The researchers asked whether a self-supervised model could learn biologically useful representations directly from mitochondria moving through three dimensions over time—and whether the extra spatial and temporal information improved drug-mechanism and cell-state predictions beyond conventional 2D snapshots.
The research question and why it matters
The researchers asked whether a self-supervised model could learn biologically useful representations directly from mitochondria moving through three dimensions over time—and whether the extra spatial and temporal information improved drug-mechanism and cell-state predictions beyond conventional 2D snapshots.
High-content screens have long used static or two-dimensional cell images, while lattice light-sheet microscopy made gentler volumetric time-lapse imaging possible. Earlier work also linked mitochondrial form with metabolism and disease. This study connects those lines by training a label-free representation model on a much larger 4D collection and directly testing how much information is lost when time or depth is removed.
What researchers found
Using full 4D movies, the model grouped and distinguished drug mechanisms with 75% accuracy, compared with 56% using 2D images. A regression probe predicted mitochondrial membrane potential with R² = 0.91. Representation quality improved from 2D to 3D to 4D, and the learned space transferred without retraining to unseen perturbations and to cells from human lung organoids at different developmental stages.
Key results from the tested systems
single-cell movies
The training library captured mitochondrial shape and movement in three dimensions over time.
drug grouping accuracy
Full 4D inputs outperformed matched flat 2D images.
functional prediction
A probe predicted mitochondrial membrane potential from the learned representation.
training compounds
The drugs were selected to perturb mitochondria through different mechanisms.
How the research worked
The team used lattice light-sheet microscopy to record mitochondria in living drug-treated cells as three-dimensional volumes over time. MitoSpace learned similarities without hand-labeled shape categories, after which the researchers tested whether its representations separated drug mechanisms, predicted mitochondrial membrane potential and generalized to perturbations and lung-organoid cells that were not part of training. Dimensional ablation compared 2D, 3D and full 4D inputs.
How to interpret this design
The design determines what kind of conclusion the evidence can support. Direct measurement strengthens the reported observation, while generalization beyond the tested subjects, material, place or conditions requires additional evidence.
Cells and organoids allow close study of molecular and developmental processes, but they do not reproduce a complete human body, immune system, metabolism or lived environment. The result is mechanistic evidence, not a demonstrated treatment effect in people.
What strengthens or limits the finding?
The model was trained on a large, information-rich imaging library, beat a matched 2D comparison, predicted an independently measured cellular property and transferred to unseen perturbations and organoids. Performance still comes from controlled laboratory systems, and commercial and patent interests require disclosure when judging future claims.
The result is meaningfully informative, but identifiable limitations could alter the size, reach or causal interpretation of the finding.
Funding and disclosure context
The recorded funding source is: National Institutes of Health grant DP2 GM150022; National Science Foundation grant NAIRR240423; Hartwell Foundation; and W.M. Keck Foundation. The recorded conflict information is: UC San Diego filed for patent protection on the technology, and corresponding author Johannes Schöneberg started a company based on it. Funding or a disclosed relationship does not by itself invalidate a result, but it is relevant when judging design choices, analysis and the need for independent replication.
What it means
The result shows that cellular motion and depth contain useful information discarded by flat screening images. An open representation model could help researchers rank compounds, compare cell states and design smaller follow-up experiments. Its immediate value is as a research tool: every promising prediction still needs biological validation, and clinical usefulness would require patient-relevant studies.
Deeper analysis
Four dimensions add biology, not just prettier images
Mitochondria continually divide, fuse and move. A single plane can make the same network look different depending on where it is cut, while a time series can reveal whether a shape is stable or transient. The ablation experiment is important because it tests that added information rather than merely assuming 4D must be better.
Prediction and mechanism are different achievements
A model can accurately associate a visual pattern with membrane potential without explaining the molecular steps that connect them. That makes the representation useful for screening while leaving causal biology to targeted experiments.
Zero-shot transfer is promising but bounded
Performance on unseen perturbations and lung organoids suggests the model did not simply memorize drug labels. Yet two successful transfer settings are not evidence of universal portability across tissues, diseases and imaging systems.
