Explainable AI prioritizes serum microRNAs for pancreatic cancer detection

A retrospective computational study of 1,101 serum samples has identified circulating microRNAs that may help distinguish pancreatic cancer from healthy controls, reporting high performance from a Bayesian stacked ensemble coupled with explainable artificial intelligence. The analysis combined five public Gene Expression Omnibus datasets generated between 2015 and 2019, corrected technical differences between cohorts, and used differential expression analysis to find 148 dysregulated microRNAs. A model trained on these features achieved an area under the receiver operating characteristic curve of 0.97 on a held-out test set of 221 samples, with 94% sensitivity, 99% specificity, 97.28% accuracy and 97% precision. These are computational estimates from archived study cohorts, not results from a prospective diagnostic trial, and the candidates still require independent experimental and clinical validation.

The study drew from GEO series GSE59856, GSE106817, GSE112264, GSE113486 and GSE124158, which together contained pancreatic cancer and healthy-control serum profiles measured with 3D Gene human miRNA microarray platforms. The sample distribution was not uniform across studies: GSE59856 contained 150 controls and 100 pancreatic cancer samples, GSE106817 115 controls and 200 cancer samples, GSE112264 50 controls and 41 cancer samples, GSE113486 40 controls and 100 cancer samples, and GSE124158 30 controls and 275 cancer samples. Such variation motivated the researchers to harmonize identifier conventions, impute missing values, align features across platforms and apply ComBat batch correction, an empirical-Bayes procedure designed to reduce nonbiological differences among datasets. The analysis started with normalized expression matrices already processed by the original studies rather than raw fluorescence measurements.

After log transformation, the researchers used the limma statistical package to compare cancer and control samples. The researchers used a limma differential expression workflow, with feature selection based on a false discovery rate threshold below 0.05 and an absolute log2 fold-change threshold of at least 2. The researchers treated the 148 resulting differentially expressed microRNAs as the input features for predictive modeling. Principal component analysis showed separation between the two groups, with the first two components accounting for 17.5% and 14.6% of variance, respectively. The researchers presented this as evidence of discriminatory information in the integrated expression profiles, but the retrospective design, harmonization of archival data and lack of uniformly available clinical variables limit how directly that separation can be interpreted in an unselected patient population.

For classification, the researchers partitioned the combined data into an 80% training set and a 20% test set, retaining proportional representation of cancer and control samples. The training set was further divided into stack-training and validation subsets. Five base algorithms—Extreme Gradient Boosting, Random Forest, Support Vector Classifier, Naïve Bayes and k-nearest neighbors—generated probabilities for the cancer class. The researchers combined those outputs through a Bayesian stacking ensemble, estimating non-negative weights by minimizing binary cross-entropy with the Sequential Least Squares Programming algorithm and constraining the weights to form a convex combination. Synthetic Minority Over-Sampling Technique was applied only within the training partitions to address the moderate class imbalance, while the test set was not oversampled. The researchers also introduced controlled label noise into 3% of training labels as an auxiliary robustness exercise; they emphasized that this was not an estimate of the real-world diagnostic misclassification rate.

On the held-out set, the ensemble’s receiver operating characteristic curve rose steeply at low false-positive rates, consistent with the reported AUC of 0.97. Its sensitivity of 94% indicates that the model identified 94% of the cancer samples in that test set, while its specificity of 99% reflects the proportion of controls correctly classified. The reported F1 score was 96%. These figures were obtained from samples drawn from the same five archived studies used to build the broader dataset, however, rather than from a strictly independent external cohort. The researchers explicitly noted a potential source of information leakage because differential expression analysis was performed before the train-test split. Although the authors characterized this as a biological feature-discovery step rather than predictive selection, they said future work should use nested cross-validation or fold-wise feature selection to test robustness more rigorously.

To make the model more transparent, the researchers applied SHAP, or Shapley Additive Explanations, to the XGBoost component of the ensemble. SHAP decomposes individual predictions into feature contributions and can rank variables by their average absolute influence across samples. The analysis used 200 samples from the SMOTE-balanced training distribution as a reference background and calculated values for all features in the test set. The resulting ranking highlighted 20 microRNAs, including hsa-miR-125a-3p, hsa-miR-6839-5p, hsa-miR-1290, hsa-miR-614, hsa-miR-4436a and hsa-miR-497-5p. hsa-miR-125a-3p ranked first and was downregulated in cancer samples, with a log2 fold change of −2.944 and an adjusted p value below 0.001. Low expression of this microRNA was associated with a stronger contribution toward a cancer classification.

