scMorphJEPA applies spatial masked prediction with SIGReg regularization to learn cell morphology representations from multi-channel fluorescence microscopy images, without EMA, without augmentations, and with only two loss terms.
On the Severin PBMC dataset (113,564 five-channel immune cell images, 8 cell types):
| Method | Training | k-NN (k=5) | k-NN (k=10) | k-NN (k=20) |
|---|---|---|---|---|
| DINO ViT-S/16 (zero-shot baseline) | None | 53.2% | 55.0% | 56.1% |
| scMorphJEPA (1K images, 30 epochs) | ~15 min, 1xT4 | 67.6% | 67.3% | 66.9% |
| scMorphJEPA (40K images, 8 epochs) | ~100 min, 1xT4 | 63.8% | 63.6% | 62.9% |
scMorphJEPA improves over the zero-shot pretrained baseline with only 1,000 training images, demonstrating strong data efficiency. Note: baseline and model numbers above were measured under their original preprocessing; a normalization-matched comparison and full-scale runs are in progress (see Ongoing Work).
Holding data size, epochs, and objective fixed and varying only the input normalization, so the comparison isolates a single factor:
| Normalization | k-NN (k=5) | k-NN (k=10) | k-NN (k=20) |
|---|---|---|---|
| per_channel | 67.0% | 67.4% | 67.1% |
| per_channel_percentile | 60.5% | 61.3% | 61.3% |
| per_image | in progress | in progress | in progress |
Robust percentile clipping (per_channel_percentile) lowers k-NN here, consistent with the brightest pixels carrying real marker signal rather than artifact. The VICReg objective on the two strongest normalizations will be added next.
Self-supervised vision transformers (scDINO, Cell-DINO) have shown excellent performance on cell phenotyping tasks using DINO-based self-distillation. However, these methods rely on:
- Exponential Moving Average (EMA) teacher networks
- Multi-crop augmentation strategies designed for natural images
- Complex multi-component training objectives
scMorphJEPA replaces all of this with a simpler framework inspired by LeWorldModel (Maes et al., 2026):
| scDINO | Cell-DINO | I-JEPA | scMorphJEPA | |
|---|---|---|---|---|
| Learning objective | Self-distillation | Self-distillation | Masked prediction | Masked prediction |
| Collapse prevention | EMA | EMA | EMA | SIGReg |
| Augmentations | Multi-crop, color | Multi-crop, color | None | None |
| EMA required | Yes | Yes | Yes | No |
| Loss terms | Multiple | Multiple | 2 + EMA | 2 only |
- Encode: A ViT-S/16 encoder (initialized from DINO-pretrained ImageNet weights, adapted to 5 fluorescence channels) processes cell images into 196 patch embeddings
- Mask: 60% of patch embeddings are randomly masked
- Predict: A lightweight transformer predictor reconstructs masked patch embeddings from visible context
- Regularize: SIGReg enforces an isotropic Gaussian distribution on embeddings, preventing collapse without EMA
Total loss = MSE(predicted, target) + lambda * SIGReg(embeddings)
One hyperparameter (lambda). No EMA. No augmentations. No multi-term loss.
We use the Deep Phenotyping PBMC Image Set (Severin et al., 2022):
- 113,564 five-channel fluorescence microscopy images
- 50x50 pixels, resized to 224x224 for ViT input
- 5 channels: Alexa Fluor 647, Brightfield, DAPI, FITC 488, PE 594
- 8 immune cell classes: T4, T8, T0, M0, DC, Nk, B, Negs
- Train/test split: 89,564 / 24,000
From source (recommended while the package is evolving):
git clone https://github.com/simo1946/scMorphJEPA.git
cd scMorphJEPA
pip install -e .Or directly:
pip install git+https://github.com/simo1946/scMorphJEPA.gitRequires Python 3.11+, PyTorch, and timm (installed automatically).
