Abstract B095: A Novel Machine Learning Technique Accurately Predicts Drug Response in Colorectal Microorganosphere Droplets

Abstract BACKGROUND: Predicting system therapy response is of paramount importance in the treatment of patients with colorectal cancer. We have previously reported development and application of MicroOrganoSphere® (MOS®) droplet technology, an automated platform that produces consistent biomimetic 3D functional precision medicine models that can assay drug response and predict clinical outcomes. Here we report the novel application of a machine learning-based imaging analysis that accurately replicates fluorescence-based dose response curves from a colorectal cancer clinical trial, utilizing brightfield microscopy images alone. We hypothesize this machine learning approach will enable development of novel precision therapies unencumbered by the limitations of current fluorescent based imaging by combining a scalable ex-vivo model with a streamlined AI-enabled drug response readout. METHODS: We trained our model using feature embeddings from the DINOv2 foundation model with an input dataset of brightfield images and EpCAM based live signal from 480,000 MOS droplets derived from 37 colorectal cancer tumors of patients enrolled in a retrospective clinical trial where patients received standard of care chemotherapy in a neoadjuvant setting. Tumor-derived MOS were treated with escalating doses of therapeutic agents mirroring clinical treatments, and Epithelial Cell Adhesion Molecule (EpCAM) based dose response curves correlated strongly with patient clinical response. Our model predicted EpCAM-based live signal utilizing only brightfield images. Dose response curves were regenerated based on model predicted EpCAM live signal intensity. Modeling was further applied to endpoint live (Cell Adhesion Molecule, CAM) and dead (Ethidium Homodimer, EthD) dyes to confirm generalizability of the approach. Modeling and evaluation were conducted in Python with packages including PyTorch, transformers, scikit-learn, pylance, and pandas. RESULTS: Predictive performance was achieved for the primary live signal (EpCAM, R2 > 0.72). Strong performance was also obtained for CAM (R2 >0.64) and EthD (R2 >0.45) dyes. The negative logarithmic expression of each drug's half maximal molar inhibitory concentration (pIC50) was utilized as the EpCAM-based dose response curve metric and closely replicated the clinical study's EpCAM based results (R2 > 0.85 across all treatments; mean R2 > 0.6 within drug groups). CONCLUSIONS: This study demonstrates a precision medicine-based response prediction, wherein prospective treatment selection utilizing patient-derived models and brightfield imaging can be accomplished. This machine learning approach abrogates dependence on time consuming, expensive and inconsistent dye-based methods. MOS droplet technology enables model training with a modest patient cohort, illustrating its predictive power. The ability to predict multiple immunofluorescent markers suggests general applicability and a capacity to discern and expand beyond multi-marker panels for higher complexity cancer phenotyping with brightfield-based modeling. Citation Format: Michael Rutenberg Schoenberg, Pavan Ramkumar, Roán Gobits, Else Driehuis, Carlton Barnett, Gavin Oliver. A Novel Machine Learning Technique Accurately Predicts Drug Response in Colorectal Microorganosphere Droplets [abstract]. In: Proceedings of the AACR-NCI-EORTC International Conference on Molecular Targets and Cancer Therapeutics; 2025 Oct 22-26; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2025;24(10 Suppl):Abstract nr B095.