Rethinking Microscopy Generation: Co-Designed Diffusion for Biologically Interpretable Single-Cell Synthesis
Authors: Shumpei Takezaki, Duway Nicolas Lesmes-Leon, Ryoma Bise, Gillian Lovell, Bianca Migliori, Andreas Dengel, Sheraz Ahmed
Organizations: Kyushu University, Fukuoka, Japan · University of Kaiserslautern-Landau (RPTU), Kaiserslautern, Germany · German Research Center for Artificial Intelligence (DFKI), Kaiserslautern, Germany · Sartorius, BioAnalytics, Royston, United Kingdom
Generative models for microscopy can already produce realistic whole-image fields of view, but offer no principled way to control how each individual cell looks. We address this gap by recasting microscopy generation at the single-cell level and designing the input controls and the model's internal representation jointly, so that both are organized along the same biologically meaningful directions. On the input side, we aggregate a long list of standard cell measurements into a small set of semantic axes (size, elongation, brightness, texture, etc.) that are each named, measurable, and traversable. On the model side, we introduce BioVAE, an encoder-decoder whose latent space is explicitly aligned with the same set of measurements. A key finding is that the benefit of biologically structuring the latent emerges only when conditioning is also moved to the semantic axes, showing that the two ends must be designed jointly. Experiments on LIVECell show that semantic-axis conditioning substantially improves generative fidelity over high-dimensional conditioning, a blind study with biologists confirms the realism of generated cells, and per-axis traversals produce stable and interpretable phenotypic changes.
Figures & tables
Figure 1: Rethinking phenotype control in microscopy generation. (a) Scene-level methods generate the whole field of view and cannot directly control individual cells. (b) Existing single-cell methods do not jointly design the conditioning signal and the latent representation, so phenotype control is unreliable or out of reach. (c) Our framework co-designs both around biologically meaningful semantic axes, enabling stable and interpretable per-cell control.
Figure 2: Overview of the controllable single-cell latent diffusion framework. (a) A concept of single-cell generation. (b) Single-cell extraction from the LIVECell [ Edlund et al.(2021)Edlund, Jackson, Khalid, Bevan, Dale, Dengel, Ahmed, Trygg, and Sjögren ] dataset and calculation of morphodynamic descriptors ( v ) and semantic axes ( s ). (c) The Biologically Regularized VAE (BioVAE) architecture, which predicts v from the latent space to enforce biological structure. (d) The conditional single-cell latent diffusion model guided by morphodynamic descriptors v or semantic axes s for controllable synthesis.
Method
MSE \downarrow\
SSIM ↑
PSNR ↑
rFID ↓
Morph. MSE ↓
VAE [ Rombach et al.(2022)Rombach, Blattmann, Lorenz, Esser, and Ommer ]
0.805
0.993
46.6
0.151
6.748
BioVAE
0.600
0.994
49.8
0.161
6.816
BioVAE w/o LBio
0.795
0.987
46.7
0.191
6.699
BioVAE w/o LEntropy
0.630
0.993
49.3
0.341
6.617
Table 1: Quantitative ablation of BioVAE loss components. Metrics include MSE ( ×10−4 ). Bold and underlined values indicate the best and second-best results, respectively. BioVAE attains the best pixel-level reconstruction (MSE, SSIM, PSNR) while remaining comparable on rFID; the entropy term trades a small increase in descriptor (Morph.) MSE for improved reconstruction and latent structure (Table 2 ).
Model
K=5
K=10
K=20
SVM Acc
Chance
0.00003
0.00006
0.00012
17.8%
VAE
0.0056
0.0040
0.0029
70.5%
BioVAE
0.0058
0.0041
0.0031
71.2%
Rel. gain
+3.6%
+2.5%
+6.9%
+1% ( p<0.001 )
Table 2: BioVAE latent structure. kNN neighborhood preservation against the morphodynamic space and cell type classification accuracy.
Regimen
FID ↓
MMD ↓
Prec. ↑
Recall ↑
Morph. MSE ↓
VAE + Morph.
2.04
0.0047
0.872
0.843
0.864
BioVAE + Morph.
2.12
0.0047
0.887
0.834
0 .876
VAE + Sem. Ax.
0 .974
0 .0015
0.875
0.86
0.944
BioVAE + Sem. Ax. (Ours)
0.933
0.0014
0 .880
0 .854
0.925
Table 3: Quantitative assessment of single-cell diffusion under different conditional regimes.
