Principal component analysis (PCA)[14,15] represents the objects in a new reference system characterized by new variables called principal components. The first principal component accounts for the maximum variance contained in the original dataset, while subsequent components account for the maximum residual variance. They are calculated hierarchically, so that systematic variations (i.e., information) are explained in the first components while experimental noise and random variations are contained in the last ones. The components are linear combinations of the original variables and are orthogonal to each other, thus containing independent sources of information. They are often used for dimensionality reduction by considering a smaller number of significant components containing only relevant information.

The graphical representation of the scores (the co-ordinates of the samples in the new reference system) in the space of the principal components allows the identification of groups of samples showing a similar behavior (samples close one to the other in the graph) or different characteristics (samples far from each other). The corresponding loading plot (representing the loadings, i.e., the coefficients of the linear combination describing each principal component) identifies the variables that are responsible for the analogies or the differences detected for the samples in the score plot.

Cluster analysis techniques[14-16] allow the identification of groups of samples or of variables in a dataset by investigating the relationships between the samples or the descriptors. Agglomerative hierarchical methods[14,15] are the most widespread, grouping the samples on the basis of their similarity. The most similar samples or groups are linked first. The final result is a graph, called a dendrogram, where the samples are represented on the X axis and are connected at decreasing levels of similarity along the Y axis. The groups can be identified by applying a horizontal cut of the dendrogram and identifying the number of vertical lines crossed by the horizontal cut. Figure 1 reports a dendrogram where cutting at level 4000 produces only two clusters (the 2 different tumor cell lines), while cutting at level 1500 produces 4 clusters (PACA and T2M4 cell lines, treated and untreated with trichostatin A). The results of hierarchical clustering strongly depend on the measure of similarity and on the linking method adopted. Clustering techniques can be applied to the original variables or to the relevant principal components[16].

Open in New Tab   Full Size Figure   Download Figure

Figure 1 Example of a dendrogram built using Ward’s linkage method and euclidean distances. Data refer to samples from two pancreatic cancer cell lines treated or not with trichostatin A.

Multidimensional scaling (MDS)[17,18] is aimed at dimensionality reduction and graphical representation of the data. Given a set of n objects and a measure of their similarity, MDS searches for a low dimensional space in which the objects are represented by points in the space so that the distances between the points match as much as possible with their original similarities[17]. There are several different approaches to MDS depending on the measure of the similarities matching, on the metrics, on the method used to compute the similarities and on the way the samples configuration is obtained[19,20]. Shepard[21,22] and Kruskal[23] provided an extension of classical MDS to the study of nonparametric similarities.

The soft independent model of class analogy (SIMCA) method[24-27] is based on the independent modelling of each class by means of PCA. Each class is described by its relevant principal components. The samples belonging to each class are contained in the so-called SIMCA boxes, defined by the relevant components of each class. The classification of each sample with SIMCA is not affected by experimental uncertainty and random variations since each class is modelled only by its relevant components. This method is also useful when more variables than objects are available since it performs a substantial dimensionality reduction.

The identification of the candidate biomarkers by SIMCA can be accomplished by the analysis of the discrimination power (DP), a measure of the ability of each variable to discriminate between two classes at a time. The greater the DP, the more a variable weighs on the classification of an object in one of the two classes compared.

Ranking PCA is a ranking method proposed by Marengo et al[7], Robotti et al[28] and successively applied by Polati et al[29], based on the description of the original data by means of principal components. The use of PCA in the field of identification of biomarkers in the omics sciences is particularly effective since it allows the relationships between the variables to be taken into consideration, providing sets of correlated biomarkers with a similar function and the possibility of solving problems where the number of variables is larger than the number of samples.

PCA is used here to describe the data coupled to a ranking procedure of the candidate biomarkers in a forward search: one variable is added at each cycle. The first variable selected is the one providing the best separation between the classes on the first principal component. The addition of another discriminating variable further improves the distance between the two classes on the first principal component. If a non-discriminating variable is successively added, the two classes will not be further separated on the same component. Sometimes, more than one component could be necessary for class separation: in this case, different independent sources of information related to the class structure are present in the data and the subsequent principal component accounting for class separation will be included in the model.

The proposed method allows the ranking of the variables according to their discrimination ability, thus assuring the exhaustiveness of the results. The result can be presented in graphical form (Figure 2), where the classification performance or the class distance are reported as a function of the variables added to the model.

Open in New Tab   Full Size Figure   Download Figure

Figure 2 Example of the graphical representation of the results from Ranking-PCA. Trend of Ind₂vs the increasing number of variables added to the model (A); trend of the distance between the two class centroids vs the number of variables added to the model (B). Variables on the X-axis are reported in the order in which they are included in the model. Both calibration and cross-validation results are reported.

Partial least squares (PLS)[14,15,30] establishes a relationship between one or more dependent variables (Y) and a group of descriptors (X). X and Y variables are modeled simultaneously to find the latent variables (LVs) in X that will predict the LVs in Y. These LVs are calculated hierarchically, as for PCA. PLS was originally set up to model continuous responses but it can be applied even for classification purposes (PLS-DA) by establishing an appropriate Y related to the association of each sample to a class. The regression is then carried out between X-block variables and the Y just established. Orthogonal PLS (OPLS)[31] is a modification of PLS developed for highly decorrelated datasets.

Classification trees[24] are built by subsequent divisions (splits) of subgroups of the original data D in two descending subgroups with the aim of classifying the data in homogeneous groups as much as possible, one different from the others. It is possible to derive a tree diagram where, starting from the root node (where the data D are not separated), a series of nodes and branches separate; each node h represents a subgroup of D. Nodes not undergoing a further split are called terminal nodes: a mode for y is associated with each terminal node.

Starting from the root node h₁, data are separated in a series of splits: in each node, the split giving the most homogeneous division of the data in the two descendent nodes is selected.

Random forests[32] is an extension of classification trees and is structured to grow many classification trees. To classify a new object, the new object is first classified by each independent tree in the forest. The forest chooses the most recurrent classification (over all the trees in the forest).

The error rate depends on the correlation between any two trees in the forest (increasing the correlation increases the forest error rate) and the strength of each individual tree in the forest (inversely correlated to the tree error rate).

The Kaplan-Meier method[33] is used to provide survival functions and can also be effectively applied to censored data. It provides a survival curve where time is reported on the X-axis, while the cumulative survival probability is on the Y-axis. The time corresponding to the point where the curve crosses 50% survival is the estimate of median survival. Kaplan-Meier curves can be compared across groups by mainly applying two non-parametric tests (the log rank test[34] and the generalized Wilcoxon test[35]). The generalized Wilcoxon test[35] is a weighted alternative where early time points weigh more than late ones; this is preferred when the effect of an experimental condition vanishes with time. The log rank test instead is preferred to detect differences during all of the follow-up period.

Cox regression[36] is applied to evaluate the effect of several risk factors on survival, defining the hazard as the probability of the endpoint. The hazard is modeled as:

ln [H(t)/H₀(t)] = b₁X₁ + b₂X₂ + ... + b_kX_k

where H(t)/H₀(t) is the so-called hazard ratio (HR), X₁ ... X_k are predictor variables and H₀(t) is the baseline hazard at time. In general, the HR may assume only positive values: if it equals 1, the groups show a not statistically different survival; if it is smaller than 1, a subject with a higher value for X has lower risk than a subject with a lower value for X; the opposite behavior is obtained if the HR estimate is larger than 1.

The Cox model works under the so-called proportional hazards assumption: the ratio between the hazards of two patient groups remains constant over the complete follow-up period. Since a HR is calculated in Cox regression, this estimate should apply to all death times: this simplification is only justified if the group difference remains constant over the whole range of follow-up time.

If the monovariate Cox regression is extended to include more than one X variable (multivariate Cox regression), the effect of the interaction between different factors can be evaluated.

Prediction analysis for microarrays (PAM) performs sample classification from gene expression data and survival outcomes, exploiting the nearest shrunken centroid approach[37]. A standardized centroid for each class is computed. The nearest centroid classification compares the gene expression profile of a new sample to each class centroid: the class whose centroid is closest in squared distance is the predicted class for the new sample.

Nearest shrunken centroid classification[37] is a modification of this approach. When PAM is applied to survival outcomes, supervised principal components analysis[38] is performed, where, instead of using all the genes to perform PCA, only a subset of genes is used, i.e., those highly correlated with survival.

