Prognostic biomarkers for pancreatic ductal adenocarcinoma: A contemporary systematic review and modelling study

Abstract

BACKGROUND

Despite extensive biomarker discovery efforts in pancreatic ductal adenocarcinoma (PDAC), few biomarkers have been implemented clinically, and outcomes in clinical trials remain modest.

AIM

To evaluate prognostic transcriptomic and proteomic biomarkers in PDAC, assess clinical trial integration, and develop an exploratory molecular prognostic signature.

METHODS

A systematic review of studies published between 1 January 2025 and 6 March 2026 was conducted. Eligible studies evaluated tumor-derived transcriptomic or proteomic biomarkers in multivariable survival analyses or clinical trials. In parallel, Clinical Proteomic Tumor Analysis Consortium PDAC data were analyzed to identify candidate proteins. The corresponding genes were evaluated using least absolute shrinkage and selection operator (LASSO) Cox regression in The Cancer Genome Atlas pancreatic adenocarcinoma cohort to develop a prognostic signature, externally validated in GSE71729.

RESULTS

A total of 71 records met the predefined eligibility criteria, comprising 57 biomarker studies and 14 clinical trials. All 57 biomarker studies reported independent prognostic associations in adjusted multivariable models. None of the 14 clinical trials incorporated molecular biomarkers to determine eligibility, stratify randomization, or define a prespecified endpoint, indicating a substantial translational gap. The resulting 5-gene LASSO signature showed a strong association with overall survival in the derivation cohort but only modest discrimination. Its prespecified continuous risk score failed external validation. Although median-based risk stratification in the validation cohort was nominally significant, the association did not remain significant after correction for threshold selection.

CONCLUSION

The 5-gene signature in this study failed external validation. Contemporary PDAC biomarker discovery has not yet produced clinically reproducible prognostic tools.

Key Words: Pancreatic ductal adenocarcinoma; Prognostic biomarkers; Proteomics; Survival analysis; Transcriptomics

Core Tip: This study integrates a contemporary systematic review with exploratory modelling to examine why prognostic biomarker discovery in pancreatic ductal adenocarcinoma has not translated into clinical practice. Although numerous tumor-derived transcriptomic and proteomic biomarkers showed independent survival associations, robust external validation and incorporation into clinical trials remained uncommon. An exploratory 5-gene molecular signature showed a strong survival association but only modest discrimination in the derivation cohort, and its continuous risk score failed external validation. These findings highlight reproducibility, prespecified validation, and biomarker-guided trial design as priorities for clinically useful pancreatic ductal adenocarcinoma risk stratification.

INTRODUCTION

Pancreatic ductal adenocarcinoma (PDAC) remains a highly lethal malignancy, with persistently poor survival rates despite advances in systemic therapy and surgical management[1,2]. This unfavorable prognosis is largely attributable to late-stage diagnosis, limited responsiveness to treatment, and substantial tumor heterogeneity[3,4]. In recent years, clinical trials in PDAC have yielded only modest survival benefits, suggesting that treatment effects are unlikely to be consistently clinically meaningful or statistically significant across diverse patient populations[5,6].

Routine PDAC management continues to rely on a single circulating marker, carbohydrate antigen 19-9 (CA19-9). However, its clinical utility is limited by suboptimal sensitivity and specificity, false elevations in inflammatory and biliary diseases, and an inability to capture the biological heterogeneity of PDAC[7]. An additional limitation is related to Lewis antigen status. CA19-9 synthesis depends on Lewis antigen expression, and approximately 5%-10% of individuals who are Lewis antigen-negative cannot produce CA19-9, even in the presence of advanced disease. Consequently, a substantial tumor burden may remain undetected, potentially delaying disease recognition and further limiting the clinical utility of this biomarker[8]. Moreover, the highly immunosuppressive tumor microenvironment of PDAC may limit the clinical utility of immune-based biomarkers in unselected patient populations, while currently available peripheral blood markers remain inadequate for accurate diagnosis, biological stratification, and disease monitoring. These limitations highlight a substantial unmet need for tissue-derived molecular biomarkers that more faithfully reflect underlying tumor biology. Proteomic biomarkers may provide complementary insights into tumor phenotype, pancreatic differentiation status, and exocrine dysfunction, while capturing broader biological processes implicated in PDAC progression, including dedifferentiation, stromal remodeling, and metabolic reprogramming[9]. Accordingly, transcriptomic, proteomic, and multi-omics studies have identified a number of candidate prognostic biomarkers associated with survival in patients with PDAC[10]. However, few of these biomarkers have undergone external validation or been translated into clinical practice, underscoring a substantial unmet clinical need[9,11].

Despite these advances, evidence supporting the prognostic value of molecular biomarkers in PDAC remains heterogeneous and is frequently limited by retrospective study designs, inconsistent approaches to multivariable analysis, inadequate external validation, and variability in reporting quality[12,13]. Consequently, the robustness, reproducibility, and clinical applicability of published biomarkers remain uncertain. A systematic appraisal of methodological quality and validation strategies is therefore needed to identify biomarkers with the greatest potential for clinical translation[4,12].

Against this background, this study aimed to systematically evaluate the independent prognostic value and methodological robustness of tumor-derived molecular biomarkers in PDAC. The review was designed as a contemporary appraisal of recent biomarker discovery pipelines and was restricted to studies published between 1 January 2025 and 6 March 2026. The primary review question was: ‘Among adults with PDAC, which molecular biomarkers are independently associated with survival after multivariable adjustment, and how reproducible are these associations across independent cohorts?’ The review focused on transcriptomic, proteomic, and integrative multi-omics biomarkers, with particular attention to multivariable adjustment, external validation, risk of bias, and certainty of evidence. To place these findings in a translational context, contemporary investigational PDAC trials were also reviewed to assess the extent to which molecular biomarkers are currently incorporated into clinical trial design, thereby illustrating the gap between biomarker discovery and clinical implementation. Finally, an exploratory prognostic model was developed and externally evaluated using independent public datasets to determine whether a data-driven molecular signature could address the principal weakness identified by the systematic review: The limited reproducibility of published prognostic biomarkers.

MATERIALS AND METHODS

This systematic review was conducted in accordance with the preferred reporting items for systematic reviews and meta-analyses (PRISMA) 2020 guidelines; the completed checklist is provided in the file “PRISMA-2009-Checklist”. The review protocol was not prospectively registered in PROSPERO or another systematic review registry. The primary objective was to systematically evaluate whether tumor-derived transcriptomic and proteomic biomarkers measured in PDAC tissue are independently associated with survival outcomes. A second objective, specified after completion of the evidence synthesis, arose from the appraisal of contemporary interventional trials. The eligibility framework is reported in full in Table 1, and the study selection process is summarized in Figure 1. Details of the included studies and model development results are provided in the file “PRISMA-2009-Checklist”.

Figure 1 Preferred reporting items for systematic reviews and meta-analyses 2020 flow chart. The included studies evaluated transcriptomic and proteomic prognostic biomarkers and related clinical trials in pancreatic ductal adenocarcinoma. PDAC: Pancreatic ductal adenocarcinoma; AI: Artificial intelligence.

Table 1 Population, intervention, comparison, outcome, and study design framework for study eligibility.

Criteria	Review stage	Inclusion	Exclusion
Study design	Abstract/full article review	Original human studies evaluating the prognostic significance of tumor-derived transcriptomic and/or proteomic biomarkers in PDAC. Eligible designs include prospective or retrospective cohort studies, nested biomarker studies, translational correlative studies embedded within clinical trials, and secondary analyses of publicly available human molecular datasets (e.g., TCGA, ICGC, CPTAC). Eligible only if the study demonstrated an independent prognostic biomarker, i.e. the tumor transcriptomic/proteomic biomarker remained significantly associated with the survival outcome in an adjusted multivariable time-to-event model including clinicopathological covariates	In vitro studies, cell line studies, animal studies, narrative reviews, systematic reviews, meta-analyses, editorials, letters, commentaries, protocols, case reports, case series with insufficient outcome data, and conference abstracts without full-text data
Population	Abstract/full article review	Adults (≥ 18 years) with histologically confirmed PDAC, including resectable, locally advanced unresectable, and metastatic disease, provided that survival or other prognostic outcome data were available	Studies focused exclusively on non-PDAC malignancies, including pancreatic neuroendocrine neoplasms, periampullary cancers, ampullary cancers, cholangiocarcinoma, acinar cell carcinoma, or mixed pancreatic tumor populations where PDAC-specific results cannot be disaggregated
Treatment/intervention	Abstract/full article review	Tumor molecular profiling using transcriptomic or proteomic methods, including RNA sequencing, gene expression microarrays, qPCR-based tumor expression profiling, mass spectrometry-based proteomics, RPPA, or molecular signatures derived from transcriptomic/proteomic datasets	Studies based solely on clinical variables, histopathology without molecular profiling, circulating biomarkers only, imaging biomarkers only, radiomics without molecular tumor data, genomic/DNA-only biomarkers without transcriptomic or proteomic readout, or studies lacking explicit biomarker assessment
Comparator	Abstract/full article review	Comparisons across biomarker-defined strata, such as high vs low expression, biomarker-positive vs biomarker-negative status, molecular subtype classification, or low- vs high-risk prognostic groups derived from transcriptomic/proteomic signatures	Studies with no biomarker-based stratification or no analysable comparison across molecular groups
Outcomes	Abstract/full article review	Primary outcome: OS. Secondary outcomes: PFS, DFS, RFS, CSS, PFI, and other clinically relevant time-to-event prognostic outcomes	Studies not reporting survival, recurrence, or other prognostic time-to-event outcomes; studies reporting only treatment response without survival correlation
Biomarker modality and sample type	Abstract/full article review	Eligible only if the biomarker was a transcriptomic (RNA) or proteomic readout measured in PDAC tumor tissue (resection specimen, biopsy, or FFPE/fresh-frozen archival tissue), including tumor tissue profiled by RNA-seq, microarray, qPCR, NanoString, mass spectrometry, RPPA, or IHC-validated protein expression where the protein readout was the tested prognostic marker. Multimodal studies are eligible if tumor transcriptomic/proteomic results are reported separately and can be disaggregated	Not eligible: (1) Liquid biopsy/circulating analytes as the sole readout (ctDNA, cfDNA, CTCs, circulating or exosomal miRNA/mRNA, plasma or serum proteomics, CA19-9 and other serum markers); (2) Histopathology alone (grade, stage, margin status, lymphovascular or perineural invasion, tumor budding, stroma scoring) without a molecular readout; (3) Genomic/DNA-only markers (KRAS, TP53, SMAD4, or CDKN2A mutations; copy number; methylation; TMB; MSI) with no transcript- or protein-level measurement; (4) Imaging/radiomics markers; and (5) Germline variants. These modalities were recorded at screening but excluded from data extraction unless accompanied by a tumor transcriptomic or proteomic result

PDAC: Pancreatic ductal adenocarcinoma; TCGA: The Cancer Genome Atlas; ICGC: International Cancer Genome Consortium; CPTAC: Clinical Proteomic Tumor Analysis Consortium; RPPA: Reverse-phase protein array; qPCR: Quantitative polymerase chain reaction; OS: Overall survival; PFS: Progression-free survival; RFS: Recurrence-free survival; DFS: Disease-free survival; CSS: Cancer-specific survival; PFI: Progression-free interval; FFPE: Formalin-fixed paraffin-embedded; RNA-seq: RNA sequencing; IHC: Immunohistochemistry; ctDNA: Circulating tumor DNA; cfDNA: Cell-free DNA; miRNA: MicroRNA; mRNA: Messenger RNA; CTC: Circulating tumor cell; CA19-9: Carbohydrate antigen 19-9; MSI: Microsatellite instability; TMB: Tumor mutational burden.

Inclusion and exclusion criteria

Eligibility was defined a priori using the population, intervention, comparison, outcome, and study design (PICOS) framework, as summarized in Table 1. Eligible studies comprised original human studies evaluating tumor-derived molecular biomarkers in PDAC. The 2025-2026 eligibility restriction was applied to the publication date of the full-text study report and not to the year of trial initiation, registration, patient recruitment, or follow-up. This restriction applied to the bibliographic evidence stream only; registered trials were identified through a separately searched ClinicalTrials.gov contextual stream without a date restriction and included if they had been completed and had results available at the time of the search, as described under Search strategy. Eligible biomarker classes were restricted to transcriptomic readouts (single-gene and multigene signatures, including machine-learning-derived scores) and proteomic readouts (single-protein and multiprotein scores) measured in PDAC tumor tissue. Genomic/DNA-only markers, tissue-based histopathological markers, imaging or radiomic markers, germline variants, and circulating (liquid-biopsy) markers were not independently eligible; they were recorded during screening but excluded from data extraction unless accompanied by a tumor transcriptomic or proteomic readout that was itself the prognostic marker being tested and was evaluated in a multivariable survival model. Eligible designs included prospective and retrospective cohort studies, nested biomarker studies, translational analyses within clinical trials, and secondary analyses of publicly available molecular datasets; e.g., The Cancer Genome Atlas (TCGA), International Cancer Genome Consortium (ICGC), and Clinical Proteomic Tumor Analysis Consortium (CPTAC). Adult patients (≥ 18 years) with histologically confirmed PDAC at any disease stage were eligible, provided that survival or other prognostic outcomes were reported. Studies were required to assess tumor-based molecular biomarkers and compare biomarker-defined groups (e.g., high vs low expression or risk groups). Primary and secondary outcomes comprised overall survival (OS) and other time-to-event endpoints, including progression-free survival, disease-free survival, and recurrence-free survival. Studies were excluded if they involved non-human or in vitro research, were reviews, lacked molecular profiling, or did not report survival outcomes. Registered interventional PDAC trials reporting survival or response outcomes were retained for contextual appraisal. These trials were not required to evaluate any of the reviewed prognostic biomarkers; instead, they were included to characterize the contemporary treatment landscape into which biomarker-based stratification would need to be introduced and to quantify the frequency with which molecular biomarkers are incorporated into trial design. They did not contribute to the prognostic evidence synthesis or evidence-level grading. Studies entered the evidence synthesis if they fitted an adjusted multivariable time-to-event model in which the nominated tumor transcriptomic or proteomic biomarker was included alongside clinicopathological covariates, regardless of whether the adjusted estimate reached statistical significance. Studies reporting a null or non-significant adjusted association were retained and reported as null results, with the adjusted effect estimate and confidence interval recorded; eligibility was determined by whether the evaluation was performed, rather than by the direction or P value of the result. Studies in which no adjusted multivariable survival model was fitted for the nominated biomarker, or in which the presence of such a model could not be established from the available text, were excluded from the synthesis denominator. These exclusions were based on reporting completeness rather than study outcomes; statistical non-significance of an adjusted estimate was not grounds for exclusion because any such post-extraction filter would constitute outcome-dependent selection and inflate pooled effect sizes.

Search strategy

The search was designed to identify studies reporting prognostic biomarkers in PDAC together with survival outcomes. Three sources were searched: PubMed, ClinicalTrials.gov, and Europe PMC. All searches were conducted on 6 March 2026. The complete search string for each source, the filters applied, and the number of records retrieved from each source are provided in the file “PRISMA-2009-Checklist” (PubMed, n = 258; ClinicalTrials.gov, n = 413; Europe PMC, n = 105).

The PubMed search strategy combined Medical Subject Headings (MeSH) and free-text terms relating to PDAC, omics technologies, biomarkers, and prognostic outcomes. The following search string was used: (‘pancreatic neoplasms’ [MeSH] OR ‘pancreatic ductal adenocarcinoma’ OR PDAC OR ‘pancreatic cancer’) AND (‘transcriptome’ [MeSH] OR ‘proteomics’ [MeSH] OR transcriptomic OR RNA-seq OR proteomic OR ‘gene expression profiling’) AND (‘biomarkers, tumor’ [MeSH] OR biomarker OR ‘molecular subtype’ OR heterogeneity OR signature) AND (‘prognosis’ [MeSH] OR ‘survival analysis’ [MeSH] OR survival OR prognostic OR outcome). Filters restricted the results to human studies published in English between 1 January 2025 and 6 March 2026. This 14-month window was selected deliberately rather than for convenience. Discovery practices in this field have changed rapidly: Machine-learning-derived risk scores, single-cell and spatial-resolution approaches, and integrative multi-omics designs have only recently become prevalent. Thus, a synchronic sample of the current literature keeps analytical platforms, annotation resources, and reporting conventions broadly comparable across the included studies while avoiding the pooling of methodologically incommensurable generations of research. The window also ends on the search date and begins at a point when these contemporary pipelines are consistently represented, allowing the appraisal to reflect methodologies currently in use rather than superseded practices. The consequence of this approach is explicitly acknowledged: This review provides an appraisal of contemporary discovery pipelines and their methodological quality rather than an exhaustive historical review of PDAC prognostic biomarkers. Established markers first reported before 2025 are therefore outside its scope, and no claim of cumulative coverage is made. In addition, ClinicalTrials.gov was searched to identify completed interventional PDAC trials reporting survival or response outcomes and to assess the extent to which molecular biomarkers are incorporated into trial design, stratification, or endpoints. The search combined condition-specific terms with biomarker- and omics-related keywords across study fields (e.g., brief title, official title, summary, and outcome measures), with an emphasis on survival and prognostic endpoints. Europe PMC was also searched using title- and abstract-based terms combining PDAC-related terminology with transcriptomic, proteomic, biomarker, and survival-related keywords. The combined searches across all sources yielded 776 records.

The ClinicalTrials.gov search was conducted as a separately searched contextual stream, distinct from the 2025-2026 bibliographic evidence stream. Because the purpose of this stream was to characterize how molecular biomarkers are incorporated into the design, stratification, and endpoints of PDAC trials with available outcome data, rather than to identify recently reported prognostic biomarkers, eligibility was defined by the availability of posted results and not by registration, completion, or publication date. The registry was therefore searched without a date restriction, and prespecified filters restricted records to studies with an overall status of ‘completed’ and with posted results (‘with results’) and to the conditions ‘pancreatic ductal adenocarcinoma’, ‘pancreatic cancer non-resectable’, or ‘pancreatic cancer’.

Selection of studies

All retrieved records were exported to EndNote. Duplicates were removed through automated matching of title, first author, and publication year, followed by manual verification, reducing the initial 776 records to 699 unique records. Screening then proceeded in 2 stages: Title and abstract screening followed by full-text assessment against the PICOS criteria. Study selection was performed independently by 2 reviewers. Where judgements differed, the reviewers discussed the record and reached a joint decision; a third reviewer was available for arbitration but was not required. Because only the final consensus decision was retained in the screening log, individual reviewer judgements were not preserved and an inter-rater agreement statistic (Cohen’s kappa) could not be calculated.

The database searches of PubMed and Europe PMC identified 363 bibliographic records, while the ClinicalTrials.gov search retrieved 413 trial records, resulting in 776 records overall. After removing 77 duplicates, 699 unique records remained. These comprised 286 bibliographic records and 413 ClinicalTrials.gov records and were evaluated through two parallel screening streams.

The 413 ClinicalTrials.gov records were filtered to identify completed trials with posted results. This process excluded 399 trial records and retained 14 eligible trials. In parallel, the 286 bibliographic records were independently screened by two reviewers at the title and abstract level, resulting in the exclusion of 159 records. The full texts of the remaining 127 articles were assessed for eligibility, with particular consideration of whether the studies demonstrated independent prognostic value. Seventy articles that did not satisfy this criterion were excluded, leaving 57 bibliographic studies for inclusion.

Consequently, 71 studies fulfilled the eligibility criteria and were included in the final evidence synthesis. The complete study-selection process is presented in Figure 1. Reasons for exclusion are reported at the relevant final screening stage in the PRISMA flow diagram, while the initial ClinicalTrials.gov step represents registry filtering only.

Data extraction and synthesis

Study information was extracted into a standardized Microsoft Excel spreadsheet according to a predefined framework. Key study characteristics, methodological features, and survival outcomes were recorded for each included study, with particular emphasis on identifying biomarkers showing independent prognostic significance in multivariable analyses. Studies were subsequently synthesized and stratified by strength of evidence to facilitate critical appraisal, using the grading scheme defined beneath Table 2[14-70], which considers multivariable adjustment, external validation, and the availability of primary experimental data. This grading scheme was developed specifically for the present review. Study-level risk of bias was also assessed using validated instruments: Quality in prognosis studies (QUIPS) for prognostic-factor studies, prediction model risk of bias assessment tool (PROBAST) for prognostic model development studies, and the Cochrane Risk of Bias 2 tool or risk of bias in non-randomized studies of interventions for interventional trials, with each domain evaluated according to prespecified criteria applied consistently across records. Certainty of evidence was assessed using the grading of recommendations assessment, development and evaluation (GRADE) framework for each evidence stream. Because effect estimates were reported using non-comparable scales, neither quantitative pooling nor a formal assessment of small-study effects was feasible. Domain-level judgements for each record are provided in the file “PRISMA-2009-Checklist”. Interventional trials retained in the separate contextual stream were not included in this evidence-grading framework. Instead, they were summarized descriptively according to disease stage, study design, reported outcomes, and biomarker content; no pooled or comparative synthesis of efficacy was undertaken.

Table 2 Prognostic strength of biomarkers in pancreatic ductal adenocarcinoma (57 biomarker studies; 14 interventional trials are presented separately in Table 3).

