Specialized DNA Arrays for the Differentiation of Pancreatic Tumors

Materials and Methods

Tissue and biopsy samples. For cytologic analysis of FNAB samples, aspirated material was expelled onto microscope slides and smeared. The slides were air-dried, stained with May-Grunwald-Giemsa stain, and evaluated by an expert cytologist. Samples were classified as “malignant” if cell aggregates showing clear signs of malignancy were detected on at least two different slides from the same biopsy, “benign” if evaluable material without any signs of malignancy were detected, or “nondiagnostic” if the results were inconclusive (e.g., the cellularity of the sample was too low).

Surgically resected pancreatic adenocarcinoma and chronic pancreatitis tissues were provided by the surgery departments at the Universities of Ulm and Homburg/Saar. Normal pancreas samples were obtained from healthy areas at the borders of chronic pancreatitis resectates. Informed consent was obtained from all patients prior to using tissue or biopsy samples. The study was approved by the local ethics committees at the Universities of Ulm (Germany), Homburg/Saar (Germany), and Verona (Italy).

RNA isolation and linear amplification. Snap-frozen surgical samples were ground on dry ice with a mortar and pestle, suspended in RLT buffer and total RNA were isolated using the RNEasy Mini Kit (Qiagen, Hilden, Germany). Fine needle biopsy in the routine diagnostic workup of pancreatic tumors was done with transabdominal ultrasound or endoscopic ultrasound guidance. FNAB sample material was recovered by flushing the needle and syringe with RLT buffer after material for cytologic analysis had been removed. Total RNA was then isolated using the RNEasy Mini Kit (Qiagen). In both cases, the total RNA was finally dissolved in water and quality-checked on a BioAnalyzer Lab-on-a-Chip system (Agilent, Waldbronn, Germany). In order to obtain sufficient material for hybridization, the complete FNAB RNA samples were subjected to one round of T7 RNA polymerase-based linear amplification using the MessageAmp Kit (Ambion, Huntingdon, Great Britain). To avoid data bias, all surgical samples were treated likewise by linearly amplifying 0.5 μg of total RNA prior to hybridization.

Array production and hybridization. cDNA fragments were PCR-amplified using vector primers and contact-printed in duplicate on nylon membranes (Nytran N+, Schleicher and Schuell, Germany). For radioactive labeling, the complete amplified RNA samples were labeled with ³³P-dATP using the StripEZ-RT Kit (Ambion) and hybridized overnight to nylon membrane arrays in ULTRArray hybridization buffer (Ambion) at 50°C. Radioactive signals were detected using a STORM phosphorimaging system (Amersham Biosciences, Feiburg, Germany) and quantified with the ArrayVision software (InterFocus, Haverhill, Great Britain). Signal intensities were normalized to the mean signal intensity of all features on an individual array.

Construction of the classifier. All equations used are listed in Panel 1. Details of the analysis and complete data sets are available as part of the Supplementary Data.⁸ The nylon array data set was filtered to include only featfures (genes) for which normalized intensities exceeded a value of 0.8 in at least 10 samples to remove uniformly low (and thus uninformative) signals. Control spots were excluded as well. The remaining 169 features were used in the construction of a linear classifier based on the analysis of the 42-sample training set (Table 1).

During the first step of the analysis, a principal component analysis (21) was done on the training data set. The first 30 principal components of the training data set, which represented 99.9% of the total variation within the data, were then used for a linear discriminant analysis to search for combinations of principal components facilitating complete separation of the tumor from the control tissue samples in the training set. All possible combinations of up to 7 out of the 30 principal components were tested for their performance in the linear separation of the diagnostic classes. Evaluation of the feature set combinations was done by measuring the area under the receiver operating characteristic curve (22–24). All combinations producing an area under the receiver operating characteristic curve of ≥0.95 were subjected to a stochastic search to add additional discriminative principal components until perfect separation of the diagnostic classes was achieved. Out of all combinations producing perfect linear separation, we selected the set that resulted in the greatest margin between tumor and control samples when plotting the samples according to their relative distances to the separating hyperplane. The resulting linear classifier was then evaluated using the independent 20-sample test set (Table 2).

