Reproducible and Accurate PD-L1 Immunohistochemistry Biomarker Quantification Using Virtual Multiplex Immunofluorescence Restaining

Sloan-Kettering Inst Can Research, New York NY

Investigators

Abstract

Project Summary Groundbreaking immunotherapy (IO) drugs offer durable response and improved survival in cancer patients who previously had limited treatment options. The semi-quantitative assessment of PD-L1 protein expression on tumor and/or immune cells (lymphocytes, macrophage) by a certified pathologist (“positive”, “negative”, “low”, “medium”, “high”) is the most prevalent biomarker for guiding clinical decision-making in IO. However, PD-L1 expression is difficult to score on standard immunohistochemistry (IHC) slides. The disagreement among pathologists for immune cell PD-L1 scoring is greater than 50%, which can lead to a high percentage of patients receiving IO when they are unlikely to benefit. This discordance could explain why most cancer patients do not benefit from these groundbreaking yet expensive Ios (costs > $200K / year per patient). Thus, there is an urgent unmet clinical need to identify patients who will not benefit from IO. As opposed to standard single-plex IHC, multiplex immunofluorescence (mpIF) staining, though expensive, provides the opportunity to examine panels of several markers (including tumor- and immune-specific markers) individually or simultaneously as a composite while permitting stain standardization, objective scoring, and cut-offs for all the markers; mpIF also has higher sensitivity and diagnostic prediction accuracy than IHC PD-L1 scoring. This opens up unique opportunities to leverage mpIF-stained images (with objective ground truth tumor/immune cell annotations and absolute PD-L1 intensities) coupled with recent deep learning methods to improve the explainability and interpretability of the conventional IHCs broadly. In previous work, restained and co-registered IHC/mpIF whole-slide images (WSIs) were used to create a deep learning virtual mpIF restaining algorithm, DeepLIIF (Deep Learning Inferred IF), for scoring IHC Ki67 and other nuclear markers. DeepLIIF is the only IHC scoring globally available as a public/free cloud-native platform with a user-friendly web interface and AI-ready IHC/mpIF datasets; it has been extensively validated in low- and high-resource settings. Recently, DeepLIIF was extended for more reproducible and accurate visual IHC PD-L1 scoring in tumor cells for lung cancer. The work proposed under this NIH R01 will further improve DeepLIIF PD-L1 scoring by incorporating mpIF immune cell (lymphocytes and macrophage) markers, whole-cell (rather than simple nuclear) segmentation, and large/diverse datasets across lung and bladder cancers. Additionally, the team will (1) validate DeepLIIF PD-L1 tumor and immune cell scoring on thousands of IHC PD-L1 whole-slide images (with manual readouts and ground truth IHC/mpIF for a subset) spanning different antibodies, platforms, and scoring systems, and (2) validate DeepLIIF-derived spatial biomarkers and PD-L1 scores on lung and bladder cancer datasets with clinical outcomes. Successful validation will establish DeepLIIF as an interpretable, tissue non-destructive, and cost-efficient solution to accurate IO patient stratification.