UC San Diego Creates 4D Virtual Cell to Predict Drug Impact
UC San Diego researchers developed the MitoSpace AI model and physics-based digital twins using 4D imaging, achieving 75% accuracy in drug grouping.
Stock photo for illustration only, not from the actual event
- UC San Diego researchers built 4D virtual cells using AI and physics laws
- The MitoSpace model learns from 40,000 cancer cell videos without manual labeling
- Achieved 75% accuracy in grouping drugs by mechanism compared to 56% in 2D images
- Ultimate goal is simulating entire human tissues to inform clinical treatments
Researchers at the University of California San Diego have pioneered a novel approach to cellular biology by feeding 40,000 video clips of living, moving cancer cells in three dimensions into an artificial intelligence model named MitoSpace. The resulting model successfully classified drugs based on their mechanisms of action with 75% accuracy, compared to just 56% achieved when trained on two-dimensional static images commonly utilized in large-scale drug screening today, and accomplished this without any human labeling indicating which drug was applied to each cell.
This study coincides with another paper published simultaneously in the journal Cell, detailing the creation of a digital twin of actual cells driven by physical laws. Both methodologies share the overarching objective of constructing a virtual cell—a digital model that replicates the biological processes of living cells in constant motion. Both approaches rely on the same foundational dataset captured via Lattice Light-Sheet microscopy, which records the movement of intracellular structures in three dimensions continuously over time, forming what is known as four-dimensional data.
Stock photo for illustration only, not from the actual event
Textbook depictions typically portray mitochondria as bean-shaped objects scattered throughout a cell, but in reality, they form an interconnected network spanning the entire cell, rapidly dividing and fusing while being transported to areas requiring high energy. Their primary role is converting nutrients into energy. Because mitochondrial network morphology shifts in response to cellular health, researchers utilize them as disease indicators and testing grounds for new therapeutics.
The historical bottleneck has been that most mitochondrial imaging yields two-dimensional static snapshots, stripping away both depth and movement, thereby hindering efforts to understand how these morphological changes impact cellular functions. Johannes Schöneberg, associate professor of pharmacology at the UC San Diego School of Medicine and lead author of both studies, summarized this limitation.
"A cell is a four-dimensional object: it has depth and it never stops moving. Virtual cells need to be built on data that captures that fact."
Johannes Schöneberg
The development of 4D virtual cells marks a major leap forward in computational biology and drug discovery. Traditional experimentation methods often face severe time, financial, and ethical constraints regarding animal testing. By leveraging AI and physical modeling to simulate sub-cellular behaviors, researchers can screen and predict drug impacts with high precision without requiring physical trials for every single variable, aligning well with regulatory pressures from bodies like the FDA to adopt alternative testing methods.
The research team began by exposing cancer cells to 25 compounds known to disrupt mitochondria via distinct mechanisms, recording 40,000 single-cell 4D clips to serve as a deep learning training repository. A key distinction of MitoSpace is its ability to autonomously uncover patterns without manual image tagging, mapping clusters of cells reacting similarly without knowing their specific drug exposure, and accurately predicting cellular energy states solely based on mitochondrial morphology and movement across 26 drug conditions.
Source: Techsauce
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment