Predicted Prognosis of Patients with Pancreatic Cancer by Machine Learning

Despite improvements in diagnostic tools and treatments, patients with pancreatic ductal adenocarcinoma (PDAC) have a poor clinical outcome. At the time of diagnosis, most patients with PDAC are in advanced stages because the anatomic location of the pancreas and lack of specific symptoms hamper early detection. Easy infiltration to the surrounding organs and early distant metastasis, even from a small primary tumor <2 cm in diameter, enhances the stage progression (1). Moreover, PDAC is sometimes difficult to distinguish from other pancreatic diseases, such as chronic pancreatitis, even when endoscopic, ultrasound-guided, fine-needle aspiration is performed (2–4). Indolent tumors such as intraductal papillary mucinous neoplasms (IPMN) also occur in the pancreas and sometimes transform to invasive cancer with a poor outcome (5–8). Currently, IPMNs are the most common cystic neoplasm of the pancreas and are classified into gastric, intestinal, pancreatobiliary, and oncocytic types (9, 10). A recent study has demonstrated that the morphologic subtype of IPMN is an independent prognostic factor (8). The overall 5-year survival rate for all diagnosed patients with PDAC is presently only 13% in Japan. Unfortunately, only 10%–20% patients present with resectable disease at diagnosis of PDAC (11). Five-year survival with tumor removal alone is generally less than 10%. After resection, the use of adjuvant chemotherapy doubled 5-year survival to around 16%–21% (12); however, it can be increased to 38.6% after a successful resection at stage IIA (13–16). Therefore, it is essential to identify effective biomarkers to enable early diagnosis and precisely predict the prognosis to recommend additional neoadjuvant and adjuvant therapies

Mucins (MUC) play crucial roles in carcinogenesis and tumor invasion in pancreatic tumors. The MUC gene product is posttranslationally modified, most likely, through extensive O-glycosylation. MUC1 and MUC4 are large membrane-bound glycoproteins that are translated as single polypeptides. These mucins undergo intracellular autocatalytic proteolytic cleavage into two subunits that form stable noncovalent heterodimers that are transported to the cell surface. MUC2 is gel-forming secretory mucin that is expressed in many organs, including the colon, small intestine, and respiratory tract (17–19). MUC1 contributes to oncogenesis by promoting the loss of epithelial cell polarity, promoting growth and survival pathways, activating receptor tyrosine kinase signaling pathways, and conferring resistance to the stress-induced cell death pathway (20, 21). MUC1 cytoplasmic domain has been implicated in the regulation of the Wnt–β-catenin, p53, and NF-κB pathways, all of which are linked to tumor progression (19, 22). MUC4 plays an important role in epithelial cell proliferation and differentiation by inducing specific phosphorylation of ERBB2 and enhancing the expression of the cyclin-dependent kinase inhibitor p27, which inhibits cell-cycle progression (23, 24). The loss of MUC2 might compromise signaling that contributes to epithelial differentiation and proliferation through contact with membrane-bound mucins or alters the differentiation program of the intestinal mucosa, resulting in an increased probability of tumor formation (25, 26). In our previous studies, we used pancreatic tissue samples and several cancer cell lines to demonstrate that MUC1, MUC2, and MUC4 expression of mRNA and/or protein are regulated by hypoxia and/or DNA demethylation (27–30). However, it has not been reported how that DNA methylation of promoter region effect for intracellular localization of these three mucin proteins. Furthermore, MUC1 and MUC4 hypomethylation status are statistically associated with the development of distant metastasis, tumor stage, and overall survival (OS) for patients with PDAC (31–34).

In machine learning, state-of-the-art classification algorithms, such as support vector machines (SVM), neural networks (NNET), or multinomial prediction models, are used for classification and regression analysis (35–37). Recently, several studies shows that in breast cancer, nasopharyngeal carcinoma, and non–small cell lung cancer, several supervised learning methods, such as decision trees using data of cDNA or tissue microarray refine prognosis (38–41). However, it was not reported whether these machine learning models could use DNA methylation status to predict the outcome of patients with pancreatic tumors.

