Researchers at UC San Diego have built artificial intelligence models and “digital twins” of human cells that can predict how mitochondria will respond to drug treatment. The advance could substantially accelerate the discovery of new treatments for cancer, diabetes, Alzheimer’s disease, and other conditions by reducing dependence on time-consuming laboratory experiments.
The team developed two complementary digital approaches using 4D lattice light-sheet microscopy, an advanced imaging technique that captures how cell structures move in three dimensions over time. One model, a deep-learning AI called MitoSpace, learns patterns from thousands of cell movies. The other is a physics-based virtual cell that simulates the actual movement and behavior of real mitochondria according to established biological rules.
Mitochondria are tiny cellular structures that convert nutrients into energy. Scientists have long known that the shape of mitochondrial networks can indicate whether a cell is healthy or diseased. But traditional research has relied on flat, still images rather than dynamic three-dimensional views, making it difficult to understand how mitochondrial changes affect overall cell function.
Training An AI Model On Thousands Of 4D Cell Movies
Researchers treated cancer cells with 25 different compounds known to affect mitochondria through different mechanisms, producing 40,000 single-cell 4D movies. They used this library to train MitoSpace, which automatically found patterns without requiring humans to manually label each image.
The results were striking. When trained on the 4D movies, MitoSpace distinguished between drugs and grouped them by their mechanism with 75 percent accuracy. That vastly outperformed the 56 percent accuracy achieved when the same model was trained on flat 2D images, which are standard in most large-scale drug screens today.
The model proved it could predict the energetic state of a cell solely from the shape and movement of its mitochondria across 26 different 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 Johannes Schöneberg, the lead researcher and an associate professor in the departments of Pharmacology and Biochemistry and Molecular Biophysics at UC San Diego School of Medicine.
MitoSpace also demonstrated adaptability by organizing drugs it had never encountered before and by sorting human lung organoid cells by their developmental stage without being retrained. This suggests the model could eventually become a general-purpose tool in cell biology.
Building A Virtual Cell From Physics And Microscopy Data
In a parallel approach, the research team created a physics-based “digital twin” of a real cancer cell. Using specialized image-analysis software, they mapped the positions of mitochondria and the microtubule tracks they travel on, then added the motor proteins that transport them according to established biological rates.
The team adjusted the virtual cell’s parameters until its mitochondrial behavior matched that of the actual cell. They then tested the digital twin by asking it to predict how mitochondria would respond when microtubules were partially broken down by a drug called nocodazole.
Without changing any of the model’s parameters, the digital twin reproduced the reduced motion and the altered fusion and fission rates observed in real cells treated with the drug. This success suggests that digital twins could be used to test drug effects, disease mutations, or cellular engineering designs, saving experimental time and accelerating research across multiple disease areas.
Combining Models Into A Complete Virtual Cell
Looking ahead, Schöneberg’s team plans to integrate MitoSpace and the physics-based digital twins into a single workflow. The AI model would identify patterns in vast amounts of data, while digital twins would examine the physical reasons behind those patterns.
The researchers also intend to incorporate other cellular organelles into these models, working toward a complete virtual human cell. “The ultimate future is not one cell but multiple cells acting as tissues,” Schöneberg said. “Modeling whole tissues will let us simulate more realistic human biology and eventually inform clinical treatment.”
The two studies were published in the journal Cell and were funded in part by the National Institutes of Health, the National Science Foundation, the Hartwell Foundation, and the W.M. Keck Foundation. UC San Diego has filed for patent protection on the technology, and Schöneberg has started a company based on the approach.
