Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers, largely because it is caught late and spreads quickly. Doctors need reliable biological markers to tell them which patients are at risk of the cancer spreading to other parts of the body, but so far no dependable markers for metastasis have been established.
Machine learning has been widely used in cancer research, but a major problem is that results often cannot be reproduced when tested on new patients from different hospitals. This study set out to create a rigorous, reproducible pipeline that could consistently identify gene-based biomarker candidates for PDAC metastasis.
The team gathered RNA sequencing data from five large public cancer databases, pooling samples from PDAC patients to maximize statistical power. Patients with lymph node or distant metastasis formed the 'metastasis group,' while those without formed the 'non-metastasis group.' Technical differences between datasets were corrected using batch-removal methods before analysis.
Variable selection used a three-step process within 10-fold cross-validation: LASSO regression to narrow down candidates, followed by Boruta and varSelRF algorithms to refine the list. Only genes selected in at least 80% of 100 models and across at least five folds were considered robust enough to move forward.
A Random Forest model was then built using the selected genes and tested on completely independent validation datasets. The researchers also compared performance against XGBoost models and randomly chosen gene sets to confirm their findings were not due to chance.
The pipeline consistently selected 15 genes across all training folds and algorithms. When tested on the independent validation dataset, the Random Forest model built from these 15 genes achieved an ROC-AUC of 0.839 and, when re-calibrated on validation data, improved to 0.941, showing strong discriminative ability between metastatic and non-metastatic PDAC patients.
Compared to models using 15 randomly chosen genes, the selected 15 performed dramatically better across nearly all evaluation metrics. The Random Forest model also marginally outperformed XGBoost, though both showed comparable results, lending further confidence to the biological signal in the selected genes.
Pathway analyses showed the 15 genes were linked to immune response, extracellular matrix remodeling, metabolic reprogramming, and angiogenesis — all biological processes known to drive cancer metastasis.
A detailed literature review of the 15 selected genes revealed they interact across immune, metabolic, and structural pathways implicated in PDAC progression. Several are associated with drug targets or have been detected at the protein level in cancer tissue, making them feasible candidates for future laboratory testing.
The researchers emphasize that their pipeline is generalizable to other diseases and omics data types, not just PDAC gene expression. Following ML best practices — including proper batch correction, cross-validation, and separate validation sets — is essential to avoid the reproducibility problems that have plagued previous biomarker studies.
This study demonstrates that combining multiple variable selection algorithms, large pooled datasets, and rigorous cross-validation can produce robust and reproducible biomarker candidates. The 15-gene composite biomarker for PDAC metastasis is a promising starting point for laboratory validation and eventual clinical application.
By relying entirely on publicly available data and sharing all code on GitHub, the authors ensure their work can be independently verified and extended. This approach represents a sustainable, environmentally mindful model for future biomarker discovery in cancer research.