Openness makes the result more testable
The authors made the model, dataset and an interactive explorer public. Independent groups can therefore probe failure modes, compare alternative models and test whether the learned space predicts new experiments rather than only reproducing the published benchmarks.
What it does NOT prove
- It does not show that MitoSpace can predict whether a drug is safe or effective in patients.
- It does not create a complete digital twin of a cell; the model represents mitochondrial images and associated states.
- It does not establish that mitochondrial appearance causes the energetic states the model predicts.
- It does not show that 75% classification accuracy is sufficient for autonomous drug selection or regulatory decisions.
- It does not eliminate wet-lab experiments, animal studies or clinical trials.
Important limitations
- The main training data came from controlled cell-culture experiments using compounds chosen because they perturb mitochondria.
- Generalization to lung organoids is encouraging but does not cover the diversity of tissues, diseases, patient backgrounds or imaging platforms.
- Self-supervised representations can encode technical artifacts as well as biology, and downstream probes do not make every learned feature interpretable.
- The 2D comparison quantifies the benefit under this experimental pipeline; other microscopes, sampling schedules and feature sets may yield different gaps.
- Patent protection and a company founded by the corresponding author create a disclosed commercial interest in the technology's future value.
How this fits with previous research
High-content screens have long used static or two-dimensional cell images, while lattice light-sheet microscopy made gentler volumetric time-lapse imaging possible. Earlier work also linked mitochondrial form with metabolism and disease. This study connects those lines by training a label-free representation model on a much larger 4D collection and directly testing how much information is lost when time or depth is removed.
Questions still unanswered
- Will the reported performance reproduce across laboratories, microscopes and cell lines?
- Can prospectively selected compounds based on MitoSpace produce genuinely new biological discoveries?
- Which learned mitochondrial features carry causal information rather than correlation or imaging artifacts?
- How well will the model transfer to primary patient cells and complex multicellular tissues?
- Can the workflow become fast and inexpensive enough for routine large-scale screening?
Relevant U.S. government resources
These resources serve different purposes. A registry can verify what researchers planned, a repository can locate government-funded work, and an agency page can supply authoritative background. None automatically proves that this paper's conclusion is correct.
OSTI.GOV research search ↗
DOE's research repository is used to locate related national-laboratory reports, accepted manuscripts and funding-linked technical work.
Reuse note: Facts and discoveries are summarized here in original language. We link to government material instead of copying it wholesale, and we do not reuse agency logos, photographs, charts or third-party material unless the specific reuse rights are verified.
A 4D model read mitochondrial drug responses better than flat images
This review was developed from the source record below and, when separately available, the primary paper or government report. The summary and analysis on this page are original editorial writing.
- Source organization
- University of California San Diego
- Source type
- University
- Authors
- Dhruv Agarwal, Zichen Wang, Eric Arkfeld, Andre Modolo, Parth Natekar, Hiroyuki Hakozaki, Mehul Arora, Gillian McMahon, Siddharth Nahar, Manav Doshi and Johannes Schöneberg
- Journal / report
- Cell
- Publication date
- September 10, 2026
- DOI
- 10.1016/j.cell.2026.08.028
- PMID
- 42721963
- Institution
- Department of Pharmacology and Department of Biochemistry and Molecular Biophysics, University of California San Diego
- Funding
- National Institutes of Health grant DP2 GM150022; National Science Foundation grant NAIRR240423; Hartwell Foundation; and W.M. Keck Foundation
- Conflicts
- UC San Diego filed for patent protection on the technology, and corresponding author Johannes Schöneberg started a company based on it
- Open access
- Yes
- Reuse approach
- Methods and results summarized independently from UC San Diego's institutional report, PubMed and the peer-reviewed paper record; no source text, microscopy, figures, tables, code or illustrations reproduced.
AI-assisted editorial process: AI tools helped organize sources and draft this review. The linked research records—not AI output—are the evidence. Publication standards and corrections are publisher-directed. Read our AI transparency policy.