An AI-Powered tissue-agnostic cellular morphometrics biomarker for risk assessment in patients with pan-gastrointestinal precancerous lesions and cancers

PURPOSETissue-agnostic biomarkers that capture the commonality in cancer biology, may provide a new avenue for treatment development and optimization across cancer types. Here, we aimed to evaluate and validate the clinical value of a tissue-agnostic cellular morphometrics biomarker (CMB) signature, which was discovered by artificial intelligence (AI) from H&E-stained whole-slide images (WSI) of diagnostic slides of colon cancers, in pan-gastrointestinal (pan-GI) pre-cancer lesions and cancers. METHODSWe discovered CMBs from WSI using our well-established CMB-ML pipeline and established a CMB risk score (CMBRS) using multivariate regression models. Based on CMBRS, we assigned individual patients from The Cancer Genome Atlas Colon Adenocarcinoma Cohort (TCGA-COAD) (n=430) to CMB risk groups (CMBRG). We then extensively evaluated tissue-agnostic clinical value of CMB signature, CMBRS and CMBRG in multi-cohorts with different types of GI cancer (n=2,219) and risk assessment of precancerous lesions (n=1,016). We unraveled each CMB-related biological function using bulk RNA-sequencing, single-cell RNA-sequencing (scRNA-seq) and opal multiplex immunohistochemistry (IHC) techniques. RESULTSFrom the TCGA-COAD cohort, we developed a 13-CMB signature and constructed CMBRS/CMBRG that predict prognosis of colon cancer patients. Importantly, this 13-CMB signature proved prognostic and predictive values for TCGA patients with rectal, gastric and esophageal cancer independent of traditional clinical factors. These findings were independently validated using multiple cohorts from Drum Tower Hospital. Moreover, 13-CMB signature exhibited the power for risk stratification of colon adenoma and early esophageal neoplastic lesion patients for predicting cancer progression. In addition, we demonstrated and validated independent prognostic impacts of gene signatures and CMB signatures and a significant increase in predictive power by integration of CMB signature, gene signature and clinical factors. Correlations between CMBs and gene expression levels revealed the association of each CMB with biological functions including cell proliferation, epithelial-to-mesenchymal transition and immune microenvironment. The association of CMBs with the immune microenvironment was prospectively validated by scRNA-seq and was further confirmed by Opal multiplex IHC staining in colon cancer. CONCLUSIONThis study demonstrates the clinical value of tissue-agnostic AI-empowered CMB signature from WSI with defined biological functions, which can be used in clinical settings to assess risk, diagnose disease, and guide clinical interventions. Tissue-agnostic CMBs potentially provide a new avenue for a rapid, robust and cost-effective cross-cancer prediction that is essential for developing common treatment strategy for multiple cancers.