Abstract
Pancreatic cancer (PC) is the fourth leading cause of cancer death in the United States due to its five-year survival rate of 10%. Late diagnosis, affiliated with the asymptomatic nature in early 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%) diagnostic accuracy, 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 in PC 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 PC subjects at varying stages to ensure validity. Naive Bayes, Neural Network and K-Nearest Neighbors models achieved the highest testing accuracy of around 92.6%. Second, the biological implication of the nineteen genes was investigated using bioinformatics tools. It was found that these genes were significantly involved in regulating the cytoplasm, cytoskeleton and nuclear receptor activity in the pancreas, specifically in acinar and ductal cells. Our novel tool is the first in the literature that achieves a PC diagnostic accuracy of above 90%, having the potential to significantly improve the detection of PC in the background of diabetes and increase the five-year survival rate.
The Clinical Challenge: Late Diagnosis and Low Survival
Pancreatic cancer (PC) remains one of the most lethal malignancies, primarily because it is notoriously difficult to detect in its early, treatable stages. The pancreas is located deep within the abdominal cavity, and early-stage tumors are often asymptomatic. Consequently, the majority of patients are diagnosed only after the disease has metastasized (Stage IV), contributing to a dismal five-year survival rate of approximately 10%.
A significant clinical hurdle in diagnosis is the complex relationship between PC and diabetes. Approximately 80% of PC patients present with diabetes, and nearly 20% of PC cases are initially misdiagnosed as simple diabetes. Prior genetic screening attempts have struggled with low accuracy (70-75%) because they failed to distinguish the subtle gene expression differences between these two conditions. This research seeks to bridge that gap by identifying a gene signature that remains robust even in the presence of diabetes.
Methodology: A Dual-Phase Approach
The study employed a two-phase approach combining computational predictive modeling with biological validation.
Phase 1: Machine Learning Classification
The researchers evaluated five classifiers: Logistic Regression, K-Nearest Neighbors (KNN), Random Forest, Naive Bayes, and Neural Networks. The models were trained on 398 samples retrieved from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO). To ensure the tool could overcome the "noise" of diabetes, data was stratified into four groups: subjects with both conditions, subjects with only diabetes, subjects with only PC, and healthy controls. The models focused on the expression levels of 19 specific genes identified as being overexpressed in PC.
Phase 2: Bioinformatics and Morphological Validation
To understand the biological drivers of the model, tools like String-DB and GeneCards were used to map pathways. Additionally, physiological analysis using tissue microarrays and imaging software compared cell area, cytoplasmic area, and nuclear morphology across different stages of the disease.
Results: High-Accuracy Early Detection
The results represent a significant breakthrough in PC diagnostics. KNN, Naive Bayes, and Neural Network models all achieved a testing accuracy of approximately 92.6%. This is the first instance in the literature of a PC diagnostic tool exceeding the 90% threshold.
Crucially, the performance remained high for early-stage detection. While Logistic Regression and Random Forest were effective for late-stage diagnosis, their accuracy dipped significantly for early-stage PC. In contrast, the top-performing models maintained high sensitivity across all stages, proving their utility as an early screening tool.
Biological Insights and Future Directions
The 19-gene signature (including the ACT and ALDH families) was found to be primarily involved in regulating the cytoplasm, actin cytoskeleton, and nuclear receptor activity. This provides a biological basis for the model's success: these genes control the physical structure of the cell. Physiological measurements confirmed that as PC progresses, there is a measurable increase in cell and nucleus area and a decrease in cytoplasmic area, providing a strong ground truth for the genetic data.
While the tool is excellent at binary classification (PC vs. No PC), future work will focus on refining the model to distinguish between specific cancer stages. The authors propose applying these algorithms to larger, real-world clinical datasets and using in situ hybridization to further validate high-impact genes like ACTA2 and ABHD12.
Summary updated