MitoSpace

Cell biologists have described mitochondrial shape along a single axis for decades: fragmented at one end, hyperfused at the other. A network that stays elongated but stops moving has nowhere to sit on that axis, and neither does the gap between two fusion inhibitors that visibly produce different phenotypes. MitoSpace replaces the axis with a 2,048-dimensional space learned from four-dimensional (x, y, z, and time) movies of living cells, using no labels and no predefined morphological descriptors.

The model was built in the Schöneberg lab at UC San Diego and published in Cell in 2026. Its training corpus is 26 TB of lattice light-sheet microscopy: Cal27 head-and-neck cancer cells under 25 mitochondria-targeting compounds plus a DMSO control, spanning fission and fusion inhibitors, uncouplers, ionophores, electron-transport-chain inhibitors, DNA-damage inducers, ROS modulators, and cytoskeletal disruptors. Cells were stained with MitoTracker Green and imaged at one-minute intervals for 20 minutes across ten regions per condition; roughly 40,000 single-cell 4D movies were cropped using Cellpose-SAM masks. A 2023 preprint from the same lab introduced MitoSpace as a 2D model over about 100,000 confocal images; the published model is that line carried into volumes and time.

The encoder is frozen after pretraining, so MitoSpace is used the way other self-supervised microscopy representation models such as SubCell are used: embeddings are computed once, then read out by lightweight probes, nearest-neighbour retrieval, or UMAP. What distinguishes it from fixed-cell screening corpora is the input: organelle motion is in the tensor, not averaged away.

#Key Features

• Function inferred from form: A TMRM membrane-potential channel was co-acquired during imaging but withheld from self-supervised training. A regression probe on the frozen embeddings predicts it with an R² of 0.91, at median errors of roughly 0.5%–5% across all 26 conditions.
• Zero-shot generalization: Trained on only 10 of the 26 conditions, top-1 k-NN accuracy drops from 74.67% to 68.37% on held-out drugs, and the unmodified model separates human lung organoid developmental stages at 62.9% despite never seeing organoid data.
• Dimensionality is load-bearing: The same framework trained at 2D, 3D, and 4D reaches 56.12%, 68.30%, and 74.67% top-1 accuracy respectively, with 4D best in 22 of 26 conditions.
• Interpretable structure: The leading principal components correlate with MitoGraph and MitoTNT network features, and probes recover segment length, motility, and fission/fusion rates at median errors of about 2%–8%.
• Stain and day robustness: It separates perturbations imaged with MitoBright LT Green and MitoTracker Deep Red, dyes absent from training, and holds accuracy across replicate days without retraining.

#Technical Details

Training runs in two stages. A 3D convolutional autoencoder compresses each volumetric frame 16-fold under a composite reconstruction loss, which makes distributed training tractable: an uncompressed single-cell, single-channel sample is about 150 MB. The 4D encoder then trains on the compressed data, with a 3D residual network embedding each time point and a bidirectional LSTM integrating those per-frame representations across time. The objective is SimCLR's InfoNCE loss (τ = 0.07) over two augmented views, where augmentation combines 3D geometric transforms with random time-slice masking to force temporal sensitivity. The input tensor is (20, 60, 256, 256; t, z, y, x) and the output is a 2,048-dimensional vector per frame. The released checkpoint holds 68.4M parameters and trained for 300 epochs with Adam (lr 3e-4) on 15 nodes of four NVIDIA V100 GPUs at SDSC, about three days of wall time. As a baseline, k-NN classification on 13 predefined image-processing features reaches 23.81% top-1 accuracy against the embedding's 74.67%.

#Applications

The immediate use is 4D phenotypic screening: embedding a compound's cells and retrieving their neighbours ranks mechanistic similarity without any target annotation, which also surfaces off-target behaviour — rotenone, a canonical complex I inhibitor, lands among the microtubule disruptors, consistent with its reported tubulin affinity. Because membrane potential is recoverable from morphology and dynamics alone, a fluorescence channel otherwise spent on a potential-sensitive dye is freed for another marker. The organoid results extend the same frozen encoder to developmental staging.

#Impact

MitoSpace makes the case that volumetric, time-resolved imaging carries phenotypic signal that flat snapshots discard, and it supplies a quantitative link between mitochondrial form and bioenergetic function at single-cell resolution. Its scope is narrow by construction: one cell line, 25 compounds, and a 60-second frame interval over a 20-minute window that resolves bulk motility and remodelling rates but not fast transients such as calcium flashes or pearling. The input geometry is fixed at 20 frames of 60 z-planes at 256 × 256, and inference requires a CUDA GPU. Release terms lag the publication — the paper states that the model, dataset, and interactive explorer are publicly available, but as of September 2026 both the GitHub repository and the Hugging Face weight repository still carry the review license the UC Regents issued for evaluation of the Cell submission, the checkpoint sits behind a peer-review-only access agreement, and the training corpus is in a requester-pays S3 bucket that anonymous clients cannot list.