Computationally-Guided Development of a Stromal Inflammation Histologic Biomarker in Lung Squamous Cell Carcinoma

Abstract

The goal of this study is to use computational pathology to help guide the development of human-based prognostic H&E biomarker(s) suitable for research and potential clinical use in lung squamous cell carcinoma (SCC). We started with high-throughput computational image analysis with tissue microarrays (TMAs) to screen for histologic features associated with patient overall survival, and found that features related to stromal inflammation were the most strongly prognostic. Based on this, we developed an H&E stromal inflammation (SI) score. The prognostic value of the SI score was validated by two blinded human observers on two large cohorts from a single institution. The SI score was found to be reproducible on TMAs (Spearman rho = 0.88 between the two observers), and highly prognostic (e.g. hazard ratio = 0.32; 95% confidence interval: 0.19-0.54; p-value = 2.5 × 10^-5 in multivariate analyses), particularly in comparison to established histologic biomarkers. Guided by downstream molecular/biomarker correlation studies starting with TCGA cases, we investigated the hypothesis that epithelial PD-L1 expression modified the prognostic value of SI. Our research demonstrates that computational pathology can be an efficient hypothesis generator for human pathology research, and support the histologic evaluation of SI as a prognostic biomarker in lung SCCs.

Figure 1

Study Overview. Note: the image of the human brain is in the Public Domain and was obtained from the Wikimedia Commons (https://upload.wikimedia.org/wikipedia/commons/8/88/PSM_V46_D167_Outer_surface_of_the_human_brain.jpg). The image of the human eye was modified from the original image (“A blue iris. A human eye.” created by user 8thstar at the English language Wikipedia) obtained from the Wikimedia Commons (https://upload.wikimedia.org/ wikipedia/commons/8/84/A_blue_eye.jpg) under the Creative Commons (CC) BY-SA 3.0 license (https://creativecommons.org/licenses/by-sa/3.0/). According to the terms of this license, this particular Fig. 1 is also distributed under CC BY-SA 3.0 terms with no additional restrictions. Abbreviations: tissue microarray (TMA), The Cancer Genome Atlas (TCGA).

Figure 2

Computationally-guided histologic hypothesis generation. (A) Left: unprocessed H&E tissue microarray (TMA) image of lung squamous cell carcinoma. Middle-left: using labeled examples of tumor stroma and epithelia provided by one author, the software learned to divide regions of all TMA images into epithelia (orange) or stroma (blue). The computational classification was correct in many instances, but did have trouble distinguishing inflamed epithelium from inflamed stroma (typically calling all such areas stroma). Middle-right: epithelial objects (yellow = small; orange = medium; brown = large objects); while nearly every tumor nuclei is correctly accounted for, the cytoplasmic borders of some tumors were not appropriately captured, thereby underestimating the extent of at least some tumor cells (i.e. in solid tumor nests [orange areas], there should few to no gaps between epithelial objects). Right: stromal objects. 768 epithelial and 768 stromal features were quantified by the software for each image. After combining with clinical data, four computationally-measured stromal (i.e. from the blue regions) features were found to be significantly associated with overall survival at a cut-off false discovery rate of <0.05. No features from the epithelia (orange region) was significant at this cutoff. (B) The four significant features and representative images from the highest and lowest ranked cases (for illustration, only one of the two images for each case) is shown. Three of the four significant features were associated with the amount of stromal lymphoplasmacytic inflammation by visual review, where more inflammation was associated with better prognosis. A more complete manual review of the four features is available in the Supplementary Data (for Features 1–4).

Figure 3

Human validation of the histologic hypothesis as a prognostic biomarker. (A) Example of manual scoring of SI from observers (Obs) 1 and 2. Patient 1 was deceased and had a low SI score; patient 2 was living and had a high SI score. (B) Kaplan Meier survival analysis for the full TMA dataset. Cases were divided into high SI (red; >median SI score; n = 195) and low SI (black; ≤median SI score; n = 229) groups. Survival was significantly better for the high SI group in comparison to the low group (median survival of 65.0 vs 33.3 months, respectively; log rank p-value = 4.6 × 10⁻⁵). The results for observers 1 (this Figure) and 2 (see Supplementary Figure 2) were similar.

Figure 4

The relationship between SI, PD-L1 expression, and overall survival. (A and B) PD-L1 expression was not strongly associated with inflammation. (A) By gene expression profiling, CD274 (PD-L1) RNA levels did not correlate strongly with RNA levels of genes expressed by immune cells in lung SCC TCGA cases. The numbers in each box are the Spearman rho values for the expression levels of the gene-pair combinations (red = high correlation; white = low correlation). (B) PD-L1 protein expression by immunohistochemistry did not correlate strongly with histologic SI scores in lung SCC TMA cases (Spearman rho = 0.20). (C and D) The prognostic value of the SI score is modified by epithelial PD-L1 expression. Cases from the TMA cohorts were separated by PD-L1 expression (high versus low). (C) SI and survival when PD-L1 expression was low. (D) SI and survival when PD-L1 expression was high. The interaction term for this Cox proportional hazards model was trending towards significance (interaction p-values = 0.056). The results from observer 1 (this Figure) and observer 2 (Supplementary Figure 4) were similar.