CytoGAN: Generative Modeling of Cell Images
Peter Goldsborough, Nick PawlowskiiD, Juan Carlos CaicedoiD, Shantanu SinghiD, Anne Elizabeth CarpenteriD
Abstract
We explore the application of Generative Adversarial Networks to the domain of morphological profiling of human cultured cells imaged by fluorescence microscopy. When evaluated for their ability to group cell images responding to treatment by chemicals of known classes, we find that adversarially learned representations are superior to autoencoder-based approaches. While currently inferior to classical computer vision and transfer learning, the adversarial framework enables useful visualization of the variation of cellular images due to their generative capabilities.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 63
- Organism
- Assay
- Research Focus
- Approach
Abstract summary
Claude Opus 5.5 · xhigh effort · 18 Oct 2026 · from the abstract
Adversarially learned features group fluorescence images of chemically treated human cells better than autoencoder features, though classical computer vision and transfer learning still do better.
Morphological profiling turns microscopy images of treated cells into feature vectors, and the open question is which representation puts compounds of the same class together. This work places generative adversarial networks beside autoencoders, classical computer vision features and transfer learning on that task, and notes that being generative lets the model show what varies across cell images.
Key results
SuggestedRepresentations learned adversarially outperform autoencoder representations at grouping treated-cell images by compound class.
Reported for human cultured cells imaged by fluorescence microscopy and treated with chemicals of known classes; the abstract gives no grouping metric, no dataset size and no model configurations. · Abstract
SuggestedThe adversarial representations remain behind classical computer vision features and transfer learning on the same grouping task.
Classical computer vision features and transfer learning are both stated to remain ahead of the adversarial approach on the same grouping task, again without numbers. · Abstract
ClaimedBecause the model generates images, it can be used to visualize how cellular appearance varies.
Offered as a qualitative benefit of the generative framework; no example interpretations or validation of the visualizations are described in the abstract. · Abstract
Methods, data and reproducibility
Generative adversarial networks are applied to fluorescence microscopy images of cultured human cells to learn morphological profiles, and the resulting representations are scored by how well they group images of cells treated with chemicals from known classes, against autoencoder, classical computer vision and transfer-learning alternatives.
- fluorescence microscopy images of cultured cells treated with chemicals belonging to known classes · human
- Methods & tools
- generative adversarial network (GAN), autoencoder, fluorescence microscopy
- Reproducibility gaps
- The abstract points to no code, image set or trained model, and names neither the compound collection nor the metric used to judge grouping.
Limitations (3)
- The authors state that the adversarial representations are currently inferior to classical computer vision and transfer learning for this profiling task. · stated by the authors
- No quantitative result is given: no clustering or classification score, no dataset size, and no specification of the baselines compared against. · noted in this summary
- Evaluation is limited to recovering known chemical classes, so whether the representations help on other profiling questions is untested here. · noted in this summary
Who should read:Read this if you build image-based morphological profiles and are weighing generatively learned representations against pretrained-network or handcrafted features.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Generative adversarial networks in cell microscopy for image augmentation. A systematic review
Duway Nicolas Lesmes-Leon, Andreas R. Dengel, Sheraz Ahmed·bioRxiv·Aug 25, 2023
- Using conditional Generative Adversarial Networks (GAN) to generate de novo synthetic cell nuclei for training machine learning-based image segmentation
Mehmet İlyas Coşacak, Çağhan Kızıl·bioRxiv·Nov 12, 2022
- Information-Distilled Generative Label-Free Morphological Profiling Encodes Cellular Heterogeneity
M. C. K. Lo, D. M. D. Siu, … K. K. Tsia·bioRxiv·Nov 6, 2023
- Cellular structure image classification with small targeted training samples
Dali Wang, Zheng Lu, … Zhirong Bao·IEEE AccessPeer-reviewed·Jan 1, 2019
- Semi supervised GAN for smart microscopy, fast and data efficient cell cycle classification
R. Manick, Youssef El Habouz, … Jacques Pécréaux·bioRxiv·May 29, 2026
- CellFluxV2: An Image Generative Foundation Model for Virtual Cell Modeling
Yuhui Zhang, Yuchang Su, … Serena Y. Yeung·bioRxiv·Jan 20, 2026