A Novel Tool for the Accurate and Affordable Early Diagnosis of Pancreatic Cancer via Machine Learning and Bioinformatics
AA Novel Tool for the Accurate and Affordable EarlyDiagnosis of Pancreatic Cancer via Machine Learningand Bioinformatics
Siya Goel [email protected]
Precollege Research Opportunities, Purdue UniversityClark Gedney [email protected]
Biological Sciences, Purdue University Jean Honorio [email protected]
Computer Science, Purdue University
Abstract
Pancreatic cancer (PC) is the fourth leading cause of cancer death in the United States dueto its five-year survival rate of 10%. Late diagnosis, affiliated with the asymptomatic nature inearly stages and the location of the cancer with respect to the pancreas, makes current widely-accepted screening methods unavailable. Prior studies have achieved low (70-75%) diagnosticaccuracy, possibly because 80% of PC cases are associated with diabetes, leading to misdiagnosis.To address the problems of frequent late diagnosis and misdiagnosis, we developed an accessible,accurate and affordable diagnostic tool for PC, by analyzing the expression of nineteen genes inPC and diabetes. First, machine learning algorithms were trained on four groups of subjects,depending on the occurrence of PC and Diabetes. The models were analyzed with 400 PCsubjects at varying stages to ensure validity. Naive Bayes, Neural Network and K-NearestNeighbors models achieved the highest testing accuracy of around 92.6%. Second, the biologicalimplication of the nineteen genes was investigated using bioinformatics tools. It was found thatthese genes were significantly involved in regulating the cytoplasm, cytoskeleton and nuclearreceptor activity in the pancreas, specifically in acinar and ductal cells. Our novel tool is thefirst in the literature that achieves a PC diagnostic accuracy of above 90%, having the potentialto significantly improve the detection of PC in the background of diabetes and increase thefive-year survival rate.
Pancreatic cancer (PC) is a disease in which malignant acinar and ductal cells are formed inthe pancreatic tissue [24]. The pancreas is a gland located behind the stomach and in front ofthe spine [5]. The pancreas has two key functions: (i) helping with digestion (exocrine) and (ii)regulating blood sugar (endocrine) [20]. More specifically, in the exocrine function, the pancreaticduct secretes enzymes which help in breaking down fats, carbohydrates and proteins [54, 4]. Theendocrine function of the pancreas consists of producing insulin which lowers blood glucose levelsand glucagon which raises blood glucose levels [32]. Maintaining this blood sugar level is importantin the functioning of key organs such as the brain, liver and kidneys [9].Even though only 56,770 Americans are diagnosed with PC (3% of all cancers), 47,740 peopledie from it, thus making PC the fourth leading cause of cancer death, causing 9% of deaths in cancerpatients [16]. The main reason for this high mortality rate is the difficulty of PC diagnosis [34].The majority of people suffering from PC are diagnosed at stage IV, i.e., the stage when the disease1 a r X i v : . [ q - b i o . Q M ] D ec igure 1: The five-year diagnosis rate of different cancers [7].has metastasized and there are signs of symptoms [37]. In fact, compared with many other cancers,the combined five-year survival rate for PC (i.e., the percentage of all patients who are living fiveyears after diagnosis) is relatively low at just 10%, compared to the survival rate of lung cancer aswell as breast cancer (Figure 1) [38].PC is difficult to diagnose early because the pancreas is located deep inside the abdomen [16].Further, patients usually do not have any symptoms until the cancer has reached later stages orhas already spread to other organs. As a result, early-stage tumors cannot be seen or felt by healthcare providers during routine physical exams [29]. In addition, there are no specific, cost-effectivescreening tests that can easily and reliably find early-stage PC in people who are asymptomatic [51].Patients usually do not have any symptoms until the cancer has reached later stages or has alreadyspread to other organs. Studies suggest that if the cancer were detected at an early stage (StageI or II) when surgical removal of the tumor was possible, the five-year survival rate would be40% [19]. However, only 15-20% of people are diagnosed at an early stage [51]. According to