UC San Diego Publishes Virtual Cells That Pair AI With Physics

University of California San Diego researchers have published two Cell studies that build virtual versions of cells around a fast-moving target: mitochondria, the structures that help convert nutrients into energy. One system, MitoSpace, learns from 4D microscopy movies; the other is a physics-based digital twin whose simulated drug response matched observations in real cells.

Mitochondria form an interconnected network that moves, splits and fuses inside cells. A flat image can miss that changing behavior. The team used lattice light-sheet microscopy, which captures structures in three dimensions over time, to test whether those records could better characterize drug-treated cells.

One method finds patterns; the other recreates behavior

MitoSpace is the pattern-finding half of the project. It was trained on 40,000 4D movies of cancer cells treated with 25 compounds that disrupt mitochondria in different ways. Without drug labels, the deep-learning model learned differences among mitochondrial networks and predicted a cell’s energetic state from their shape and movement across 26 drug conditions.

A simulation tested against a specific disruption

The companion study starts from a different premise: model the machinery behind mitochondrial movement. Researchers mapped mitochondria and the microtubule tracks they travel on, then modeled transport by motor proteins in a cancer-cell digital twin. They adjusted the simulation until its mitochondrial behavior matched the observed cell.

For a test, the team simulated nocodazole treatment, which partially breaks down microtubules. Without changing model parameters, the twin reproduced the reduced mitochondrial movement and changed fusion and fission rates seen in treated cells. MitoSpace sorts patterns across many images; the twin asks whether a defined model can recreate the result of one intervention.

Early signs MitoSpace may extend beyond its training set

• The researchers said it organized drugs the model had not previously seen.
• It also sorted human lung organoid cells by developmental stage without retraining.

The team plans to combine the two approaches: MitoSpace would identify patterns in large image collections, while digital twins would examine their physical basis. They also plan to add other organelles and eventually model multiple cells acting as tissues. For now, the studies make a narrower case: dynamic mitochondrial data can support virtual-cell systems that complement laboratory experiments.