Biomarker	Type	Outcome	Multivariable analysis	Independent prognostic association	Validation	Internal descriptive category	Ref.
UFM1	Single protein	OS/PFS	Yes	Yes	Limited	B	Wang et al[14]
9-gene signature	Multi-gene	OS	Yes	Yes	External	A-	Wang et al[15]
3-gene MRS	Multi-gene	OS	Yes	Yes	External	A	Li et al[16]
LMRGS	Multi-gene ML	OS	Yes	Yes	Multi-cohort	A	Ge et al[17]
PCLM-BM signature	Multi-gene	OS	Yes	Yes	External	A-	Zhou et al[18]
MCMBP	Single gene	OS	Yes	Yes	External	A	Zhang et al[19]
APOBEC score	Genomic	OS	Yes	Yes	External	A	Yu et al[20]
JUP	Single gene	OS	Yes	Likely	Limited	B	Yang et al[21]
Manganese metabolism signature	Multi-gene	OS	Yes	Yes	External	A	Xiong et al[22]
ALDH family	Multi-gene	OS	Yes	Likely	Limited	B	Wu et al[23]
EFFscore/P2RY6	Multi-gene	OS	Yes	Yes	External	A	Wang et al[24]
mCAF signature	Multi-gene	OS	Yes	Yes	Multi-cohort	A	Wang et al[25]
18-gene immune signature	Multi-gene	OS	Yes	Yes	External	A	Wang et al[26]
PCD signature	Multi-gene	OS	Yes	Yes	Multi-cohort	A	Wang et al[27]
ICD signature (AJM1)	Multi-gene	OS	Yes	Yes	External	A	Ma et al[28]
MDRG/CITED4	Multi-gene	OS	Yes	Likely	Limited	B	Liu et al[29]
IGFBP signature	Multi-gene	OS	Yes	Yes	External	A	Guan et al[30]
A2ML1 subtype model	Multi-gene	OS	Yes	Yes	Multi-cohort	A	Ge et al[31]
BM-related signature (LAMA3/ITGA3/ITGB6)	Multi-gene	OS	Yes	Yes	External	A	Ge et al[32]
ER stress-related TF signature	Multi-gene	OS	Yes	Yes	External	A	Gao et al[33]
PLAAT2	Single gene	OS	Yes	Yes	External	A	Chen et al[34]
LY6D (stem-like, gastric-like cell-state marker)	Single gene	OS	Yes	Yes	Limited	B	Shi et al[35]
Disulfidptosis-related gene signature; key genes: RPN1, PML; Model genes: NDUFA11, MYH14, TRIP6, ACTN1, ANP32E, PML, CD2AP, RPN1	Multi-gene signature	OS	Yes	Yes	External	A-	Pei et al[36]
Palmitoylation-related gene signature (7 genes: DCBLD2, CXCL5, ASPH, LAMC2, KRT6A, TOP2A, NPNT); Upstream regulators: ZDHHC family (palmitoyl transferases)	Multi-gene signature	OS	Yes	Yes	External	A-	Lu et al[37]
7-gene SASP prognostic signature: ANGPTL4, DKK1, EREG, IGFBP3, IL18, MMP13, PLAU; Plus SASP molecular subtypes	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Kuang et al[38]
SEMA3 family genes; 9-gene prognostic signature derived from SEMA3-related DEGs; key driver: SEMA3C	Multi-gene signature	OS	Yes	Yes	External	A-	Jin et al[39]
Six-gene mitophagy-related signature (PAPPA, NBPF12, CXCL11, CKLF-CMTM1, CCDC6, AHNAK)	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Cai et al[40]
Six ICD-associated lncRNA prognostic signature: ZEB2-AS1, AC099792.1, LINC01091, LINC02613, AC005304.3, LINC00705	Multi-gene signature	OS	Yes	Yes	External	A-	Zhu et al[41]
11-gene ITH-related prognostic signature	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Zhao et al[42]
Ferroptosis-related 3-gene signature (NOTCH2, KRT18, H1-2); key single biomarker: NOTCH2	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Zhang et al[43]
16-gene DNA damage and repair-related prognostic risk score model; key experimental gene: LY6D	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Zhang et al[44]
STING signalling pathway-derived 6-gene prognostic signature: DTX4, IFI16, MAVS, PRKDC, ZDHHC1, IFIT2	Multi-gene signature	OS	Yes	Yes	External	A	Yu et al[45]
SERPINE1	Single gene	OS	Yes	Yes	External	A	Wang et al[46]
KRAS protein level, KRAS mRNA expression, KRAS mutation status/subtype	Single gene	OS	Yes	Yes	Limited	B	Stålberg et al[47]
RAD51AP1 (DNA repair-associated protein)	Single gene	OS	Yes	Yes	External	A-	Rezagholizadeh et al[48]
Mitochondrial 12-gene prognostic signature; mechanistic emphasis on CAMK2A and THEM4	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Nie et al[49]
Innate immune barrier gene signature (84 genes); key genes: UBASH3B, ITGB6, COL17A1, MMP28, DIAPH3 (focus on UBASH3B)	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Luo et al[50]
P4HA1, CCDC148, UCA1, TMPRSS11E + tumour stage	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Lin et al[51]
KRAS mutational status as key biomarker; integrated molecular classifier of KRAS wild-type vs KRAS-mutant classical vs KRAS-mutant basal-like; exploratory HRD and other somatic mutations	Multi-gene signature	OS	Yes	Yes	Internal	B+	Lansbergen et al[52]
PAUF (ZG16B-related protein product)	Single gene	OS	Yes	Yes	Limited	B	Kim et al[53]
Six BA-related prognostic genes: SERPINB5, CALU, TFRC, LY6D, SFRP1, GBP2	Multi-gene signature	OS	Yes	Yes	External	A-	Kang et al[54]
TIPARP (PARP7) expression	Single gene	OS	Yes	Yes	Limited	B	Jiang et al[55]
Microenvironmental, stromal, and immune markers, including MFAP4, EGR3, KRT8, monocyte infiltration, CD68+ macrophages, CD3/CD4/CD8 T-cell infiltration	Multi-gene signature	OS	Yes	Yes	Limited	B	Han et al[56]
Hypoxia score; 13-gene signature (including LYZ, PLAU, INSIG2, and KRTCAP2)	Multi-gene signature	OS/PFS	Yes	Yes	External	A	Ge et al[57]
CSCLPI genes (SHCBP1, CEP55, GMNN, CDCA5, FAM111B)	Multi-gene signature	OS	Yes	Yes	Limited	B	Fang et al[58]
23 ISGS-based features in risk score model	Multi-gene signature	OS	Yes	Yes	External	A	Fang et al[59]
Cuproptosis-related lncRNA prognostic model; key lncRNAs: LINC00900 and AC078820.1	Multi-gene signature	OS	Yes	Yes	Partial	B+	Duo et al[60]
ST3GAL2, SLCO1B3, PTGES, PRKACA	Multi-gene signature	OS/PFS	Yes	Yes	External	A-	Deng et al[61]
DCBLD2, OCIAD2, SAMD9	Multi-gene signature	OS	Yes	Yes	External	A	Cui et al[62]
Disulfidptosis-related prognostic signature (8-gene signature: SLC3A2, HMGA1, S100A2, DUSP5, SPRR1B, MET, CREB5, MACROD2)	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Cui et al[63]
M2 macrophage-related 6-gene prognostic signature: CCL2, CD9, FBP1, LIMS1, LYZ, TGFBI; M2/TAM biology and TNFRSF12A also highlighted	Multi-gene signature	OS	Yes	Yes	Partial	B+	Chen et al[64]
20-protein diagnostic panel; 4 proteomic prognostic subtypes; 18-protein prognostic risk score (PURB, SDCBP2, CD2BP2, GALM, SERPINA3, OAS3, FAN1, ZPR1, KRT2, NUDT2, SMNDC1, SERPINA4, CUTA, WDR36, POSTN, CLEC11A, PEX14, PI4KA)	Multi-protein signature	OS	Yes	Yes	External	A	Aref et al[65]
Complement pathway genes (C1R, C1S, C3, C4B, C7) in tumour microenvironment	Multi-gene signature	OS	Yes	Yes	Partial	B+	Zhang et al[66]
13-gene autophagy-related signature (EIF4EBP1, HDAC6, GABARAP, CASP4, ATG16 L2, BCL2 L1, BAG3, ATG4B, TNFSF10, PTK6, ITGA3, BIRC5, BAK1)	Multi-gene signature	OS	Yes	Yes	External	A-	Wang et al[67]
Fibroblast-related 14-gene prognostic signature	Multi-gene signature	OS	Yes	Yes	Multi-cohort	A	Wang et al[68]
ANP32E	Single gene	OS	Yes	Yes	External	A-	Liu et al[69]
Cuproptosis-related signature score (MIR3677, AC068620.2, UPK2, LDHAP7, FAM83A)	Multi-gene signature	OS	Yes	Yes	External	A-	Ge et al[70]

The categories below were developed by the authors for this review. Categories were defined as follows: A+: Multivariable significance with robust external validation across multiple independent cohorts and prospective validation; no included study met these criteria, and the category was therefore unawarded. A: Multivariable significance with external validation in more than 1 independent cohort, or in a single independent cohort with no recorded limitations. A-: Multivariable significance with external validation in a single independent cohort where limitations were recorded (e.g., dataset overlap, limited cohort diversity, or methodological concerns). B+: Moderate evidence; multivariable significance with internal validation or partial external validation. B: Multivariable significance without adequate validation. Studies without a statistically significant multivariable result are shown as not applicable because every category presupposes a significant result. MRS: Mitochondrial unfolded protein response-related gene risk score; LMRGS: Lactate metabolism-related gene signature; PCLM-BM: Pancreatic cancer liver metastasis and basement membrane; EFFscore: Efferocytosis-related score; mCAF: Myofibroblastic cancer-associated fibroblast; PCD: Programmed cell death; ICD: Immunogenic cell death; MDRG: Metastatic dissemination-related gene; BM: Basement membrane; TF: Transcription factor; ER: Endoplasmic reticulum; zDHHC: Zinc finger DHHC-type palmitoyl transferase; SASP: Senescence-associated secretory phenotype; DEGs: Differentially expressed genes; lncRNA: Long noncoding RNA; mRNA: Messenger RNA; ITH: Intratumoral heterogeneity; HRD: Homologous recombination deficiency; BA: Brown adipocyte; CD: Cluster of differentiation; ISGS: Immunologic signature gene sets; TAM: Tumor-associated macrophage; OS: Overall survival; PFS: Progression-free survival.

Prognostic model development

Proteomic profiling data were obtained from the CPTAC-PDAC cohort (accession phs001287) through the University of California Santa Cruz (UCSC) Xena platform (downloaded 25 March 2026). The dataset initially comprised 140 primary tumor samples and 140 adjacent-normal samples. After duplicated adjacent-normal sample identifiers were removed and the dataset was restricted to matched tumor-normal pairs, 80 paired samples were retained for differential expression analysis. Median-normalized protein intensities were analyzed in R using the linear models for microarray data package, with paired linear models and empirical Bayes moderation. Before analysis, proteins were retained if they were quantified in at least 70% of both tumor and adjacent-normal samples. Differentially expressed proteins were defined using a Benjamini-Hochberg false discovery rate (FDR) < 0.05 and an absolute log₂ fold change ≥ 1.