priorstudies, up to 80% of PC patients have diabetes which increases the risk of PC by 8-fold [36], whichsuggests that diabetes could be a precursor of PC [2]. However, because of inaccuracies of thecurrent diagnostic methods, 20% of PC cases are mistakenly diagnosed as diabetes [43]. Therefore,in order to aid in the detection of PC tumors at early stages, we developed a diagnostic tool forPC by distinguishing between gene expression in PC and diabetes. A potential solution leading toa more accurate diagnosis could be to discover differentially and similarly expressed genes betweendiabetes, PC, and normal cells [5]. To the best of our knowledge, our work is the first machine learning study to focus on early diagnosisand to compare PC and diabetes for gene signatures. That is, none of the prior works that wediscuss in this section pertain to early diagnosis.Image diagnosis is one of the two main methods used to diagnose PC [11]. Unfortunately,image diagnosis is considered ineffective, due to image blur, resolution and the appearance oftumors and visual change in later stages, making early diagnosis difficult [10]. This makes PCimaging inaccurate, expensive and unattainable [30].A study done by [6] examined 82 images of pancreatic segmentation for diagnosis. The images2btained were axial CT slices of the pancreas, which were analyzed by a variety of convolutionalneural networks. However, these algorithms only achieved 70% accuracy in diagnosis [6]. Anotherstudy done by [53] focused on pancreatic segmentation through pictures obtained from MRI scansfrom 78 subjects. Two convolutional neural networks were made to distinguish the differencebetween normal and cancerous tissue, achieving a 73.2% accuracy [53]. A third study done by [12]took a different approach and classified the four most common cyst types. The model consisted oftwo parts (i) a probabilistic random forest classifier, which analyzes manually selected quantitativefeatures and (ii) a convolutional neural network trained to discover high-level imaging features fora better differentiation. The data used contained 134 abdominal CT scans, achieving an accuracyof 83.6% [12].Besides imaging methods, another method used to diagnose PC is through genetic expressionanalysis [1]. According to a recent survey conducted by [47], only three studies have used machinelearning for gene expression in PC. All of these studies consisted of a limited amount of samples(i.e., 175). The accuracy of the three machine learning studies were 77%, 80% and 83% [47]. Itis important to note that the above studies were not accounting for the misdiagnosis of PC withdiabetes, as gene expression between the two groups are similar. In addition, in all of the abovestudies, 2,000 parameters or genes were used to predict the binary output label (i.e., whetherthe subject has PC or not) which might be affected by overfitting. We argue that in order toavoid overfitting, only a subset of genes need to be considered in order to provide an accuratediagnosis [23].
Our aim is to develop an early diagnostic tool of PC in an accessible, accurate, timely and affordablemanner by analyzing expression of few (tens of) genes in PC and diabetes. More specifically, ourgoal is to: • obtain an accuracy of over 90% for the diagnosis for PC in both early and late stage PC • understand the difference between the four groups of subjects depending on the occurrenceof PC and diabetes • provide a simple and accessible way to diagnose PC in order to make it more available to thepublic by creating a simple yet accurate algorithm • provide a cheap and affordable diagnosisOur design consists of the two phases described below: Phase 1.
400 samples (i.e., subjects) from open-source, annotated datasets were used in order toensure the validity of our study over different gene expression levels. A series of machine learningclassifiers (i.e., Logistic Regression, K-Nearest Neighbors, Random Forest, Naive Bayes, NeuralNetwork) were applied on four groups of subjects, depending on the occurrence of PC and diabetes.Nineteen genes were chosen from independent studies.
Phase 2.
We found the biological implication between the nineteen genes analyzed and significantgenetic function in PC such as cytoplasmic structure, formation of actin cytoskeleton and nuclearreceptor activity.In what follows, we describe the above two phases in detail.3able 1: Summary of data retrieved for this study.
