CODNet models enable real-time assessment of colorectal cancer organoids and uncover the integral role of LTBP2 in the diffuse organoid subtype.

To achieve precise classification of colorectal cancer (CRC) at the patient-derived organoids (PDOs) level, we constructed a real-time CRC PDO assessment system through artificial intelligence (AI) assisted model using an optimized object detection algorithm. We constructed the CODNetV1 model for the identification, feature extraction and drug sensitivity detection of CRC PDOs (mAP of 88.05%). And we optimized CODNetV2 to classify four morphological subtypes of CRC PDOs (mAP of 85.65%). CODNetV1 and CODNetV2 achieved real-time, automated, multi-parameter detection of CRC PDOs, directly facilitating the detection of organoid drug sensitivity and subtype classification. By Integrating analyses of the RNAseq and scRNAseq of different subtype CRC PDOs, we found that LTBP2 was highly expressed in cancer-associated fibroblasts (CAFs) of Diffuse CRC PDOs. As a key regulatory factor, LTBP2 modulates intercellular communication between tumor cells and CAFs. On one hand, CAFs-derived LTBP2 enhances the invasive capacity of CRC PDOs by activating the integrin a6ß4-AKT signaling pathway, and simultaneously increases CXCL11 expression. On the other hand, diffuse CRC PDOs-derived CXCL11 activates the YAP/TEAD2-LTBP2 pathway in CAFs, thus forms a feedback activation loop centered on LTBP2. Collectively, molecular characterization of CRC PDO subtypes at the organoid level identified the important clinical value of LTBP2 in the diagnosis and targeted therapy of highly invasive CRC.