Therefore, the aim of this study was to develop a machine learning–based prognostic classifier to predict OS with pancreatic cancers by integrating multiple DNA methylation statuses of three mucin genes. In this study, we increased the number of patients that were evaluated from that of our previous study and performed a sequential analysis of mucin promoter–associated CpGs by methylation-specific electrophoresis (MSE) analysis.

Human pancreatic cancer cell lines BxPC3, HPAF2, and Panc1; human colon adenocarcinoma cell lines Caco2 and LS174T; and human lung adenocarcinoma cell lines A427 and NCI-H292 were obtained from the ATCC. HPAF2, LS174T, and Caco2 cells were cultured in Eagle Minimum Essential Medium (Sigma); PANC1 and A427 cells were cultured in DMEM (Sigma); and BxPC3 and NCI-H292 cells were cultured in RPMI1640 medium (Sigma). The media were supplemented with 10% FBS (Invitrogen), 100 U/mL of penicillin (Sigma), and 100 μg/mL streptomycin (Sigma). Hypoxic culture conditions were achieved with a multigas incubator containing a gas mixture of 94% N₂, 5% CO₂, and 1% O₂ (ASTEC).

We obtained 300 surgically resected tissues (approximately 2 × 2 × 2 mm in size) from neoplastic and nonneoplastic areas of 191 patients. Table 1 summarizes the clinicopathologic features of the 114 neoplastic samples and 186 nonneoplastic samples (including 109 paired samples). We collected 125 patient samples (48 neoplastic samples and 120 nonneoplastic samples, including 43 paired samples) from Kagoshima University (Kagoshima, Japan), from August 2007 to May 2014 and 66 patient samples (all nonneoplastic and neoplastic paired samples) from Ulm University (Ulm, Germany), from February 2001 to February 2013. The nonneoplastic tissues were collected around the resection stump. These samples were checked for pathologic diagnosis using intraoperative frozen and formalin-fixed paraffin-embedded (FFPE) tissue sections. On the other hand, neoplastic tissues were macroscopically collected, processed for FFPE tissue sections, and diagnosed by a board-certified pathologist. The clinical features used in this study were TNM, age, American Society of Anesthesiologists (ASA) physical status classification score, the presence or absence of comorbidities, and preoperative chemotherapy. Almost all patients had not undergone radiotherapy before surgery; therefore, we removed the information about radiotherapy in statistical analysis.

tRNA was extracted from the cell lines, human pancreatic tissues, and pancreatic juices using an RNeasy Mini Kit (Qiagen). Then, the tRNA (1 μg) was reverse transcribed with a High-Capacity RNA-to-cDNA Kit (Applied Biosystems), and real-time reverse transcription-PCR was performed on a Roche LightCycler 96 System using FastStart Essential DNA Green Master (Roche). Gene expression was normalized to the β-actin mRNA level in each sample. The data were normalized using the NCI-H292 cell line, and the A427 cell line was used as a negative control. Primer sets are shown in previously study (31, 33, 34).

DNA from the cell lines, pancreatic tissues, and pancreatic juices was extracted using a DNeasy Tissue System (Qiagen). The bisulfite modification of the genomic DNA was carried out using an EpiTect Bisulfite Kit (Qiagen). The purification of PCR products was carried out using a Wizard SV Gel and PCR Clean-Up System (Promega KK; refs. 31, 33, 34).

MSE analysis was performed using previously described methods. Briefly, the target DNA fragments were amplified by nested PCR using bisulfite-treated DNA with the primer sets detailed in previously study (31–34). Then, the PCR products were run on a polyacrylamide gel with a linear denaturant gradient at 60°C at 70 V for 14 hours using a D-Code Universal Mutation Detection System (Bio-Rad Laboratories). Band intensity was quantified by Image J Software (NIH, Bethesda, MD). The demethylation index was calculated as the proportion of the highest band intensity/total band intensity of the sample. Subsequently, the demethylation index for each sample was normalized using data from a hypomethylated and hypermethylated cell line. Cell lines with hyper- and hypomethylation of MUC1 and MUC4 (Caco2 and LS174T, respectively) were used as control standards. We performed MSE analysis on duplicate samples.

