Artificial intelligence (AI)-powered tumor microenvironment (TME) analysis to identify potential biomarkers for ICIs with or without bevacizumab in hepatocellular carcinoma (HCC).

Abstract

549 Background: While immunotherapies have been approved for use in HCC patients, there is a need for validated predictive biomarkers that correlate with treatment outcomes. We investigated whether AI-powered spatial analysis of non-cancerous cells within the TME can be potential biomarkers for ICIs with or without bevacizumab in advanced HCC. Methods: Analysis of images of H&E-stained slides was conducted by an AI model, Lunit SCOPE IO in pre-treatment tumor samples of 163 HCC patients treated with atezolizumab plus bevacizumab as first-line ( n=82), or monotherapies of nivolumab or pembrolizumab as ≥ second-line ( n=81) at CHA Bundang Medical Center or Samsung Medical Center. We analyzed the correlation between clinical outcomes after the treatment and AI-powered TME-related variables, including TILs and endothelial cells, within intratumoral or stromal areas. Inflamed immune phenotype (IIP) was defined as cases exhibiting enrichment of intratumoral TILs. Results: Baseline characteristics, including Child-Pugh liver classification, Barcelona Clinic Liver Cancer stage, and hepatitis B virus, were well-balanced between treatment regimens. IIP was predictive of longer progression-free survival (PFS) of nivolumab or pembrolizumab monotherapy (median PFS 4.7 months for IIP vs. 2.2 months for non-IIP; hazard ratio [HR] 0.50; 95% confidence interval [CI] 0.25-0.99; p=0.042), but not PFS of atezolizumab plus bevacizumab (median PFS 6.8 months vs. 6.2 months; HR 0.92; 95% CI 0.50-1.69; p=0.762). PFS of atezolizumab plus bevacizumab was significantly longer in cases harboring intratumoral endothelial cell density in the highest quartile (median PFS 6.7 months for upper 25% vs. 3.9 months for lower 75%; HR 0.51; 95% CI 0.27-0.97; p=0.037), while there was no significant difference in PFS of pembrolizumab or nivolumab monotherapy (median PFS 2.3 months vs. 2.8 months; HR 1.02; 95% CI 0.59-1.77; p=0.935). Conclusions: AI-powered TME analysis shows IIP is predictive of longer PFS with ICI monotherapies, whereas intratumoral endothelial cell density is specifically associated with PFS of atezolizumab plus bevacizumab in advanced HCC. The latter finding may suggest efficacy of combined VEGFRi and ICI activity is dependent on the amount of baseline tumor vasculature.