Loading…
Latent diffusion model that paints high-resolution Cell Painting images of cells responding to a chemical compound or an over-expressed gene.
No providers recorded yet. Browse all providers
Cell Painting assays read out how a cell's morphology changes under an intervention: fluorescent dyes label compartments such as DNA, actin, mitochondria and RNA, and the resulting multichannel images encode a phenotypic fingerprint of a compound or a genetic edit. The assay is cheap relative to most functional readouts but still costs plate time, and the space of possible interventions is vastly larger than any screen. Predicting the image rather than measuring it is therefore an appealing shortcut — but earlier generative attempts worked on cropped single-cell patches, losing the cell density and intercellular context that make a full field of view informative.
MorphoDiff generates full-resolution Cell Painting images conditioned on a perturbation. It adapts the Stable Diffusion latent-diffusion pipeline by keeping the image autoencoder and replacing the CLIP text encoder with a perturbation encoder. That substitution is the whole idea: a model that normally conditions on a text prompt instead conditions on a numerical description of what was done to the cells, and because that description is a continuous embedding rather than a categorical label, it can be queried with interventions never seen in training.
The work is by Zeinab Navidi, Bo Wang and colleagues at the University of Toronto, University Health Network, the Vector Institute, the Broad Institute Imaging Platform, and the Structural Genomics Consortium. It was posted to bioRxiv in December 2024 and presented at ICLR 2025.
The base model is Stable Diffusion v1-4; its VAE, trained on natural images, is retained unchanged and compresses a 512×512 RGB image into a four-channel 64×64 latent, and the UNet denoiser is fine-tuned on Cell Painting data with the perturbation embedding supplied where the text embedding would otherwise go. Each dataset's fluorescence channels are mapped down to three RGB channels first. Three public datasets were used: RxRx1, with 1,108 perturbations across four cell types, run both across all HUVEC batches and constrained to a single batch; BBBC021, comprising 13,200 images of MCF7 breast cancer cells, run over all compounds and over a 14-compound subset with mechanism-of-action labels; and the Rohban dataset of U2OS cells carrying 323 over-expressed genes, run on 5-gene and 12-gene subsets. Fidelity is measured with Fréchet Inception Distance and Kernel Inception Distance against a conditional Stable Diffusion baseline. On RxRx1 across all batches MorphoDiff reaches FID 0.78 and KID 0.05 against 1.15 and 0.11; on BBBC021 across all compounds, 1.99 and 0.21 against 3.84 and 0.47; on the Rohban 5-gene setting, 2.51 and 0.33 against 3.26 and 0.45. Checkpoints are released per source dataset and experiment rather than as one universal model.
The target use is in-silico triage of a perturbation library. A screening group can render the predicted morphology for candidate compounds or gene over-expressions, cluster the generated images by CellProfiler features, and prioritize which interventions merit plate time — particularly for compounds structurally similar to ones already assayed, where the generalization results are strongest. The recovered mechanism-of-action structure also makes the model a hypothesis generator: an unexpected predicted phenotype is itself a result worth checking. Because the released checkpoints are dataset-specific, applying the model to a new imaging setup means training on that setup's images.
MorphoDiff shows that the text-to-image diffusion stack transfers cleanly to perturbational microscopy once the conditioning signal is replaced, and that the transfer preserves enough biology for downstream feature extraction to recover mechanism-of-action structure. Weights are public on Hugging Face and code on GitHub, both under Apache-2.0, though the repository documents training and generation workflows rather than a one-line path from released checkpoint to sampled image. Other conditioning signals have since been explored — MorphDiff, a separate model with a similar name, drives generation from measured L1000 expression profiles instead. The paper is candid about limits: checkpoints are per dataset, fidelity metrics are computational rather than a wet-lab confirmation that predicted phenotypes match observed ones, and generalization was assessed on held-out compounds chosen for structural similarity to the training set.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.