A robust predictive marker of clinical outcomes to immune-checkpoint inhibition in advanced squamous head-and-neck cancer calculated from histopathological slides

Background: Immune checkpoint inhibitors, namely programmed death-1 (PD-1) inhibitors, prolong survival in advanced head and neck squamous cell carcinoma (HNSCC). The combined positive score (CPS) guides treatment selection but has limited predictive value, underscoring the need for development of accurate and practical biomarkers in this space. ENLIGHT-DP predicts clinical outcomes to targeted and immune therapies directly from hematoxylin and eosin (H&E) slide scans. We previously showed that ENLIGHT-DP is predictive of outcomes in HNSCC treated with PD-1 inhibitors and in other indications. Here, we train and test an ENLIGHT-DP biomarker on a cohort of HNSCC cases using cross-validation (CV), and further validate it on two independent cohorts. Methods: We obtained high resolution scans of pre-treatment tumor H&E slides of 89 advanced HNSCC patients from National Yang Ming Chiao Tung University (NYCU) treated with first-line PD-1 inhibitors +/- chemotherapy. Response (RECIST v1.1) was available for all cases, and long-term follow up for 65 of them. Using this dataset, we trained a model to predict response, in leave-4-out-CV. The model consists of (1) splitting the slides into tiles of 256x256 pixels, (2) extracting tile features using a deep learning model, (3) predicting tile states using a multi-layered perceptron, (4) pooling the tile states into a whole slide state using attention-based multiple instance learning, and (5) Predicting the response from the pooled state using a linear layer. We report the CV results on this dataset, as well as the performance of the same model on two previously reported cohorts from Hadassah Medical Center (Hadassah, n=25, Oral Oncology 2025) and Princess Margaret Cancer Centre (BIO2, n=14, JITC 2019, ASCO 2025). Results: The model is predictive of ORR in the NYCU cohort with ROC AUC of 0.67 and was able to stratify progression free survival (PFS, HR: 0.943, p=0.017). Applying the model to the Hadassah cohort resulted in ROC AUC of 0.76, and stratification of PFS (HR: 0.89, p=0.023), while CPS, which was available for that cohort, exhibited no predictive power (AUC=0.47, insignificant PFS association). Applying the model to the BIO2 cohort also resulted in ROC AUC of 0.76 and showed an insignificant trend for PFS prediction. CPS exhibited a much weaker effect for response prediction, with an AUC of 0.56, and insignificant association with PFS. Conclusion: The ENLIGHT-DP IO biomarker predicts response to immunotherapy directly from whole H&E slide image scans and demonstrates high predictive power for ORR to PD-1 inhibitors +/- chemo in HNSCC. Although the model was trained on a relatively small cohort and using only response labels, it achieves significant results in terms of both response and PFS and generalizes well to two independent cohorts, on which it outperforms the commonly used PD-L1 IHC marker.