Protein and lipid MALDI profiles classify breast cancers according to the intrinsic subtype

Collecting and processing clinical material and protein MALDI MS analysis

Thirty-four pairs of breast cancer and adjacent normal tissue samples were collected at the time of surgery, with informed consent and institutional review board approval, from breast cancer patients undergoing surgery at National Cancer Center in Korea from 2001 to 2010, and stored in liquid nitrogen until analysis. Eleven samples were excluded by spectra quality filter of ClinProTools (version 2.2, Bruker Daltonics). Additional 21 samples were excluded because of inadequate (< 50%) tumor content. These excluded samples were not different from samples analyzed in this study in patient age or intrinsic subtype. Breast cancer intrinsic subtypes were classified according to an immunohistochemistry (IHC) surrogate panel [20, 21] as follows: luminal [estrogen receptor (ER) positive and/or progesterone receptor (PR) positive], HER2+ [HER2+ regardless of hormone receptor status], and triple-negative [ER-, PR-, and HER2-]. A cut-off value of 1% or more of positively stained nuclei was used to define ER and PR positivity. HER2 was scored as 0-3+ according to the method recommended for the HercepTest (Dako, Glostrup, Denmark). The cases with IHC scores of 3+ or ERBB2 gene amplification by fluorescence in situ hybridization (FISH) were considered positive for HER2.

Thin (10 μm) sections were obtained from the frozen tissues using a cryostat (Leica CM 3050S, Leica Microsystems Inc., Bannockburn, IL). Multiple (three to seven) serial sections were obtained from each tissue. One section was affixed to a standard glass slide, and then stained with hematoxylin and eosin (H&E). The other sections were thaw-mounted onto an indium tin oxide (ITO) slide (HST Inc., Newark, NJ), desiccated in vacuum for 20 min, and washed with graded ethanol solutions (70%, 90%, and 95% ethanol for 30 sec each) for subsequent protein MALDI MS analysis. Sinapinic acid (SA) was used as the protein MALDI matrix and prepared as a 20 mg/ml solution in 50:50 acetonitrile: 0.1% trifluoroacetic acid (TFA). Using a micropipette, 250 nl of the matrix was manually deposited twice per spot. The H&E-stained serial section was then evaluated by a pathologist, who confirmed that entire H&E section of tumor samples had at least 50% tumor content. Mass spectra were acquired using an Autoflex III (Bruker Daltonics, Bremen, Germany) MALDI TOF equipped with a SmartBeam laser (Nd:YAG, 355 nm) and run using a linear-mode acquisition method optimized for 2-30 kDa, a laser frequency of 200 Hz, and a delay time of 7 ns. A total of 400 spectra were acquired at each spot position.

Lipid MALDI MS analysis

For lipid MALDI MS analysis, the binary matrix solution was prepared by dissolving 7 mg each of 2, 5-dihydroxybenzoic acid (DHB) and α-cyano-4-hydroxycinnamic acid (CHCA) in 1 mL of 70% methanol plus 0.1% TFA and 1% piperidine [19]. Under the guidance of the same H&E-stained adjacent section, the binary matrix (250 nl × 2) was manually deposited on the tumor-rich area of cryosections that were thaw-mounted onto ITO slides. Mass spectra were acquired both in positive- and negative-ion reflector modes, using an Autoflex III (Bruker Daltonics) TOF equipped with a SmartBeam laser and run using a linear-mode acquisition method optimized for 500-1,200 Da, a laser frequency of 200 Hz (3,000 consecutive laser shots), and a delay time of 0 ns. Before each data acquisition, an external calibration was conducted using lipid-mixed calibration standards with m/z ranges of 674-834 Da (positive ion mode) and 564-906 Da (negative ion mode).

MALDI LIFT (MS/MS) analysis was directly performed on the tissue section after MALDI MS. LIFT data and some lipid databases http://lipidsearch.jp or http://www.lipidmaps.org were used to facilitate and confirm the assignment of phospholipid species.

