Mitochondria — tiny structures that convert nutrients into energy — are often depicted as discrete kidney bean-shaped objects. But in reality, they form a dynamic, interconnected network throughout the entire cell, rapidly splitting and fusing as they're transported to where energy is needed most.
Because mitochondrial networks change shape based on cellular health, they can be used as markers of disease or to test new treatments. However, understanding how these changes affect cell function has been difficult because mitochondria have mostly been imaged as flat, still snapshots.
Now, researchers at University of California San Diego have employed two different approaches to quantify these morphological changes by creating "virtual cells" — digital models that mimic the dynamic biological processes of real cells. Both approaches use 4D lattice light-sheet microscopy, an advanced technique that captures how mitochondria and other structures move in three dimensions over time.
One approach trained a deep-learning artificial intelligence (AI) model on 40,000 4D movies of drug-treated cells to predict cellular health from mitochondrial shape alone.
The other built a "digital twin" of a living cell from a 4D movie by defining a set of rules about how its organelles (tiny internal structures) behave and implementing those rules in a physics-based model. They found that the way virtual mitochondrial networks responded to drugs closely matched real cells.
Together, these studies, both published in Cell, could reduce dependence on time-consuming lab experiments and accelerate drug discovery for a variety of diseases including cancer, diabetes, Alzheimer's, and pediatric mitochondrial disorders.
MitoSpace: a deep-learning AI model
The researchers treated cancer cells with 25 different compounds known to perturb mitochondria through different mechanisms, producing 40,000 single-cell 4D movies. They used this library to train a deep-learning model called MitoSpace . Unlike most AI models that require humans to manually label images, MitoSpace found patterns on its own, learning what makes the mitochondria of one cell distinct from that of another.
The results were striking: without knowing which drug was used on each cell, the model produced an organized map that grouped cells that respond in similar ways together. Furthermore, the model was able to predict the energetic state of the cell solely based on the shape and movement of its mitochondria across 26 drug conditions.
"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," said corresponding author Johannes Schöneberg, PhD, Roger Tsien Chancellor's Faculty Fellow and associate professor in the Department of Pharmacology at UC San Diego School of Medicine and in the Department of Biochemistry and Molecular Biophysics.
When trained on the 4D movies, the model distinguished between drugs and grouped them by mechanism with 75% accuracy, compared to only 56% accuracy when trained on the flat 2D images common in large-scale drug screens today.
"A cell is a four-dimensional object: it has depth and it never stops moving," said Schöneberg. "Virtual cells need to be built on data that captures that fact."
Using MitoSpace with 4D movies could potentially speed up the discovery of new treatments for disease and reveal new uses for existing drugs. It also demonstrated adaptability by organizing drugs it had not seen before and by sorting human lung organoid cells by their developmental stage without retraining. This suggests that the model could become a general-purpose tool in cell biology.
Physics-based "digital twins"
In the other Cell study, the researchers created a physics-based "digital twin" of a real cancer cell. Using specialized image-analysis software, they mapped the positions of mitochondria and of the microtubule tracks they travel on, then added the motor proteins that transport them according to previously established rates.
This virtual cell incorporated the laws of motion, and Schöneberg's team adjusted its parameters until mitochondrial behavior matched that of the real cell.
"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," said Schöneberg.
To test the model, the team asked it to predict how mitochondria would respond when microtubules were partially broken down by a drug called nocodazole. Without changing a single parameter, the digital twin reproduced the reduced motion and the fusion and fission rates in real cells treated with the drug. Schöneberg thinks such digital twins could be used to test drug effects, disease mutations, or cellular engineering designs, saving experimental effort and accelerating research.
"Cal27 is a head and neck cancer line where mitochondrial regulation is an active question," said Schöneberg. "By understanding mitochondrial regulation, we can ultimately find new ways to treat that cancer by using digital twins."
Impact on cell science
Looking ahead, Schöneberg's team plans to combine MitoSpace and digital twin virtual cells into a single workflow. The AI model would find patterns in vast amounts of data, and digital twins would examine the physical reasons behind those patterns. By integrating other cell organelles into these models, the researchers hope to build a complete virtual human cell.
"The ultimate future is not one cell but multiple cells acting as tissues," said Schöneberg. "Modeling whole tissues will let us simulate more realistic human biology and eventually inform clinical treatment."
Read the full studies: " 4D spatiotemporal landscape of mitochondrial phenotypes across cellular states unlocked through representation learning ," and " Whole-cell particle-based digital twin simulations from 4D lattice light-sheet microscopy data. "
Additional co-authors on the first study include: Dhruv Agarwal, Zichen Wang, Eric Arkfeld, Andre Modolo, Parth Natekar, Hiroyuki Hakozaki, Mehul Arora, Gillian McMahon, Siddharth Nahar and Manav Doshi at UC San Diego. The study was funded, in part, by the National Institutes of Health (NIH) (grant DP2 GM150022), the National Science Foundation (grant NAIRR240423), the Hartwell Foundation and the W.M. Keck Foundation.
Additional co-authors on the second study include: Eric Arkfeld, Zichen Wang and Hiroyuki Hakozaki at UC San Diego. The study was funded, in part, by the NIH (grant 1R01GM148765-01). Disclosures: UC San Diego has filed for patent protection on the technology described. Schöneberg has started a company based on the technology.
YouTube videos:
- 4D landscape of mitochondrial phenotypes across cell states unlocked through representation learning
- Whole Cell Digital Twin Simulations from 4D Lattice Light-Sheet Microscopy Data