Dataset Number of Samples
The Cancer Genome Atlas: project PAAD S11: 63 samplesS01: 91 samplesGene Expression Omnibus: datasets GSE22309 S00: 72 samplesS10: 32 samplesGene Expression Omnibus: datasets GSE15932,GSE16515, GSE14245 and GSE49515 S00: 34 samplesS10: 22 samplesS01: 51 samplesS11: 33 samples
Total Count
398 samplesS00: 106 samplesS10: 54 samplesS01: 142 samples (early stage PC: 47,late stage PC: 95)S11: 96 samples (early stage PC: 34,late stage PC: 62)
Five different classifiers: Logistic Regression, K-Nearest Neighbors (KNN), Random Forest, NaiveBayes and Neural Network, were applied to observe whether the accuracy of diagnosis of early PCcould be improved. The dataset retrieved for this study is comprised of four groups of subjects: • S11: subjects with Diabetes and PC • S10: subjects with Diabetes, without PC • S01: subjects without Diabetes, with PC • S00: subjects without Diabetes, without PCIn the above dataset, one subject corresponds to one sample or observation. Thus, in thismanuscript, we interchangeably refer to subjects or samples. A total of 398 samples were retrievedand consisted of 106 samples from S00, 54 samples from S10, 142 samples from S01, and 96 samplesfrom S11. The sources of data included The Cancer Genome Atlas (project PAAD) and the GeneExpression Omnibus (datasets GSE22309, GSE15932, GSE16515, GSE14245 and GSE49515) (Ta-ble 1). The Cancer Genome Atlas samples were filtered to be the ones that (i) mentioned whetheror not the patient had diabetes and (ii) mentioned the stage of PC. The Gene Expression Omnibussamples were chosen based on the following keywords: “pancreatic cancer” and “pancreatic ade-nocarcinoma”. The “organism” filter was set to “Homo sapiens” and the “study type” filter wasset to “expression profiling by array”. The datasets were chosen to be the ones that (i) mentionedwhether or not the patient had diabetes, and (ii) mentioned the stage of PC.Nineteen genes were found to be significantly overexpressed in PC through literature and wereselected as features for the five classifiers.Through literature, it was found that nineteen genes play a significant role in monitoring actinproduction and acinar and ductal cell function, and tend to be overexpressed in PC. Thus, theywere selected as features for the five classifiers. The nineteen genes are shown in Table 2.We used the Python programming language with libraries such as NumPy, Panda and SciKit-Learn. The data from sets S00 and S01 were named “No Diabetes” and the data from sets S10 and4able 2: Nineteen genes selected as features from the literature.
Gene Name Gene ID Reference
ABHD12 ENSG00000100997 [22]ABHD14B ENSG00000114779 [8]ABHD2 ENSG00000140526 [48]ACTA2 ENSG00000107796 [25]ACTB ENSG00000075624 [15]ACTN1 ENSG00000072110 [39]ACTN4 ENSG00000130402 [17]ACTR1A ENSG00000138107 [46]ACTR2 ENSG00000138071 [40]ADAR ENSG00000160710 [3]ADPRHL2 ENSG00000116863 [50]ADRA2A ENSG00000150594 [26]ADRM1 ENSG00000130706 [27]ALDH1A1 ENSG00000165092 [21]ALDH9A1 ENSG00000143149 [21]ALKBH5 ENSG00000091542 [13]ALKBH7 ENSG00000125652 [13]ANXA11 ENSG00000122359 [44]APOC1 ENSG00000130208 [49]
S11 were named “Diabetes”. Both datasets “Diabetes” and “No Diabetes” contain input features(i.e., the expression of the nineteen genes) and a binary output label (i.e., whether the subject hasPC or not).We tested five different classifiers and considered different hyperparameters to be tuned foreach of them. For the Logistic Regression classifier, we considered the hyperparameter C (inverseof regularization strength) to be either of ten values between 0.01 and 1000 in a geometric scale.For the Random Forest classifier, we considered the hyperparameter n estimators (number of trees)to be either 1, 2, 3, 5, 8, 10, 15, 20, 25 or 30. For the K-Nearest Neighbors, we considered thehyperparameter n neighbors (number of neighbors) to be either 10, 20, 30 or 40. For the NaiveBayes classifier, we considered the hyperparameter var smoothing (variance added for calculationstability) to be either of ten values between 10-7 to 10 in a geometric scale.In what follows, we describe the procedure that was applied independently for each classifier.For cross-validation, the “No Diabetes” dataset was randomly split into an 80% training datasetand a remaining 20% validation dataset. We trained the classifier for different values of the hy-perparameter on the training set. We then computed the accuracy on the validation set. We kepttrack of the value of the hyperparameter that performed best, i.e., the value that led to the highestaccuracy in the validation set. This process was then iterated 50 times, and we chose the value ofthe hyperparameter that performed best most of the 50 times. The classifier with the best hyper-parameter value chosen above was trained on the entire “No Diabetes” dataset. We then computedthe following performance metrics on our test dataset, i.e., the “Diabetes” dataset: • Accuracy = TP + TNTP + FP + TN + FN • Precision = TPTP + FP 5igure 2: Involvement of the genes in acinar and ductal cells in the cytoplasmic structure (red),actin cytoskeleton formation (blue) and nuclear receptor activity (green).Figure 3: Representative 10x images from different stages of pancreatic cancer in our dataset. • Recall = TPTP + FN • F2 Score = 5 × Precision × Recall4 × Precision + Recall • AUC is the area under the receiver-operator-characteristics curve, i.e., the area under thecurve defined by Recall and Specificity = TNTN + FP .In the above formulas, TP is the number of “true positives”, i.e., subjects with PC, predicted assuch. TN is the number of “true negatives”, i.e., subjects without PC, predicted as such. FP is thenumber of “false positives”, i.e., subjects without PC, predicted as having PC. FN is the numberof “false negatives”, i.e., subjects with PC, predicted as not having PC.The entire process described above was repeated 100 times, which allowed us to assess the meanand standard error of each of the performance metrics.
