Scientists at UC San Diego have created AI-driven virtual cells and physics-based digital twins to model mitochondrial behavior, potentially streamlining the development of new medical treatments.
Key facts
- •The MitoSpace AI model achieved 75% accuracy in identifying drug mechanisms, compared to 56% for models using 2D images.
- •Researchers used 40,000 4D movies of cancer cells to train the deep-learning model.
- •The digital twin model successfully predicted mitochondrial responses to the drug nocodazole without parameter adjustments.
- •The studies were published in the journal Cell in 2026.
- •The research team is led by corresponding author Johannes Schöneberg, an associate professor at UC San Diego.
Researchers at the University of California San Diego have developed two new methods to create virtual cells that mimic the dynamic biological processes of real cells. By utilizing 4D lattice light-sheet microscopy, the team captured the movement of mitochondria in three dimensions over time. These models, recently published in the journal Cell, aim to reduce the need for time-consuming laboratory experiments and accelerate drug discovery for diseases such as cancer, diabetes, and Alzheimer's.
By the numbers
AI-Powered Mitochondrial Analysis
The research team trained a deep-learning model called MitoSpace on 40,000 4D movies of cancer cells treated with 25 different compounds. Unlike traditional models that rely on manual labeling, MitoSpace autonomously identified patterns in mitochondrial shape and movement. The model successfully grouped cells based on their response to specific drugs and predicted the energetic state of cells without prior knowledge of the treatment used. When tested, MitoSpace categorized drugs by their mechanism with 75% accuracy, significantly outperforming the 56% accuracy achieved by models trained on standard 2D images.
Physics-Based Digital Twins
In a separate study, researchers built a physics-based 'digital twin' of a living cancer cell. By mapping the positions of mitochondria and the microtubule tracks they use for transport, the team implemented rules governing organelle behavior. The model was validated using the drug nocodazole, which breaks down microtubules; the digital twin accurately reproduced the reduced motion and fusion rates observed in real cells treated with the same substance. This approach allows researchers to simulate the effects of drugs or disease mutations before testing them in a laboratory setting.
Future Research Directions
The UC San Diego team intends to integrate MitoSpace and digital twin technology into a unified workflow. This combined approach would allow the AI model to identify patterns in large datasets, while the digital twins provide physical explanations for those observations. The researchers aim to eventually expand these models to include other cellular organelles and simulate whole tissues, which they believe will provide a more accurate representation of human biology for clinical research.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by Phys.org.
