MorphoGenie

A CellProfiler run describes one cell with roughly 1,700 hand-crafted measurements; a deep encoder replaces those with an embedding of several hundred dimensions. Both separate cell states, and neither tells a biologist which property of the cell actually changed. MorphoGenie compresses a single-cell image into ten latent dimensions and then earns the right to name them: sweeping one dimension while the other nine are held fixed produces a series of reconstructed cells, and measuring how each of 35 explicitly defined morphological descriptors — size and dry-mass density, global texture, local texture statistics — varies along that sweep reveals what the dimension encodes. Clustering those variances yields what the authors call an interpretation heatmap, which turns an unsupervised embedding into a profile a wet-lab reader can reason about.

It was developed by Kevin Tsia's group at The University of Hong Kong with the Advanced Biomedical Instrumentation Centre, posted to bioRxiv in September 2024 and published in Nature Communications in December 2025. Much of its imagery comes from the group's own ultrafast quantitative phase imaging (QPI) flow cytometers.

What makes it a reusable model rather than a per-experiment pipeline is that a checkpoint trained on one dataset can be applied to unseen datasets acquired in a different image contrast without retraining: a model trained on fluorescence Cell Painting images of adherent cells classifies label-free QPI images of suspension lung cancer cells. Its generative half serves that goal rather than image synthesis for its own sake, unlike MorphoDiff, where conditioned image generation is the product.

#Key Features

• Ten-dimensional disentangled latent space: Each image reduces to ten factors, against 512 for an adversarial autoencoder, 1,024 for a VQ-VAE and roughly 1,700 CellProfiler features.
• Interpretation by latent traversal: The interpretation heatmap maps latent dimensions onto a hierarchy of bulk, global-texture and local-texture descriptors, so a shift in the profile can be attributed to a named morphological property.
• Cross-modality transfer: Frozen checkpoints carry to new datasets across QPI, fluorescence, bright-field and phase-contrast imaging, and across suspension and adherent cell formats.
• High-fidelity reconstruction: The GAN stage recovers subcellular detail the VAE alone blurs, which is what makes the traversal images interpretable.
• Discrete and continuous readouts: One representation serves both cell-state classification and pseudotime trajectory inference.

#Technical Details

Training proceeds in two stages on 256×256 single-cell crops. A FactorVAE first learns the disentangled ten-dimensional representation, its total-correlation penalty giving a better disentanglement-versus-reconstruction trade-off than a vanilla VAE or β-VAE in the authors' comparisons. That inference model is then transferred to a GAN trained with an information-distillation objective in the manner of ID-GAN, so the generator adds entangled nuisance detail on top of the disentangled prior instead of a plain Gaussian. Preprocessing segments plated images with Cellpose 2.0 and flow-cytometry images by intensity threshold, then centers each cell to keep position out of the representation.

The training corpora are four modest datasets rather than a web-scale collection: in-house multi-ATOM QPI of lung cancer cell lines, a Clobetasol Propionate subset of the public BBBC022 Cell Painting assay, a TGF-β-induced EMT time-lapse of A549 cells expressing vimentin-RFP, and FACED-based QPI of MDA-MB-231 cell-cycle progression. Generalization is quantified by F1 score on held-out datasets; the journal version extends this to eight checkpoints across eight imaging scenarios, including primary human CD4⁺/CD8⁺ T-cell activation, multi-batch lung cancer subtyping, bright-field drug response, and the LIVECell phase-contrast collection. Training ran on two consumer GPUs, an RTX 1080 for the VAE stage and an RTX 4090 for the GAN.

#Applications

The framework suits laboratories doing image-based phenotypic screening or high-throughput imaging cytometry who want an annotation-free profile they can still explain. Demonstrations include classifying histological subtypes of lung cancer from label-free images, tracking morphological response to a glucocorticoid receptor agonist in a Cell Painting assay, ordering cells along cell-cycle and epithelial-mesenchymal transition trajectories, and separating resting from activated primary T cells. Because profiles from different studies share one compact latent space, they can be compared across experiments.

#Impact

MorphoGenie makes an argument that matters beyond this model: a small, disentangled representation can be both more transferable and more interpretable than a large one, provided the interpretation is established empirically rather than asserted. Its scope should be read honestly. Pretraining used a handful of datasets from a small number of labs, generalization is shown among those and a few public collections rather than against a broad external benchmark, and there is no single universal checkpoint — the released models are per-dataset, and the transfer claim is that any of them works on the others. The authors list 3D imaging, batch-effect correction and QPI-to-fluorescence translation as future work. Code and pre-trained weights are released under GPL-3.0 and the reproduction datasets sit on Figshare, with the EMT time-lapse available only by permission from its original authors.