Rethinking Microscopy Generation: Co-Designed Diffusion for Biologically Interpretable Single-Cell Synthesis

• Whole-scene generation lacks direct control over individual-cell attributes such as size or brightness, although phenotype is defined at the cell level.Existing single-cell approaches also leave gaps: shape masks do not control brightness or texture, and raw-feature controls can be redundant or poorly reflected in generated images.Source: §1 · p. 2
• The framework jointly designs the conditioning space and latent representation around shared biological phenotype directions.It aggregates standard measurements into named, measurable, traversable semantic axes and aligns BioVAE’s latent space with the same measurements.Source: §1 · pp. 2–3
• The paper’s contribution is the co-design of controls and representation, not single-cell synthesis itself, and it treats single-cell and whole-image generation as complementary.Whole-image models capture cell arrangements, while single-cell generation offers fine-grained control over individual-cell appearance.Source: §1 · p. 3
• On LIVECell, semantic-axis conditioning improves generation quality over raw measurements, while biologist review and traversals support realism and stable, interpretable per-axis changes.The paper also reports that BioVAE’s benefit emerges when conditioning uses semantic axes rather than the full measurement vector.Source: §1 · p. 3
• BioVAE aligns its compressed representation with biological measurements using an auxiliary regression objective and an anti-collapse entropy term.This makes the model’s internal representation correspond to the biological measurements underlying its controls.Source: §1 · p. 3
• Semantic axes provide named controls for recognizable phenotype attributes and work better than conditioning on the raw measurement list.Examples include size, elongation, brightness, and texture.Source: §1 · p. 3

2.1 Whole-Image Microscopy Generation

• Whole-image models capture global cell density and spatial arrangement but cannot specify how an individual cell should look.Their conditioning describes the field of view rather than each cell.Source: §2.1 · pp. 3–4

2.2 Single-Cell Modeling and Synthesis

• Earlier single-cell methods offer limited phenotype control: some lack conditioning, some use correlated raw measurements, and mask-based synthesis leaves appearance uncontrolled.The paper identifies co-designing the user’s controls and internal representation as the gap its framework targets.Source: §2.2 · pp. 4–5

2.3 Controllable and Conditional Diffusion

• Standard conditional-diffusion mechanisms specify how signals enter a generator, but not what the signal should represent or how it should relate to the latent space.The paper reports that feeding redundant raw measurements directly to a FiLM-conditioned diffusion model yields unstable generation and unreliable per-axis control.Source: §2.3 · p. 5

2.4 Disentangled and Structured Latent Spaces

• Unlike statistical disentanglement, the paper defines directions from cell measurements and preserves their natural correlations because forcing independence could erase real patterns.The goal is to match real cell variation rather than require strictly isolated changes.Source: §2.4 · p. 5

2.5 Biological Descriptors in Computational Cell Biology

• Cell biology routinely uses measurements of cell appearance, and prior work found that high-dimensional measurements can be summarized into a few meaningful patterns.This paper uses such measurements as supervision and conditioning signals for image generation.Source: §2.5 · pp. 5–6

3 Controllable Single-Cell Latent Diffusion Framework

• The framework aligns semantic-axis controls and BioVAE’s latent representation around shared biological directions to support stable, interpretable changes in generated cells.The semantic axes condition a latent diffusion model, while BioVAE aligns its latent with the same measurements.Source: §3 · p. 6

3.1 Single-Cell Extraction and Feature Aggregation

• The pipeline extracts 64 × 64 LIVECell crops and computes 61 measurements, aggregated into 15 named, measurable, traversable semantic axes.The axes cover morphology, pixel intensity, texture, and spatial distribution; they remain partially correlated to preserve real patterns in cell appearance.Source: §3.1 · pp. 6–7

3.2 BioVAE: Biologically Regularized VAE

• BioVAE trains a predictor from latent codes to the full 61-dimensional measurement vector, while an entropy-based dispersion term counters collapse and the VAE KL term bounds latent spread.The full vector retains fine-grained measurement structure; the 15 axes are reserved as a compact control surface.Source: §3.2 · pp. 7–8

