An integrated platform for high-throughput phenospace learning of 3D multilineage organoid systems

Complex multilineage organoid systems lack quantitative phenotyping methods preserving spatial architecture at high throughput. Current approaches compromise biological complexity, spatial resolution, or robust homogeneous multilineage assembly. We establish an integrated experimental-computational platform for high-throughput spatial phenotyping of multilineage organoids through developing a modular tumoroid culture system incorporating pancreatic ductal adenocarcinoma (PDAC) cells and cancer-associated fibroblasts (CAFs) in 384-well format with multiplexed whole-mount imaging. We developed Phenocoder, a machine learning framework combining conditional variational autoencoders with spatial graph analysis to extract multiscale organoid features. Rigorous validation demonstrates robust performance in PDAC tumoroids. The platform identifies pathway modulators that disrupt the fibrotic microenvironment and discovers stroma-dependent vulnerabilities, undetectable in monocultures. Extending to immuno-competent tumoroids, we assess fibrosis modulators in combination with T cell bispecific antibodies, identifying treatments that enhance immune cell proliferation and infiltration inducing cancer cell death, validated in patient-derived explants. This platform establishes a generalizable framework for multilineage organoid phenotyping.