Self-identified African ancestry and pathomics-derived tumor-infiltrating lymphocyte scores in colon and rectal adenocarcinoma whole-slide images: an exploratory multicohort whole-slide image study.

Introduction: AI assisted computed higher tumor infiltrating lymphocyte% (TIL%) scores from hematoxylin and eosin (H&E) stained whole slide images (WSI) may serve as a useful biomarker for stratifying patients for treatment with immune checkpoint inhibitors. Methods: To test the hypothesis that self-identified African ancestry (AA) vs. European ancestry (EA) was associated with higher TIL% scores, WSI were assembled from three diverse cohorts of formalin-fixed paraffin embedded (FFPE) colon and rectal adenocarcinoma (COAD-READ) tumor tissues: 1.) 242 AA vs. EA WSI downloaded from The Cancer Genome Atlas (TCGA)-COAD-READ database; 2.) 97 independent US self-identified AA vs. EA WSI assembled from three US medical centers; 3.) 48 (of 51) Nigerian WSI assembled from a single Nigerian medical center. There were 33 self-identified AA WSI suitable for analysis in the TCGA cohort and 49 self-identified AA WSI in the US cohort. For the TCGA cohort, 241 whole transcriptome RNA-sequence data were generated from parallel frozen tumor samples collected alongside the FFPE samples. For the US cohort, 87 RNA-seq enriched by exome capture data were generated from the same FFPE blocks as WSI. Results: No difference in median TIL% scores was detected between the combined TCGA and US AA cohorts and the Nigerian cohort. Exploratory multiple regression analysis of the combined TCGA and US cohorts revealed that AA had a higher TIL% score (Estimated difference = 3.01%, 95% CI (0.54%, 5.49%), P-value = 0.0170), while controlling for MMR/MSI status and other covariates. However, this result needs to be interpreted with caution because of differences between the TCGA and US cohorts with respect to representation of self-identified AA. After adjustment for cohort in the model, the association of AA with a higher TIL% score was no longer significant (Estimated difference = 2.05%, 95% CI(-0.59%, 4.70%), P-value = 0.1277). Moderate correlations were detected in the TCGA and US cohorts analyzed separately between computed TIL% scores and CIBERSORTx estimates of lymphocyte and T-cell abundance. Moderate correlations were also detected between TIL% scores and CXCL10 and CCL5 gene expression values. Conclusions: These preliminary results underscore the need for expanding the representation of minority populations in publicly accessible datasets.