Summary
A medRxiv preprint reports deep-learning models that predict HPV/p16 status and molecular subtypes of vulvar squamous cell carcinoma from routine tissue slides. The HPV-status model performed better than the model distinguishing all three subtypes.
A medRxiv preprint describes two deep-learning models that predict human papillomavirus (HPV) status and molecular subtype in vulvar squamous cell carcinoma (VSCC) from routine histopathology slides. In testing on two independent Dutch cohorts, the HPV-status model performed more strongly than the model tasked with distinguishing all three subtypes.
VSCC includes HPV-associated, HPV-independent/p53-wildtype and HPV-independent/p53-abnormal forms, which have different prognoses. Current classification uses p16 and p53 immunohistochemistry—staining tissue to detect particular proteins. The researchers note that this testing can be incomplete in retrospective patient cohorts, and it uses tissue that may be limited.
How the models were tested
The team developed HERA to predict HPV/p16 status and HERMES to classify the three molecular subtypes. Both used features extracted from whole-slide images stained with haematoxylin and eosin (H&E), a routine method that makes tissue structures visible under a microscope. The models used attention-based multiple-instance learning, an approach that combines information from regions across a slide.
The models were developed using data from 644 patients in three Dutch VSCC cohorts, then tested on 286 patients from two independent cohorts. In this study, the reported area under the receiver operating characteristic curve (AUC) measures how well a model distinguishes between the target groups; it does not by itself describe the accuracy of every individual prediction.
HPV status was easier to predict than all three subtypes
In independent testing, HERA had an AUC of 0.84 (95% confidence interval 0.75–0.91). Its specificity was 0.93 and negative predictive value was 0.94, while sensitivity was 0.68. The model’s predicted groups also showed prognostic separation, similar to HPV testing (log-rank p<0.001). The sensitivity result is important when interpreting the model’s potential: it did not identify every HPV-positive case in these cohorts.
HERMES had a macro-average AUC of 0.70. It distinguished HPV-associated tumours more effectively (AUC 0.80) than the two HPV-independent groups: AUC 0.68 for p53-abnormal and 0.63 for p53-wildtype. The predicted classes showed a corresponding pattern in five-year recurrence rates: 14.0% for HPV-associated, 31.1% for HPV-independent/p53-wildtype and 43.3% for HPV-independent/p53-abnormal tumours (p=0.0051). These results show separation of outcome groups in the study cohorts; they do not mean the model changes patients’ recurrence risk.
The findings suggest that routine H&E slides contain features useful for estimating HPV status and may help researchers analyse large retrospective cohorts where additional staining is missing. The authors say HERA’s discrimination and calibration need external validation across scanners and populations before such use. The study is a preprint, and HERMES’s weaker distinction between the two HPV-independent subtypes remains a central limitation.