Significantly differentially expressed proteins were mapped to their corresponding genes and carried forward as candidate predictors for development of the prognostic model. Transcriptomic and clinical data were obtained from the TCGA pancreatic adenocarcinoma (TCGA-PAAD) cohort through UCSC Xena. After harmonization with the clinical dataset and application of the predefined eligibility criteria, 177 patients with complete transcriptomic and OS data were retained for analysis. The optimal penalty parameter (λ) was selected using 10-fold cross-validation, after which the coefficient-weighted risk score was calculated as the weighted sum of the selected gene expression values.

The prognostic signature was externally validated without modification in the independent Gene Expression Omnibus (GEO) GSE71729 cohort, which included 125 patients with complete OS information. The continuous risk score was prespecified as the primary prognostic analysis and evaluated using univariable and multivariable Cox proportional hazards regression. Median-based dichotomization using the cohort-specific median risk score was conducted as a secondary exploratory analysis using Kaplan-Meier survival analysis and log-rank testing. Model discrimination was assessed using Harrell’s concordance index (C-index). Confidence intervals for the C-index were non-parametric bootstrap percentile intervals based on 2000 cohort resamples with model coefficients held fixed; therefore, they describe the sampling precision of the discrimination estimate rather than an optimism-corrected value, and the reported C-indices are apparent estimates. Calibration was assessed using the calibration slope, defined as the Cox coefficient of the linear predictor in the cohort being evaluated, and overall prediction accuracy was evaluated using inverse-probability-of-censoring-weighted Brier scores at 1 year, 2 years, and 3 years, with an integrated score over 0-3 years. The censoring distribution was estimated using the Kaplan-Meier method, and baseline survival was estimated using the Breslow method at the cohort mean risk score. Bootstrap internal validation with optimism correction of the derivation C-index was not performed and could not be reconstructed from the retained data because a valid correction would require the entire model-building procedure, including least absolute shrinkage and selection operator (LASSO) selection of the 5 constituent genes, to be repeated within each resample, whereas only the final per-patient risk score was retained. Candidate selection and model fitting were confined to the derivation data. The resulting model is referred to throughout as a 5-gene signature: Although the candidate set was identified from proteomic measurements, the coefficients were estimated using, and are applied to, gene expression values. The 28 candidate genes were carried forward from the paired CPTAC tumor-vs-adjacent-normal proteomic contrast, which did not use survival information from any cohort. Before LASSO selection, an initial biologically derived signature was constructed from the 28 genes corresponding to proteins retained after mapping the significantly differentially abundant CPTAC proteins to the TCGA-PAAD gene-expression dataset: SLC16A10, PRSS3, DHRS2, KLK1, AQP12B, PRSS1, GSTA2, CPB1, CPA1, PNLIP, SERPINI2, CELA3A, CTRB2, CELA2A, GP2, RBPJL, SYCN, PLA2G1B, PNLIPRP1, CPA2, AMY2A, AQP8, CEL, FGL1, PAIP2B, AMY2B, SLC43A1, and FKBP11. The resulting signature was evaluated as a continuous predictor of OS using univariable and clinically adjusted multivariable Cox regression.

Penalized selection of the 5 constituent genes and estimation of their coefficients were performed solely in TCGA-PAAD using 10-fold cross-validated LASSO Cox regression (seed fixed, λ selected at the cross-validation minimum). No expression, clinical, or survival data from GSE71729 were examined during candidate selection, penalization, coefficient estimation, or threshold definition; the external cohort was accessed only after the 5 coefficients had been fixed, and these coefficients were applied without modification or recalibration. Exploratory cut-point optimization was additionally performed on the continuous risk score using maximally selected rank statistics. The standardized log-rank statistic was evaluated across all candidate thresholds between the 10^th and 90^th percentiles of the score in each cohort, and the P value at the selected optimum was corrected for the multiplicity of this search; quartile-based analyses were also performed to assess departure from linearity. These analyses were exploratory, and both uncorrected and corrected P values are reported. P values for the primary prognostic analyses were additionally adjusted for multiple testing using the Benjamini-Hochberg FDR procedure across the family of 8 tests reported for the 2 signatures in the derivation and validation cohorts. To assess threshold transportability, the exact numerical cut-points derived in TCGA-PAAD were applied without recalibration to the GSE71729 risk scores, and the resulting dichotomization was evaluated using Cox regression and the log-rank test. Multivariable models in the derivation cohort were adjusted for age, sex, pathological stage, and resection status. In the external validation cohort, in which these clinical covariates were not consistently available, multivariable models were adjusted for Moffitt tumor subtype and stromal subtype. Patients with missing survival information were excluded, and missing values were not imputed. All statistical analyses were conducted using R version 4.5.1 (R Foundation for Statistical Computing, Vienna, Austria).

Cross-platform data were handled as follows. The derivation transcriptomic data comprised the TCGA-PAAD IlluminaHiSeq (HiSeqV2) gene-level dataset obtained through UCSC Xena, consisting of RNA-sequencing by expectation-maximization (RSEM)-normalized counts on a log₂ (x + 1) scale indexed by gene symbol. The dataset comprised 20530 genes across 178 tumor samples, of which 177 were matched to a clinical record. Candidate predictors comprised the 29 proteins that were significant in the paired CPTAC contrast, 28 of which were present in the TCGA matrix (ERO1B was absent). The penalized Cox model was fitted to these untransformed log₂ expression values; no standardization was applied to the model matrix, and the risk score was calculated as the coefficient-weighted sum of these values.

External data were obtained from the GSE71729 series matrix retrieved using GEO query (19749 gene-level features across 357 samples) and were used as deposited by the original submitters, indexed by gene symbol and without further transformation. Because the deposited matrix was already collapsed to the gene level, no probe aggregation was required. Gene mapping consisted of trimming whitespace and converting symbols to uppercase, after which all 5 signature genes were matched exactly, with no unmapped or ambiguous symbols. The analysis was restricted to the 125 primary tumors with complete OS information, with survival recorded in months. The 5 coefficients derived in TCGA-PAAD were applied unchanged to the GSE71729 expression values.

No cross-platform normalization, batch correction (ComBat, quantile normalization, or rank transformation), or per-cohort standardization of the risk score was performed at any stage, and the coefficients were not recalibrated. The score was therefore calculated using each cohort’s native expression scale RSEM log₂ (x + 1) counts in the derivation cohort and the deposited series-matrix values in the validation cohort. These values are not on a common numerical scale: The derivation risk scores are entirely negative (-2.16 to -0.62), whereas the validation scores include both positive and negative values and are centred near zero (-1.99 to 1.11). This difference is stated explicitly because it directly accounts for the threshold-transportability failure reported below; the scale difference and threshold-transportability results should therefore be interpreted together.

RESULTS

Results of the literature research

The literature search identified 776 records, comprising 363 records from bibliographic databases and 413 records from ClinicalTrials.gov. After 77 duplicate records were removed, 699 unique records remained. Of these, 399 registry trials records were removed by the search filters and 159 were excluded at title and abstract screening, leaving 127 reports for full-text assessment, of which 70 were excluded as PRISMA flow diagram shows (Figure 1). Overall, 71 studies met the eligibility criteria and were included in the review. These comprised 57 prognostic biomarker studies that fitted adjusted multivariable survival models for the nominated biomarker, all of which reported an independent prognostic association, together with 14 interventional clinical trials included for contextual appraisal of biomarker incorporation into PDAC trial design.

Across the 57 included biomarker studies, the biomarker class for each study is reported in Table 2. All 57 studies reported multivariable-adjusted survival analyses, with OS as the primary endpoint in all but 2 studies, which reported progression-free survival together with OS and progression-free interval, respectively. An independent prognostic association was reported in all 57 studies. External validation in a single independent cohort was reported in 29 studies and external validation across multiple cohorts in 14; validation was partial in 3 studies, internal only in 1, and limited in 10. According to the internal descriptive categories, 31 studies were categorized as A, 12 as A-, 4 as B+, and 10 as B. The highest category, A+, was not awarded because no included study combined multivariable significance with robust validation across multiple independent cohorts and prospective validation. Ten previously reported categories changed under the single stated rule, with 7 upgraded and 3 downgraded, while 32 studies were categorized for the first time in this revision. Analyzed cohorts were generally modest in size: Among the 38 studies for which an analyzed patient total could be established, the median was 178 (interquartile range, 147-225; range, 41-582), with 23 studies including between 100 and 200 patients, 4 including fewer than 100, 10 including 201-500, and 1 including more than 500.

The 14 interventional trials retained for contextual appraisal (Table 3)[71-84] were predominantly small, single-arm studies (10 of 14), with only 3 randomized comparisons and 1 non-randomized dose-escalation study; enrolment ranged from approximately 12 patients per arm to 284 patients overall. The trials spanned the full disease spectrum, from perioperative treatment of resectable or borderline-resectable disease to locally advanced and metastatic settings, and generally showed modest activity. In the advanced and metastatic trials, median OS was approximately 4-10 months and objective response rates were generally below 20%, whereas the perioperative studies achieved higher completion or conversion-to-surgery rates (83.1% and 45.5%, respectively) and correspondingly longer survival. Both larger randomized trials failed to meet their primary endpoint, including the trial evaluating the addition of panitumumab to gemcitabine and erlotinib (median OS: 4.2 months vs 8.3 months, P = 0.36) and the trial evaluating the addition of pamrevlumab to standard chemotherapy in locally advanced disease [17.25 months vs 17.94 months; hazard ratio (HR) = 1.08, 95% confidence interval (95%CI): 0.83-1.41][78]. Where molecular characterization was performed, it was descriptive rather than decisional: Trials reported CA19-9 or carcinoembryonic antigen dynamics, SMAD4 immunohistochemistry, or generic biomarker sample collection, and none used a molecular biomarker to determine eligibility, stratify randomization, or define a prespecified endpoint.

Table 3 Characteristics and reported outcomes of interventional trials included for clinical context in pancreatic ductal adenocarcinoma.

