Micellangelo

Micellangelo is a generative model that predicts how cells respond to the physical microstructure of the surfaces they grow on. The design of biomaterials for implants, tissue scaffolds, and regenerative medicine depends heavily on surface topography, which can steer cell shape, cytoskeletal organization, and signaling. Experimentally screening the vast space of possible micro-topographies is slow and expensive, motivating computational surrogates that can anticipate cellular responses before fabrication and imaging.

Developed by Nikita Konshin, Koen Minartz, Jan de Boer, and Vlado Menkovski at Eindhoven University of Technology and posted to bioRxiv in November 2025, Micellangelo is a flow-matching generative model conditioned on surface micro-topographies. Given a topography pattern, it synthesizes high-resolution fluorescence images of human dermal fibroblasts across three subcellular stains: DNA (nucleus), F-actin (cytoskeleton), and YAP (a mechanotransduction transcriptional regulator). The result is a "digital twin" that allows in-silico exploration of cell-material interactions.

Micellangelo sits at the intersection of generative image modeling and mechanobiology. Rather than predicting a structure or a scalar readout, it generates the full imaging phenotype a researcher would observe under the microscope, making the relationship between a surface design and its cellular consequences directly visualizable.

#Key Features

• Topography-conditioned generation: The model takes a surface micro-topography as conditioning input and generates the corresponding fluorescence appearance of fibroblasts grown on that surface.
• Multi-channel subcellular imaging: A single model jointly produces DNA, F-actin, and YAP channels, capturing nuclear, cytoskeletal, and mechanotransduction readouts simultaneously.
• Flow-matching backbone: Built on flow matching, a continuous-time generative framework related to diffusion that transforms Gaussian noise into realistic single-cell images with high sample quality.
• Subcellular perturbation from a fixed checkpoint: Once trained, the model supports in-silico perturbation of subcellular features without retraining, enabling cheap exploration of cell-material design hypotheses.

#Technical Details

Micellangelo is a conditional flow-matching generative model. Flow matching learns a velocity field that transports a simple Gaussian prior to the data distribution of fluorescence images, conditioned on the surface topography. The model was trained on paired data of human dermal fibroblasts imaged across three fluorescence stains (DNA, F-actin, YAP) on a defined set of engineered surface micro-topographies. The reported scope is deliberately narrow: roughly ten distinct topographies, three stains, and a single cell type (dermal fibroblasts). Within that domain, the model reproduces realistic single-cell morphology and the topography-dependent variation in cytoskeletal and YAP signal. The authors do not report released weights or code at preprint time, and parameter counts and full hyperparameters are not summarized in the brief.

#Applications

Micellangelo is aimed at biomaterials and mechanobiology researchers who design surface topographies for implants, scaffolds, and cell-instructive substrates. By generating expected fluorescence phenotypes for a candidate topography, it can help prioritize designs for physical fabrication, support hypothesis generation about how specific surface features drive cytoskeletal remodeling or YAP localization, and provide a fast in-silico stand-in for imaging assays during early-stage screening.

#Impact

Micellangelo is an early demonstration that flow-matching generative models can serve as digital twins of cell-material interactions, generating full imaging phenotypes rather than summary statistics. Its main limitation is scope: it is trained and validated only on dermal fibroblasts across a small panel of topographies and three stains, so generalization to other cell types, stains, or material surfaces is unproven, and the absence of released weights or code constrains immediate reuse. Even so, it points toward generative imaging as a practical tool for accelerating biomaterial design.