OrgaCast: A Trustworthy Spatiotemporal Diffusion Model for Fluorescence Organoid Forecasting

Area of Science:

• Biomedical Engineering
• Computational Biology
• Developmental Biology

Background:

• Accurate forecasting of biological system dynamics from microscopy time-series is crucial for drug discovery.
• Existing generative models struggle with the complex spatiotemporal patterns and irregular morphology of organoids.
• Human pluripotent stem cell (hPSC)-derived cardiac organoids present unique challenges for predictive modeling.

Purpose of the Study:

• Introduce OrgaCast, a novel multimodal conditional diffusion model for high-fidelity organoid forecasting.
• Address limitations of current models in capturing intricate organoid developmental dynamics.
• Enhance the interpretability and trustworthiness of biological forecasts.

Main Methods:

• Developed OrgaCast, a multimodal conditional diffusion model.
• Conditioned the model on historical image sequences, numerical metadata, and text captions.
• Implemented a post-hoc uncertainty quantification method for confidence maps.

Main Results:

• OrgaCast achieved high visual accuracy and biological plausibility in forecasting organoid development.
• The model demonstrated superior performance over baseline methods on a cardiac organoid dataset.
• Quantified prediction uncertainty using confidence maps, improving interpretability.

Conclusions:

• OrgaCast offers a robust solution for biological forecasting, accelerating research and drug discovery.
• The multimodal approach effectively captures complex spatiotemporal dynamics.
• Uncertainty quantification enhances the reliability of AI-driven biological predictions.