NCT	Study design	n	Intervention	Stage	Primary outcome	Result	Secondary outcomes	Key finding	Ref.
NCT01726582	Single-arm	130	Personalised perioperative treatment (chemo/CRT/surgery)	Resectable/borderline resectable	Completion of surgery (feasibility)	108/130 (83.1%)	OS: 45 months vs 11 months; PFS: 36/130 (27.7%); Biomarker: 93/130 (71.5%); CR: 0/107; Near CR: 17/107; PR: 44/107; Poor: 42/107	The study demonstrated a high rate of completion of surgery (83.1%)	Tsai et al[71]
NCT00536874	Single-arm	35	Gem + oxaliplatin (neoadjuvant)	Resectable	18-month OS rate	63% (95%CI: 45-76)	OS: 39.6 months; PR: 4/35; SD: 28/35; PD: 3/35; CEA Δ: -5.1%; CA19-9 Δ: -25.6%	The 18-month OS rate was 63% (95%CI: 45%-76%)	O’Reilly et al[72]
NCT01926197	Randomised	13/12	mFOLFIRINOX vs mFOLFIRINOX + SBRT	Locally advanced (unresectable)	PFS (median)	6.5 months vs 8.4 months	Local PFS: 6.5 vs 8.4; PFS 1 year: 5/13 vs 2/12; MFS: 12.9 vs 10.8; OS: 12.9 vs 13.4; GI toxicity events: 16 vs 19	Median PFS was numerically higher in the SBRT arm (8.4 months vs 6.5 months)	Vitzthum[73]
NCT00810719	Single-arm	30	Gem + erlotinib	Advanced/metastatic	PFS (median)	2.07 months (1.87-5.50)	ORR: 11%; OS: 5.67 months (2.83-11.87)	The observed median PFS of 2.07 months indicated limited disease control	Semrad et al[74]
NCT00769483	Phase I/II, non-randomised, multi-arm	Approximately 15/arm	MK-0646 + Gem ± erlotinib	Stage IV (metastatic)	MTD + PFS (time to progression/death)	MTD: 10 mg/kg (Gem), 5 mg/kg (Gem + erlotinib); PFS: Approximately 1.8-2.0 months	OS: 5.7-10.4 months; ORR: Approximately 13%-17%; Toxicity: High	The study achieved its dose-escalation objective (MTD), but showed limited efficacy with median PFS of approximately 2 months	Abdel-Wahab et al[75]
NCT00365144	Single-arm	36	Bevacizumab + erlotinib	Metastatic, Gem-refractory	OS rate (6 months) + safety	8/36 (22.2%)	ORR: 1/36 (2.8%); TTP: 40 days (35-41); CA19-9 decline ≥ 25%: 4/26 (15.4%)	Low 6-month OS rate (22.2%)	Ko et al[76]
NCT00550836	Randomised	46/46	Gem + erlotinib vs + panitumumab	Metastatic	OS (median; arm A vs arm B)	4.2 months vs 8.3 months; P = 0.36	ORR: 8.7% vs 6.5%; PFS: 2.0 months vs 3.6 months, P = 0.419; TTF: 2.0 months vs 3.8 months, P = 0.36; G3+ AE: 41/46 vs 41/46; G4+ AE: 15/46 vs 14/46	Primary endpoint was not met	Halfdanarson et al[77]
NCT03941093	Randomised, double-blind	143/141	Pamrevlumab + G/nab-P or FOLFIRINOX vs placebo + G/nab-P or FOLFIRINOX	Locally advanced, unresectable	OS	17.25 months vs 17.94 months; HR 1.08 (0.83-1.41); P = 0.5487	EFS: 5.72 months vs 5.78 months; PFS: 9.36 months vs 9.40 months; ORR: 43/143 (30.1%) vs 64/141 (45.4%)	Primary endpoint not met	Picozzi et al[78]
NCT00408564	Single-arm	38-39	Gem + oxaliplatin + cetuximab, reassessment, surgery or RT/capecitabine	Locally advanced, unresectable, nonmetastatic	6-month PFS rate	82% (70%-95%)	G3-4 AE: 9/39 (23.1%); OS rate: 49% (36%-68%); ORR: 5/38 (13.2%)	High 6-month PFS rate	Medical University of South Carolina[79]
NCT00521404	Single-arm	61	CS-1008 (anti-DR5) + Gem	Unresectable/metastatic (chemo-naive)	16-week PFS rate	52.5% (39.3%-64.1%)	PFS: 3 months 59%, 6 months 34.4%; OS: 6 months 55.7%, 12 months 24.6%; ORR: 5/61 (8.2%); SD: 28/61 (45.9%)	Modest activity	Sankyo[80]
NCT00417976	Single-arm	39-40	Bevacizumab + Gem + infusional 5-FU	Stage III-IV	6-month PFS rate	49%	ORR: 12/40 (30%); SD: 18/40 (45%); PD: 10/40 (25%)	Moderate activity	Bekaii-Saab[81]
NCT00499486	Single-arm	31	Sirolimus monotherapy	Advanced PDAC, post-Gem (refractory)	6-month OS rate	26%	CR + PR: 0%. SD: 4 patients (approximately 13%). ORR = 0%	mTOR inhibition alone is insufficient in PDAC	Garrido-Laguna et al[82]
NCT01560949	Single-arm	33	mFOLFIRINOX, Gem-based CRT, surgery	Borderline resectable/high-risk resectable	Resectability rate	15/33 (45.5%)	R0: 10/15 (66.7%); DFS: 11.1 months (5.8-NA); OS: 24.0 months (16.2-29.6); SMAD4+ pre: 4/15 (26.7%); SMAD4+ post: 4/15 (26.7%); Local recurrence: 0/15; Distant recurrence: 10/15 (66.7%)	Moderate conversion to surgery; distant failure predominates	M.D. Anderson Cancer Center[83]
NCT01146054	Single-arm	49	SBRT + Gem	Locally advanced, unresectable, nonmetastatic	Late GI toxicity ≥ grade 2 at 1 year	5/49 (10.2%)	Acute GI toxicity: 14/49 (28.6%); PFS: 7.8 months (5.8-10.2); OS: 13.9 months (10.2-16.7); FFLP: 0.78 (0.60-0.89)	Low late GI toxicity; local control > systemic control	Koong[84]

NA: Not available; NCT: National clinical trial; CRT: Chemoradiotherapy; Gem: Gemcitabine; SBRT: Stereotactic body radiotherapy; FOLFIRINOX: Fluorouracil, leucovorin, irinotecan, and oxaliplatin; 5-FU: 5-Fluorouracil; DR5: Death receptor 5; RT: Radiotherapy; PDAC: Pancreatic ductal adenocarcinoma; MTD: Maximum tolerated dose; GI: Gastrointestinal; OS: Overall survival; PFS: Progression-free survival; CI: Confidence interval; HR: Hazard ratio; CR: Complete response; CEA: Carcinoembryonic antigen; PR: Partial response; SD: Stable disease; PD: Progressive disease; MFS: Metastasis-free survival; CA19-9: Carbohydrate antigen 19-9; TTP: Time to progression; ORR: Objective response rate; EFS: Event-free survival; TTF: Time to treatment failure; AE: Adverse event; FFLP: Freedom from local progression; R0: Microscopically margin-negative resection; DFS: Disease-free survival; mTOR: Mammalian target of rapamycin.

Study-level risk of bias was assessed using QUIPS for all 57 biomarker studies (overall risk of bias moderate in 29 and high in 28; none low) and additionally using PROBAST for the 46 studies that developed a risk score, signature, or nomogram. Among these 46 model-development studies, the overall PROBAST risk of bias was rated as high in 21 and unclear in 25, with none rated low. The principal concerns related to the retrospective analysis of public or convenience cohorts without consecutive, prospectively defined enrolment; absent, limited, or internal-only validation; data-derived cut-points; inadequate reporting of calibration; and insufficient justification of the number of events per predictor parameter. Certainty of evidence was evaluated using GRADE across the prespecified evidence streams: Multi-gene or multi-protein signatures (n = 45); Single-gene or single-protein markers (n = 10); Pathway, and multimodal markers (n = 2); and registered therapeutic trials (n = 14) retained for the clinical-trial integration analysis. Certainty was rated as very low across all streams because of concerns regarding risk of bias and imprecision arising from small analyzed cohorts. Indirectness was judged on population and the test assessed: It was rated as not serious for the three biomarker streams, because every study enrolled a PDAC population and assessed the marker as a prognostic test, whereas the trial stream was downgraded by two levels for indirectness because the trials addressed treatment effect rather than prognosis. The heterogeneous and non-comparable effect measures precluded quantitative pooling and formal assessment of small-study effects. Domain-level QUIPS and PROBAST judgements for each eligible study, together with the domains driving high ratings, are provided in the file “PRISMA-2009-Checklist”.

Results of proteomic analysis

Differential expression analysis identified 29 proteins that were significantly dysregulated between matched PDAC tumor and adjacent-normal tissues (FDR < 0.05, |log₂ fold change| ≥ 1) (Figure 2). All 29 proteins were downregulated in tumor tissue relative to adjacent-normal pancreas.

Figure 2 Differential protein expression between pancreatic ductal adenocarcinoma tumors and matched adjacent-normal tissues in the Clinical Proteomic Tumor Analysis Consortium pancreatic ductal adenocarcinoma cohort. Labeled proteins met both significance criteria. CPTAC: Clinical Proteomic Tumor Analysis Consortium.

Hierarchical clustering of the top differentially expressed proteins showed clear separation between PDAC tumor and matched adjacent-normal tissues based on protein abundance profiles. Most of the significant proteins showed consistently lower abundance in tumor samples, including PLA2G1B, CPA1, CPA2, CPB1, CTRB2, PRSS1, PNLIP, CELA2A, CELA3A, and AMY2A, consistent with the loss of normal exocrine pancreatic differentiation (Figure 3). Proteins with similar abundance profiles clustered together, suggesting coordinated biological regulation.

Figure 3 Heatmap of the top differentially expressed proteins identified in the Clinical Proteomic Tumor Analysis Consortium pancreatic ductal adenocarcinoma cohort. Rows represent proteins, and columns represent individual samples. Protein expression values are displayed as row-wise standardized (z-score) expression levels, with brown indicating relatively high expression and blue indicating relatively low expression. CPTAC: Clinical Proteomic Tumor Analysis Consortium; PDAC: Pancreatic ductal adenocarcinoma.

Gene Ontology (GO) Biological Process enrichment analysis of the differentially expressed proteins identified significant over-representation of pathways associated with exocrine pancreatic function. The most significantly enriched term was digestion (GO: 0007586; FDR = 5.72 × 10^-5), involving 6 proteins (CTRB2, AMY2A, CEL, PNLIP, PRSS1, and PRSS3). Other enriched biological processes included hormone metabolic process, lipid digestion, intestinal lipid absorption, triglyceride and lipid catabolic processes, protein processing, organic cation transport, and water transport. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis similarly showed that the differentially expressed proteins were predominantly involved in exocrine pancreatic function. The most significantly enriched pathway was pancreatic secretion (hsa04972), comprising 14 proteins (fold enrichment = 58.50, adjusted P = 2.04 × 10^-21). Significant enrichment was also identified for protein digestion and absorption, fat digestion and absorption, glycerolipid metabolism, and carbohydrate digestion and absorption. Taken together, these findings suggest that the differentially expressed proteins primarily represent digestive enzymes and other proteins involved in pancreatic secretory function, consistent with the loss of normal exocrine pancreatic differentiation in PDAC.

Multivariable Cox regression analysis was performed for each candidate gene identified through differential expression analysis, with adjustment for clinical covariates. After correction for multiple testing, none of the candidate genes showed an independent association with OS (Figure 4).