Monte Carlo methods[39] are computational algorithms that rely on repeated random sampling to obtain numerical results; simulations are run several times to calculate probabilities. They are useful for simulating systems with several degrees of freedom and modeling systems characterized by large uncertainty in inputs.

The Metropolis algorithm[40] is used to generate a series of numbers, X₁, X₂, .., X_n with a distribution fixed a priori. The method is based on the generation of numbers that are accepted or rejected to obtain the selected type of distribution.

The Metropolis algorithm can be implemented in Monte Carlo simulations to perform random sampling. The Monte Carlo optimization can be used in biomarkers search to determine the coefficients of model containing the relevant biomarkers.

The studies based on serum and tissue biomarkers determined by non-omics techniques usually exploit monovariate and multivariate Cox regression to evaluate the effect played by different factors on time to progression (TTP) and overall survival (OS). They will be presented here, divided into prognostic/predictive biomarkers and diagnostic biomarkers.

Recent studies about prognostic and/or predictive biomarkers (Table 2) determined in serum or plasma regard the determination of glycoproteins, both alone or associated with other markers, and the determination of circulating factors of the insulin-like growth factor. Three recent studies have been published based on CA19-9, one of the most debated biomarkers for PC. The first study, by Boeck et al[41], includes 115 patients with histologically confirmed advanced PC treated with first-line therapy. The novelty of this study is the modelling of the effect of CA 19-9 kinetics by treating it as a time-varying covariate. For CA 19-9 kinetics during chemotherapy, data from 69 patients (TTP) and 84 patients (OS) were available. The proposed approach allowed the modeling of the effect of log (CA 19-9) measured during therapy on the event (TTP or OS) rather than modeling the effect of the pretreatment value of log (CA 19-9).

In the second study, Haas et al[42] pooled pre-treatment data on CA 19-9, carcinoembryonic antigen (CEA), C-reactive protein (CRP), lactate dehydrogenase (LDH) and bilirubin from two multicenter randomized phase II trials and prospective patient data. Marker levels were assessed before the start of palliative first-line therapy for advanced PC and during treatment (for CA 19-9 only).

In the third study, Boeck et al[43] evaluated pre-treatment (palliative first-line chemotherapy) values and weekly values of cytokeratin 19-fragments (CYFRA 21-1), CA 19-9 and CEA in blood samples from patients with PC. CYFRA 21-1 are biomarkers for different epithelial diseases but their role in PC has not been investigated yet.

In Boeck et al[41] pre-treatment log (CA19-9) proved to be significantly associated with TTP and OS. Moreover, log (CA 19-9) kinetics after the start of treatment was found to be a significant predictor for both TTP and OS. Similar results were found by Haas et al[42] where pre-treatment CA 19-9, LDH and CEA levels were significantly associated with TTP. Regarding OS, baseline CA 19-9, LDH, CRP and bilirubin were significant. Boeck et al[43] found that CYFRA 21-1 and CA 19-9 showed a high correlation with TTP and OS, while in multivariate analysis, only CYFRA 21-1 and performance status were independent predictors for OS.

McCaffery et al[44] assessed the predictive nature of baseline circulating factors of the insulin-like growth factor (IGF) axis on the treatment effect of ganitumab plus gemcitabine in metastatic PDAC. Baseline levels of IGFs/IGF binding proteins were analyzed in serum or plasma while mutations and gene expression were analyzed in archival samples. Ganitumab was associated with improved OS vs placebo. The treatment effect on improved OS was larger in patients with higher levels of IGF-1, IGF-2 or IGFBP-3, or lower levels of IGFBP-2. Interaction between treatment and IGFs/IGFBPs showed predictive potential for IGF-2 and IGFBP-2.

The studies about prognostic and/or predictive biomarkers determined in tissue samples, usually by immunohistochemistry or tissue microarray analysis (TMA), instead include the determination of glycoproteins, other proteins and enzymes.

Fong et al[45] investigated the expression of TROP2 (human trophoblast cell-surface) antigen, a glycoprotein found to be strongly expressed in a variety of human epithelial cancers, and the expression of junctional adhesion molecule A (JAM-A) antigen[46], a type I transmembrane glycoprotein, which has been recently shown to affect the prognosis of several malignancies. The two studies involved 197 and 186 patients with PDAC respectively. TROP2 overexpression was observed in 55% of patients, while low expression of JAM-A was observed in 42% of samples; both markers were significantly associated with decreased OS. They were both correlated with lymph node metastasis and tumor grade. In the subgroup of patients surgically treated with curative intent, TROP2 and low expression of JAM-A correlated with poor progression-free survival.

