Supporting data for "Agentic and generative AI for biological imaging and strcuture prediction"

The raw data from the agent measurement for statistical analysis and plotting in all experiments are provided.<br><br>Micro-Copilot implements fluorescence microscopy as a human-in-the-loop acquisition-analysis-reasoning cycle. A reward-based navigation agent controls the microscope hardware, explores the specimen, and centres informative fields.

An analysis agent performs real-time cell segmentation, organelle detection, feature extraction, and colocalization analysis. A retrieval-grounded co-scientist integrates population statistics with experimental metadata and literature to propose mechanistic hypotheses. The platform quantified tens of thousands of organelles within hours across five lysosome studies and, through iterative discovery cycles, uncovered a previously unknown mechanism in which the GPX4 inhibitor RSL3 drives lysosomal hyperacidification through autophagy-dependent V-ATPase assembly.<br></p>