Radiogenomics for predicting p53 status, PD-L1 expression, and prognosis with machine learning in pancreatic cancer

Pancreatic ductal adenocarcinoma (PDAC) is highly lethal, with an overall 5-year survival rate around 6% and limited improvement even after curative surgery. Genomic studies identify frequent alterations in KRAS (>90%), and tumor suppressors p53 (≈50–70%), CDKN2A, and SMAD4. p53 mutations have been linked to worse prognosis, while PD-L1 expression is considered a poor prognostic factor and a biomarker for immune checkpoint inhibitor therapy. Radiomics quantifies imaging phenotypes from CT/MRI; radiogenomics integrates radiomics with genomic information to non-invasively infer tumor molecular profiles. The study’s purpose was to evaluate whether CT-based radiogenomics can predict p53 mutation status and PD-L1 expression in PDAC and relate these to prognosis, offering a simple, inexpensive tool to support precision medicine.

Prior work shows KRAS mutations in >90% of PDAC and common alterations in p53, CDKN2A, and SMAD4. p53 mutations have been associated with poor outcomes in PDAC. PD-L1 expression has been reported as a poor prognostic factor in several cancers, with some reports indicating worse prognosis in PD-L1-high PDAC. Immune checkpoint inhibitors are effective in cancers such as lung cancer and are under consideration in PDAC. Radiogenomics has successfully predicted clinically relevant molecular profiles in breast cancer, lung cancer, and glioblastoma from imaging data. However, before this study, radiogenomic prediction of genetic information such as p53 status and PD-L1 expression in pancreatic cancer had not been reported.

Design and cohort: Retrospective single-institution study of 107 PDAC patients (January 2013–December 2017) who underwent surgery without preoperative chemotherapy and had postoperative follow-up. Ethics approval obtained and informed consent provided. Clinical variables included demographics, tumor markers (CEA, CA19-9), operative details, pathological factors (lymph nodes, margins, cytology, lymphatic/neural/vascular invasion, differentiation), and TNM stage (UICC 8th). For survival analyses, continuous variables were dichotomized at cohort medians. Immunohistochemistry (IHC): p53 detected with mouse monoclonal DO7 (Nichirei), PD-L1 with rabbit monoclonal SP263 (Ventana). Formalin-fixed paraffin-embedded sections (5 μm) processed on VENTANA BenchMark ULTRA with OptiView DAB; heat-induced antigen retrieval (CC1) as specified. Two pathologists quantified percentage of positive tumor cells. p53 status: abnormal defined as either nuclear overexpression (accumulation) or complete absence (null pattern); wild-type resembled adjacent normal weak nuclear staining. For analyses, p53-positive defined as >20% nuclei stained or complete absence. PD-L1 status: threshold set at ≥1% membranous staining of tumor cells; an 80% inter-pathologist agreement criterion was used with consensus resolution for disagreements. CT acquisition and segmentation: 128-detector CT (Siemens SOMATOM Definition Flash). Parameters: 120 kVp, 160 mAs, pitch 0.6, voxel size 0.68×0.68×5 mm. Contrast-enhanced two-phase imaging at 40 s (early) and 120 s (late) post-injection with weight-based dosing; saline flush followed. A diagnostic radiologist (15 years) and a surgeon (7 years) delineated tumor VOIs (VOIpc) on early and late phases. Peritumoral expansion VOIs (VOI+4mm) were generated by 4 mm axial expansion around VOIpc to include peritumoral regions. Radiomics feature extraction: Using PyRadiomics v2.2.0 with absolute rescaling (−150 to 500 HU) and discretization to 64 gray levels. From original images and from Laplacian of Gaussian-filtered and wavelet-transformed images, morphology, histogram, and texture features were extracted. A total of 1,037 features per VOI per phase were obtained; combining early and late phases yielded 2,074 features per VOI type (VOIpc and VOI+4mm analyzed separately). Statistics and machine learning: Group differences for p53/PD-L1 status vs clinicopathologic variables assessed by χ2, Fisher’s exact, or Mann-Whitney U tests. Overall survival (OS) defined from surgery to last follow-up; Kaplan–Meier with log-rank tests; Cox proportional hazards for multivariable analysis. Machine learning pipeline: two-step feature selection—(1) univariate Mann–Whitney U to retain significant features; (2) recursive feature elimination using a random forest. Models constructed with XGBoost using nested cross-validation: inner 5-repeat 5-fold CV for feature selection; outer 10-repeat 5-fold CV for model training/evaluation. For ROC analysis, mean predicted probabilities across 10 repeats were used. Cutoffs chosen from ROC (closest to top-left) to dichotomize predicted status for survival analyses.