Data were analyzed using the “R” computing environment version 3.5.2 (42). The normality of the data distribution was evaluated using the Kolmogorov–Smirnov test. Differences between groups were analyzed using Welch t test. A nonparametric test of group difference was performed using the Mann–Whitney U test. Survival rate analysis was evaluated using Cox proportional hazards model. Hierarchical cluster analysis on a set of dissimilarities was performed using the reproduce package of R. A P < 0.05 was considered statistically significant.

We constructed the SVM classifier using the kernlab package, including the ksvm function (37). Methylation analyzed data mining, such as scaling and centering, were performed using the scale function. C-classification with the linear vanilladot kernel, Gaussian radial basis function, polynomial kernel, hyperbolic tangent kernel, Bessel function of the first kind kernel, Laplace radial basis kernel, and ANOVA radial basis kernel were used in training and predicting. In this study, we used default hyperparameters and cost of constraints violation for preparing the prediction models. The quality of the model was assessed using 5-fold cross-validation of the training data.

Neural networks provide a flexible, nonlinear extension of multiple logistic regressions to perform classification, pattern recognition, and prediction modeling. The nnet package, including the nnet function, was used to construct an NNET classifier (43). The classifier parameters were set with a weight decay of 0.1, 2 units in the hidden layer and 100 as the maximum number of iterations. To construct the multinomial log-linear model (MU) via the NNET classifier, the multinom function was employed.

This study was conducted in accordance with the guiding principles of the Declaration of Helsinki. The ethical committees of both Kagoshima University Hospital (Kagoshima, Japan) and Ulm University Hospital (Ulm, Germany) approved sample collection, and informed written consent was obtained from each patient. All studies using human materials in this article were approved by the Ethical Committee of Kagoshima University Hospital (Kagoshima, Japan; revised 20–82, revised 22–127, and 26-145).

Unsupervised hierarchical clustering analysis was performed for three mucin gene expression datasets including neoplastic and nonneoplastic pancreatic regions (Fig. 1A). The samples were divided into two clusters according to the clustering results. Cluster 1 showed a significantly higher expression level of the three mucin genes as MUC1, MUC2, and MUC4 compared with cluster 2 (all P < 0.001; Fig. 1B). Cluster 1 also demonstrated a significantly poorer prognosis than cluster 2 (HR = 0.52; P < 0.001; Fig. 1C). As shown in Table 2, this clustering analysis showed the same result in the nonneoplastic tissue (P < 0.001). In nonneoplastic region, cluster 1 showed a significantly poorer prognosis than cluster 2 (HR = 0.49; P = 0.012; Supplementary Fig. S1A). Similarly, cluster 1 showed a significantly poorer prognosis than cluster 2 in the neoplastic region (HR = 0.53; P = 0.032; Supplementary Fig. S1B).

Unsupervised hierarchical clustering analysis was performed for three mucin gene methylation datasets including neoplastic and nonneoplastic pancreatic regions (Fig. 2A). According to the clustering results, the samples were divided into four clusters. Cluster 1 showed a significantly higher hypomethylation status of MUC1 and MUC4 genes than the other clusters (P = 0.003 and P < 0.001, respectively; Fig. 2B). Cluster 1 also demonstrated a significantly higher expression level of MUC1, MUC2, and MUC4 than the other clusters (P = 0.004, P = 0.002, and P = 0.015, respectively; Supplementary Fig. S2). Furthermore, cluster 1 demonstrated a significantly poorer prognosis than the other clusters (HR = 0.33; P = 0.012; Fig. 2C). As shown in Table 2, this clustering analysis did not identify a significant difference in the nonneoplastic ratio between Cluster 1 and the other clusters (P < 0.357). In the nonneoplastic regions, cluster 1 showed a significantly poorer prognosis than the other clusters (HR = 0.49; P = 0.010; Supplementary Fig. S3A). Similarly, in the neoplastic regions, cluster 1 also showed a significantly poorer prognosis than the other clusters (HR = 0.48; P = 0.014; Supplementary Fig. S3B).