# Severin PBMC dataset
wget -O severin_pbmc.zip "https://www.research-collection.ethz.ch/bitstreams/8689d69b-d916-4c8e-9b3f-2981c512b70b/download"
unzip -q severin_pbmc.zip -d severin_data
# DINO pretrained ViT-S/16 weights
wget -O dino_vits16.pth "https://dl.fbaipublicfiles.com/dino/dino_deitsmall16_pretrain/dino_deitsmall16_pretrain.pth"python -m scmorphjepa.cli train \
--data_dir severin_data/DeepPhenotype_PBMC_ImageSet_YSeverin \
--checkpoint dino_vits16.pth \
--epochs 50 --batch_size 24 --n_images 0 # 0 = use all
python -m scmorphjepa.cli evaluate \
--model output/best_model.pt \
--data_dir severin_data/DeepPhenotype_PBMC_ImageSet_YSeverin \
--checkpoint dino_vits16.pthfrom scmorphjepa.models.builder import build_scmorphjepa
from scmorphjepa.models.cell_jepa import ScMorphJEPAConfig
from scmorphjepa.data.datasets import SeverinDataset
from scmorphjepa.training.trainer import Trainer, TrainConfig
model = build_scmorphjepa("dino_vits16.pth", ScMorphJEPAConfig(in_channels=5))
train_ds = SeverinDataset("severin_data/.../Training", normalize="per_image")
val_ds = SeverinDataset("severin_data/.../Test", normalize="per_image")
cfg = TrainConfig(
batch_size=24, epochs=50,
regularizer="sigreg", # sigreg | vicreg | koleo | barlow | none
n_images=0, # 0 = use all
output_dir="output",
)
Trainer(model, train_ds, val_ds, cfg).train()Evaluate frozen embeddings with k-NN:
from scmorphjepa.evaluation.evaluate import extract_embeddings, knn_evaluate
ref = extract_embeddings(model, train_ds, batch_size=64)
query = extract_embeddings(model, val_ds, batch_size=64)
acc = knn_evaluate(ref["cls_tokens"], ref["labels"], query["cls_tokens"], query["labels"],
k_values=[5, 10, 20])
print({k: f"{v:.1%}" for k, v in acc.items()})Training is resumable: re-running the same configuration continues from the last completed epoch. Set drive_checkpoint_dir to mirror checkpoints to Google Drive for Colab.
scMorphJEPA/
├── README.md
├── pyproject.toml
├── scmorphjepa/ # installable package
│ ├── models/ # encoder, predictor, builder, baselines
│ ├── data/ # dataset loaders (Severin + generic folder)
│ ├── training/ # trainer + regularizers (sigreg, vicreg, koleo, barlow)
│ ├── evaluation/ # k-NN, linear probe, clustering metrics
│ ├── analysis/ # interpretability, channel attribution
│ └── cli.py # command-line train / evaluate / analyze
├── configs/ # run configs
├── tests/
├── figures/
└── results/
└── results_log.txt
Both prediction loss and SIGReg loss decrease consistently during training, confirming the model learns meaningful spatial structure:
| Epoch | Pred Loss (train) | SIGReg (train) | Pred Loss (val) |
|---|---|---|---|
| 1 | 5.70 | 15.78 | 3.19 |
| 10 | 1.38 | 6.92 | 1.36 |
| 20 | 0.81 | 5.33 | 0.79 |
| 27 | 0.27 | 4.53 | 0.26 |
- Normalization study (per_image vs per_channel vs per_channel_percentile) at matched budget
- Objective ablation (SIGReg vs VICReg vs KoLeo vs Barlow)
- Full scaling curve (1K to 89K images) with consistent training
- Normalization-matched comparison with scDINO (reproduction on same dataset)
- Cross-channel latent-predictive JEPA and channel predictability map
- Extension to Cell Painting datasets (BBBC021)
scMorphJEPA is Stage 1 of the CellAgora research program, a multi-stage framework for AI-driven cell biology. Future stages extend to transcriptomic pairing, perturbation prediction, and spatial cell-cell interaction modeling.
@article{bonaccorsi2026scmorphjepa,
title={scMorphJEPA: Self-Supervised Cell Morphology Learning via Spatial
Joint-Embedding Predictive Architecture},
author={Bonaccorsi, Simone},
year={2026},
note={Preprint in preparation}
}This work builds on:
- LeWorldModel (Maes et al., 2026), SIGReg regularization for stable JEPA training
- scDINO (Pfaendler et al., 2023), self-supervised ViTs for cell microscopy
- I-JEPA (Assran et al., 2023), image-based JEPA framework
- Severin PBMC dataset (Severin et al., 2022)
MIT