In the study by Zong et al[47], the expression of the T-box transcription factor 4 (TBX4) was investigated in 77 stage II PDAC tumors. 62.3% of cases expressed TBX4 at a high level. Significant correlation was only detected between TBX4 expression and tumor grade and liver metastasis recurrence. The survival with TBX4-high expression was significantly longer.

Applying tissue microarray (TMA) analysis, Schäfer et al[48] correlated heat shock protein 27 (HSP27) expression status with clinicopathological parameters in PDAC from 86 patients. HSP27 expression was found in 49% of tumor samples. A significant correlation was found with OS. The authors also assessed the impact of HSP27 on chemo- and radio-sensitivity directly in PC cells. HSP27 expression emerged as an independent prognostic factor and correlated inversely with nuclear p53 accumulation, indicating protein interactions between HSP27 and p53 or TP53 mutation-dependent HSP27-regulation. HSP27 overexpression rendered HSP27 low-expressing PL5 PC cells more susceptible to treatment with gemcitabine, while HSP27 protein depletion in HSP27 high-expressing AsPC-1 cells caused increased gemcitabine resistance.

Maréchal et al[49] identified deoxycytidine kinase (dCK), a recombinant enzyme, to be associated with prolonged survival after adjuvant gemcitabine administration for resected PDAC. The study involved 45 patients. The lymph node (LN) ratio and dCK protein expression were significant predictors of DFS and OS in univariate analysis. On multivariate analysis, a step-down procedure based on the likelihood ratio test, dCK protein expression was the only independent prognostic factor.

Having previously reported that Notch3 activation appeared to be associated with more aggressive PC disease, Mann et al[50] examined components of this pathway (Notch1, Notch3, Notch4, HES-1, HEY-1) in resectable and non-resectable tumors compared to uninvolved pancreas. All three Notch family members were significantly increased in tumor tissue, with expression maintained within matched lymph node metastases. Significantly higher nuclear expression of Notch1, -3 and -4, HES-1 and HEY-1 was noted in locally advanced and metastatic tumors compared to resectable cancers. Nuclear Notch3 and HEY-1 expression were significantly associated with reduced OS and DFS following tumor resection.

Regarding diagnostic biomarkers (Table 2), the most recent studies are based on the determination of protein panels in serum or plasma, exploiting ELISA or proximity ligation assay (PLA) for their determination. PLA is a highly sensitive technique for multiplex detection of biomarkers in plasma with little interfering background signal. Some of the studies proposed the determination of both diagnostic and prognostic markers. All studies presented here involve CA19-9 as a potential biomarker in association with other markers.

In the first study, Chang et al[51] applied PLA to the identification of plasma levels of 21 biomarkers in 52 locally advanced PDAC cases and 43 age-matched controls. The optimal diagnostic biomarker panel was computed using a combination of the PAM algorithm and logistic regression modeling.

In the second study, Brand et al[52] investigated 83 circulating proteins in sera of patients diagnosed with PDAC, benign pancreatic conditions and healthy controls. Samples from each group were split randomly into training and blinded validation sets prior to analysis. A Metropolis algorithm with Monte Carlo simulation (MMC) was used to identify discriminatory biomarker panels in the training set. Identified panels were evaluated in the validation set and in patients diagnosed with colon, lung and breast cancers.

In the third study, Schultz et al[53] tested the hypothesis that high plasma YKL-40 and IL-6 are associated with PC and short OS. 559 patients with PC from prospective biomarker studies were studied. Plasma YKL-40 and IL-6 were determined by ELISA and serum CA 19.9 by chemiluminescent immunoassay.