Results and Discussion

In order to develop the pancreatic cancer diagnostic cDNA array, we extensively analyzed the results of various studies on differential gene expression in pancreatic cancer done in our own group (25, 26) as well as information obtained from the SAGE (http://www.ncbi.nlm.nih.gov/SAGE/) and Digital Differential Display (http://www.ncbi.nlm.nih.gov/UniGene/info_ddd.shtml) gene expression databases and reports from the literature to identify genes with the potential to differentiate between malignant and nonmalignant tumors of the pancreas. In order to allow for robust normalization of the hybridization results, we have designed the array to comprise a sufficiently high total number of features (558), including balanced numbers of up- and down-regulated genes. In addition, important genes were represented by multiple cDNA clones and control spots of mixed cDNA clones were included to facilitate grid alignment (for a complete list of features, see Supplementary Information).

In the present study, 16 FNAB samples of pancreatic adenocarcinoma and benign pancreatic tumors for which clinical patient follow up with a definitive diagnosis was available were analyzed both by conventional cytology and diagnostic array hybridization. Cytologic analysis of the 16 FNAB samples correctly identified a malignant process in 9 out of 10 adenocarcinoma cases (90%) and a benign process in 3 out of 6 chronic pancreatitis and pseudocyst cases (50%). The remaining adenocarcinoma case as well as the three benign cases were nondiagnostic due to the absence of evaluable intact cells. The resulting overall diagnostic accuracy of 75% is well in agreement with the numbers reported in the literature (1–4).

Residual material from the same biopsy samples which were used for cytologic analyses were subjected to expression profiling analyses using the diagnostic arrays. In order to ensure an adequate representation of different tumor stages in the data sets used for the development and evaluation of the classification procedure, we analyzed an additional 27 samples of histopathologically well-defined surgically resected ductal adenocarcinomas as well as 19 surgically resected control samples of chronic pancreatitis or normal pancreas. The samples were arbitrarily divided into a 42-sample training set (Table 1) and a 20-sample test set (Table 2), such that both sets contained equal proportions of malignant and benign samples as well as FNAB's. The training set was subsequently used to develop the classification system for the distinction between malignant and benign samples (see below), which was independently evaluated using the test sample set. The complete process is schematically outlined in Fig. 2 .

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Flowchart outlining the process of construction and evaluation of the linear classifier using the independent training and test sample sets.

Even though the number of genes featured on the diagnostic array was low compared with large-scale arrays, it still far exceeded the number of tissue and biopsy samples available for training of the classifier in the training set. It was therefore of paramount importance to first reduce the number of features used for classification and thus the dimensionality of the data set in order to avoid overadaption of the classifier to this specific set of data (27). Instead of omitting individual genes from the analysis to achieve this purpose, we opted to apply principal component analysis (21) to the data, resulting in a reduced set of combined features (principal components) representing weighted combinations of all genes in the data set. Tissue or biopsy samples can be mapped to the principal components (or a subset thereof), effectively creating a coordinate system of uncorrelated variables which replace the high dimensional space that individual gene expression values fall into. Principal component analyses thus serves to greatly reduce the dimensionality of the data whereas preserving its general structure, resulting in reduced sensitivity to outliers or hybridization artifacts in individual diagnostic samples.

For the construction of the classifier, we opted to perform linear discriminant analysis using the first 30 principal components of the training data set, which represented 99.9% of the total variation within the data. Because linear discriminant analysis assumes a relatively simple model of sample distribution, it is far less prone to overadaption to a specific data set than nonlinear methods, again increasing the robustness of the classification procedure. We identified a total of 429,917 different combinations of principal components producing perfect linear separation of tumor and control samples in the training data set. Out of these, a set of 23 principal components which provided the maximum margin between tumor and control samples was selected to define the linear classifier (Fig. 3A ; see also Supplementary Information).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Separation of the adenocarcinoma (red circles) and control (blue circles) samples in the training and test data sets using the selected combination of 23 principal components. The samples are plotted according to their relative distances to the separating hyperplane. Dotted lines, cutoff calculated by linear discriminant analysis. FNAB samples are identified by triangles next to the corresponding circles. A, perfect linear separation of the diagnostic classes (including surgically resected tissue as well as FNAB samples) in the training set. B, 95% diagnostic accuracy using the linear classifier on the independent test set. One surgically resected chronic pancreatitis sample is misclassified; all FNAB samples are correctly classified.