The study protocol has been approved by the Internal Review Board of Purdue University, andinformed consent was obtained from participants. To test the biological implication behind ourfindings, the nineteen genes were put into String-DB (Figure 2). The connections between thegenes were analyzed including the significance of the gene in the morphology and functionality ofthe nucleus, cytoplasm and cell membrane, as well as expression in the pancreas, specifically in https://string-db.org/ database which contains data fromEntrez Gene, UniProtKB, PathCards and BioGPS.To understand the physiological difference of the key components the nineteen genes were in-volved in, we used a PA804a tissue microarray from US Biomax. A Nikon E600 microscope was thenconnected to a Nikon D500 camera. Pictures of cells of each group in the microarray were taken,and acinar and ductal cells were identified [28]. From an initial set of 300 images, we excluded theones with less than 10% occurrence of acinar and ductal cells, using the ImageJ software. Theremaining thirty images consisted of 5 tissue samples without PC but with diabetes, 5 samples ofwithout PC neither diabetes), 5 samples of early stage PC with diabetes, 5 samples of early stagePC without diabetes, 5 samples of late stage PC with diabetes, and 5 samples of early stage PCwithout diabetes. We compared the cell area and aspect ratio (major axis/minor axis) of the abovethirty images (Figure 3). 7igure 5: Area of the cell, cytoplasm and nucleus for samples with and without diabetes and fordifferent stages of PC. (Each image contained 80-100 acinar or ductal cells.) Overall accuracy, accuracy in late stage PC subjects, accuracy in early stage PC subjects, precision,recall, AUC and F2 scores of the KNN, Naive Bayes and Neural Network models were all > < > ± https://imagej.nih.gov/ij/ > All classifiers except random forest achieved the goal of obtaining an accuracy above 90% to distin-guish between the four groups of subjects, proving that the algorithms were optimal in diagnosingPC. All the classifiers achieved an accuracy greater than 90% for late stage diagnosis; however, thelogistic regression and random forest did not do so for early diagnosis. Therefore, the best classifierswere the KNN, Naive Bayes and Neural Network due to their high performance in both late andearly diagnosis. The logistic regression and random forest classifiers behaved poorly possibly dueto the simplicity of logistic regression and high model complexity of random forests.One of the main reasons why these classifiers performed well is because the nineteen genes wereidentifiers of both early diagnosis and misdiagnosis. The aforementioned genes may control the area9igure 7: Genes involvement in the strings’ activity on the GeneCards database. The resultssuggest that the difference in expression of the nineteen genes may cause the changes shown inFigures 5 and 6.in the cytoplasm, the nucleus, and the acinar or ductal cell which change based on the presenceof PC and/or diabetes. The nucleus grows as cancer progresses because there is an increase in theneed for ribosomes and polysomes which have proteins necessary for the cell growth process, a chainupregulated in cancer [18]. The proteins thus cause the cell to grow rapidly and the cytoplasmdecreases due to nuclear enlargement [45]. This in turn causes the aspect ratio to get larger, ascell growth generates irregularity in shape [14]. The same changes occur in type 2 diabetes as thecell and nucleus grow to produce leptin, a hormone that maintains fat content [31]. However, thisleptin production is less rapid compared to the uncontrollable cell growth in cancer, causing lessdrastic changes in cellular, nuclear, and cytoplasmic areas and aspect ratio [52]. These differencessuggest (i) that the cellular, nuclear, and cytoplasmic changes started early on in the pancreas and(ii) that differences in these areas and aspect ratios are different depending on the occurrence ofPC and/or diabetes.Literature supports the involvement of these nineteen genes with PC and diabetes as PC sampleshave been affiliated with higher expression levels in the ACT genes, the main gene groups consideredfor our machine learning algorithms. The ACT genes are found to be upregulated in 80% of