PD-L1 expression assessed by the combined positive score (CPS) is an established biomarker used to predict responses to immune checkpoint blockade in gastric cancer. Automated, quantitative digital pathology approaches have potential to standardise and refine PD-L1 assessment and to extract additional features of the tumour microenvironment that may improve prediction of immunotherapy benefit.
The authors aimed to develop a deep learning–based model that uses PD-L1 (clone 28-8) immunohistochemistry on whole-slide images to generate a computational positive cell ratio (cPCR) and to test whether cPCR can predict response to nivolumab in gastric carcinoma. A secondary objective was to determine whether combining cPCR with other histopathological features could yield a risk-score model that improves prediction of immunotherapy outcomes.
A cell-detection neural network was trained using 1,927 image patches derived from 88 whole-slide images. The trained network produced a quantitative measurement termed the computational positive cell ratio (cPCR), representing the proportion of PD-L1–positive cells detected by the algorithm on PD-L1 28-8–stained slides.
Predictive performance of cPCR was evaluated in an independent clinical cohort of 147 patients who received nivolumab in combination with chemotherapy. The study compared ORR (objective response rate) and discrimination metrics between conventional PD-L1 CPS and the automated cPCR measure. The authors also applied stepwise variable selection to identify additional histopathological features to combine with cPCR for a multivariable risk-score model.
In the evaluation cohort (n = 147), the overall ORR was 49.66%.
When patients were stratified by PD-L1 CPS using a threshold of 5, ORR was 56.12% in patients with CPS ≥ 5 and 36.73% in those with CPS < 5. Using the computational measure, ORR was 56.03% for cPCR ≥ 5 and 25.81% for cPCR < 5.
Discriminative performance quantified by area under the receiver operating characteristic curve (AUC) was 0.586 for CPS and 0.601 for cPCR; these AUCs did not differ significantly. However, a net reclassification analysis indicated superior predictive performance for cPCR with a reported p value of 0.0269.
The authors used stepwise variable selection to integrate cPCR with the tumour–stroma ratio and thereby constructed a risk-score model. This combined model improved the ability to predict benefit from nivolumab-containing therapy compared with CPS or cPCR alone, as assessed by the analyses reported in the article.
The study applied stepwise selection procedures to identify complementary histopathological variables that, together with the automated cPCR, enhance prediction of immunotherapy response. The tumour–stroma ratio was selected and incorporated into a risk-score framework. The combined score produced better classification of responders versus non-responders than the PD-L1 CPS alone, according to the net reclassification and model comparison metrics reported.
The article includes several illustrative figures that document the development and evaluation pipeline:
Reported numeric performance highlights include the ORR values by CPS and cPCR thresholds, AUCs for CPS (0.586) and cPCR (0.601), and a statistically significant net reclassification improvement favoring cPCR (p = 0.0269).
The authors conclude that an automated cPCR-based model derived from PD-L1 28-8 digital pathology can forecast response to nivolumab in gastric carcinoma. While AUC differences between CPS and cPCR were small and not statistically different, net reclassification analyses and the risk-score model that combines cPCR with tumour–stroma ratio provided improved prediction of immunotherapy benefit. The study presents the cPCR measurement and a risk-scoring approach as promising tools to optimise patient selection for nivolumab-based regimens in gastric cancer.
The article states that all relevant data related to the study are included within the article or in the supplementary materials. The authors note that further data will be provided upon reasonable request to the corresponding author.
The paper cites prior work on artificial intelligence in histopathology, PD-L1 as a biomarker in gastric cancer, and computational assessment of tumour immune microenvironment, among other background literature. Specific numbered references and their details are provided in the original article.