Letter to the Editor: Explainable artificial intelligence helps early cancer diagnosis via extracellular vesicle long RNA

Abstract

Pancreatic ductal adenocarcinoma is a highly aggressive malignancy characterized by poor prognosis and a lack of effective early diagnostic methods. While liquid biopsy has emerged as a promising non-invasive strategy, many existing models rely on “black-box” machine learning algorithms that lack biological transparency. This work discussed and reviewed a methodological advancement by Liu and Zhang, who developed an interpretable two-layer machine learning framework utilizing plasma extracellular vesicle long RNA as a biomarker source. While limitations and challenges still exist, by combining the stability of extracellular vesicle long RNA with an interpretable computational approach, Liu and Zhang’s work, published in the recent issue of World Journal of Gastrointestinal Oncology, represents a significant step forward in developing transparent and highly accurate tools for the early detection and differential diagnosis of pancreatic ductal adenocarcinoma.

Key Words: Pancreatic ductal adenocarcinoma; Liquid biopsy; Extracellular vesicles; Long RNA; Explainable artificial intelligence; Machine learning

Core Tip: Interpretable machine learning applied to extracellular vesicle long ribonucleic acid profiles is a promising direction for earlier pancreatic cancer detection and for differentiating pancreatic ductal adenocarcinoma from chronic pancreatitis. To accelerate clinical readiness, reported discrimination should be benchmarked against standard diagnostic pathways, validated across multiple centers with standardized pre-analytical processing, and accompanied by robustness checks that confirm explanation stability under realistic perturbations.

TO THE EDITOR

Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal malignancies, primarily arising from the pancreatic ductal epithelium and acinar cells. This cancer is highly aggressive, and its insidious onset makes early diagnosis particularly difficult. The disease progresses rapidly, and the average survival time is short. Consequently, PDAC represents a significant clinical challenge characterized by extremely low early diagnosis rates and poor prognosis. Liquid biopsy offers a non-invasive approach capable of capturing tumor-derived molecules from biofluids and has emerged as a promising strategy for early PDAC detection. Various analytes have been explored, circulating free DNA (cfDNA), cell-free RNA, and extracellular vesicles (EVs). Among these, EVs are particularly attractive because they are stable carriers of RNA and protein cargo reflecting the molecular state of tumor cells.

COMMENTARY ON THE PUBLISHED STUDY

In the research study, Yu et al[1] first demonstrated that plasma EV long RNAs could discriminate PDAC from healthy controls with high accuracy [area under the receiver operating characteristic curve (AUC) = 0.94]. Subsequent studies extended this finding: (1) Han et al[2] showed that plasma EV mRNA profiles capture diagnostic and prognostic information in PDAC; and (2) He et al[3] identified plasma-derived exosomal lncRNAs as novel non-invasive biomarkers. In parallel, strong diagnostic performances have been achieved using cfDNA methylation and fragmentomic signatures[4], as well as cell-free RNA-based liquid biopsies[5], underscoring the rapidly growing potential of multi-analyte approaches for early cancer detection.

Within this evolving landscape, the study by Liu and Zhang[6], published in the recent issue of World Journal of Gastrointestinal Oncology, developed a highly sensitive and specific non-invasive diagnostic tool for the early detection and differential diagnosis of PDAC. The authors used plasma extracellular vesicle long RNA (EvlRNA) as the biomarker source and constructed an interpretable two-layer machine learning model. The first layer performed feature selection by comparing various dimensionality-reduction methods (MDA, SVD-PCA, FastHCS Index), while the second layer conducted classification using multiple algorithms (random forest, support vector machine, deep learning, XGBoost). Their multimodal model demonstrated superior performance, particularly in distinguishing PDAC from chronic pancreatitis (CP), a clinically challenging comparator. For instance, in the PDAC vs CP task, the MDA-DL model achieved an accuracy of 89.21% and an AUC of 0.9493.

To aid clinical interpretation, we encourage explicit benchmarking against standard diagnostic components for PDAC, including carbohydrate antigen 19-9 and imaging-based evaluation (e.g., computed tomography/magnetic resonance imaging and endoscopic ultrasound). Reporting how the model’s AUC, sensitivity, and specificity compare to these references would clarify its incremental value, particularly in clinically realistic comparators such as CP, ideally within matched cohorts[7].

Comparable studies have also leveraged machine learning for PDAC liquid-biopsy diagnostics, although most rely on black-box models. For example, Liu et al[8] developed a small-EV mRNA-based diagnostic score using deep learning that achieved high accuracy but offered limited interpretability. Similarly, cfDNA methylation-based early-detection frameworks employing neural networks or ensemble learners have reported AUCs above 0.90[4], yet they provide minimal biological transparency. These examples highlight how the interpretability of Liu and Zhang’s framework[6] represents a meaningful methodological advance.

The study exhibits notable strengths and innovations. First, it employs a cutting-edge biomarker source plasma EvlRNA which is both stable and reflective of tumor-microenvironment interactions, offering a rich reservoir for non-invasive PDAC biomarkers[1]. Second, the two-layer machine-learning architecture, integrating systematic comparisons across feature-extraction and classifier levels, enhances robustness. Importantly, its focus on interpretability enables biological insight into which EvlRNA features drive diagnostic power – an aspect often lacking in “black-box” artificial-intelligence applications[9]. Third, the model’s high AUC (approximately 0.95) demonstrates that combining EvlRNA biomarkers with interpretable machine learning yields substantial diagnostic potential.

While the results are encouraging, certain challenges remain before clinical translation. As the authors acknowledge, a persistent limitation of machine-learning models is overfitting and lack of robustness. The study reports strong internal validation but lacks independent external validation cohorts from other centers or time points. Broader validation across diverse populations and clinical settings is necessary to confirm generalizability and reliability[10]. Several limitations warrant more direct discussion. Performance may be sensitive to cohort spectrum (healthy vs symptomatic controls) and to pre-analytical variability in extracellular vesicle workflows, for which standardized reporting has been strongly recommended. Moreover, interpretability claims would be strengthened by reporting the stability of selected features/attributions under resampling, alongside prospective multi-center external validation to support generalizability.

CONCLUSION

Collectively, recent advances in liquid biopsy–based diagnostics – including EV-derived long RNAs, lncRNAs, and cfDNA methylation – are converging toward more accurate and minimally invasive early detection of PDAC. The work by Liu and Zhang[6] represents an important step forward by coupling interpretability with state-of-the-art machine learning, addressing a key gap in the field. Rather than merely echoing existing trends, their framework bridges the divide between biological insight and computational performance. We encourage the authors and the field at large to pursue multi-center validation and multi-analyte integration, which could further advance the clinical readiness of EV-based diagnostics for PDAC.