In the leave-one-out cross-validation (LOOCV) test, the SVM classifier model by C-classification with ANOVA RBF kernel function using the dataset for DNA methylation of three mucins, including both nonneoplastic and neoplastic regions, showed good classification for prognosis after surgery (Fig. 3A). In addition, ANOVA RBF kernel as a nonlinear kernel showed the best classification than the other kernels, including linear kernel (Supplementary Table S1). This SVM classification showed no significant differences in the nonneoplastic ratio between high-risk positive and negative groups (Supplementary Table S2). In nonneoplastic regions, the high-risk group indicated by SVM showed a significantly poorer prognosis after surgery than the negative group (Fig. 3B). Similarly, in neoplastic regions, the high-risk group indicated by SVM also showed a significantly poorer prognosis after surgery than the negative group (Fig. 3C). In univariate analysis, patients who were classified as positive by the SVM model were associated with a significantly poorer OS. Multivariate Cox regression analysis after adjustment for clinicopathologic variables, such as ASA score, preoperative chemotherapy, comorbidities, and TNM stage, revealed that the SVM classifier remained a powerful and independent prognostic factor for OS in the LOOCV test (Supplementary Table S3). Following several k-fold cross-validation tests, we performed 3-, 4-, and 10-fold cross-validations in this study. The SVM classifier showed good classification for prognosis after surgery in both regions (Supplementary Fig. S4; Table 3). Multivariate and/or univariate analysis revealed that SVM classifier was an independent prognostic factor (Supplementary Table S3).

In the LOOCV test, the NNET classifier model using the dataset for DNA methylation of three mucins, including both nonneoplastic and neoplastic regions, showed good classification for prognosis after surgery (Fig. 3D). This NNET classification showed no significant difference in the nonneoplastic ratio between high-risk positive and negative groups (Supplementary Table S2). In nonneoplastic regions, the high-risk group indicated by NNET showed a significantly poorer prognosis after surgery than the negative group (Fig. 3E). Similarly, in neoplastic regions, the high-risk group indicated by NNET also showed a significantly poorer prognosis after surgery than the negative group (Fig. 3F). In univariate analysis, patients who were classified as high-risk positive by the NNET model were associated with a significantly poorer OS. Multivariate Cox regression analysis after adjustment for clinicopathologic variables, such as ASA score, preoperative chemotherapy, comorbidities, and TNM stage, revealed that the NNET classifier remained a powerful and independent prognostic factor for OS in the LOOCV test (Supplementary Table S3). After several k-fold cross-validation tests, the NNET classifier showed a tendency toward good classification for prognosis after surgery in both nonneoplastic and neoplastic regions (Supplementary Fig. S4; Supplementary Table S3). Multivariate and/or univariate analysis showed that NNET classifier was independent from prognostic factors (Supplementary Table S3).

In the LOOCV test, the MU classifier model using the dataset for DNA methylation of three mucins, including both nonneoplastic and neoplastic regions, did not show good classification for prognosis after surgery (Fig. 3G). However, this MU classification revealed a significant difference in the nonneoplastic ratio between the high-risk positive and negative groups (P < 0.043; Supplementary Table S2). In neoplastic regions, the negative group indicated by MU showed a significantly poorer prognosis after surgery than the high-risk group (Fig. 3H), but not in nonneoplastic regions (Fig. 3I). The multivariate Cox regression analysis after adjustment for clinicopathologic variables such as TNM stage revealed that the MU classifier was an independent prognostic factor for OS in the LOOCV test. However, in univariate analysis, patients who were positive according to the MU model were not associated with significantly poorer OS (Supplementary Table S3). In k-fold cross-validation test, the MU classifier demonstrated good classification ability for prognosis after surgery in the neoplastic region but not in nonneoplastic region (Supplementary Fig. S4; Supplementary Table S3).

