AI edits cell-nucleus images while separating gene and shape signals

Why the slice matters

Spatial transcriptomics measures gene activity while retaining information about where cells sit in tissue. Some platforms also capture fluorescence images of cell nuclei, allowing researchers to pair gene-expression measurements with numerical descriptions of nuclear appearance, known as pathomics features.

The preprint’s authors highlight a complication: a tissue section is a two-dimensional slice through three-dimensional cells. A cut near a cell’s middle can expose more cytoplasm than one near its edge, potentially capturing more messenger RNA. Nuclear measurements such as area and perimeter also depend on the section being imaged.

The researchers argue that this shared geometry can create apparent relationships between gene expression and morphology that reflect tissue preparation rather than an underlying biological connection. Their study investigates how to edit images under separate gene and morphology controls despite these relationships.

Editing an existing image

The team applies an existing framework called Manifold-Aware Diffusion, or MAD. Its task differs from generating a cell image from scratch: it starts with an existing nucleus image and a target change in one feature group, aiming to preserve the other group.

MAD combines a conditional diffusion model, which produces images guided by numerical features, with a variational autoencoder, or VAE. The VAE learns patterns in the observed combinations of gene-expression and pathomics measurements.

At editing time, the requested feature combination is adjusted through this learned representation before guiding image production. The authors describe this step as a way to respect relationships within feature groups and reduce unintended interference between them. For example, changing a nuclear-area target should not automatically require changing the gene-expression settings.

These are computational image edits. They do not alter actual cells or establish what would happen biologically if a gene’s activity changed.

Testing on kidney tissue

The researchers used CosMx Spatial Molecular Imager data comprising 16 kidney-tissue sections from two samples: eight sections classified as diabetic nephropathy and eight as normal. They selected podocytes using positive expression of the WT1 marker.

Fourteen sections were used for training and two for testing, with one diabetic and one normal section in the test set. Each image was a 128-by-128-pixel grayscale patch centered on a nucleus, taken from the DAPI fluorescence channel.

The conditioning information comprised three genes—ACTN4, PODXL and VIM—and five image-derived features. All features were standardized using the training data.

Importantly, the observed relationships between the two groups were weak: cross-group Pearson correlations had absolute values no greater than 0.14. Some within-group relationships were much stronger, including a correlation of 0.95 between area and perimeter. The findings therefore concern maintaining separation between groups with weak measured linear associations, not correcting a demonstrated strong confound in this dataset.

Better control, with a trade-off

For pathomics editing, the researchers measured how closely features extracted from generated images matched their requested targets. MAD achieved a mean absolute error of 0.6636, compared with 0.9158 for CCDM, a closely matched model without the VAE, and 1.901 for an adapted Stable Diffusion model. Lower values indicate better target matching.

MAD also recorded the lowest Fréchet Inception Distance among the compared methods: 25.19, versus 31.15 for CCDM and 63.21 for Stable Diffusion. This metric compares generated and real image distributions; it does not establish biological correctness.

There was a trade-off. Stable Diffusion preserved more similarity to the starting image, scoring 0.8415 on the structural similarity metric compared with MAD’s 0.7964. The authors interpret its stronger preservation alongside poor target accuracy as evidence of insufficient editing.

What the evidence establishes

A synthetic experiment provided a more controlled test because independence between feature groups was built into the data. When editing area, MAD produced an average unintended change of 0.112 in the other group, compared with 0.271 for CCDM. A reverse-direction test also favored MAD. These results support the claimed contribution of the VAE under the simulated conditions.

The real-data gene-editing results were qualitative: different ACTN4 settings produced visually distinct MAD outputs, but the paper did not report an equivalent quantitative validation of those edits.

The evidence is thus strongest for controlled morphology editing and synthetic independence preservation. Testing was limited to podocytes from two kidney samples on one platform. The authors identify other cell types, additional measurement modalities and spatial-neighborhood information as future directions.