PCsubjects and are significantly upregulated in diabetes [17]. This is because these genes change theshape of the cytoskeleton due to their involvement with alpha actins [42]. These actins are a diversegroup of cytoskeletal proteins, including the alpha and beta spectrins and dystrophins, which areinvolved in binding the actin to the membrane [35]. Thus, our analysis suggests that the ACTgenes are linked to changes in the cellular shape of both PC and diabetes, in acinar and ductalcells.Our novel tool is the first in the literature that achieves a PC diagnostic accuracy of above 90%.Together with the belief that the five-year survival rate would be 40% if the cancer were detectedat an early stage [19], we conclude that our tool can potentially increase the five-year survival rateto 36% (i.e., 40% × Concluding Remarks
The additional applications of the project include public use, affordability and continuous improve-ment. Our results allow for many people to use our diagnostic tool due to affordable and easytechniques to extract gene expression, like RT-qPC. This diagnostic tool is also more affordablecompared to the current CT and MRI scans used to diagnose PC. These algorithms will improvefurther if applied to the real world as more data will be available, making better conclusions.The bioinformatics analysis conducted provides a better understanding of the connection betweenthe nineteen genes and their functionality. This genetic understanding could give rise to preci-sion medicine in order to control the expression of the nineteen genes, leading to specialized PCtreatment.Currently, the classifiers cannot tell the difference between early and late diagnosis. To helpimprove practicality and provide patients for an accurate treatment, the classifiers can be trainedto predict not just whether the patient has PC or not, but also the stage of PC. Further, to solidifythe significance of these genes in PC, specific genes out of the nineteen can be analyzed using insitu hybridization. Since String-DB shows that all these genes are directly correlated, we conjecturethat differences in fluorescence in a couple of genes (e.g., ACTA2 and ABHD12) will prove thatthese genes are differentially expressed in the four different groups.
Acknowledgements.
The authors would like to thank Prof. Brittany Allen-Petersen for thevery helpful feedback, and the PreCollege Research Opportunities Program at Purdue University.
References [1] I. Alldinger, D. Dittert, M. Peiper, A. Fusco, G. Chiappetta, E. Staub, M. L¨ohr, R. Jesnowski,G. Baretton, D. Ockert, H.-D. Saeger, R. Gr¨utzmann, and C. Pilarsky. Gene expression analysisof pancreatic cell lines reveals genes overexpressed in pancreatic cancer.
Pancreatology , 5(4-5):370–379, 2005.[2] D. K. Andersen, . Andren-Sandberg, E. J. Duell, M. Goggins, M. Korc, G. M. Petersen,J. P. Smith, and D. C. Whitcomb. Pancreatitis-diabetes-pancreatic cancer: Summary of anNIDDK-NCI workshop.
Pancreas , 42(8):1227–1237, 2013.[3] S. Aravindan, S. K. Ramraj, S. T. Somasundaram, T. S. Herman, and N. Aravindan.Polyphenols from marine brown algae target radiotherapy-coordinated EMT and stemness-maintenance in residual pancreatic cancer.
Stem Cell Research & Therapy , 6(1):182, 2015.[4] N. Ballian and F. C. Brunicardi. Islet vasculature as a regulator of endocrine pancreas function.
World Journal of Surgery , 31(4):705–714, 2007.[5] N. Bardeesy and R. A. DePinho. Pancreatic cancer biology and genetics.
Nature ReviewsCancer , 2(12):897–909, 2002.[6] J. Cai, L. Lu, Z. Zhang, F. Xing, L. Yang, and Q. Yin. Pancreas segmentation in MRI usinggraph-based decision fusion on convolutional neural networks.
International Conference onMedical Image Computing and Computer-Assisted Intervention , 9901:442–450, 2016.[7] M. Carpelan-Holmstr¨om, S. Nordling, E. Pukkala, R. Sankila, J. L¨uttges, G. Kl¨oppel, andC. Haglund. Does anyone survive pancreatic ductal adenocarcinoma? A nationwide studyre-evaluating the data of the Finnish cancer registry.
Gut , 54(3):385–387, 2005.118] J. C. Carr, E. A. Boese, P. M. Spanheimer, F. S. Dahdaleh, M. Martin, D. Calva, B. Schafer,D. M. Thole, T. Braun, T. M. O’Dorisio, M. S. O’Dorisio, and J. R. Howe. Differentiationof small bowel and pancreatic neuroendocrine tumors by gene-expression profiling.