Chang et al[51] found that three markers (CA19-9, OPN and CHI3L1) have superior sensitivity for PC vs CA19-9 alone (93% vs 80%) and two markers (CEA and CA125) proved to have a prognostic significance for survival of PC (P < 0.003) when measured simultaneously. Brand et al[52] found that the panel of CA 19-9, CEA and TIMP-1 discriminated PDAC patients from benign subjects with an SN/SP of 76%/90% in the training set and of 71%/89% in the validation set. The CA19-9, intercellular adhesion molecule 1 (ICAM-1), OPG panel is selective for PDAC and does not recognize breast (SP = 100%), lung (SP = 97%) or colon (SP = 97%) cancer. Schultz et al[53] instead showed that high preoperative IL-6 and CA 19.9 were independently associated with short OS. High pre-treatment levels of each biomarker were independently associated with short OS in non-operable patients.

Tissue and serum biomarkers in proteomics are usually determined by: SDS-PAGE followed by LC-MS for identifying the most up- or down-regulated proteins; matrix-assisted laser desorption ionization time-of-flight (MALDI-TOF) mass spectrometry; and surface-enhanced laser desorption/ionization time-of-flight (SELDI-TOF) mass spectrometry. The exploitation of instrumental techniques providing a high amount of information makes the application of multivariate methods necessary in order to detect panels of biomarkers with the best predictive ability.

In the first study by McKinney et al[54], matched pairs of tumor and non-tumor pancreas from patients undergoing tumor resection were treated to obtain cytosol, membrane, nucleus and cytoskeleton cellular protein fractions. The fractions were analyzed by SDS-PAGE followed by LC-MS/MS to identify 2393 unique proteins. The spectral count data were compared using a power law global error model (PLGEM) to identify statistically significant protein changes between non-tumor and tumor samples[55,56]. Among the 104 proteins significantly changed in cancers, four (biglycan (BGN), pigment epithelium-derived factor (PEDF), thrombospondin-2 (THBS-2) and TGF-β induced protein ig-h3 precursor (βIGH3)) were further validated and proved to be up-regulated in cancer and have potential for development as minimally-invasive diagnostic markers.

Kojima et al[57] differentiated pancreatic neoplasia from non-neoplastic pancreatic disease. Samples from 50 patients [15 healthy (H), 24 cancer (Ca), 11 chronic pancreatitis (CP)] were collected. A high-throughput method was applied, using high-affinity solid lipophilic extraction resins, enriched low molecular weight proteins for extraction with a high-speed MALDI-MS. Multivariate analysis was carried out by MDS as in Mobley et al[58]. Using eight serum features, Ca were differentiated from H (SN = 88%, SP = 93%), Ca from CP (SN = 88%, SP = 30%) and Ca from both H and CP combined (SN = 88%, SP = 66%). In addition, nine features obtained from urine differentiated Ca from both H and CP, combined with high efficiency (SN = 90%, SP = 90%). Interestingly, the plasma samples did not show significant differences.

Ehmann et al[59] and Hauskrecht et al[60] instead applied SELDI-TOF mass spectrometry. In the first study[59], 96 serum samples from patients undergoing cancer surgery were compared with 96 controls. Samples were fractionated by anion exchange chromatography. Data analysis, involving Mann-Whitney U test and classification and regression tree (CART) analysis, identified 24 differentially expressed protein peaks, 21 of which were under-expressed in cancer samples. The best single marker can predict 92% of controls and 89% of cancer samples. The best model with a set of 3 markers showed a sensitivity of 100% and a specificity of 98% for the training data and a sensitivity of 83% and a specificity of 77% for test data. Apolipoprotein A-II, transthyretin and apolipoprotein A-I were identified as markers (decreased in cancer sera). Hauskrecht et al[60] instead proposed a feature selection method that extracts useful feature panels from high-throughput spectra. 57 PC samples were compared to 59 controls. The results clearly show the improved classification performances when the method is compared to standard strategies.

The last study makes use of protein microarrays to explore whether a humoral response to PC-specific tumor antigens has utility as a biomarker of PC. To determine if such arrays can be used to identify novel autoantibodies in the sera from PC patients, Patwa et al[61] resolved proteins from a PDAC cell line (MIAPACA) by 2-D liquid-based separations and then arrayed them on nitrocellulose slides. The slides were probed with sera from a set of patients diagnosed with PC and compared with age- and sex-matched normal subjects. To account for patient-to-patient variability, a non-parametric Wilcoxon rank-sum test was used in which protein biomarkers were identified. Classification by the PAM algorithm showed 86.7% accuracy, with a SN and SP of 93.3% and 80%, respectively. The identified candidate autoantibody biomarkers were validated using an independent sample set of 238 samples. Phosphoglycerate kinase-1 and histone H4 were noted to elicit a significant differential humoral response in cancer sera.

