Immune cell clustering is commonly observed in histopathology images. As the frequency and nature of immune cell clustering may represent biological phenomena critical to an immunotherapy response, they are import features to measure to differentiate immune phenotypes, which may predict patient responses to immunomodulating therapies. Computational pathology datasets and unsupervised machine-learning approaches are capable of measuring immune cell clustering using a variety of methods; however, it has not been clear how such measurements might be applied to generate a validated computational pathology score that truly captures the immune phenotype. This work explores a methodology for the development and analytical validation of digital pathology scores for immune cell clustering
derived from the application of unsupervised learning to digital pathology datasets.
Materials and Methods
Computational pathology data derived using Flagship’s Computational Tissue Analysis (cTA®)
platform from 12 non–small cell lung cancer (NSCLC) biopsy samples stained with a validated CD8–Ki-67 duplex immunohistochemistry (IHC) assay were used to create a virtual library of scores that described the clustering of CD8-positive (CD8+) stained cells. The library was created using different combinations of clustering
methods, parameters, and scoring schemes. Scores with high distinguishability measured via a 2-way intraclass correlation coefficient, interrun precision measured via the coefficient of variation, and dynamic range were considered analytically validated. Principal component analysis and hierarchical clustering of the analytically validated scores were used to further optimize selection of the most informative subset of clustering scores from the library.
Results and Conclusions
A total of 17 clustering scores passed the validation criteria. Of these 17 scores, 3 appear to be relatively uncorrelated with each other and capture unique information about the spatial relationships of CD8+ cells. A further analysis of these scores demonstrates the ability to distinguish different immune cell clustering profiles in samples that contain similar biomarker expression levels using the common scoring methods of both overall percentage of positive cells and percentage of positive cells per tissue area.
This process for screening and analytically validating a virtual library of computational biomarker scores appears to hold much promise for bringing cluster-derived computational pathology scores into future clinical applications in a way that is analogous to the analytical validation of traditional IHC assays in support of oncology drug development.