Digital pathology turns PD-L1 staining into a more usable predictor for nivolumab in gastric cancer
A deep learning model built from routine pathology slides may help separate gastric cancers that respond to nivolumab from those that do not. In a study published in the British Journal of Cancer, investigators trained a cell-detection network on 1,927 patches from 88 whole-slide images and then tested the resulting computational positive cell ratio, or cPCR, in an independent cohort of 147 patients treated with nivolumab plus chemotherapy.
The appeal is obvious at the bedside. Gastric cancer immunotherapy still depends on imperfect tissue biomarkers, and conventional PD-L1 combined positive score can be difficult to interpret consistently across specimens and observers. A computational readout that extracts immune-staining information directly from digitized pathology slides offers a way to standardize part of that decision-making process without waiting for new assays or more invasive sampling.
The cohort-level response data showed the expected enrichment of benefit in PD-L1-positive disease, but also exposed the limitations of current thresholds. Overall objective response rate was 49.66%. Among patients with PD-L1 CPS ≥ 5, the objective response rate reached 56.12%, compared with 36.73% in those with CPS below 5.
A similar separation emerged with the image-derived measure. Objective response rate was 56.03% in patients with cPCR ≥ 5 and 25.81% in those with cPCR below 5. The pattern suggests that the slide-based model is capturing a biologically relevant immune context rather than merely echoing the original scoring system. In practical terms, tumors with more abundant PD-L1–associated cellular signal may represent a microenvironment already engaged in immune signaling, making them more vulnerable to checkpoint blockade.
One sentence captures the clinical tension in this paper: the new model did not simply outperform the standard marker on raw discrimination alone.
The area under the curve for CPS was 0.586, and the AUC for cPCR was 0.601, a difference that was not statistically significant. Yet net reclassification analysis favored cPCR, with p = 0.0269, indicating that the computational measure improved how patients were sorted into clinically meaningful response categories. In other words, the machine learning output did not need to dominate CPS outright to add value; it only needed to reorganize risk more accurately at the margins where treatment decisions are often made.
Risk scoring extends beyond PD-L1 alone
The study went one step further by combining cPCR with the tumour-stroma ratio through stepwise variable selection. That approach produced a risk-score model that improved prediction of immunotherapy benefit. The logic is straightforward and biologically plausible. Tumor cells do not exist in isolation, and the balance between malignant glands and surrounding stroma can shape immune trafficking, antigen exposure, drug penetration, and the overall suppressive tone of the tumor bed.
Digital pathology may be particularly well suited to capture this interaction because it can quantify features that human eyes recognize only impressionistically. A pathologist may see a tumor as crowded, fibrotic, inflamed, or immune-excluded. A computational model can turn those impressions into repeatable numerical inputs, then combine them with staining patterns to produce a response probability that is more portable across cases.
The result is not a replacement for conventional pathology. It is a tighter computational layer on top of it. For gastric cancer, where checkpoint inhibitor benefit remains variable and patient selection remains challenging, even modest improvement in enrichment could matter if it reduces exposure to treatment unlikely to help.
The model also sits within a broader movement in oncology toward digitized biomarker development. The relevant clinical question is no longer whether artificial intelligence can recognize patterns in histology. The question is whether those patterns improve treatment selection in a way that translates into better response prediction at the point of care. This study addresses that question in a focused immunotherapy setting rather than in abstract validation only.
Still, the evidence remains centered on prediction rather than survival advantage. The abstract provides response data and model performance metrics, along with a progression-free survival figure, but it does not report a definitive practice-changing survival endpoint in the text available here. The sample was also modest, with 147 treated patients in the independent validation cohort. Those features make the findings useful and credible, but they also keep the model in the translational category rather than the ready-for-guideline category.
Even so, the direction of travel is clear. Gastric cancer care increasingly depends on biomarkers that can keep pace with the complexity of tumor immunity, and digital pathology is beginning to offer that possibility without requiring new tissue acquisition. The most useful next step would be external validation across broader populations, scanners, staining platforms, and pathology workflows, because a model that performs well in one setting still has to survive the variability of real-world laboratories before it can influence treatment selection with confidence.