TP53 is the gene encoding the p53 protein, often called the "guardian of the genome" because it monitors for DNA damage and can trigger cell death when damage is irreparable. Mutations in TP53 are found in approximately 70% of pancreatic ductal adenocarcinoma (PDAC) cases, making p53 dysfunction one of the most common molecular alterations in this cancer.
PD-L1 (Programmed Death Ligand 1) is a protein expressed on the surface of some cancer cells that acts as a brake on the immune system. By binding to the PD-1 receptor on T cells, PD-L1 prevents the immune system from attacking the tumor. PD-L1 overexpression is a mechanism by which pancreatic tumors evade immune destruction, and it predicts response to immunotherapy drugs called immune checkpoint inhibitors.
Both p53 mutation status and PD-L1 expression are independent prognostic factors in PDAC - they predict how aggressively the cancer will behave regardless of other factors. However, determining these biomarker statuses currently requires surgical biopsy, which is invasive and not always feasible. A non-invasive method to predict these statuses would be highly valuable for treatment planning.
Radiogenomics (also called imaging genomics) is a field that seeks to identify relationships between quantitative features extracted from medical images and the molecular or genetic characteristics of the underlying tumor. The core hypothesis is that tumor genetics shape how tumors look on imaging - their shape, density, texture, and structure - and that machine learning can discover these subtle correlations.
If radiogenomic relationships are robust enough, they could allow clinicians to infer a tumor's molecular profile from a standard CT scan obtained during routine clinical care, without the need for a tissue biopsy. This would be particularly impactful in cancers like PDAC where obtaining adequate biopsy tissue is technically difficult.
The field draws on radiomics - the extraction of hundreds of quantitative image features that describe texture, shape, and intensity patterns in a way that goes far beyond what the human eye can perceive. Combining radiomics with machine learning creates a pipeline capable of identifying molecular correlates in imaging data.
The study enrolled 107 patients with confirmed pancreatic ductal adenocarcinoma, all of whom had undergone preoperative CT imaging and subsequent surgical resection. This allowed the researchers to match CT scan features directly to the molecular results obtained from the actual tumor tissue after surgery.
For each patient, the CT scan was used to identify and delineate the tumor volume - a process called tumor segmentation. The segmented tumor region was then analyzed using the open-source PyRadiomics software library, which extracted hundreds of quantitative radiomic features from within the tumor boundary.
Molecular status was determined from the resected tumor specimens. p53 mutation status was assessed by immunohistochemistry (looking for aberrant protein accumulation, which correlates with TP53 mutation), and PD-L1 expression was measured by PD-L1 immunohistochemical staining. These tissue-based results served as the ground truth labels for training and validating the machine learning models.
PyRadiomics extracted a large set of quantitative features from each patient's CT scan within the tumor region. These features fell into categories including: first-order statistics (intensity distributions), shape features (tumor geometry and surface complexity), and texture features derived from methods like GLCM, GLRLM, and GLSZM - approaches that quantify how pixel intensities are spatially arranged within the tumor.
After extracting hundreds of features, the researchers applied feature selection to identify the most predictive subset, reducing the risk of overfitting. The selected features were then used to train an XGBoost (eXtreme Gradient Boosting) model - a powerful tree-based ensemble algorithm that has become one of the most widely used methods in medical machine learning due to its high performance and robustness.
Separate XGBoost models were built for predicting p53 mutation status and PD-L1 expression level. The models were evaluated using receiver operating characteristic (ROC) analysis, which measures a classifier's ability to distinguish between positive and negative cases across all possible decision thresholds.
For predicting p53 mutation status, the XGBoost radiomic model achieved an area under the ROC curve (AUC) of 0.795. An AUC of 0.5 represents no predictive ability and 1.0 represents perfect prediction, so 0.795 indicates moderately strong discrimination between mutant and wild-type p53 tumors based on CT imaging features alone.
For predicting PD-L1 expression level, the model achieved an AUC of 0.683. PD-L1 prediction is a more challenging task because PD-L1 expression is more variable and heterogeneous within tumors than p53 mutation, which may explain the lower but still above-chance performance.
Both models performed significantly better than random chance, demonstrating that there are genuine CT imaging correlates of these molecular biomarkers. The p53 prediction model in particular showed clinically meaningful performance, suggesting that radiomic features may capture structural consequences of p53 loss - such as changes in tumor heterogeneity or infiltrative growth patterns.
The study confirmed that both p53 mutation and elevated PD-L1 expression were independently associated with worse overall survival in the 107-patient cohort. Patients with p53-mutant tumors had significantly shorter survival compared to those with wild-type p53, consistent with p53's fundamental role in controlling cell death and genomic stability.
Patients with high PD-L1 expression also showed worse survival, likely reflecting both the intrinsic aggressiveness of PD-L1-overexpressing tumors and their ability to suppress anti-tumor immune responses. However, PD-L1 high tumors are also more likely to respond to checkpoint inhibitor immunotherapy, which creates a nuanced clinical picture.
Multivariate analysis confirmed that both biomarkers remained independent prognostic factors even after adjusting for other clinical variables, including tumor stage, resection margin status, and lymph node involvement. This underscores their biological importance and potential clinical utility in guiding treatment decisions.
If radiogenomic prediction models can be sufficiently validated, they could transform the clinical management of PDAC. Currently, determining a tumor's molecular profile requires tissue - either from surgical resection or biopsy - and many patients are not surgical candidates. A reliable CT-based prediction of p53 and PD-L1 status would enable molecular-informed treatment planning for all patients.
For PD-L1 in particular, predicting expression from CT could help identify which PDAC patients are most likely to benefit from immunotherapy. While immune checkpoint inhibitors have largely failed in unselected PDAC cohorts, enriching trial populations for PD-L1-high tumors using radiogenomic selection could improve response rates.
The authors acknowledge that the 107-patient dataset is relatively small for robust machine learning validation, and external validation in independent patient cohorts is needed before clinical application. Prospective studies will also need to address challenges of imaging protocol standardization across different CT scanners and institutions.
The standard radiomics workflow begins with medical image acquisition (CT in this case), followed by tumor segmentation - drawing boundaries around the tumor volume of interest. Segmentation can be performed manually by radiologists or semi-automatically with computer assistance, and its accuracy heavily influences all downstream results.
After segmentation, a large number of radiomic features are calculated within the segmented region. PyRadiomics, used in this study, is an open-source Python library that standardizes this process and implements the Image Biomarker Standardisation Initiative (IBSI) guidelines, helping ensure reproducibility across different institutions and studies.
Feature selection is then performed to reduce dimensionality - selecting the most informative features while discarding redundant or noisy ones. This step is critical in radiomics because extracting hundreds of features from a relatively small patient cohort creates a high risk of spurious correlations. Proper feature selection and cross-validation are essential safeguards against overfitting.