Saad Nadeem, PhD
Assistant Attending Computer Scientist
Memorial Sloan Kettering Cancer Center
Objective PD-L1 Immunohistochemistry Scoring Using Virtual Multiplex Immunofluorescence Restaining
Background: The reported PD-L1 score is a semi-quantitative calculation such as a visual estimate of the percent positive cells or the 'Combined Positive Score' which is an estimate of the number of PD-L1 staining cells (tumor cells, macrophages, lymphocytes) divided by the total number of viable tumor cells and multiplied by 100. Standard clinical chromogenic PD-L1 IHC staining has a limited dynamic range and does not allow for multiple marker assessments on the same slide needed to reliably determine which cells are expressing PD-L1, such as tumor or immune cell subsets; the hematoxylin counterstain in standard IHC slides can also mask weak positivity, causing misclassification of highly-prognostic PD-L1-low as PD-L1-negative. Moreover, the disagreement among pathologists for immune cell (IC) scoring of PDL1 on IHC has been reported to be above 50% and for macrophages it is above 80%, which means patients are not being stratified to the most effective treatment. Thus, current research/commercial PD-L1 scoring solutions relying on manual cell-phenotype annotations from individual pathologists to train state-of-the-art deep learning algorithms are highly error-prone.
Methods: To address this unmet clinical problem, we have developed a new deep learning algorithm that utilizes co-registered same-slide (single-plex) IHC and high-dynamic-range multiplexed immunofluorescence (mpIF) data to virtually translate low-cost/prevalent IHC images to more informative mpIF-like representations, creating a quantifiable Deep-Learning-Inferred ImmunoFlourescence (DeepLIIF) image. In essence, our DeepLIIF model performs virtual cell phenotyping (by inferring mpIF CD3 for lymphocytes, CD68 for macrophages, and Sox10/PanCK for tumor cells) and allows PD-L1 expression quantification in specific cells for more accurate and objective cell-based biomarker derivation. Building on the success of our published (Nature Machine Intelligence'22 & CVPR'22) DeepLIIF cloud-native platform (https://deepliif.org) for IHC Ki67/ER/PR and CD3/CD8 scoring, we extend the DeepLIIF pltaform functionality to handle IHC PD-L1 scoring (PD-L1 support will be enabled few days before Pathology Visions'22 conference).
Results: We showed that the inferred DeepLIIF images allowed more consistent and reproducible IHC PD-L1 scoring by multiple pathologists for urothelial carcinoma and melanoma patients.
Conclusions: Denovo IHC and mpIF stained slides provide the most accurate/objective training data for deep learning algorithms. DeepLIIF translated/phenotyped mpIF representations can improve the consistency and interpretive capacity of standard clinical IHC and in particular, improve the predictive value of PD-L1 IHC as a biomarker for checkpoint blockade therapy.
Dr. Saad Nadeem is an Assistant Professor in the Departments of Medical Physics and Pathology at Memorial Sloan Kettering Cancer Center (MSKCC). He completed his PhD in Computer Science from Stony Brook University in 2017 and his postdoc from MSKCC in 2019 before transitioning to Assistant Professor position. His lab develops advanced mathematical and machine learning techniques for analyzing patient data at multiple scales (macro: radiology/radiation oncology/surgery/endoscopy, meso: pathology, and micro: molecular - genomics / proteomics / transcriptomics / metabolomics) to improve patient outcomes. The lab is specifically focused on building user-friendly tools that seamlessly fit into the clinical workflows and facilitate accurate and timely diagnosis/prognosis/decision-making while aiding in novel biomarker discovery.