CellART segments and labels cells in one pass across four spatial transcriptomics platforms

Spatial transcriptomics now resolves detail smaller than a cell. Even so, the CellART paper on nature.com starts from a problem that resolution has left in place ^[14]. According to its abstract, current platforms either capture sparse transcript counts in each spot or measure only a limited number of genes, and both limits make comprehensive single-cell information hard to extract ^[1].

CellART's response is to pool the evidence. It is built as one framework meant to run across diverse high-resolution platforms ^[2], and it pairs deep learning with probabilistic modelling so that cell boundaries and cell types are estimated together ^[4][5]. In principle, that lets a candidate cell outline be judged partly by whether the transcripts inside it resemble a type in the reference. For a lab with an existing downstream stack, compatibility matters most: the authors say CellART's outputs work with widely used community tools, and they describe the method as "efficient, generalizable and robust" ^[6].

Two of the datasets are the kind of control I like to see. The lung and colorectal cancer samples were each measured on both Xenium and Visium HD ^[8]. With the same tissue on two technologies, agreement between the two sets of calls would suggest the method is reading the biology and not a platform artefact. The abstract does not report accuracy figures, runtimes or a comparison against separate segmentation and labelling tools.

Labelling depends on the reference. A reference matched to the sample is the easier case, and the paper uses both kinds. Breast and colorectal cancer came with matched single-cell references. Lung used a reference from the CZI CELLxGENE portal, and mouse brain an Allen Brain Atlas Smart-seq dataset of cortex and hippocampus ^[11][12].

Anyone who wants numbers can generate them. The processed inputs used by CellART are on Zenodo ^[9], and the Python code is public with documentation and worked examples ^[10]. Simulated data is the only setting in this design where true cell boundaries are known in advance. That makes the released simulation scripts the first thing a sceptical lab can rerun against its current pipeline.

I think the one-step approach is a plausible replacement for chained tools on Xenium and Visium HD, provided a decent reference exists for the tissue. The study spans four platforms from three vendors ^[1], but every human sample sits on 10x Genomics hardware ^[3]. For human tissue on MERFISH or Stereo-seq, the paper's evidence is mouse brain ^[2].