Researchers at the University of California, San Diego have proposed using “virtual cells” to speed the early search for new drugs, combining artificial intelligence with a physics-based model of cellular behaviour. The two approaches focus on mitochondria, the structures responsible for cellular energy production, whose shape, movement and interactions can change when cells are under stress or exposed to a drug.
The work, published in two papers in Cell, uses 4D lattice light-sheet microscopy to record mitochondrial activity over time. One system, called MitoSpace, finds patterns in large sets of cell films. The other is a digital twin that aims to reproduce the physical processes behind those patterns. The researchers said the methods currently complement laboratory experiments rather than replace them.
MitoSpace was trained on 40,000 4D films of cells treated with drugs. The experiments involved cancer cells exposed to 25 different compounds that disrupt mitochondrial function through different mechanisms. Rather than relying on people to label the images by hand, the artificial intelligence system identified patterns independently.
The model predicted a cell’s energy state from the shape and movement of mitochondria alone across 26 drug exposure conditions. When trained on 4D films, it distinguished drugs and grouped them by mechanism of action with 75% accuracy. Its accuracy was 56% when it was trained on flat 2D images.
The result suggests that recording how mitochondria behave over time can provide information that conventional still images miss. MitoSpace also found patterns among drugs that had not been included in its training data. In a separate test, it sorted human lung organoid cells by developmental stage without retraining.
The second approach creates a digital twin from 4D film of a living cell. Researchers first established rules for organelle behaviour and then built those rules into a physical model. They mapped mitochondria, as well as the microtubule tracks along which the structures travel, and modelled the transport produced by motor proteins.
The digital twin includes mitochondrial networks and microtubule networks, together with dynein and kinesin motors, the plasma membrane and the nucleus. Its model covers fusion and fission remodelling, diffusion, and active transport along microtubules driven by motors.
Without changing any parameters, the model reproduced a reduction in mitochondrial movement and in the frequency of fusion and division that had been observed in real cells. That ability to simulate a cellular response offers a different form of evidence from MitoSpace’s pattern recognition, and could help researchers examine the physical causes of changes found in imaging data.
The teams plan to combine the methods in one workflow. In that arrangement, artificial intelligence would search large datasets for meaningful patterns, while digital twins would be used to investigate the physical processes that may explain them. The researchers also hope to add other organelles over time, with the longer-term aim of modelling a complete virtual human cell and then whole tissues.
Such models could make drug testing more realistic while reducing some time-intensive laboratory work and costs during early-stage drug development. The researchers see possible applications in work on cancer, Alzheimer’s, diabetes and paediatric mitochondrial disorders. Important questions remain, including how the systems would perform across other types of cells, whether they can assess drug combinations, and whether they can predict toxicity or off-target effects.
Comments (0)