Pancreatic ductal adenocarcinoma (PDAC) has a five-year survival rate below 8%. A major reason for poor outcomes is the lack of good laboratory models for testing how individual patients' tumors will respond to treatment. Traditional 2D cell cultures do not accurately reflect the 3D complexity of real tumors.
Patient-derived organoids (PDOs) are three-dimensional 'mini-tumors' grown in a gel from a patient's own tumor tissue. They preserve the genetic diversity, architecture, and drug resistance patterns of the original tumor. Early studies have shown that organoid drug responses predict how that patient's tumor will actually respond in the clinic.
However, most organoid systems test chemotherapy only — not immunotherapy. Immunotherapy requires immune cells to be present, and immune cells are difficult to incorporate into organoid cultures without them dying or losing function. This study developed both a new co-culture system and an AI tool to analyze the results.
OrganoIDNet is a deep learning algorithm trained to detect and analyze pancreatic organoids in bright-field (no fluorescent dye required) microscopy images. It was trained on both murine (mouse) and human patient-derived PDAC organoids imaged over time.
The algorithm tracks organoids across time points to measure: total organoid number and area, organoid size distribution, eccentricity (shape irregularity, which changes when organoids are stressed by drugs), and healthy versus unhealthy status. This provides a much richer dataset than traditional endpoint-only assays like CellTiter-Glo.
For immunotherapy testing, the team developed a sandwich co-culture protocol where organoids were placed between two layers of matrix, with peripheral blood mononuclear cells (PBMCs) — the patient's own immune cells — added above and below. This stable configuration allowed continuous live imaging over several days without the organoids drifting out of focus.
For chemotherapy testing, OrganoIDNet detected that gemcitabine (a standard pancreatic cancer drug) preferentially killed larger organoids while smaller ones remained relatively protected. This size-specific drug response would have been missed by traditional endpoint assays that measure only total cell viability. The AI-derived measurements matched results from the CellTiter-Glo endpoint assay, validating OrganoIDNet's accuracy.
For immunotherapy testing, organoids were co-cultured with HLA-matched PBMCs and treated with Atezolizumab, a PD-L1 checkpoint inhibitor. OrganoIDNet tracked organoid shrinkage and death over multiple days. The immune cell tumor-killing effect was enhanced when Atezolizumab was added, and responses varied between individual patient-derived organoids — exactly the type of patient-to-patient variability that personalized medicine needs to capture.
The platform also measured eccentricity as a marker of organoid response to therapy: unhealthy organoids changed shape as they underwent cell death, and OrganoIDNet could detect these morphological changes automatically across thousands of images per experiment.
The ability to test immunotherapy in patient-derived organoids co-cultured with the patient's own immune cells is a significant step toward individualized treatment planning. If a patient's organoids respond to Atezolizumab in the lab, there is reason to believe the patient may benefit from it clinically.
The organoid variability seen between different patients — some showed strong immune-mediated tumor killing, others did not — mirrors the clinical reality where checkpoint inhibitors only benefit a subset of patients. The platform could help identify which patients are likely responders before committing to an immunotherapy course.
Because the assay uses live imaging that doesn't require killing the cells, the same organoids can be monitored continuously and even used for sequential drug testing, maximizing information from limited patient tissue.
OrganoIDNet represents a significant advance in pre-clinical cancer modeling by enabling real-time, AI-analyzed monitoring of organoid responses to both chemotherapy and immunotherapy. This moves organoid-based testing from a static endpoint measurement to a dynamic, longitudinal platform.
The combination of the sandwich co-culture protocol (for stable immune-organoid interactions) and the deep learning analysis tool (for automated quantification) addresses two major technical barriers that have limited organoid-based immunotherapy testing.
Future clinical validation studies should determine whether organoid drug responses in this system predict actual patient outcomes, establishing this platform as a tool for personalized treatment selection in pancreatic cancer.