3.3 Biological Conditioning for Latent Diffusion Models

• Latent diffusion uses a fixed BioVAE encoder and conditions denoising on either the full 61-dimensional measurement vector or the 15 semantic axes.Because the latent and axes share measurement structure, traversing an axis at conditioning time produces a consistent, interpretable change.Source: §3.3 · p. 8

4.1 Experimental Setup

• LIVECell provided 64×64 grayscale cell crops, partitioned into 928k training, 166k validation, and 411k test images; evaluation covered reconstruction, latent structure, and generation metrics.Reconstruction metrics were MSE, PSNR, SSIM, and rFID; latent metrics were k-NN preservation and cell-type SVM accuracy, while generation metrics included FID, MMD, Precision, and Recall.Source: §4.1 · pp. 8–9, §4 · p. 8

4.2 Implementation Details

• BioVAE used a Stable Diffusion VAE architecture and biological and entropy loss weights; the conditional LDM used FiLM conditioning and DDPM sampling.Both models were trained for 500k iterations with batch size 64; crops were feathered by 2 pixels to discourage mask-boundary shortcuts, with no generated boundary artifacts observed.Source: §4.1 · p. 9, §4.2 · p. 9

4.3 Evaluation of BioVAE

• BioVAE improved reconstruction (MSE 0.60 vs 0.81; PSNR 49.8 vs 46.6), while entropy slightly raised Morph. MSE (6.82 vs 6.62) but supported improved latent structure.Neighborhood preservation increased across k, and linear classification accuracy increased by ∼1% (p < 0.001); absolute neighborhood values remained small on the evaluation set of over 160k cells.Source: §4.3 · pp. 9–10, Table 1 · p. 10, Table 2 · p. 10

4.4 Generated Quality

• Semantic-axis conditioning approximately halved FID (0.93–0.97 vs 2.04–2.12), and BioVAE + semantic axes achieved the best performance (FID 0.93, MSE 0.925).Figure 3 shows more faithful morphological preservation with BioVAE LDM than VAE LDM, alongside challenging examples with low organelle preservation.Source: §4.4 · p. 10, Table 3 · p. 11, Figure 3 · p. 11

4.5 Blind Study with Biologists

• Experts' real/synthetic discrimination was 50.9% (AUC = 0.509; p = 0.58), and realism-score distributions did not differ significantly between real and synthetic cells (p = 0.41).Only 20.2% of synthetic cells were correctly identified as synthetic, compared with 81.7% of real cells correctly classified as real.Source: §4.5 · p. 11, Figure 4 · p. 12

4.6 Controllability via Axis Traversal

• Size, shape complexity, and symmetry had strong target control; elongation, irregularity, and contrast had lower gain, while compactness and granularity were weakly controllable.The authors interpret traversals as semantic modulation within a correlated phenotype manifold, not fully disentangled control; correlated descriptors also complicated isolated manipulation.Source: §4.6 · pp. 12–13, Figure 5 · p. 13

The unit of generation and its scope

• The framework does not model neighboring-cell interactions or larger tissue context, and evaluation is limited to LIVECell phase-contrast images from eight cell lines.The authors identify fluorescence, brightfield, 3D imaging, and multicellular scenes as directions for future work.Source: §5 · p. 13

Scope of the semantic axes

• The 15 semantic axes do not explicitly represent fine-grained organelle structure, sub-cellular spatial patterns, or biological state transitions.Their coverage also depends on the standard measurements from which the axes were manually constructed.Source: §5 · p. 14

Why the input and latent representation should be designed together

• BioVAE does not improve generation with the full 61-dimensional measurement vector; its benefit emerges when conditioning also uses the 15 semantic axes.The authors therefore describe controllability as a property of the overall pipeline rather than a single component.Source: §5 · p. 14

External comparison

• Evaluation relies mainly on internal ablations rather than direct comparisons with external single-cell generators because compatible LIVECell metrics, feature extractors, and protocols are unavailable.The authors identify a common benchmark for controllable single-cell generation as a community-level direction.Source: §5 · p. 14