Several other leading candidates were upregulated in the integrated data. hsa-miR-1290 had a log2 fold change of 2.774, hsa-miR-614 had 2.250 and hsa-miR-128-2-5p had 2.522, each with an adjusted p value below 0.001. hsa-miR-497-5p was among the downregulated features, with a log2 fold change of −2.665 and an adjusted p value below 0.001. The overlap with previously reported pancreatic-cancer markers was interpreted as supporting biological plausibility: earlier work had associated serum miR-1290 with early pancreatic cancer and miR-125a-3p with pancreatic and biliary cancer. The researchers also emphasized the limits of that comparison. A high SHAP value describes how strongly a feature influences this particular model; it does not demonstrate that the microRNA causes cancer, acts upstream in tumor development or would independently improve patient outcomes.

The team next examined predicted or reported targets for the 20 SHAP-ranked microRNAs using DIANA-miRPath v4.0, miRDB, miRTarBase 2022, miRBase v22.1 and KEGG resources. Twelve microRNAs had DIANA-associated targets connected to pathways relevant to pancreatic cancer, including genes involved in cell-cycle control, MAPK signaling, apoptosis, DNA-damage responses, angiogenesis and inflammation. The highlighted targets included CDK4, CDK6, CCND1, CDKN1A, MAPK1, RAF1, MAP2K1, TP53, BCL2L1, E2F3, VEGFA and RELA. The researchers reported that hsa-miR-497-5p targeted several components of the MAPK pathway, while hsa-miR-125a-3p was linked to CDK4 and VEGFA. Six candidates not represented in DIANA were examined through miRDB, where their predicted targets were associated with cytoskeletal remodeling, ubiquitination, transcriptional regulation and metabolic adaptation. These database relationships generate mechanistic hypotheses but do not experimentally verify the proposed interactions.

Pancreatic cancer remains difficult to detect early because it can progress without obvious symptoms, while benign pancreatic disorders can resemble it clinically and radiologically, and definitive diagnosis may involve invasive tissue sampling. That context makes minimally invasive blood markers attractive, but the new results do not establish a ready screening test. The study used a case-control design comparing pancreatic cancer samples with healthy controls rather than a prospective screening population, and it did not evaluate performance against benign pancreatitis, other cancers or patients with other pancreatic diseases. Disease stage, tumor subtype, demographic characteristics, treatment history and other clinical variables were not consistently available across the five public datasets, so the resulting signature cannot be interpreted as specific to a stage, subtype or treatment group. Its reported diagnostic performance may therefore be optimistic in routine care, where disease prevalence, spectrum and sample handling differ from the research setting.

Translation would also require a reproducible laboratory assay, clinically meaningful decision thresholds and prospective testing in cohorts that represent the intended screening population. The 20-microRNA panel may need to be reduced to a smaller set to limit cost and complexity, and its performance should be compared with, or combined with, established markers such as CA19-9 rather than assumed to replace existing testing. Orthogonal methods, including RT-qPCR or digital PCR, would be needed to confirm that the prioritized microRNAs can be measured reliably in serum. Independent cohorts, standardized collection and processing, assays against benign pancreatic disease and functional experiments would then test whether the candidates retain their predictive value and whether their proposed target relationships have biological meaning. The study therefore offers a detailed and reproducible computational shortlist rather than a validated pancreatic-cancer diagnostic: its strongest contribution is the combination of multi-cohort data, transparent model attribution and pathway-based context, while its clinical promise remains to be tested.

Subject of Research: Computational identification of circulating microRNA candidates for pancreatic cancer detection

Article Title: Explainable artificial intelligence identifies circulating microRNA candidate biomarkers for liquid biopsy-based pancreatic cancer detection

Article References: Mandal, S., Ansari, M. A., & Chanu, W. P. (2026). Explainable artificial intelligence identifies circulating microRNA candidate biomarkers for liquid biopsy-based pancreatic cancer detection. Frontiers in Artificial Intelligence, 9, Article 1810551. https://doi.org/10.3389/frai.2026.1810551

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1810551

Keywords: pancreatic cancer, liquid biopsy, circulating microRNAs, explainable artificial intelligence, machine learning, SHAP, biomarker discovery, serum diagnostics, Explainable, artificial, intelligence, identifies