Figure 4 Multivariable Cox regression analysis of candidate genes in The Cancer Genome Atlas pancreatic adenocarcinoma cohort. The vertical dashed line represents a hazard ratio of 1, indicating no association with survival. CI: Confidence interval; PDAC: Pancreatic ductal adenocarcinoma.

LASSO Cox regression generated a continuous risk score for each patient in the TCGA-PAAD derivation cohort. Risk scores ranged from approximately -2.2 to approximately -0.6 and followed a gradual, continuous distribution, with no evidence of distinct patient clusters. For exploratory Kaplan-Meier analysis, patients were subsequently stratified into low- and high-risk groups using the cohort-specific median risk score. This dichotomization was used solely for visualization and descriptive survival analysis, whereas the continuous risk score remained the primary predictor in all Cox regression models (Figure 5). Positive coefficients for PRSS3 and SLC16A10 were associated with increased risk, whereas negative coefficients for FKBP11, SLC43A1, and AMY2B contributed to lower predicted risk.

Figure 5 Distribution of the least absolute shrinkage and selection operator-derived risk score in The Cancer Genome Atlas pancreatic adenocarcinoma cohort. Each point represents an individual patient ranked according to the continuous coefficient-weighted risk score derived from the 5-gene signature (PRSS3, SLC16A10, FKBP11, SLC43A1, and AMY2B). LASSO: Least absolute shrinkage and selection operator; TCGA-PAAD: The Cancer Genome Atlas pancreatic adenocarcinoma.

Using the cohort-specific median LASSO risk score, patients were stratified into low- and high-risk groups. Kaplan-Meier analysis showed significantly poorer OS in the high-risk group than in the low-risk group (log-rank P = 0.0016) (Figure 6A).

Figure 6 Kaplan-Meier overall survival curves. A: The Cancer Genome Atlas pancreatic adenocarcinoma cohort according to the least absolute shrinkage and selection operator-derived 5-gene risk signature; B: The GSE71729 cohort according to the least absolute shrinkage and selection operator-derived 5-gene risk signature. PDAC: Pancreatic ductal adenocarcinoma; TCGA-PAAD: The Cancer Genome Atlas pancreatic adenocarcinoma.

Before LASSO selection, an initial biologically derived signature was constructed from the 28 genes corresponding to proteins identified in the paired CPTAC tumor-vs-adjacent-normal proteomic analysis. This prespecified score was evaluated in TCGA-PAAD using univariable and clinically adjusted multivariable Cox regression. The initial biologically derived signature was not associated with OS in either univariable (HR = 1.13, P = 0.239) or multivariable Cox regression (HR = 1.08, P = 0.480), and its discrimination was not distinguishable from chance (C-index = 0.522, 95%CI: 0.472-0.579 in univariable analysis and 0.563, 95%CI 0.498-0.626 after adjustment; both intervals included 0.50). By contrast, the LASSO-derived 5-gene signature showed a strong independent association with OS in both univariable (HR = 6.41, P < 0.001) and multivariable analyses (HR = 6.13, P < 0.001), although discrimination improved only modestly (C-index = 0.587, 95%CI: 0.526-0.634 in univariable analysis and 0.599, 95%CI: 0.534-0.653 after adjustment) (Table 4). Calibration, overall prediction accuracy, and baseline survival for both signatures in both cohorts are summarized in Table 5. These findings suggest that penalized feature selection identified a combination of molecular features with greater prognostic relevance in the derivation cohort than the initial biologically selected candidate signature.

Table 4 Comparison of prognostic performance between the initial biologically derived signature and the least absolute shrinkage and selection operator-derived 5-gene signature in The Cancer Genome Atlas pancreatic adenocarcinoma cohort.

Prognostic model	Cox model	Hazard ratio	P value	Harrell’s C-index	95%CI
Initial biologically derived signature	Univariable	1.13	0.239	0.522	0.472-0.579
	Multivariable	1.08	0.480	0.563	0.498-0.626
LASSO-derived 5-gene signature	Univariable	6.41	< 0.001	0.587	0.526-0.634
	Multivariable	6.13	< 0.001	0.599	0.534-0.653

CI: Confidence interval; LASSO: Least absolute shrinkage and selection operator.

Table 5 Calibration, overall prediction accuracy and baseline survival for the initial biologically derived signature and the least absolute shrinkage and selection operator-derived 5-gene signature in The Cancer Genome Atlas pancreatic adenocarcinoma derivation cohort and the Gene Expression Omnibus dataset GSE71729 validation cohort.

Cohort	Model	Calibration slope	Brier score at 1/2/3 year	Integrated Brier score, 0-3 years	Baseline survival at 1/2/3 years (%)
TCGA-PAAD derivation cohort (n = 177, 93 events)	LASSO-derived 5-gene signature, univariable	1.861	0.183/0.203/0.171	0.168	76.9/42.0/35.4
	LASSO-derived 5-gene signature, multivariable	1.817		0.163
	Initial biologically derived signature, univariable	0.124	0.185/0.240/0.225	0.186	75.5/40.5/35.2
	Initial biologically derived signature, multivariable			0.173
GSE71729 validation cohort (n = 125, 84 events)	LASSO-derived 5-gene signature, univariable	0.154	0.234/0.213/0.163	0.189	64.0/33.3/22.2
	LASSO-derived 5-gene signature, subtype-adjusted	0.223		0.177

GSE71729: Gene Expression Omnibus dataset GSE71729; LASSO: Least absolute shrinkage and selection operator; TCGA-PAAD: The Cancer Genome Atlas Pancreatic Adenocarcinoma.

Values in the derivation cohort, as well as all Brier scores and baseline survival estimates, are apparent, in-cohort estimates: Predicted survival was calculated using the Breslow baseline hazard fitted in the same cohort in which the model was scored. Brier scores were calculated using Graf inverse-probability-of-censoring weighting, with the censoring distribution estimated using the Kaplan-Meier method; the integrated score represents the trapezoidal average over 0-3 years. A Brier score of approximately 0.25 corresponds to an uninformative model. The calibration slope is the Cox coefficient of the linear predictor in the cohort being assessed, with an ideal value of 1.0. Although the five-gene coefficients were estimated in the TCGA-PAAD derivation cohort, the derivation calibration slope is not necessarily exactly 1.0 because the coefficients were obtained using penalized LASSO Cox regression, which introduces coefficient shrinkage. Baseline survival was estimated at the cohort mean risk score at 365 days, 730 days, and 1095 days in the derivation cohort and at 12 months, 24 months, and 36 months in the validation cohort. A fully external Brier score and calibration plot applying the derivation baseline hazard to GSE71729 could not be computed because the derivation baseline was not retained and the 2 cohorts differ in assay platform and time units. The external calibration slope of 0.154 indicates that the effect of the score in GSE71729 is approximately 1/6 of the strength implied by its derivation weights. In the subtype-adjusted validation model, the slope for the score was 0.223, compared with 0.664 for tumor subtype in the same model. The initial biologically derived signature had a calibration slope of 0.124 in its own derivation cohort; that is, it was more poorly calibrated in the data in which it was constructed than the LASSO score was during external testing. No single slope is reportable for the initial-signature multivariable model because its linear predictor combines the score with clinical covariates.

The prognostic performance of the 5-gene signature was evaluated in the independent GSE71729 cohort, comprising 125 primary PDAC tumors with OS data (84 deaths and 41 censored patients). The continuous risk score, prespecified as the primary validation analysis, was not significantly associated with OS in either univariable (HR = 1.17, 95%CI: 0.88-1.55; P = 0.292; C-index = 0.548) or multivariable Cox regression adjusted for tumor and stromal subtypes (HR = 1.25, 95%CI: 0.93-1.66; P = 0.135; C-index = 0.602).

Within the same multivariable model, basal-like tumor subtype remained an independent predictor of poorer survival compared with classical subtype (HR = 1.93, 95%CI: 1.20-3.12; P = 0.007), whereas neither stromal subtype was significantly associated with outcome (normal-like vs low, HR = 0.77, P = 0.508; activated vs low, HR = 1.29, P = 0.468). Thus, the modest improvement in model discrimination (C-index = 0.602) primarily reflected the contribution of tumor subtype rather than that of the molecular signature itself (Figure 7).

Figure 7 External validation of the least absolute shrinkage and selection operator-derived prognostic signature in the GSE71729 cohort. Forest plot showing hazard ratios with 95% confidence intervals for the continuous risk score, median-based high- vs low-risk groups, Moffitt tumor subtype, and stromal subtype.

As a secondary exploratory analysis, patients were dichotomized using the cohort-specific median risk score (Figure 8). Median-based stratification resulted in modest separation of the Kaplan-Meier survival curves (log-rank P = 0.044; HR = 1.55, 95%CI: 1.01-2.39), with shorter survival in the high-risk group.

Figure 8 Distribution of the least absolute shrinkage and selection operator-derived risk score in the external GSE71729 cohort. Patients are ranked by increasing risk score and stratified into low- and high-risk groups based on the median cut-off.

Exploratory cut-point optimization using maximally selected rank statistics identified an optimal threshold of -0.457, which was nearly identical to the cohort median of -0.455 and produced essentially the same partition (62 patients vs 63 patients). The same log-rank statistic was the maximum across 101 candidate thresholds and corresponded to a multiplicity-corrected P value of 0.45. No dichotomization of the risk score remained significant in this cohort after accounting for threshold selection; the nominal median-split result should therefore be considered hypothesis-generating only (Figure 6B). None of the individual candidate genes was significantly associated with OS after FDR correction across the 28 genes tested (lowest adjusted P = 0.40 for PRSS3; adjusted P values for the 5 constituent signature genes were 0.40 for PRSS3; 0.69 for FKBP11, SLC16A10, and SLC43A1; and 0.74 for AMY2B). The prognostic signal of the signature therefore derives from the weighted combination of its components rather than from any individually significant constituent gene, while the signature itself remains significant after FDR correction at the model level in the derivation cohort. The validation result also did not survive correction for multiple testing: Across the family of 8 primary prognostic analyses, Benjamini-Hochberg adjustment left only the 2 derivation-cohort LASSO analyses significant (adjusted P = 0.0009 for univariable and 0.0032 for multivariable analysis), while all validation-cohort tests were non-significant, including the median split (adjusted P = 0.092) and optimized cut-point (adjusted P = 0.092). Applying the TCGA-derived thresholds to GSE71729 without recalibration confirmed that they were not transportable: The locked median yielded an HR of 1.35 (95%CI: 0.83-2.22, P = 0.230), and the locked optimal threshold yielded an HR of 1.24 (95%CI: 0.70-2.21, P = 0.460). The reason for this failure was structural rather than statistical. Derivation risk scores were entirely negative (-2.16 to -0.62), whereas validation scores ranged from -1.99 to 1.11, and 60% of patients in GSE71729 had scores higher than those of every patient in TCGA-PAAD. Consequently, the locked median classified 72% of the validation cohort as high risk, while the locked optimum classified 80% as high risk. Because the 2 cohorts were assayed on different platforms, the scores were not on a common numerical scale, preventing a fixed numerical threshold from partitioning the validation cohort in the same manner as the derivation cohort.