Surgery ,152(6):998–1007, 2012.[9] K. Chan, S. M. Busque, M. Sailer, C. Stoeger, S. Br¨oer, H. Daniel, I. Rubio-Aliaga, and C. A.Wagner. Loss of function mutation of the SLC38A3 glutamine transporter reveals its criticalrole for amino acid metabolism in the liver, brain, and kidney.
Pfl¨ugers Archiv: EuropeanJournal of Physiology , 468(2):213–227, 2016.[10] S. T. Chari. Detecting early pancreatic cancer: Problems and prospects.
Seminars in Oncology ,34(4):284–294, 2007.[11] D. L. Clarke, S. R. Thomson, T. E. Madiba, and C. Sanyika. Preoperative imaging of pancreaticcancer: A management-oriented approach.
Journal of the American College of Surgeons ,196(1):119–129, 2003.[12] K. Dmitriev, A. E. Kaufman, A. A. Javed, R. H. Hruban, E. K. Fishman, A. M. Lennon, andJ. H. Saltz. Classification of pancreatic cysts in computed tomography images using a randomforest and convolutional neural network ensemble.
International Conference on Medical ImageComputing and Computer-Assisted Intervention , 10435:150–158, 2017.[13] T. Fujii, K. Shimada, S. Anai, K. Fujimoto, and N. Konishi. ALKBH2, a novel AlkB ho-mologue, contributes to human bladder cancer progression by regulating MUC1 expression.
Cancer Science , 104(3):321–327, 2013.[14] M. Gaviraghi, P. Tunici, S. Valensin, M. Rossi, C. Giordano, L. Magnoni, M. Dandrea, L. Mon-tagna, R. Ritelli, A. Scarpa, and A. Bakker. Pancreatic cancer spheres are more than justaggregates of stem marker-positive cells.
Bioscience Reports , 31(1):45–55, 2011.[15] C. Guo, S. Liu, J. Wang, M.-Z. Sun, and F. T. Greenaway. ACTB in cancer.
Clinica ChimicaActa , 417:39–44, 2013.[16] M. Hidalgo. Pancreatic cancer.
New England Journal of Medicine , 362(17):1605–1617, 2010.[17] K. Honda. The biological role of actinin-4 (ACTN4) in malignant phenotypes of cancer.
Cell& Bioscience , 5(1):41, 2015.[18] L. Huang, A. Holtzinger, I. Jagan, M. BeGora, I. Lohse, N. Ngai, C. Nostro, R. Wang, L. B.Muthuswamy, H. C. Crawford, C. Arrowsmith, S. E. Kalloger, D. J. Renouf, A. A. Con-nor, S. Cleary, D. F. Schaeffer, M. Roehrl, M.-S. Tsao, S. Gallinger, G. Keller, and S. K.Muthuswamy. Ductal pancreatic cancer modeling and drug screening using human pluripo-tent stem cell and patient-derived tumor organoids.
Nature Medicine , 21(11):1364–1371, 2015.[19] T. Kamisawa, L. D. Wood, T. Itoi, and K. Takaori. Pancreatic cancer.
The Lancet ,388(10039):73–85, 2016.[20] S. K. Kim. Intercellular signals regulating pancreas development and function.
Genes &Development , 15(2):111–127, 2001.[21] S. K. Kim, H. Kim, D.-h. Lee, T.-s. Kim, T. Kim, C. Chung, G. Y. Koh, H. Kim, and D.-S.Lim. Reversing the intractable nature of pancreatic cancer by selectively targeting ALDH-High, therapy-resistant cancer cells.
PLoS ONE , 8(10):e78130, 2013.1222] L. Kind and P. Kursula. Structural properties and role of the endocannabinoid lipases ABHD6and ABHD12 in lipid signalling and disease.
Amino Acids , 51(2):151–174, 2019.[23] J. Ko, N. Bhagwat, T. Black, S. S. Yee, Y.-J. Na, S. A. Fisher, J. Kim, E. L. Carpenter, B. Z.Stanger, and D. Issadore. miRNA profiling of magnetic nanopore-isolated extracellular vesiclesfor the diagnosis of pancreatic cancer.
Cancer Research , 78(13):3688–3697, 2018.[24] D. Li, K. Xie, R. Wolff, and J. L. Abbruzzese. Pancreatic cancer.
The Lancet , 363(9414):1049–1057, 2004.[25] Y. Li, T. Song, Z. Chen, Y. Wang, J. Zhang, and X. Wang. Pancreatic stellate cells activationand matrix Metallopeptidase 2 expression correlate with lymph node metastasis in pancreaticcarcinoma.
