ETD

• artificial intelligence
• biomarker
• colorectal cancer
• immunotherapy
• pMMR

Immune checkpoint inhibitors (ICIs) provide limited benefit in proficient mismatch repair (pMMR) metastatic colorectal cancer (mCRC), underscoring the need of predictive biomarkers. Artificial Intelligence (AI) methods may enable the extraction of such biomarkers from tumor hematoxylin & eosin (H&E) whole-slide images (WSIs).
We aimed to develop an AI-powered prediction score (AI-PS) from H&E WSIs, by leveraging the Lunit SCOPE IO platform, able to quantify the density of lymphocytes, fibroblasts, macrophages, tumor, endothelial and mitotic cells in cancer area and stroma. The AI-PS was developed in pMMR mCRC patients enrolled in the experimental arm of the AtezoTRIBE trial (FOLFOXIRI/bevacizumab + atezolizumab; training cohort), then tested as a predictive marker in the overall AtezoTRIBE population (FOLFOXIRI/bevacizumab ± atezolizumab; application cohort) and validated in the phase II AVETRIC trial (mFOLFOXIRI/cetuximab/avelumab; validation cohort). To further substantiate the relevance of this work, the model was externally tested in two exploration cohorts - TGCA and CPTAC-COAD open datasets - including ICI-untreated patients.
Additionally, the correlation of the AI-PS with gene expression profiles from exploration cohorts allowed for a deeper investigation of the underlying biological rationale, thereby strengthening the relevance of the developed predictive model. As one of the first scores trained specifically on pMMR mCRC samples, the AI-PS is able to provide a comprehensive characterization of the tumor microenvironment by integrating both tumoral and stromal components. This approach lays the foundation for refining patient selection for ICI-based investigational treatments in pMMR mCRC.