Researchers at the University of California San Diego have developed virtual cells—digital models that replicate dynamic biological processes—using two complementary AI-driven approaches published in Cell. These models leverage 4D lattice light-sheet microscopy to accelerate drug discovery for cancer, diabetes, Alzheimer’s, and pediatric mitochondrial disorders.
Mapping the Cellular Frontier in Four Dimensions
Mitochondria, however, form dynamic, interconnected networks that constantly split, fuse, and travel where energy is required within the cell. Traditional flat snapshots failed to capture this morphological complexity.
Unsupervised Deep Learning With MitoSpace
To overcome this limitation, a research team led by corresponding author Johannes Schöneberg, PhD, developed a deep-learning model named MitoSpace. Schöneberg serves as the Roger Tsien Chancellor’s Faculty Fellow and associate professor in the Department of Pharmacology at UC San Diego School of Medicine. MitoSpace was trained on 40,000 single-cell 4D movies of cancer cells treated with 25 different compounds.
“For a century we have believed that mitochondrial form reflects function; this shows the relationship is strong enough that a model can learn it without ever being shown the answer,” Schöneberg explained regarding the unsupervised learning architecture.
Unlike conventional models requiring manual data labeling, MitoSpace discovered patterns independently. When tested on 4D movies, the deep-learning model distinguished between drugs and grouped them by mechanism with 75% accuracy. In contrast, flat 2D images common in current large-scale drug screens achieved only 56% accuracy.
Constructing a Physics-Based Digital Twin
The second study took a mechanistic approach by constructing a physics-based digital twin of a real cancer cell. Researchers mapped mitochondrial positions and microtubule tracks using specialized image-analysis software. They then integrated motor proteins that transport mitochondria at established rates, incorporating fundamental laws of motion until the virtual cell’s behavior matched live biological recordings.
“We have built a physics‑based virtual cell and can compare it side‑by‑side to the actual 3D microscopy movie, something that has never been possible before,” Schöneberg noted.
To validate the integrity of the digital twin, the team tested how mitochondria responded when microtubules were partially broken down using the drug nocodazole. Without altering underlying parameters, the digital twin accurately reproduced the reduced motion and altered fusion-fission rates observed in drug-treated real cells. This capability allows researchers to test drug efficacy, disease-causing mutations, or cellular engineering designs before running physical experiments.
Broad Horizons in Preclinical Pharmacology
The implications extend far beyond initial cancer cell assays. MitoSpace demonstrated high adaptability by organizing drugs it had never encountered during its training phase. Furthermore, the deep-learning model successfully sorted human lung organoid cells by developmental stage without requiring retraining.
This versatility positions the tool for broader applications in computational biology, potentially uncovering new therapeutic uses for existing medications while speeding up the identification of novel treatments. Both virtual cell architectures target diseases where mitochondrial dysfunction plays a core pathological role, pointing toward an automated, high-throughput future for preclinical pharmacology.
