Deep learning–powered H&E whole-slide image analysis of endothelial cells to characterize tumor vascular environment and correlate treatment outcome to immunotherapy.

2578 Background: Beyond their vascular function, endothelial cells (ECs) regulate tumor growth through recently described angiocrine signaling, influencing cancer progression and treatment outcomes. Here, we applied a deep learning model, which we validated using spatial transcriptomics data, to a pan-cancer dataset to analyze the EC distribution associated with the response to immuno-oncology (IO) treatments. Methods: An AI-powered H&E analyzer, Lunit SCOPE IO, quantifies tumor microenvironment EC density and tumor-infiltrating lymphocytes (TILs) density in cancer epithelium and stroma. We validated AI-powered cell type prediction by evaluating cell-specific gene expression through spatial transcriptomics (10x Xenium). 7,467 pan-carcinoma samples from The Cancer Genome Atlas (TCGA) were analyzed for EC distribution and overall survival (OS). From a previously described multi-center, multi-national pan-cancer cohort (Pan-IO, Shen et al JITC 2024), 1,654 patients were analyzed for IO treatment response. Results: We validated our AI prediction of cell types by demonstrating consistency with known cardinal gene expressions from spatial transcriptomic results, where 77.5% of AI-predicted endothelial cells (ECs) expressed VEGFR2 (compared to 6.9% in tumor cells (TCs)), while VEGFA expression was 2.8 times higher in TCs. Consistent with previous studies, EC density was highest in RCC and HCC, while lowest in pancreatic adenocarcinoma, melanoma, and cholangiocarcinoma. While EC and TIL density were not correlated pan-cancer (r = 0.08), exceptions were head and neck cancer (r = 0.45, p < 0.001) and pancreatic cancer (r = 0.44, p < 0.001). In the TCGA cohort, the high EC density was associated with prolonged OS (HR 0.84, p < 0.001). In contrast, within the Pan-IO cohort (HR 1.26, p < 0.001) and its lung cancer subgroup (HR 1.26, p < 0.001) high EC density was associated with shorter progression-free survival (PFS) on treatment with IO monotherapy. Moreover, the predictive impact of EC density varied by TIL status. In the Pan-IO cohort, EC density predicted PFS in both TIL-high (HR 1.40, p < 0.001) and TIL-low cancers (HR 1.33, p < 0.001). Among four groups divided by median values, the EC-low and TIL-high group showed the longest PFS (median PFS of EC/TIL high/low: 2.5m, high/high: 3.2m, low/low: 3.6m, low/high: 5.6m, p < 0.001). Conclusions: EC distribution varied among cancer types, and importantly high EC content strongly correlated with poor IO monotherapy response, including NSCLC. These findings support exploring immunotherapy combination strategies that include anti-endothelial approaches, which could encompass the emerging class of PD-1/VEGFR bispecifics, for tumors with high EC content. Early evidence was observed for this for HCC (Chon et al ASCO GI 2024), with a hazard ratio of 0.62 for high EC HCC for atezolizumab/bevacizumab.