The American Journal of the Medical Sciences , 357(1):16–22, 2019.[26] X. Liu, J. Wang, H. Wang, G. Yin, Y. Liu, X. Lei, and M. Xiang. REG3A acceleratespancreatic cancer cell growth under IL-6-associated inflammatory condition: Involvement of aREG3AJAK2/STAT3 positive feedback loop.
Cancer Letters , 362(1):45–60, 2015.[27] J. Long, G. Luo, C. Liu, X. Cui, K. Satoh, Z. Xiao, B. Zhang, J. Xu, Q. Ni, M. Li, andX. Yu. Development of a unique mouse model for pancreatic cancer lymphatic metastasis.
International Journal of Oncology , 41(5):1662–1668, 2012.[28] D. S. Longnecker. Anatomy and histology of the pancreas.
Pancreapedia: Exocrine PancreasKnowledge Base , 2014.[29] R. C. G. Martin, K. McFarland, S. Ellis, and V. Velanovich. Irreversible electroporation inlocally advanced pancreatic cancer: Potential improved overall survival.
Annals of SurgicalOncology , 20(S3):443–449, 2013.[30] P. M. McMahon, E. F. Halpern, C. Fernandezdel Castillo, J. W. Clark, and G. S. Gazelle. Pan-creatic cancer: Cost-effectiveness of imaging technologies for assessing resectability.
Radiology ,221(1):93–106, 2001.[31] T. H. Meek, M. E. Matsen, M. D. Dorfman, S. J. Guyenet, V. Damian, H. T. Nguyen, G. J.Taborsky, and G. J. Morton. Leptin action in the ventromedial hypothalamic nucleus issufficient, but not necessary, to normalize diabetic hyperglycemia.
Endocrinology , 154(9):3067–3076, 2013.[32] C. P. Molina, Y. C. Lee, and S. F. Badylak. Pancreas whole organ engineering.
Transplantation,Bioengineering, and Regeneration of the Endocrine Pancreas , 2:527–536, 2020.[33] R. D. Morey, R. Hoekstra, J. N. Rouder, M. D. Lee, and E.-J. Wagenmakers. The fallacyof placing confidence in confidence intervals.
Psychonomic Bulletin & Review , 23(1):103–123,2016.[34] T. Nakazawa, H. Ohara, H. Sano, T. Ando, H. Imai, H. Takada, K. Hayashi, Y. Kitajima, andT. Joh. Difficulty in diagnosing autoimmune pancreatitis by imaging findings.
GastrointestinalEndoscopy , 65(1):99–108, 2007.[35] B. J. Perrin and J. M. Ervasti. The actin gene family: Function follows isoform.
Cytoskeleton ,67(10):630–634, 2010. 1336] R. Pezzilli. Is diabetes mellitus a risk factor for pancreatic cancer?
World Journal of Gas-troenterology , 19(30):4861–4866, 2013.[37] M. Porta, X. Fabregat, N. Malats, L. Guarner, A. Carrato, A. de Miguel, L. Ruiz, M. Jariod,S. Costafreda, S. Coll, J. Alguacil, J. M. Corominas, R. Sol, A. Salas, and F. X. Real. Exocrinepancreatic cancer: Symptoms at presentation and their relation to tumour site and stage.
Clinical and Translational Oncology , 7(5):189–197, 2005.[38] M. Quaresma, M. P. Coleman, and B. Rachet. 40-year trends in an index of survival for allcancers combined and survival adjusted for age and sex for each cancer in England and Wales,19712011: A population-based study.
The Lancet , 385(9974):1206–1218, 2015.[39] D. Rajamani and M. K. Bhasin. Identification of key regulators of pancreatic cancer progressionthrough multidimensional systems-level analysis.
Genome Medicine , 8(1):38, 2016.[40] H. E. Rauhala, S. Teppo, S. Niemel¨a, and A. Kallioniemi. Silencing of the ARP2/3 complexdisturbs pancreatic cancer cell migration.
Anticancer Research , 33(1):45–52, 2013.[41] M. Richardson, A. Prakash, and E. Brill. Beyond PageRank: Machine learning for staticranking.
