Pre-publication peer-review materials
By accessing these weights you agree to use them solely for peer review of the associated manuscript submitted to Cell. No redistribution, modification, or derivative works are permitted prior to publication.
Log in or Sign Up to review the conditions and access this model content.
MitoSpace
Self-supervised 3D-ResNet + Bi-LSTM trained with SimCLR on 4D MitoTracker lattice light-sheet volumes. Produces 2048-d spatiotemporal embeddings of mitochondrial morphology.
- Interactive atlas: mitospace.ai
- Code: github.com/schoeneberglab/MitoSpace4D
- Status: released alongside a manuscript under review at Cell
Specs
| Input | (B, T=20, C=1, Z=60, H=256, W=256), float32 in [0, 1] |
| Output | (B, T, 2048) per-frame features (typically L2-normalized) |
| Objective | InfoNCE (SimCLR), ฯ = 0.07, 2 views per sample |
| Checkpoint | ms4d โ Cal27, 26 perturbation conditions |
| Hardware (training) | 15 ร 4 V100 (SDSC), DDP + fp16, 300 epochs, โ 3 days |
| Optimizer | Adam, lr = 3e-4, wd = 1e-5, cosine annealing |
| Framework | PyTorch + PyTorch Lightning |
Files
| File | |
|---|---|
model.safetensors |
Trained weights (flat state-dict, no model. Lightning prefix) |
config.json |
Input contract + architecture summary |
LICENSE |
Review License โ see below |
mitospace.gif |
Atlas preview |
Data manifest
The processed-data manifest is a Parquet table of sample metadata and relative paths used across the pipeline (autoencoder training, SimCLR training, evaluation). Point data.manifest_path in autoencoder/config.yaml (and the corresponding field in simclr/config.yaml) at the local file after download.
S3
aws s3 cp s3://mitospace4d/processed_data/manifest.parquet manifest.parquet
Hugging Face (schoeneberglab/mitospace; needs HF_TOKEN with read access while the repo is private)
export HF_TOKEN=<read_token>
python utils/hf_checkpoint.py download --filename processed_data/manifest.parquet
Usage
Requires a CUDA GPU (MitoSpace4D.__init__ unconditionally moves its augmentation pipeline to CUDA). Clone the code repo and pip install -e . first.
import sys, torch
import torch.nn.functional as F
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
sys.path.insert(0, "/path/to/MitoSpace4D")
from simclr.model import MitoSpace4D
from utils.utils import load_config
weights = hf_hub_download("schoeneberglab/mitospace", "model.safetensors")
cfg = load_config("simclr/config.yaml")
model = MitoSpace4D(embedding_size=2048, cfg=cfg, apply_aug=False).cuda().eval()
missing, unexpected = model.load_state_dict(load_file(weights), strict=False)
assert not missing and not unexpected, (missing[:5], unexpected[:5])
model._channels = model._channels.cuda() # required for GPU fancy-indexing
x = torch.rand(1, 20, 1, 60, 256, 256, device="cuda")
with torch.no_grad():
features, _ = model(x) # (1, 20, 2048)
embedding = F.normalize(features, dim=-1)
For batch inference + k-NN evaluation over a dataset, use evaluate.py in the code repo.
Intended use and limitations
Research only. Not validated or fit for clinical or diagnostic use. Trained on MitoTracker LLSM volumes from a specific cell-line panel; performance on different microscopes, magnifications, cell types, or reporters should be evaluated empirically. Input geometry is fixed at T=20, Z=60.
Citation
Manuscript under review at Cell; BibTeX will be added on publication. For now:
@software{mitospace4d,
author = {Schoeneberg Lab},
title = {MitoSpace},
url = {https://github.com/schoeneberglab/MitoSpace4D},
year = {2026},
}
License
Released under the terms in LICENSE โ a Review License granted solely for evaluating the associated manuscript submitted to Cell. No redistribution, modification, or derivative works are permitted prior to publication. Copyright ยฉ 2026 The Regents of the University of California.
- Downloads last month
- 3