Diagnostic biomarkers in proteomics have also been recently determined from cell lines (PACA44, T3M4) or on animal models. The most exploited analytical techniques in this case are SDS-PAGE, followed by LC-MS, 2D-PAGE or 2D-LC approaches.

Marengo et al[62] identified the regulatory proteins in human PC treated with trichostatin A by 2D-PAGE maps and multivariate analysis. PCA was applied to a spot quantity dataset comprising 435 spots detected in 18 samples belonging to two different cell lines (Paca44 and T3M4) of control and drug-treated PDAC cells. PCA allowed the identification of the groups of samples present in the dataset; the loadings analysis allowed the identification of the differentially expressed spots, which characterize each group of samples. The treatment of both the cell lines with trichostatin A showed an evident effect on the proteomic pattern of the treated samples. Identification of some of the most relevant spots was also performed by MS analysis. The same authors applied different multivariate statistical tools to the same set of data to provide sets of candidate biomarkers; the first application regards the exploitation of SIMCA classification to evaluate the biomarkers characterized by a significant discriminant power[25], while the second application regards the development of ranking PCA[28]. This second application is particularly interesting for overcoming the limitations of the methods usually adopted as variable selection tools to identify only significant biomarkers: they are usually aimed at the selection of the smallest set of variables (spots) providing the best performances in prediction. This approach does not seems to be the best choice in the identification of potential biomarkers since all the possible candidate biomarkers have to be identified to provide a general picture of the ‘‘pathological state’’; exhaustivity has to be preferred to provide a complete understanding of the mechanisms underlying the pathology. Ranking PCA allowed the exhaustive identification of a complete set of candidate biomarkers.

Paulo et al[63] compared differentially expressed proteins in rat in activated and serum-starved non-proliferating pancreatic stellate cells (PaSC), emerging key mediators in chronic pancreatitis and PC pathogenesis. About 1500 proteins were identified after SDS-PAGE and LC-MS/MS. Qualitative and quantitative proteomic analysis revealed several hundred proteins to be differentially abundant between the two cell states. Significance analysis was performed using QSPEC, a recently published algorithm for determining the statistical significance of differences in spectral counting data from two sample sets[64]. This algorithm exploits the Bayes factor instead of the P-value as a measure of statistical significance[65,66]. Proteins of greater abundance in activated PaSC included isoforms of actin (e.g., smooth muscle actin) and ribosomal proteins. Proteins more abundant in non-proliferating PaSC than in activated PaSC included signaling protein MAPK-3 and Ras-related proteins. The molecular functions and biological pathways for these proteins were also determined by gene ontology analysis and KEGG pathway.

Paulo et al[67] also evaluated the endoscopic pancreatic function test (ePFT) as a method able to safely obtain pancreatic fluid for MS analysis from patients during upper endoscopy and reproducibly identify pancreas-specific proteins. The ePFT-collected pancreatic fluid from 3 individuals without evidence of chronic pancreatitis was analyzed by SDS-PAGE and GeLC-MS/MS. The SDS-PAGE analysis revealed no significant variation in protein concentration during the 1 h collection. The GeLC-MS/MS analysis identified pancreas-specific proteins previously described from endoscopic retrograde cholangiopancreatography and surgical collection methods. Gene ontology further revealed that most of the proteins identified have a molecular function of proteases.

Prognostic and predictive biomarkers in genomics (Table 4) are usually determined in tissue samples. Some of the proposed studies include both protein expression and microRNA expression profiles. In these studies, multivariate Cox regression analysis is usually exploited.

Ogura et al[68] studied the K-ras mutation status in 242 patients with unresectable PC. The authors focussed on K-ras mutation subtypes since recent reports indicate that K-ras mutation status acts as a prognostic factor. CA19-9, metastatic stage and mutant-K-ras were negative prognostic factors, indicating a reduced survival. Among the patients who had K-ras mutation subtypes, CA19-9, metastatic stage and the presence of the G12D or G12R mutations were negative prognostic factors for OS.