Data Processing and Statistical Analysis

ClinProTools was used for baseline subtraction, spectral recalibration, and spectral area calculation. A resolution of 300 was applied to the peak detection method. The Top Hat baseline with 10% minimal baseline width was used for baseline subtraction. Data reduction was performed at a factor of four. Spectra were recalibrated with a maximal peak shift of 2,000 ppm between reference and peak masses. The value of the "% Match to Calibrant Peaks" parameter was set to 20%. Spectra that were not recalibratable were excluded. All data with signal-to-noise ratios > 5 were acquired, and the peak area was used for peak calculation with zero level integration. An average peak list was set up for each tissue sample by choosing peaks on the calculated total average spectrum for each tissue sample to create one average spectrum per patient. Peaks with the highest intensity in each ion mode (m/z = 616.23 and 539.53 in positive and negative ion modes respectively) were identified as non-lipid molecules and therefore excluded from subsequent analysis. Average-normalized datasets (i.e., protein, positive mode-lipid, and negative-mode lipid datasets) were then combined into a single dataset and subjected to statistical analysis using BRB-ArrayTools (NCI, version 3.8) [22]. A principal component analysis (PCA) plot was generated using multi-dimensional scaling analysis of BRB-ArrayTools, which graphically represents correlation coefficients among samples without forcing the samples into specific clusters. The three primary principal components were used as the axes for the 3-dimensional scaling representation. Class comparison and class prediction analyses were also performed using BRB-ArrayTools. The class comparison analysis computed the F-test for each peak, and listed peaks differentially expressed among the classes at selected statistical significance level. Then, it performed random permutations of the class labels. For each random permutation, all F-tests were re-computed for each peak. The class comparison tool computed the proportion of the random permutations that gave as many peaks significant at the selected level of significance as were found by comparing the true class labels. Protein and lipid profiles of the classes were considered different if this probability (designated as Permutation P value) was calculated < 0.05.

To evaluate whether classes have different protein and lipid profiles, class prediction analyses were performed using all samples as a training set. The 0.632+ bootstrap cross-validated misclassification rates was computed for all classifier functions (compound covariate predictor, diagonal linear discriminant analysis, 1- and 3-nearest neighbors, nearest centroid, and support vector machine) in the training set. Then, class labels were randomly shuffled and the cross-validated misclassification rate was computed for each random dataset. Permutation P value, which is defined as the proportion of random datasets that give as small misclassification rate as is obtained with real class labels, was then calculated. MALDI MS profiles of the classes were considered different if this permutation P value was < 0.05.

To estimate the predictive power of discriminatory peaks, the class prediction analyses were also performed by randomly dividing the whole sample into two (training and test) subsets at 1-to-1 ratio. Randomization was performed using nQuery Advisor software (version 7.0, Statistical Solutions, Saugus, MA).

Cancer versus adjacent normal breast tissue

A principal component analysis (PCA) plot graphically demonstrated that cancer and adjacent normal samples were separately clustered in an unsupervised analysis (Figure 2A). When a class comparison analysis was performed using BRB-ArrayTools, the proportion of the random permutations that gave as many significant peaks at a feature selection of P < 10^-5 as were found by comparing the true class labels (cancer vs normal) was < 0.001, strongly suggesting that the cancer and normal tissue samples have significantly different protein and lipid profiles (Table 2). Class prediction analysis was subsequently performed after randomly dividing the entire set of samples into two groups at 1-to-1 ratio. At a feature selection P < 0.001, the median class prediction accuracy in test sets was 94.1% for all classifiers tested, in 100 random training-to-test partitions. Thus, breast cancer and adjacent normal tissues have clearly different protein and lipid profiles.

To evaluate whether patients' age affects cancer-associated protein and lipid alteration, the cancer/normal peak area ratio was compared between older (≥50 years, n = 10) and younger (< 50 years, n = 24) age groups at feature selection P < 0.05, using class comparison algorithm of BRB-ArrayTools. The probability of getting as many significant peaks by chance as was obtained with real class labels, if there are no real differences between class, was 0.78, suggesting that cancer-associated protein and lipid alteration was not significantly different between older and younger age groups.

Classification according to intrinsic subtype

We then planned to estimate the predictive power of discriminatory proteins and lipids in training-to-test partitions, although this analysis is limited by the sample size. At a feature selection P < 0.05, the median prediction accuracy in test sets ranged from 80% to 81.3% (80%, 81.3%, 80%, and 80%, for diagonal linear discriminant analysis, 1-nearest neighbor, 3-nearest neighbors, and nearest centroid, respectively), in 100 random training-to-test partitions. Discriminatory peaks are listed in Table 3.

Metaplastic carcinoma vs adenocarcinoma

Metaplastic carcinoma is a unique breast cancer subtype, characterized by distinct morphologic feature and poor prognosis [23]. To further evaluate whether protein and lipid profiles could distinguish biological feature, we compared metaplastic carcinomas (n = 4) with adenocarcinomas (n = 30) for protein and lipid profiles. According to PCA analysis, metaplastic carcinomas were clearly separated from adenocarcinomas (Figure 2C). For class prediction analysis, all 34 cancer samples were used as a training set, given the small number of metaplastic carcinomas. At a feature selection P < 0.001, permutation P values for 0.632+ bootstrap cross-validated misclassification rates were < 0.05 for all classifiers tested (0.01, < 0.01, < 0.01, < 0.01, 0.01, and < 0.01, for the compound covariate predictor, diagonal linear discriminant analysis, 1-nearest neighbor, 3-nearest neighbors, nearest centroid, and support vector machine, respectively). Although sample size is small, metaplastic carcinomas appear to have unique protein and lipid profiles.