Conclusion

• On LIVECell, co-designing semantic-axis conditioning with BioVAE substantially improves generation quality over conditioning on raw high-dimensional measurements.BioVAE aligns its latent space with the same biologically meaningful directions as the named, measurable, and traversable semantic axes.Source: §6 · p. 14
• In a blind study, biologists could not reliably distinguish generated single cells from real ones.The conclusion reports this as evidence of generated-cell realism.Source: §6 · p. 14
• Traversal experiments verify stable per-axis control over phenotypic changes in generated cells.The paper presents controllable single-cell generation as complementary to whole-image approaches.Source: §6 · p. 14

A Morphodynamic descriptors

• The morphodynamic descriptors form a 61-dimensional set that the appendix categorizes and defines.The supplied passage does not specify the categories or individual descriptors.Source: A Morphodynamic descriptors · p. 18

A.1 Morphology

• Morphology descriptors include area, perimeter, equivalent diameter, fitted-ellipse axes, shape measures, and Zernike moments for coarse global shape.The listed measures also include convex area, extent, solidity, and circularity.Source: §A.1 · pp. 18–19

A.2 Pixel Intensity

• Pixel-intensity descriptors summarize the distribution and amount of signal within the segmented object, including mean, variability, minimum, sum, median, and percentiles.The listed percentiles are the 5th, 25th, 75th, and 95th.Source: §A.2 · p. 20

A.3 Texture

• Texture descriptors include Haralick features from gray-level co-occurrences, a 10-bin local-binary-pattern histogram, and statistics of image gradients.The gradient mean and standard deviation capture local edge strength and texture variation.Source: §A.3 · pp. 21–22

A.4 Spatial distribution

• Spatial-distribution descriptors include a six-bin radial intensity profile, computed as mean intensity within concentric bins from object center toward boundary.The bins are formed by uniformly dividing the radial interval within the object mask.Source: §A.4 · p. 22

B Semantic axes; B.1 Morphology

• The 15 semantic axes aggregate normalized measurements into interpretable controls, including size, elongation, compactness, boundary irregularity, and global shape complexity.The morphology axes describe overall size, directional stretch, closeness to a circular compact shape, outline smoothness, and shape complexity beyond simple geometry.Source: B Semantic axes · p. 23, §B.1 · p. 23

B.2 Pixel Intensity

• Pixel-intensity semantic axes describe overall signal level, within-cell intensity variation, and whether signal is dominated by hotspots or evenly distributed.The paper names these axes overall signal level, intensity contrast, and intensity skewness.Source: §B.2 · p. 23

B.3 Texture; 3. Granularity: Finds fine-scale repeating patterns

• Texture semantic axes describe intracellular heterogeneity and smoothness, with edge strength listed alongside fine-scale repeating patterns under granularity.The axis descriptions characterize texture heterogeneity, texture smoothness, and internal boundaries and edges.Source: §B.3 · p. 24, §3 · p. 24

B.4 Spatial Distribution

• The spatial-distribution semantic axis, central enrichment, indicates whether signal is concentrated toward the cell center or edges.Source: §B.4 · p. 24

B.5 Spearman Correlation

• The matrix reports Spearman correlations among semantic axes in the training split.Source: Figure S1 · p. 25
• Several semantic axes exhibit structured biological dependencies.Source: Figure S1 · p. 25
• The passage attributes these dependencies to shared descriptors and correlated phenotype properties.Source: Figure S1 · p. 25

Semantic axes traversal

• Target controllability is measured by the Spearman correlation between requested and generated semantic values.Source: Table S1 · p. 26
• Control gain is estimated as the slope of a linear regression between requested and generated semantic values.Source: Table S1 · p. 26
• Selectivity is the target-axis displacement divided by the average displacement of non-target axes, measuring relative dominance of the requested response.Source: Table S1 · p. 26

D Blind Study Details

• Six biologists evaluated 960 cell instances across eight cell types in the blind study.Source: D Blind Study Details · p. 26
• Study participation was voluntary, and participants were informed that responses would be used for research purposes.Source: D Blind Study Details · p. 26
• Responses were analyzed only in aggregate, with no personally identifiable information collected.Source: D Blind Study Details · p. 26

Semantic axes traversal

• Traversal plots show measured semantic-axis responses alongside generated cell images for six validation anchor cells.Source: Figure S2 · p. 27
• Dashed red lines mark each original anchor cell's semantic value.Source: Figure S2 · p. 27