Mitochondria do not sit still for a portrait. They divide, join, travel and change shape as a cell's energy demands and stresses change. Two UC San Diego-led studies published in Cell use that motion to build different kinds of detailed virtual cells: one learns patterns from thousands of movies, while the other simulates the physical machinery moving the organelles.
The machine-learning system, called MitoSpace, was trained on about 40,000 four-dimensional movies of cancer cells exposed to 25 compounds. Three spatial dimensions plus time allowed the model to relate mitochondrial shape and movement to a cell's energetic condition. It grouped drugs by mechanism with 75% accuracy, compared with 56% for analysis based on flat two-dimensional images, the researchers reported.
That comparison supports a limited conclusion. For this dataset and task, motion and three-dimensional structure contained useful information that a single plane missed. It does not mean MitoSpace can diagnose a patient or predict every drug response. The training cells, compounds, imaging conditions and scoring method define the evidence; performance can change when biology or equipment changes.
The system also classified drugs it had not seen during training and estimated stages of lung-organoid development without being retrained, the university reported. Those tests suggest that the learned representation captured more than a lookup table of the original treatments. External laboratories will still need to test how well that transfer holds across cell types, microscopes and experimental protocols.
A companion study started from mechanisms rather than examples. Its digital twin represented mitochondria, microtubules and the motor proteins that carry organelles along the cell's internal scaffolding. Researchers tuned the model against observations, then tested whether it could predict the effect of nocodazole, a drug that disrupts microtubules. The model reproduced the response without a new round of parameter adjustment.
The two approaches answer different questions. MitoSpace can scan complex image sequences for recurring states without requiring researchers to label every pattern in advance. The physics model makes explicit claims about forces, tracks and molecular components. Agreement between data-driven classification and a mechanistic simulation would be more informative than either alone, because one can find a pattern while the other tests a possible cause.
Drug discovery is a plausible application because early screens must separate useful biological effects from toxicity and noise across many candidates. A richer cell model could prioritize compounds before slower animal or clinical studies. That would accelerate a stage of the pipeline, not replace the stages that establish safety, dosing and benefit in people. A virtual cell remains a selective model of measured processes rather than a complete human cell.
UC San Diego said a patent application has been filed for MitoSpace, and the university disclosed a company connection involving one researcher. Those interests do not invalidate the results, but they increase the importance of independent replication and accessible benchmarks. The advance here is methodological: cells can be analyzed as changing systems instead of still images. Whether that view shortens a real drug program will require prospective tests beyond the published demonstrations.