International Conference on World Wide Web , pages 707–715, 2006.[42] J.-B. Rivi`ere, B. W. M. van Bon, A. Hoischen, S. S. Kholmanskikh, B. J. O’Roak, C. Gilissen,S. Gijsen, C. T. Sullivan, S. L. Christian, O. A. Abdul-Rahman, J. F. Atkin, N. Chassaing,V. Drouin-Garraud, A. E. Fry, J.-P. Fryns, K. W. Gripp, M. Kempers, T. Kleefstra, G. M. S.Mancini, M. J. M. Nowaczyk, C. M. A. van Ravenswaaij-Arts, T. Roscioli, M. Marble, J. A.Rosenfeld, V. M. Siu, B. B. A. de Vries, J. Shendure, A. Verloes, J. A. Veltman, H. G. Brunner,M. E. Ross, D. T. Pilz, and W. B. Dobyns. De novo mutations in the actin genes ACTB andACTG1 cause Baraitser-Winter syndrome.
Nature Genetics , 44(4):440–444, 2012.[43] R. P. Sah, S. J. S. Nagpal, D. Mukhopadhyay, and S. T. Chari. New insights into pancre-atic cancer-induced paraneoplastic diabetes.
Nature Reviews Gastroenterology & Hepatology ,10(7):423–433, 2013.[44] X. Shen, L. Han, Z. Ma, C. Chen, W. Duan, S. Yu, P. Li, L. Zhang, W. Li, Q. Xu, and Q. Ma.Aspirin: A potential therapeutic approach in pancreatic cancer.
Current Medicinal Chemistry ,20(33):4153–4162, 2013.[45] J. Son, C. A. Lyssiotis, H. Ying, X. Wang, S. Hua, M. Ligorio, R. M. Perera, C. R. Ferrone,E. Mullarky, N. Shyh-Chang, Y. Kang, J. B. Fleming, N. Bardeesy, J. M. Asara, M. C. Haigis,R. A. DePinho, L. C. Cantley, and A. C. Kimmelman. Glutamine supports pancreatic cancergrowth through a KRAS-regulated metabolic pathway.
Nature , 496(7443):101–105, 2013.[46] D. Tang, Q. Wu, Z. Yuan, J. Xu, H. Zhang, Z. Jin, Q. Zhang, M. Xu, Z. Wang, Z. Dai, H. Fang,Z. Li, C. Lin, C. Shi, M. Xu, X. Sun, and D. Wang. Identification of key pathways and geneschanges in pancreatic cancer cells (BXPC-3) after cross-talk with primary pancreatic stellatecells using bioinformatics analysis.
Neoplasma , 66(05):681–693, 2019.[47] X. Yan, H. Wan, X. Hao, T. Lan, W. Li, L. Xu, K. Yuan, and H. Wu. Importance of geneexpression signatures in pancreatic cancer prognosis and the establishment of a predictionmodel.
Cancer Management and Research , 11:273–283, 2019.1448] M. Yu, L. Zhang, Y. Liu, D. Liu, and Z. Guo. Retinoic acid induces differentiation of mouseF9 embryonic carcinoma cell by modulating the miR-485 targeting of ABHD2.
InternationalJournal of Molecular Sciences , 20(9):2071, 2019.[49] N. Zelcer. Liver X receptors as integrators of metabolic and inflammatory signaling.
Journalof Clinical Investigation , 116(3):607–614, 2006.[50] J. Zhang and S. Zhang. Discovery of cancer common and specific driver gene sets.
NucleicAcids Research , 45(10):e86–e86, 2017.[51] L. Zhang, S. Sanagapalli, and A. Stoita. Challenges in diagnosis of pancreatic cancer.
WorldJournal of Gastroenterology , 24(19):2047–2060, 2018.[52] D. Zhao, S.-W. Zou, Y. Liu, X. Zhou, Y. Mo, P. Wang, Y.-H. Xu, B. Dong, Y. Xiong, Q.-Y.Lei, and K.-L. Guan. Lysine-5 acetylation negatively regulates lactate dehydrogenase A andis decreased in pancreatic cancer.
Cancer Cell , 23(4):464–476, 2013.[53] Y. Zhou, L. Xie, W. Shen, Y. Wang, E. Fishman, and A. Yuille. A fixed-point model forpancreas segmentation in abdominal CT scans.
International Conference on Medical ImageComputing and Computer-Assisted Intervention , pages 693–701, 2017.[54] G. Zoppi, G. Andreotti, F. Pajno-Ferrara, D. M. Njai, and D. Gaburro. Exocrine pancreasfunction in premature and full term neonates.