Hwang et al[69] evaluated whether expression of novel candidate biomarkers, including microRNAs, can predict clinical outcome in PDAC patients treated with adjuvant therapy. 82 resected PDAC cases were analyzed for protein expression by immunohistochemistry and for microRNA expression by quantitative real time PCR (qRT-PCR). Lower than median miR-21 expression was associated with a significantly lower HR for death and recurrence in the subgroup of patients treated with adjuvant therapy. MiR-21 expression status emerged as the single most predictive biomarker for treatment outcome. No significant association was detected in patients not treated with adjuvant therapy. The results were confirmed in an independent validation of 45 PDAC tissues.

One-fifth of patients with seemingly “curable” PDAC experienced an early recurrence and death, while some patients with advanced stage tumors are deemed “unresectable” by conventional staging criteria, yet progress slowly. Effective biomarkers that stratify PDAC based on the biological behavior are therefore needed. Building on a compendium of 2500 published candidate biomarkers in PDAC[70], Winter et al[71] constructed a survival tissue microarray (s-TMA) comprised of short-term (12 mo) and long-term survivors (30 mo) who underwent resection for PDAC. The s-TMA acts as a biological filter to identify prognostic markers. 13 putative PDAC biomarkers were identified from the public biomarker repository and tested against the s-TMA. MUC1 and MSLN were highly predictive of early cancer-specific death. By comparison, no pathological factors (size, lymph node metastases, resection margin status and grade) reached statistical significance.

Chen et al[1] detected metastasis-associated gene 2 (MTA2) expression in PDAC and related it to prognosis. MTA2 mRNA and protein expression were determined by qRT-PCR and immunohistochemistry in primary cancers and their adjacent non-cancerous tissues. MTA2 mRNA and protein expression levels were up-regulated in PC. MTA2 was correlated with poor tumor differentiation, TNM stage and lymph node metastasis. Patients with high expression levels of MTA2 showed lower OS.

Diagnostic biomarkers from genomics-based studies (Table 4) are identified both from tissue and serum samples.

To identify biomarkers for early detection, Pedersen et al[72] examined DNA methylation differences in leukocyte DNA between PC cases and controls. In phase I, methylation levels were measured at 1505 CpG sites in leukocyte DNA from 132 never-smoker PC patients and 60 never-smoker controls. Significant differences were found in 110 CpG sites. In phase II, 88 of 96 phase I selected CpG sites were tested and validated in 240 PC cases and 240 matched controls. Using penalized logistic regression, a prediction model was built consisting of five CpG sites (IL10_P348, LCN2_P86, ZAP70_P220, AIM2_P624, TAL1_P817) that discriminated cancer patients from controls. One CpG site (LCN2_P86) alone could discriminate resectable patients from controls.

Morse et al[73] used complementary assays of mRNA expression profiling of cell-surface genes to determine increased expression in PC vs normal pancreas tissues and validated protein expression by immunohistochemistry on tissue microarrays. This approach was aimed at the identification of targets for potential use in the molecular imaging of cancer, allowing for non-invasive determination of tumor therapeutic response and molecular characterization of the disease, or in the targeted delivery of therapy to tumor cells, decreasing systemic effects. Expression profiles of 2177 cell-surface genes for 28 pancreatic tumor specimens and 4 controls were evaluated. 170 unique targets were highly expressed in 2 or more of the pancreatic tumor specimens and were not expressed in controls. Two targets (TLR2 and ABCC3) were further validated for protein expression and proved to be potential for the development of diagnostic imaging and therapeutic agents for PC.

Huang et al[74] explored specific biomarkers that can differentiate PC-associated diabetes from type 2 diabetes for the early detection of PC. Peripheral blood samples were collected from 25 patients diagnosed with PC and diabetes, 27 patients with PC without diabetes, 25 patients with diabetes mellitus > 5 years, and 25 controls. 32 samples were used in microarray experiments to find differentially expressed genes specific for cancer-associated diabetes. The results were further validated by quantitative qRT-PCR for 101 blood samples. Protein expression of selected genes in serum and tissues was also detected. 58 genes were found to be unique in patients with cancer-associated diabetes (23 up-regulated; 35 down-regulated). 11 up-regulated genes were further validated by RT-PCR and 2 of these, vanin-1 (VNN1) and matrix metalloproteinase 9 (MMP9), showed the best discrimination of cancer-associated from type 2 diabetes.