To evaluate whether the prediction models constructed using test datasets could detect high-risk groups in other training dataset, we split the total dataset into two groups: training and test datasets. These two groups showed almost similar distribution in biological analysis results and clinicopathologic features (shown in Supplementary Table S4). SVM and NNET classifiers exhibited significantly good classification ability for prognosis after surgery in not only the neoplastic region but also the nonneoplastic region (Fig. 4). Multivariate and/or univariate analysis revealed that these classifiers were independent from prognostic factors (shown in Supplementary Table S3). The MU classifier demonstrated good classification ability for prognosis after surgery in the neoplastic region but not in nonneoplastic region (Fig. 4G–I).

PDAC is an aggressive malignancy with an extremely poor prognosis due to delayed diagnosis, early metastasis, and resistance to most cytotoxic agents (1, 13). Thus, it is critical to establish new diagnostic, prognostic, and therapeutic biomarkers. It has been previously demonstrated that mucin gene expression (including MUC1, MUC2, MUC3, MUC4, and MUC5AC) is regulated by DNA methylation at promoter regions in cancer cell lines (27–30). In this study, scientific computer prediction model–based SVM and NNET methods were able to classify between a good and poor prognosis after surgery on pancreatic cancers. These predicted models were constructed from the methylation status of three mucin genes (MUC1, MUC2, and MUC4), which had all previously demonstrated a significant difference in mucin expression levels in pancreatic cancers.

To construct a clinical test to predict prognosis after surgery, we used an expression dataset including nonneoplastic and neoplastic pancreatic tissues. To determine whether the expression levels of MUC1, MUC2, and MUC4 mRNA could distinguish a poor prognosis from data, we performed unsupervised hierarchical clustering analysis using RT-PCR data. These mucin mRNA clustering analyses significantly separated a poor prognosis cluster. This selected cluster had higher expression levels of the three mucins than the other cluster. In our recent histologic studies, we reported that mucin gene expression, particularly MUC4, was an independent indicator of worse prognosis in PDAC (5–9). MUC2 is reported as a tumor suppressor gene, and loss of MUC2 promotes tumor progression in colon (44). However, another study showed that MUC2 expression may have a poor prognostic value for differentiated adenocarcinomas in pancreas (45). The relationship between high expression of three mucins mRNA and poor prognosis supported these recent pathologic studies. The selected cluster showed a significant difference in the nonneoplastic: neoplastic content ratio compared with the other cluster. In the neoplastic region data, the selected cluster showed a poorer prognosis than the other cluster. Interestingly, in the nonneoplastic region data, this selected cluster also showed a poorer prognosis than the other cluster. These results suggested that scientific computer methods could identify a poor prognosis using the combined data of mucin gene expression even if there was a mixture of nonneoplastic and neoplastic tissue.

A previous study revealed that an analysis of the DNA methylation status in promoters of MUC1, MUC2, and MUC4 (MSE analysis of pancreatic juice samples) could differentiate between gastric-type IPMN, intestinal-type IPMN, other-type IPMN, and PDAC (33, 34). The correlation between hypomethylation of the promoter and high expression levels of mucin mRNA in pancreatic tissue was shown (28). Furthermore, we have proposed that aberrant methylation of MUC1 and MUC4 promoters are potential prognostic biomarkers for PDAC and suggested further MSE analysis of human clinical samples to determine its utility for the early diagnosis of pancreatic neoplasms and for stratifying patients with respect to modes of treatment (31). Thus, we used a methylation dataset including nonneoplastic and neoplastic pancreatic tissue. To establish whether the methylation levels of these three mucins could distinguish a poor prognosis, we performed unsupervised hierarchical clustering analysis using MSE analysis data. The clustering analysis of the three mucin genes' methylation data significantly separated a poor prognosis cluster. This cluster group had a higher hypomethylation level of MUC1 and MUC4 than the other clusters. Moreover, this cluster demonstrated a higher neoplastic-including ratio than the other clusters, and in neoplastic regions, this cluster showed a poorer prognosis than the other clusters. Interestingly, in the nonneoplastic region analysis, this cluster also showed a poorer prognosis than the other clusters. These results suggested that scientific computer methods could provide a model to identify a poor prognosis using the combined data of mucin gene methylation.

