Leveraging artificial intelligence to predict immune checkpoint inhibitor (ICI) efficacy in proficient MMR mCRC: Translational analyses of AtezoTRIBE and AVETRIC trials
Martina Carullo, Carlotta Antoniotti, Changho Ahn, Chiyoon Oum, Valentina Angerilli, Francesca Bergamo, Carolina Sciortino, Alessandra Boccaccino, Maria Alessandra Calegari, Jessica Gasparello, Stefano Tamberi, Rossana Intini, Mario Scartozzi, Camilla Damonte, Chiara Citterio, Lisa Salvatore, Chiara Boccaccio, Matteo Fassan, Chan-Young Ock, Chiara Cremolin
ESMO, 2025
Abstract
Background Artificial Intelligence (AI) methods may enable to extract predictive biomarkers from tumor hematoxylin & eosin (H&E) whole-slide images (WSIs). We aimed to develop an AI-driven biomarker predictive of ICIs benefit in proficient mismatch repair (pMMR) metastatic colorectal cancer (mCRC) using Lunit SCOPE IO platform.
Methods Lunit SCOPE IO quantified the density of lymphocytes (LC), fibroblasts (FB), macrophages (MP), tumor (TC), endothelial (EC) and mitotic (MTC) cells in cancer area (CA) and stroma (CS) on pre-treatment H&E WSIs from pts with pMMR mCRC enrolled in AtezoTRIBE (FOLFOXIRI/bevacizumab +/- atezolizumab [atezo]) and AVETRIC (FOLFOXIRI/cetuximab/avelumab) trials. A multivariate Cox regression model was trained using the most predictive variables for progression-free survival (PFS; average C-index > .5) in the AtezoTRIBE atezo-treated arm. A PFS-based cut-off set by maximal rank statistics dichotomized tumors as biomarker-high or low. AVETRIC served as a validation set.
Results The AI-driven analysis of WSIs from 161 patients (pts) enrolled in the AtezoTRIBE study identified a biomarker incorporating densities of TC, MTS, LC on CA and FB, MP, EC on CS. Among them,113 (70%) pts were classified as biomarker-high, characterized by older age (P=.030) and higher incidence of liver metastases (P= .023). In atezo arm, biomarker-high pts had better prognosis as compared to biomarker-low pts (PFS P= .036, Overall Survival [OS] P= .024), but not in control arm (PFS P= .564, OS P= .186). Interactions between treatment and biomarker were found in PFS (P=.114) and OS (P= .025), with biomarker-high pts but not biomarker-low ones deriving benefit from adding atezo (high, HR PFS: 0.69, 95%CI 0.45-1.04; OS: 0.54, 95% CI 0.33-0.88; low, HR PFS: 1.34, 95%CI 0.66-2.72; OS: 1.70, 95% CI 0.69-4.20). In the AVETRIC cohort, WSIs from 48 pts were analyzed; 36 (75%) cases were classified as biomarker-high, with better PFS (P= .043) and OS (P= .053) compared to biomarker-low ones.
Conclusions Our AI-derived tumour microenvironment biomarker may help to predict benefit from ICI-based treatments in pMMR mCRC, supporting further investigations of AI-powered approaches.