MorphGen

MorphGen is a generative model for high-content cell microscopy that simulates, in silico, how cells respond to interventions. Phenotypic screening platforms such as Cell Painting capture rich, multichannel fluorescent images in which different stains reveal distinct organelles, and generating realistic versions of these images could help model cellular responses to perturbations for drug discovery. MorphGen targets this problem with a diffusion model that produces controllable and morphologically plausible microscopy images across multiple cell types and perturbations.

Its central design choice sets it apart from prior work. Rather than compressing the multiple fluorescent stains into a single RGB image — a common simplification that blurs organelle-specific information — MorphGen generates the complete set of fluorescent channels jointly, preserving per-organelle structure. To keep the generated images biologically consistent, the model is trained with an alignment loss that matches its internal representations to the phenotypic embeddings of OpenPhenom, a biological foundation model for microscopy. This grounds generation in a learned representation of cell phenotype rather than pixel fidelity alone. MorphGen was developed by researchers at the Institute of Science and Technology Austria together with the Chan Zuckerberg Initiative, and released as a preprint with open code in October 2025.

The combination of joint multichannel synthesis and foundation-model alignment yields images that retain the morphological detail needed for downstream phenotypic analysis.

#Key Features

• Joint multichannel generation: Synthesizes the full set of fluorescent channels together rather than collapsing them into RGB, preserving organelle-specific detail.
• OpenPhenom alignment: An alignment loss matches generated representations to the phenotypic embeddings of the OpenPhenom foundation model, keeping outputs biologically consistent.
• Controllable across conditions: Generates images conditioned on cell type and perturbation, enabling in-silico simulation of cellular responses to interventions.
• State-of-the-art fidelity: Achieves an FID score more than 35% lower than the prior state-of-the-art model MorphoDiff.
• Public implementation: Code from the CZI AI residency project is released publicly under a custom CZI/ISTA non-commercial license.

#Technical Details

MorphGen is a diffusion-based generative model for fluorescent microscopy that conditions generation on cell type and perturbation and outputs all fluorescent channels of a Cell Painting-style image jointly. Beyond the standard diffusion objective, it is trained with an alignment loss that pulls its representations toward the phenotypic embeddings produced by OpenPhenom, a state-of-the-art microscopy foundation model, so that generated images preserve morphological and per-organelle structure rather than only matching pixel statistics. On image-quality evaluation the model reports a Fréchet Inception Distance more than 35% lower than the previous best method, MorphoDiff. The work is a preprint awaiting peer review.

#Applications

MorphGen is aimed at researchers working with high-content imaging screens who want to model or augment cellular responses to genetic and chemical perturbations. Realistic multichannel image generation can support in-silico exploration of intervention effects, data augmentation for phenotypic classifiers, and hypothesis generation about morphological changes, benefiting drug-discovery and functional-genomics groups that rely on Cell Painting-style assays. Because it preserves individual channels, downstream per-organelle analyses remain possible on the generated images.

#Impact

By generating full multichannel microscopy images and anchoring them to a phenotypic foundation model, MorphGen advances generative modeling of cell morphology beyond RGB approximations and improves image fidelity over prior diffusion baselines. It illustrates a broader pattern of coupling generative models with biological foundation models to enforce domain consistency. As a recently released preprint with open code, its adoption and independent validation in real screening pipelines are still emerging.