- Cohort: 107 PDAC patients analyzed; median age 70; surgical procedures PD (n=70), DP (n=35), TP (n=2). - IHC positivity rates: p53-positive in 75/107 (70.0%); PD-L1-positive in 36/107 (33.6%). - Clinicopathologic associations: PD-L1 positivity correlated with lymph node metastasis (P<0.001) and showed a trend toward higher TNM stage; p53 status showed a sex distribution difference (P=0.036) but no significant associations with other variables. - Prognostic associations (univariate): Higher preoperative CA19-9 associated with worse OS (P=0.004). DP/TP procedures and longer operation time associated with worse OS (both P=0.001). Venous invasion (P=0.013), neural invasion (P=0.043), lymph node metastasis (P<0.010), and T3 (UICC 8th) (P=0.018) associated with poorer OS. Both p53-positive and PD-L1-positive groups had worse prognosis (P=0.008 and P=0.013, respectively). - Multivariable analysis: Among p53, PD-L1, CA19-9, operation time, operative procedure, and lymph node metastasis, all factors except operation time remained independently associated with poorer prognosis. - Radiogenomic model performance: Using VOIpc features, AUCs were 0.705 (p53) and 0.660 (PD-L1); using VOI+4mm (including peritumoral region), AUCs improved to 0.795 (p53) and 0.683 (PD-L1). - Survival by predicted status: Groups defined by model-predicted p53 status showed significantly different OS (P=0.015), with predicted p53-positive associated with poorer survival. Predicted PD-L1 status trended toward prognostic separation but was not significant (P=0.096).

This study demonstrates that abnormal p53 expression and PD-L1 positivity by IHC are independent adverse prognostic factors in PDAC. Crucially, CT-based radiogenomics could non-invasively predict p53 status with moderate-to-good discrimination, particularly when including peritumoral features (VOI+4mm), and the predicted p53 status stratified patient survival. These findings address the central question by showing that imaging-derived features can reflect underlying tumor molecular alterations relevant to prognosis. While PD-L1 prediction was less accurate and did not significantly stratify survival, the trend suggests potential with further optimization. Overall, the results support radiogenomics as a practical, low-cost adjunct for molecular profiling and risk stratification in PDAC, with implications for selecting patients for targeted or immunotherapies and informing precision medicine when tissue sampling is limited or risky.

Radiogenomic analysis of routine contrast-enhanced CT can predict p53 status in PDAC and thereby stratify prognosis; PD-L1 prediction showed modest performance. These findings suggest a non-invasive, inexpensive approach to infer tumor genetics and support precision oncology in PDAC. Future work should focus on larger, multi-center validation cohorts, incorporation of genomic sequencing as ground truth for p53 mutations, refinement of PD-L1 predictive modeling, and prospective evaluation of clinical utility and integration into treatment decision workflows.

- Retrospective, single-center study, which may limit generalizability. - Modest sample size (n=107) with no external validation cohort; models were assessed via nested cross-validation only. - p53 mutation status was inferred from IHC patterns (overexpression/null) rather than confirmed by genomic sequencing. - PD-L1 IHC lacks standardized scoring across studies; a ≥1% threshold was used, which may impact comparability and model performance. - Imaging derived from a single CT scanner/protocol; variability across scanners and protocols was not assessed.