v) acetonitrile (Fisher Scientific) Eluted protein was stored at 4°C for no more than 1 hr prior to matrix co-crystallisation.

To facilitate reproducible co-crystallisation of protein with matrix solution, a modification of the slow crystallisation method [23] was used. Briefly, 20 ul of purified protein was mixed with 20 μl of acetonitrile containing 0.1% trifluoroacetic acid, saturated with sinapic acid (Sigma Aldrich). A 20 μl aqueous solution containing diammonium citrate (200 mM) and nitrotetracetic acid (0.1%) was added and crystal formation was allowed to proceed for 2-3 hrs. Crystallised matrix-protein samples were spotted onto a stainless steel MALDI target plate and spectra were acquired using a MALDI-TOF mass spectrometer (Reflex IV; Bruker Daltonics) with the following instrument settings: ion source 1, 20 kV; ion source 2, 16.65 kV; lens voltage, 9.5 kV; pulsed ion extraction, 200 ns. Ionisation was achieved by irradiation with a nitrogen laser (e = 337 nm) operating at 25 Hz and 20% laser power. For matrix suppression, we used a high gating factor with signal suppression up to 1500 Da. Mass spectra were detected in linear positive mode. Detector gain was set at 1600 V, sample rate at 1.0 and electronic gain at 100 mV with real-time smoothing. Spectra were acquired in duplicate from 500 laser shots delivered as 5 × 100 pulses and were internally calibrated using 'FlexAnalysis' spectral processing software (Version 2.0; Bruker Daltonics) with reference marker peaks at 2426.9Da, 6109.5 Da and 12471.6 Da. External calibration used the following reference standards: bombesin (1620.86 Da), somatostatin (3149.57 Da), insulin (5734.51 Da), ubiquitin I (8565.76 Da), cytochrome c (12,360.97 Da) and myoglobin (16,952.30 Da).

Calibrated spectra were exported as ASCII files and were digitally processed by smoothing, de-noising, baseline subtraction and normalisation (by total ion current) using the 'SpecAlign' suite of spectral computational tools [24, 25]. Validation of the reproducibility of the resulting mass spectrometry profiles and elimination of 'outliers' was accomplished as described elsewhere [26]. Duplicate spectra with a cross-correlation function of < 0.950 were discarded. From the initial cohort of specimens, representing matched tumour and adjacent normal mucosa from 36 patients, a total of 64 spectra representing 31 tumours and 33 normal mucosa were obtained (see Table 1). Of the 5 tumour and 3 mucosa specimens that were excluded from analysis, 2 tumour and one mucosa failed to yield reproducible spectra on repeated protein preparations. The remaining 3 tumour and 2 mucosa specimens consistently gave spectra of poor quality (outliers), presumably as a result of specimen deterioration. Matching peaks were aligned across spectra by using the combined Fast Fourier Transform/Peak matching method [25] and modelled peak areas for the entire set of spectra were exported as a single csv file.

Subsequent spectral analysis was implemented in the 'GenePattern' suite of software tools (Broad Institute, MIT, USA) [27]. Hierarchical clustering used Euclidean correlation as the column distance measure with pair-wise average linkage as the clustering method. Comparative Gene Marker Selection [28, 29] with either a t-test or a signal-to-noise ratio (SNR) test statistic was used to identify and rank differentially expressed marker peaks and to assign Bonferroni-corrected P and false discovery rate (FDR) values [28‐30]. The k-nearest neighbours (kNN) algorithm [29] was used to build a classification model for tumour vs normal using separate training and test datasets. For this purpose, two thirds of the spectra, comprised of a representative proportion of tumour and normal spectra, were randomly assigned to a training dataset, with the remaining third being used as an independent test dataset. Spectra were randomly assigned using the GenePattern 'SplitDatasetTrainTest' module [27]. Alternatively the kNN algorithm was used in an iterative, 'leave-one-out' cross-validation mode. Other statistical analysis used the SPSS software.

Table 1 summarises the clinico-pathological data for the 36 CRC patients from whom specimens were obtained. In most cases, spectra of adequate quality from matching pairs of tumour and adjacent normal mucosa were obtained. However, some tissue protein preparations consistently yielded spectra of poor quality or that were poorly reproducible (see Methods section); these were excluded from the analysis. The resulting 64 spectra, representing 31 tumour and 33 normal mucosa specimens, generated a total of 265 protein peaks in the mass range 1800-16000Da. Illustrative examples of raw MALDI-TOF spectral profiles are shown in additional file 1. Although the overall intensity profile of individual protein peaks was very heterogeneous across different specimens, unsupervised hierarchical cluster analysis classified most tumour and normal mucosa into distinct cluster groups (Figure 1) consistent with major differences in the tumour verses normal protein expression profiles.

Bild vergrößern

To quantitatively evaluate the differences between the protein expression profiles of tumour verses normal tissue, the Comparative Gene Marker Selection algorithm [28] was applied to the spectral data-set to determine the level of significance of difference between tumour and normal for each protein peak. Figure 2 shows the frequency distribution (occurrences) of protein peak P values (Feature P) that were binned in increments of 0.05. Above P = 0.05, the representation of protein peaks was fairly evenly distributed. However, nearly 100 peaks gave a P value < 0.05, indicating that a sizable fraction of proteins detected by MALDI-TOF mass spectrometry discriminate between tumour and normal colonic tissue. Applying a threshold of P ≤ 0.01, FDR ≤ 0.05, the expression profile of a total of 73 protein peaks was significantly different between tumour and normal tissue with 57 being up-regulated in normal tissue and 16 being up-regulated in tumour tissue. Figure 3 shows a heat-map profile of these 'marker peaks' and additional file 2 summarises their statistical features.

Bild vergrößern

Bild vergrößern

To rigorously demonstrate that tumour and normal mucosa tissue could be distinguished using their protein spectral profiles, the 64 spectra were randomly split into separate training and test datasets. The training spectra dataset was used to optimise a kNN algorithm [29] for predicting tumour or normal status. As summarised in additional file 3, the model correctly predicted the status of specimens in the independent test spectra dataset with 100% sensitivity and specificity (95% confidence interval: 67.9-99.2%).

To determine whether the protein expression profiles of tumour tissue could be used to predict individual clinico-pathological characteristics of patients (Table 1), the kNN algorithm was used to optimise a series of classification models. Since the limited numbers of datasets precluded analysis by using independent train and test spectra, the kNN algorithm was used in an iterative, 'leave-one-out' cross-validation mode. Table 2 summarises the results of this analysis. The predictive model for distinguishing poorly differentiated from well/moderately differentiated tumours gave a receiver-operator characteristics (ROC) error of 0.171, correctly classifying 10/13 poorly differentiated and 16/18 well/moderately differentiated tumours (P = < 0.001). Additional file 4 summarises the kNN algorithm results and Figure 4A shows the expression profiles of the top two ranked discriminating peaks. The kNN model for disease recurrence also gave a low ROC error (0.105 - see Table 2A). As summarised in additional file 5, the model correctly predicted 5/6 patients with recurrent disease and 22/23 who are disease-free (P = < 0.001). Figure 4A shows the expression profiles of the top two ranked marker peaks for classifying disease outcome.

Bild vergrößern

In a similar analysis of normal mucosa spectra (Table 3), only the characteristic of lymph node involvement gave a low ROC error (0.212). As shown in Table 3 and in additional file 6, the kNN algorithm correctly predicted 11/14 patients with, and 15/19 patients without lymph node involvement (P = 0.001). Figure 4B shows the expression profiles of the top two ranked marker peaks for classifying the characteristic of lymph node involvement.