Figure 3: Generated images from VAE and BioVAE LDM with semantic axes. (a) BioVAE LDM preserves morphological features more faithfully than VAE LDM. (b) Challenging examples with low organelle preservation.
Figure 4: Biology expert evaluation. (a) Confusion matrix for real versus synthetic classification. (b) Distribution of realism scores for real and synthetic images.
Figure 5: Representative semantic traversals generated from three validation anchor cells. Each line plot shows the measured semantic-axis response across the traversal range, with their corresponding generated cell images. Dashed red lines indicate the semantic value of the original anchor cell.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1: Spearman correlation matrix of the semantic axes in the train data split. Several semantic axes exhibit structured biological dependencies due to shared underlying descriptors and correlated phenotype properties.
Axis
Controllability ↑
Gain ↑
Selectivity ↑
cell_size
0.978
0.712
1.544
elongation
0.754
0.287
0.957
shape_compactness
0.006
0.027
0.448
boundary_irregularity
0.599
0.199
1.043
global_shape_complexity
0.978
0.901
0.945
symmetry
0.915
0.643
1.169
Appendix
Table S1: Quantitative controllability analysis of the 15 semantic axes. Target controllability is measured using the Spearman correlation between requested and generated semantic values. Control gain is estimated as the slope of a linear regression between requested and generated semantic values. Selectivity is defined as the ratio between the target-axis semantic displacement and the average displacement of the non-target semantic axes, quantifying the relative dominance of the requested semantic response.
Figure S2: Representative semantic traversals generated from six validation anchor cells. Each line plot shows the measured semantic-axis response across the traversal range, with their corresponding generated cell images. Dashed red lines indicate the semantic value of the original anchor cell.
Modeling single-cell gene expression across diverse biological and technical conditions is crucial for characterizing cellular states and simulating unseen scenarios. Existing methods often treat genes as independent tokens, overlooking their high-level biological relationships and leading to poor performance. We introduce SAVE, a unified generative framework based on conditional Transformers for multi-condition single-cell modeling. SAVE leverages a coarse-grained representation by grouping semantically related genes into blocks, capturing higher-order dependencies among gene modules. A Flow Matching mechanism and condition-masking strategy further enhance flexible simulation and enable generalization to unseen condition combinations. We evaluate SAVE on a range of benchmarks, including conditional generation, batch effect correction, and perturbation prediction. SAVE consistently outperforms state-of-the-art methods in generation fidelity and extrapolative generalization, especially in low-resource or combinatorially held-out settings. Overall, SAVE offers a scalable and generalizable solution for modeling complex single-cell data, with broad utility in virtual cell synthesis and biological interpretation. Our code is publicly available at https://github.com/fdu-wangfeilab/sc-save
Jiahao Li, Jiayi Dong, Peng Ye +3
College of Computer Science and Artificial Intelligence, Fudan University · Shanghai Key Laboratory of Intelligent Information Processing, Fudan University
Flow-matching generative models are increasingly used to simulate cell responses to biological perturbations. However, the design space for building such models is large and underexplored. We systematically analyse the design space of flow matching models for cell-microscopy images, finding that many popular techniques are unnecessary and can even hurt performance. We develop a simple, stable, and scalable recipe which we use to train our foundation model. We scale our model to two orders of magnitude larger than prior methods, achieving a two-fold FID and ten-fold KID improvement over prior methods. We then fine-tune our model with pre-trained molecular embeddings to achieve state-of-the-art performance simulating responses to unseen molecules. Code is available at https://github.com/valence-labs/microscopy-flow-matching
Charles Jones, Emmanuel Noutahi, Jason Hartford +1
Imperial College London · Valence Labs · University of Manchester
Building virtual cells with generative models to simulate cellular behavior in silico is emerging as a promising paradigm for accelerating drug discovery. However, prior image-based generative approaches can produce implausible cell images that violate basic physical and biological constraints. To address this, we propose to post-train virtual cell models with reinforcement learning (RL), leveraging biologically meaningful evaluators as reward functions. We design seven rewards spanning three categories-biological function, structural validity, and morphological correctness-and optimize the state-of-the-art CellFlux model to yield CellFluxRL. CellFluxRL consistently improves over CellFlux across all rewards, with further performance boosts from test-time scaling. Overall, our results present a virtual cell modeling framework that enforces physically-based constraints through RL, advancing beyond "visually realistic" generations towards "biologically meaningful" ones.