The predictive performance of the linear classifier was then evaluated by assessing its performance on the independent 20-sample test set. Our system correctly classified 19 out of the total 20 test samples, resulting in an overall diagnostic accuracy of 95% (Fig. 3B). Only one surgically resected chronic pancreatitis sample (“Infl/Norm_15”; Table 2) was misclassified, producing one false-positive result. Of note, all FNAB samples were correctly classified regardless of the outcome of the cytologic analysis. Moreover, all FNAB samples were placed well clear of the separating cutoff by the classifier (Fig. 3B).

We were thus able to show that expression profiling analyses of surgically resected tumor specimens and FNAB samples using specialized diagnostic arrays with limited numbers of highly selected genes produce reliable, reproducible, and informative results. Supplementing conventional cytologic analyses of aspiration biopsies with diagnostic array profiling thus promises to significantly improve the accuracy of preoperative diagnostics of suspect masses in the pancreas. The large number of different combinations of principal components yielding perfect linear separation of the diagnostic classes in the training set, as well as the convincing performance of the classifier on the independent test set, serve to show both the expedience of the diagnostic gene collection as well as the validity of the analytic approach. Our results confirm conclusions drawn from earlier expression profiling studies using large-scale arrays, which have shown that the number of informative genes for the classification of different types and subtypes of cancer is usually <100 (28–30), suggesting that dedicated diagnostic arrays should perform as well as whole-genome arrays in defined diagnostic settings.

Due to the use of residual material from biopsy needles for the analysis of the FNAB samples, the amount of starting material available for expression profiling analysis was extremely limited, so that we initially produced the arrays in the nylon membrane format to take advantage of the superior sensitivity of radioactive labeling and detection. Parallel hybridizations of a subset of samples to diagnostic arrays produced in the glass microarray format (see Supplementary Material), however, showed that the concept and design of the diagnostic array can readily be transferred to the glass microarray/fluorescent labeling platform as well, which may be better suited for routine clinical settings.

In the present study, we have focused on the distinction between pancreatic ductal adenocarcinoma and nonmalignant diseases of the pancreas, because pancreatic ductal adenocarcinoma is by far the most frequent malignant tumor arising in the pancreas and thus poses the clinically most relevant diagnostic problem (20). We are currently in the process of analyzing additional tumor entities, such as acinar and neuroendocrine tumors, using both the diagnostic array as well as large-scale arrays, in order to develop a multiclass classification system for the comprehensive diagnosis of different malignancies in the pancreas. In addition, we expect further development of the array in combination with careful analysis of clinical patient data to result in the recognition of distinct prognostic gene expression signatures predicting important clinical variables such as stage of disease, response to therapy, or prognosis, thus setting the stage for therapeutic regimens custom tailored to the individual patient.

Appendix. Panel 1 (Equations used in the report)

Given is a set of n d-dimensional samples x₁, …, x_n, n₁ in the subset D₁ and n₂ in the subset D₂.

The sample mean μ̂_i is estimated by:

The scatter matrices S_i and S_w are:

The projection vector w for Fisher's linear discriminant is given by:

The area under the receiver operating characteristic curve is estimated by:where n₂ is the number of pancreatic carcinoma patients, and Ξ_j (j = 1, …, n₂) are the ranks of these cases obtained by ranking all n = n₁ + n₂ values of w^Tx.

The margin m for  = 1 is calculated by: where i = 1, …, n₁, j = 1, …, n₂, and the threshold is centered in the margin.