To investigate whether the lack of independent prognostic performance reflected overlap with established molecular subtypes, the distribution of risk scores across Moffitt tumor subtypes was examined. The continuous risk score did not differ significantly between classical and basal-like tumors (Cohen’s d = 0.22; P = 0.284), and none of the 5 constituent genes showed differential expression according to subtype (all P ≥ 0.07). These findings suggest that the signature is neither a surrogate marker for the Moffitt molecular classification nor collinear with tumor subtype.

Taken together, the external validation showed that although median-based stratification produced modest separation of the survival curves, the continuous risk score did not retain independent prognostic significance in the external cohort. The 5-gene signature should therefore be considered an exploratory, hypothesis-generating model rather than a clinically validated prognostic tool.

DISCUSSION

This study combined a systematic appraisal of tumor-derived prognostic biomarkers in PDAC with an independent exploratory modelling analysis, and both components converged on the same conclusion. All 57 biomarker studies retained for detailed synthesis reported an independent prognostic association in a multivariable survival model. Yet none of the positive findings was supported by evidence of adequate methodological quality. Moreover, the exploratory 5-gene signature developed here reproduced precisely the weakness identified in the published literature: Strong performance in the cohort in which it was derived, followed by loss of that performance on independent testing. The central message of this work is therefore not the identification of a new prognostic marker, but rather the demonstration that the reproducibility gap in PDAC biomarker research is systematic rather than incidental.

The quality of the underlying evidence constrains the inferences that can be drawn from the synthesis. Of the 57 biomarker studies, 35 were recorded as retrospective analyses of public, institutional, or combined cohorts, 1 (a nationwide observational programme with integrated whole-genome and transcriptome profiling) was prospective, and the study design of 21 did not permit classification. No study achieved a low overall risk of bias according to QUIPS. Among the 46 studies assessed using PROBAST, the overall risk of bias was rated as high in 21 and unclear in 25, with none rated low. Certainty of evidence was very low across all 4 GRADE streams, with a 1-level risk-of-bias downgrade for multi-gene signatures and 2-level downgrades for the remaining streams, compounded by imprecision due to small analyzed cohorts and the reporting of effect estimates on non-comparable scales, which precluded pooling. Although numerous candidate biomarkers and multigene signatures continue to be proposed[36,46], only a minority demonstrate independent prognostic value supported by both adequate multivariable adjustment and external validation[52,65,85]. This pattern mirrors the wider biomarker literature, in which statistical significance is frequently not accompanied by reproducibility or clinical utility[74,77,78,82,86,87].

The clinical rationale for pursuing tissue-derived molecular markers nonetheless remains strong. CA19-9 is still the only circulating marker in routine use, but its imperfect sensitivity and specificity, false-positive elevations in inflammatory and biliary diseases, and inability to represent tumor heterogeneity directly support the need for molecular stratification[7]. Its dependence on Lewis antigen expression further compounds these limitations, leaving a subset of patients without a usable circulating marker even in the presence of advanced disease[8]. Transcriptomic, proteomic, and integrative multi-omics readouts can, in principle, simultaneously capture tumor phenotype, stromal remodeling, immune context, and pancreatic differentiation status[88,89]. It is this breadth, rather than any single analyte, that motivated the present analysis.

The proteomic component of this work illustrates what such readouts actually detect in PDAC tissue. Paired analysis of CPTAC-PDAC identified 29 significantly dysregulated proteins, all of which were downregulated in tumor relative to adjacent-normal pancreas. These included classical digestive enzymes (PRSS1, CPA1, CPA2, CPB1, CTRB2, PNLIP, CELA2A, CELA3A, AMY2A, and PLA2G1B) alongside transport and metabolic regulators (SLC43A1, SLC16A10, AQP8, and AQP12B). Enrichment analysis yielded concordant findings, with digestion as the most significant GO term (GO: 0007586; FDR = 5.72 × 10^-5) and pancreatic secretion as the dominant KEGG pathway (hsa04972; fold enrichment = 58.50, adjusted P = 2.04 × 10^-21). The dominant proteomic signal in resected PDAC therefore reflects progressive loss of exocrine pancreatic identity and functional collapse of the secretory apparatus, rather than activation of a single uniform oncogenic programme[10]. This finding has a practical consequence that may be easily overlooked: A tumor-vs-normal contrast of this magnitude is driven by the disappearance of normal tissue function, and abundance differences of this kind do not necessarily carry prognostic information among patients who all have established disease.

This observation bears directly on the feature-selection strategy adopted here. Candidates entered the model based on a tumor-vs-adjacent-normal proteomic contrast and were subsequently mapped to the transcript level for penalized modelling in TCGA-PAAD, an approach that embeds 2 assumptions. The first is cross-platform transferability: Protein and messenger RNA abundance correlate only moderately in PDAC tissue, with a median RNA-protein correlation of 0.35 reported in the CPTAC proteogenomic analysis[90], meaning that a protein selected on the basis of a tumor-normal differential is not necessarily represented faithfully by its transcript in an independent cohort. The second, and more consequential, assumption is that a tumor-vs-normal contrast is equivalent to a malignancy contrast. All 29 significantly dysregulated proteins were downregulated in tumor tissue and were dominated by acinar digestive enzymes (AMY2A, PRSS1, PNLIP, CTRB2, and related proteins), a pattern generated primarily by the loss of exocrine parenchyma and variable admixture of non-neoplastic tissue in bulk samples rather than by properties of the malignant cells themselves. Filtering along this axis therefore preferentially selects features that track acinar depletion and normal-tissue content, which may vary between samples and cohorts for reasons unrelated to outcome, while under-sampling the stromal, immune, and proliferative programmes that plausibly carry prognostic information among patients who already have established disease. We regard this as the most likely explanation for the pattern observed here: A strong association in the cohort in which the coefficients were fitted, near-chance discrimination, and loss of the continuous predictor on external testing. The study design cannot distinguish a weak, distributed biological signal from a signal arising from tissue composition, and we therefore do not claim that the 5 retained genes represent tumor-intrinsic biology. A discovery step anchored on within-tumor contrasts high vs low risk among tumors, tumor purity as a covariate, or deconvolution-adjusted analysis would be more appropriate for a prognostic question, and we identify this as the primary methodological revision required before further evaluation of the signature.

That expectation was borne out in the derivation cohort. After adjustment for clinical covariates, none of the individual candidate genes was independently associated with OS after accounting for multiple testing, while the initial biologically derived signature was null in both univariable (HR = 1.13, P = 0.239) and multivariable Cox regression (HR = 1.08, P = 0.480), with discrimination close to chance (C-index = 0.52-0.56). Penalized selection altered the picture only within the same dataset: The LASSO-derived 5-gene signature (FKBP11, SLC43A1, SLC16A10, PRSS3, and AMY2B) was strongly associated with OS in univariable models (HR = 6.41, P < 0.001) and multivariable models (HR = 6.13, P < 0.001), yet discrimination improved only marginally, reaching a C-index of 0.599. The dissociation between a large HR and a near-uninformative C-index is itself indicative of a weak, distributed signal, and because no optimism correction was applied, the derivation-cohort estimates should be considered optimistic. Retention of a gene by the LASSO procedure reflects statistical selection within 1 cohort and establishes neither biological relevance nor transportability. The same caution applies to the cross-platform observation that PRSS3 protein is more abundant in tumor than in normal pancreas[91], which demonstrates detectability of the analyte.

External validation in GSE71729 (n = 125) supported this interpretation. The prespecified continuous risk score was not associated with OS in either univariable analysis (HR = 1.17, 95%CI: 0.88-1.55; P = 0.292; C-index = 0.548, 95%CI: 0.482-0.627, an interval including 0.50) or after adjustment for tumor and stromal subtypes (HR = 1.25, 95%CI: 0.93-1.66; P = 0.135; C-index = 0.602, 95%CI: 0.535-0.671). Within the same model, basal-like subtype remained an independent predictor of poorer survival than classical subtype (HR = 1.93, 95%CI: 1.20-3.12; P = 0.007), whereas neither stromal subtype was informative, indicating that the modest apparent gain in discrimination was largely attributable to subtype rather than the signature. Importantly, this does not appear to represent a signature acting as a proxy for molecular subtype: The continuous risk score did not differ between classical and basal-like tumors (Cohen’s d = 0.22; P = 0.284), and none of the 5 constituent genes was differentially expressed by subtype (all P ≥ 0.07). Thus, the signature was biologically distinct from the Moffitt classification but failed to add independent prognostic information alongside it. Moreover, because no nested-model comparison against subtype, stage, grade, and resection status was undertaken, its incremental value over established prognostic factors remains formally undetermined.

Only the secondary median-dichotomized analysis retained nominal significance (log-rank P = 0.044; HR = 1.55, 95%CI: 1.01-2.39; median survival, 14 months vs 20 months), and this nominal result did not survive correction for threshold selection: The same statistic represented the maximum across 101 candidate cut-points and yielded a corrected P = 0.45. The result also failed Benjamini-Hochberg correction across the 8 primary prognostic analyses (adjusted P = 0.092) and was therefore non-significant on 2 independent grounds. Because the threshold was derived within the validation cohort itself, and because the dichotomized and continuous analyses yielded discordant results in the same patients, this group-level difference cannot be regarded as evidence of clinical relevance. The primary continuous predictor failed validation, and a surviving cut-point effect does not overcome that failure. The gradual, unimodal distribution of risk scores in the derivation cohort, with no discrete clustering, is consistent with this interpretation and with the underlying biology of PDAC, in which overlapping and adaptive transcriptional states predominate molecular classes.

The clinical trial stream places these findings in a translational context that is arguably more consequential than the modelling result itself. Across the same publication window in which 57 independently significant prognostic biomarkers were reported, none of the 14 interventional trials appraised here used a molecular biomarker to define eligibility, stratify randomization, or specify an endpoint. Where molecular content was present, it was confined to descriptive CA19-9 or carcinoembryonic antigen dynamics, SMAD4 immunohistochemistry, or generic specimen collection. Biomarker discovery and clinical trial design in PDAC therefore appear to operate as parallel, largely non-communicating activities, and further retrospective signature development alone will not close that gap.

Experimentally validated biomarkers

Experimental grounding, rather than statistical novelty, most clearly distinguishes the included studies. Although many published signatures reported significant survival associations and some form of external validation, the majority were based on retrospective transcriptomic analyses of overlapping public datasets; thus, statistical reproducibility did not necessarily correspond to biological robustness. Among the studies retained in the evidence table, a substantial minority incorporated some form of primary experimental data, but in most cases this was limited to expression-level confirmation quantitative reverse transcription polymerase chain reaction, western blotting, immunohistochemistry, or formalin-fixed paraffin-embedded tissue evaluation and only 2 studies provided integrated mechanistic validation combining functional experiments with in vivo models[20,21].

