Pancreatic cancer (PACA) has a five-year survival rate of approximately 11% and ranks among the leading causes of cancer deaths worldwide. Its prognosis has barely improved over the last decade despite advances in surgery and chemotherapy.
The dominant tool for prognosis and treatment planning remains TNM staging, which classifies patients by tumor size, lymph node involvement, and metastasis. While TNM staging provides a useful population-level framework, it fails to capture the biological heterogeneity within each stage - two patients with the same TNM stage can have dramatically different survival times due to differences in their tumor's molecular biology.
There is therefore an urgent need for molecular biomarkers that go beyond anatomy-based staging. An ideal prognostic signature would accurately predict survival, identify patients likely to respond to specific treatments, and remain valid across diverse patient populations - not just in the institution where it was developed.
The research team developed the Artificial Intelligence-Derived Prognostic Signature (AIDPS) using a systematic, unbiased approach to machine learning model selection. Rather than choosing a single algorithm based on prior preference, they tested 10 machine learning algorithms combined in 76 different configurations, evaluated on nine independent testing cohorts.
The process began with univariate Cox regression across 15,288 genes to identify 32 consensus prognostic genes (CPGs) - genes whose expression was significantly associated with survival in both the training cohort and multiple testing cohorts. This stringent selection ensures the starting gene set is genuinely prognostic, not just correlated by chance in one dataset.
The 32 CPGs were then fed into all 76 algorithm combinations, and each model's performance was evaluated using the concordance index (C-index) - a measure of how well a model ranks patients by predicted survival risk. The combination of CoxBoost and Survival-SVM with the highest average C-index (0.675) across nine testing cohorts was selected as the final AIDPS model, resulting in a nine-gene signature.
A major strength of this study is its extraordinary validation scale. AIDPS was evaluated across 13 independent multicenter cohorts drawn from diverse international datasets including TCGA-PAAD, multiple GEO datasets (GSE62452, GSE28735, GSE78229, GSE79668, GSE85916), and European cohorts (PACA-CA-Seq, E-MTAB-6134). Combined, these cohorts included 1,280 patients for development and over 290 patients in external validation.
The breadth of validation addresses the most common criticism of published cancer prognostic signatures: that they overfit to their training data and fail to generalize. By requiring consistent performance across 13 cohorts from different continents, institutions, and time periods, the authors established that AIDPS captures a genuine biological signal rather than a statistical artifact.
Beyond survival, the model was also evaluated for predicting relapse-free survival (RFS) - the time until cancer returns after treatment. Predicting recurrence is clinically distinct from predicting overall survival and provides additional value for guiding decisions about surveillance intensity and adjuvant therapy.
One of the study's most compelling analyses was a direct comparison between AIDPS and 86 previously published prognostic gene signatures for pancreatic cancer. The researchers collected all available mRNA and lncRNA signatures from the literature and re-evaluated them uniformly in the same cohorts used to validate AIDPS.
The results revealed a common problem with published signatures: most performed well only in the specific cohort used for their original development (often TCGA-PAAD), but poorly in other cohorts. This pattern is the hallmark of overfitting - the model has learned idiosyncrasies of one dataset rather than a generalizable biological signal.
AIDPS, by contrast, ranked first in four cohorts, second or third in five cohorts, and maintained C-index values above 0.65 across virtually all 13 cohorts. Only two other signatures (a 20-gene model by Demirkol CS and a 6-gene model by Stratford JK) showed comparable consistency, but both had lower average performance than AIDPS across the complete cohort set.
Beyond survival prediction, the researchers investigated the biological meaning of the AIDPS score by analyzing the immune microenvironment of high- vs. low-AIDPS tumors. Using single-sample gene set enrichment analysis (ssGSEA), they quantified the abundance of 28 immune cell types in each tumor.
Patients in the low-AIDPS group had significantly higher infiltration of immune cells including activated CD4+ T cells, CD8+ T cells, and natural killer cells. They also showed higher expression of 27 different immune checkpoint molecules including PD-L1 (CD274), CTLA-4 pathway components, and other co-stimulatory and co-inhibitory molecules.
This immune-active state in low-AIDPS tumors suggests these patients may be more likely to respond to immune checkpoint inhibitor therapy - drugs that block the checkpoint molecules and unleash the immune system against the tumor. If validated, this finding could help identify the subset of pancreatic cancer patients most likely to benefit from immunotherapy, which has so far failed in unselected PDAC populations.
The study performed comprehensive genomic analysis of the two AIDPS patient groups. Low-AIDPS tumors showed significantly higher tumor mutation burden (TMB) - the total number of mutations present in the tumor genome. Higher TMB has been associated with better response to immunotherapy in several cancer types, consistent with the immune-active phenotype observed in low-AIDPS tumors.
Common pancreatic cancer driver mutations were distributed differently between the groups. KRAS, TP53, and CDKN2A mutations were more frequent in the low-AIDPS group, while SMAD4, TTN, and RNF43 mutations were more common in the high-AIDPS group. These differences provide molecular context for why the two groups behave differently clinically.
DNA methylation analysis identified four methylation-driven genes (MDGs) whose expression levels were inversely correlated with their methylation status in PACA. Two genes (MAP3K8 and PCDH7) were hypermethylated in the high-AIDPS group, while two others (PCDHB1 and SPAG6) showed the opposite pattern. Higher MAP3K8 and PCDH7 methylation was associated with significantly longer survival.
After dimensionality reduction through the CoxBoost-based feature selection step, nine genes were retained in the final AIDPS signature. These genes were identified through the intersection of consistent prognostic association across all 10 development cohorts and optimal performance in the CoxBoost algorithm.
The nine AIDPS genes include SELENBP1 (selenium binding protein) and PLCB4 (phospholipase C beta 4), which showed positive correlation with the AIDPS score, alongside seven genes negatively correlated with AIDPS including DCBLD2, PRR11, UNC13D, EREG, ADM, TGM2, and one additional gene. Each gene has established roles in cellular signaling, metabolism, or immune function.
The signature score is calculated using Survival-SVM, a support vector machine algorithm adapted for survival time prediction. The algorithm assigns a continuous risk score to each patient based on the weighted expression of all nine genes. Patients are then classified into high- or low-AIDPS groups using the median score as a cutoff, though continuous scoring is preferred for clinical applications.
The clinical promise of AIDPS lies in its potential to move pancreatic cancer management beyond one-size-fits-all approaches. By stratifying patients into biologically distinct high- and low-risk groups, AIDPS could inform decisions about the intensity of treatment, the appropriateness of immunotherapy, and the need for aggressive surveillance.
The finding that low-AIDPS patients have an immune-active tumor microenvironment is particularly actionable. These patients may benefit most from anti-PD-1/PD-L1 checkpoint inhibitor therapy, which has shown minimal benefit in unselected PDAC patients but might perform better when restricted to those with immune-infiltrated tumors. AIDPS scoring could serve as a selection biomarker for immunotherapy clinical trials.
The authors also note that drug repurposing analysis identified panobinostat - a histone deacetylase (HDAC) inhibitor - as a potential therapeutic agent specifically for high-AIDPS patients, whose tumors were enriched for pathways sensitive to this drug class. This represents an example of how AI-derived molecular stratification can generate novel, testable therapeutic hypotheses.