To evaluate whether machine learning prediction models using state-of-the-art classification algorithms such as SVM and NNET could distinguish between a poor prognosis group and others, we constructed prediction models. In LOOCV tests and k-fold cross-validation tests to evaluate prediction ability, the SVM and NNET models could significantly judge the identified high-risk group as having a poor prognosis, but the MU model could not. Multivariate and univariate analyses showed that the prediction of high-risk by SVM or NNET model was a prognostic factor significantly independent from TNM score, ASA score, preoperative chemotherapy, and comorbidities. For the neoplastic tissue analysis data, the SVM, NNET, and MU models could identify the high-risk group. Interestingly, even in the nonneoplastic tissue analysis data, the SVM and NNET model-selected high-risk group had a significantly poorer prognosis than others, similar to the neoplastic analysis. Therefore, these results suggested that the prediction models using cytological specimens and liquid biopsy samples, which are mixture of nonneoplastic and neoplastic cells, might be applicable to high-risk screening in PDAC.

When the model has high variance and low bias, such as that showing too much optimization for the training dataset, the prediction model has low prediction performance for data that has never been learned. To evaluate whether the prediction models constructed from the test dataset could detect high-risk groups in other training datasets, we split the total dataset into two groups having almost similar distribution of biological status and clinicopathologic features. The SVM and NNET classifiers could significantly distinguish the high-risk group, which has poor prognosis in the test group that has never been learned, but the MU model could not. Multivariate and univariate analyses showed that the prediction of high-risk by SVM and NNET classifiers was a prognostic factor significantly independent from TNM score, ASA score, preoperative chemotherapy, and comorbidities. These results suggested that SVM and NNET classifiers have low variance and demonstrated the high ability to distinguish a poor prognosis. Although a prospective, much larger, and multicenter randomized trial would be necessary to validate our results, it is suggested that the SVM- or NNET-based prediction models could provide a clinical risk test to predict the prognosis after surgery using MUC1, MUC2, and MUC4 gene methylation analyses. Even though these SVM and NNET classifiers were a highly accurate predictor of OS, we are aware that other biomarkers may extend the precision and predictive value of the classifiers, and new markers are being identified and new techniques developed every year (46, 47). Thus, the SVM and NNET classifiers may be further improved by including additional markers.

In summary, this study demonstrated that machine learning prediction models, based on SVM and NNET, could accurately distinguish patients with pancreatic cancer after surgery with substantially different OS. A further study is needed to expand the clinical sample spectrum, where these classifiers based on SVM or NNET might work for decision-making regarding follow-up scheduling after surgery.

No potential conflicts of interest were disclosed.

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Conception and design: S. Yokoyama, S. Yonezawa

Development of methodology: S. Yokoyama, S. Yonezawa

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S. Yokoyama, T. Hamada, M. Higashi, K. Maemura, H. Kurahara, M. Horinouchi, T. Hiraki, M. Kornmann

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S. Yokoyama, T. Hamada, T. Sugimoto, S.K. Batra

Writing, review, and/or revision of the manuscript: S. Yokoyama, M. Higashi, M. Kornmann, S.K. Batra, A. Tanimoto

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S. Yokoyama, K. Matsuo, H. Kurahara, T. Akahane, S. Yonezawa

Study supervision: M. Higashi, S. Yonezawa, M.A. Hollingsworth, A. Tanimoto

The authors thank Orie Iwatani, Yoshie Jitoho, and Yukari Nishida for their assistance with clinical sampling and excellent technical assistance with IHC. The authors would like to thank Enago (www.enago.jp) for the English language review. This study was supported, in part, by a grant from Grants-in-Aid for Scientific Research on Scientific Research (C; 18K07019, to M. Higashi), Scientific Research (C; 18K07018, to T. Hamada), and Scientific Research (C; 18K07326, to S. Yokoyama) from the Ministry of Education, Science, Sports, Culture and Technology, Japan by the Kodama Memorial Foundation, Japan (to S. Yokoyama) and by the Pancreas Research Foundation of Japan (to S. Yokoyama).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.