Those 2 studies were also the most biologically informative in the set. One reported that apolipoprotein B messenger RNA editing enzyme catalytic polypeptide-like-associated mutagenesis contributes to genomic instability, RAS-driven tumor evolution, intratumoral heterogeneity, defective DNA damage response, and the development of an immunosuppressive microenvironment characterized by reduced cluster of differentiation (CD) 8⁺ T-cell infiltration, increased regulatory T-cell infiltration, and enrichment of CD73⁺ cancer-associated fibroblasts[20]. The second identified the FOXM1-junction plakoglobin signaling axis as a driver of epithelial plasticity, epithelial-to-mesenchymal–associated metastasis, protein kinase B/Wnt activation, and chemoresistance[21]. Neither finding can be reduced to a single stable expression threshold, which is precisely why these studies are informative: They describe biological processes that vary within and across tumors over time.

This provides a coherent explanation for the poor external transportability documented throughout this review, including in the present model. If PDAC progression reflects a dynamic continuum of genomic instability, RAS-driven evolution, epithelial plasticity, stromal remodeling, immune suppression, and progressive loss of exocrine differentiation, rather than a set of stable molecular states, then a fixed coefficient-weighted score fitted in 1 cohort would be expected to attenuate in another cohort differing in case mix, platform, and treatment era[20,21]. From this perspective, attenuation during validation is the predicted behavior of the method rather than an anomaly specific to any particular signature.

The composition of our own signature can be interpreted within that framework, although only as a hypothesis. PRSS3 has been linked to invasive and adaptive tumor behavior[92]; the amino acid transporters SLC16A10 and SLC43A1 are consistent with a metabolic and stress-adaptive phenotype[92,93], FKBP11 with secretory and endoplasmic reticulum stress responses, and AMY2B with preservation of exocrine identity the axis that dominated the proteomic contrast[94]. However, these correspondences are based solely on the identities of genes retained by penalized regression in a single cohort, without functional experiments performed in the present study; membership in the signature cannot establish that the model captures a biologically relevant process. These associations should therefore be regarded as candidate hypotheses for functional testing rather than biological validation of the model.

Methodological considerations

Several methodological considerations qualify these conclusions. The included studies were heterogeneous in biomarker class, measurement platform, patient population, cut-point definition, and analytical approach, precluding quantitative pooling and limiting direct comparison; effect estimates reported on non-comparable scales cannot be reconciled through narrative synthesis alone. The recurrent use of the same public resources (TCGA, GEO, ICGC, and CPTAC) for both discovery and validation across the literature introduces a risk of data circularity and optimistic performance estimates that cannot be excluded by any individual study. No included study reached a low risk of bias, and GRADE certainty was very low across all 4 evidence streams; the synthesis therefore rests on uniformly weak evidence. The internal descriptive categories in Table 2 were developed by the authors for this review, have not been externally validated, and do not substitute for the formal appraisal. These categories are also limited by the completeness of the extraction record rather than by the rule used to generate them: A study is categorized as A when external validation is recorded and no limitation is noted, while the absence of a recorded limitation reflects only the absence of a note in the extraction sheet rather than positive evidence of methodological quality. The resulting distribution is therefore top-heavy, and more complete recording of study limitations, rather than a stricter rule, would refine it. The screening log retained only the final consensus decision for each record; consequently, inter-rater agreement between the 2 reviewers could not be quantified, and no agreement statistic is reported. This represents a limitation in reporting transparency rather than a departure from the screening procedure itself. The review was not prospectively registered in PROSPERO or another public registry. Although a written protocol and predefined eligibility framework were applied before screening, the absence of registration reduces methodological transparency and prevents selection bias from being fully excluded. The 14-month search window further limits the synthesis to the most recent literature rather than the complete body of PDAC biomarker evidence. In 23 records, it could not be determined from the available text whether an adjusted multivariable model had been fitted for the nominated biomarker; these records are reported as a distinct ‘unclear’ category rather than being assigned to either group. Because non-significance is a plausible reason for an adjusted estimate to remain unreported, this group represents a larger potential source of outcome-dependent selection than the 3 studies excluded because the biomarker did not retain independent significance after adjustment.

Limitations

The modelling analysis has its own, partly distinct limitations. It is based on retrospective datasets with batch effects and incomplete clinical annotation, and adjustment in the external cohort was limited to tumor and stromal subtypes; residual confounding by stage, resection status, adjuvant treatment, and performance status therefore cannot be excluded. The search window represents the principal limitation on the scope of the review rather than a detail of its conduct: Restricting eligibility to studies published between 1 January 2025 and 6 March 2026 means that the synthesis characterizes current discovery practice but cannot address the historical development of PDAC prognostic biomarkers, and markers established before this period fall outside its scope. The findings should therefore be interpreted as an appraisal of contemporary pipelines rather than a cumulative account of the field. A further limitation concerns threshold transportability: The TCGA-derived cut-points do not transfer to the validation cohort because the risk scores are not on a common numerical scale across platforms, with 60% of patients in GSE71729 scoring above every patient in TCGA-PAAD. Any application of this signature based on a fixed numerical cut-off would therefore require cohort-specific recalibration, and no claim of clinical utility based on a portable threshold can be supported in the model’s current form. Feature selection was anchored on a tumor-vs-adjacent-normal contrast without adjustment for tumor purity or cellular composition; consequently, selection driven by acinar depletion and non-neoplastic tissue content cannot be excluded. Moreover, discovery was performed at the protein level and modelling at the transcript level despite only moderate protein-messenger RNA concordance in PDAC tissue. The statistical characterization also remains incomplete: Although CIs for the concordance indices, calibration slopes, and inverse-probability-of-censoring-weighted Brier scores are now reported, time-dependent discrimination was not assessed, no bootstrap optimism correction was applied to the derivation-cohort estimates, and no nested-model comparison against established prognostic factors was performed. The precision of the reported discrimination therefore cannot be fully assessed, while the incremental value of the signature remains undetermined rather than absent. Finally, none of the 14 interventional trials appraised here applied biomarker-based eligibility or stratification; thus, the trial stream provides translational context only and cannot be used to assess the clinical utility of any reviewed biomarker.

Future directions

Three priorities follow directly from these findings. First, the analytical standard for prognostic claims in PDAC should shift from significance testing towards demonstrated incremental value. For the present signature specifically, the necessary next step is a nested-model comparison against stage, resection status, adjuvant treatment, age, performance status where recorded, and tumor subtype, with reporting of likelihood-ratio tests and the change in concordance attributable to the signature, together with calibration, time-dependent discrimination, and optimism-corrected internal validation. Second, discovery should be biologically anchored rather than purely computational: The 2 mechanistically validated studies in this review generated more transportable insights than any of the retrospective signatures, and integrative multi-omics designs capturing tumor, stromal, and immune compartments in prospectively collected, well-annotated cohorts are better suited to a disease characterized by adaptive rather than stable molecular states. Third, and most consequentially, molecular biomarkers must be prospectively embedded in trial design for stratification, eligibility, or response assessment because the complete absence of biomarker-driven design across the 14 trials appraised here means that clinical utility currently cannot be tested even in principle. Standardization of analytical pipelines, cut-point definitions, and reporting frameworks would support all 3 priorities. Translating these priorities into study design yields 5 concrete requirements, each addressing a specific failure documented in this review. First, prospective cohort design with consecutive enrolment: Every study in the evidence base for which the design was recorded was retrospective apart from 1, and study participation was among the weakest domains in the appraised literature. A prognostic claim intended to change clinical practice should therefore be based on a cohort assembled prospectively for that purpose, with eligibility, sampling, and follow-up defined in advance and reported according to a registered protocol. Second, a standardized measurement platform with a documented harmonization pathway. The transportability failure in our own model was not necessarily a failure of the underlying biology: The derivation and validation cohorts were assayed on different platforms, no cross-platform normalization was applied, and 60% of validation patients had scores above those of every derivation patient; consequently, a fixed numerical threshold could not partition the external cohort as intended. Any signature intended for application beyond its derivation cohort should therefore be specified together with the assay, normalization procedure, and a defined method for mapping values onto a common scale. This mapping should be tested before the signature itself is evaluated for prognostic performance. Third, cut-off values should be defined a priori and their derivation clearly reported. Data-derived thresholds were used in the great majority of studies appraised here, and our own analysis illustrates why this is problematic: The median split in the external cohort was nominally significant, yet the same statistic represented the maximum across 101 candidate thresholds and yielded a corrected P = 0.45; it also failed correction for multiple testing across the primary analyses. A threshold selected by maximizing a test statistic in the same cohort used to evaluate it is more data-dependent than the median split it replaces, not less. Thresholds should therefore be fixed before examination of the validation cohort and reported together with the rule used to derive them. Fourth, biomarkers should be integrated with established clinical prognostic factors rather than simply presented alongside them. In our external cohort, the subtype-adjusted model appeared to discriminate better than the score alone, but this gain was attributable to tumor subtype rather than to the signature, whose coefficient was not significant. Demonstrating incremental value requires nested-model comparison against stage, resection status, adjuvant treatment, performance status, and molecular subtype, with reporting of likelihood-ratio tests and the change in discrimination attributable to the biomarker, together with calibration and optimism-corrected internal validation. Reporting an HR in isolation cannot establish whether a biomarker provides prognostic information beyond established clinical factors. Fifth, biomarker-guided trial design is required. None of the 14 contemporary interventional trials appraised here used a molecular biomarker to define eligibility, stratify randomization, or specify an endpoint, whereas the same publication window yielded 57 studies reporting independent prognostic associations. Until candidate biomarkers are prospectively embedded in trials as stratification factors, eligibility criteria, or prespecified secondary endpoints, with samples collected according to a common protocol clinical utility cannot be tested even in principle. Without this step, the discovery literature will continue to accumulate associations that no trial is designed to act upon.

CONCLUSION

This review identified 57 biomarker studies that fitted adjusted multivariable survival models for the nominated biomarker, all of which reported independent prognostic associations. None of the studies was supported by evidence of adequate methodological quality, and all fell within GRADE streams rated as having very low certainty. In parallel, 14 contemporary trials were identified in which molecular biomarkers played no role in trial design. Our exploratory 5-gene signature reproduced this pattern rather than resolving it: Although it separated risk groups after median dichotomization, discrimination was only modest in the derivation cohort (C-index = 0.599), and its prespecified continuous predictor failed external validation (HR = 1.25, 95%CI: 0.93-1.66, P = 0.135). Meanwhile, basal-like subtype remained independently prognostic in the same model (HR = 1.93, 95%CI: 1.20-3.12, P = 0.007). Although the signature is biologically distinct from the Moffitt classification, it provided no demonstrable prognostic information beyond it and should therefore be regarded as an exploratory, hypothesis-generating result requiring prospective evaluation and formal demonstration of incremental value over tumor subtype and standard clinical prognostic factors, rather than as a validated prognostic tool. On the basis of current evidence, the rate-limiting step in PDAC prognostic biomarker research is not the discovery of additional candidates but the methodological and translational infrastructure needed to test them.