Spatial transcriptomics skips cell segmentation first: SSAF-T types cells from a literature knowledge graph, then uses that map to guide segmentation

Cell-type identification in spatial transcriptomics can proceed directly from spatial gene expression without cell segmentation as a prerequisite, and the same pipeline yields cell boundaries usable by existing downstream tools. Earlier segmentation-free methods either depend on matched external single-cell reference data or require clustering or a prior estimate of the number of cell types, which itself still demands cell typing; they also produce no "cells", making them hard to plug into graph analyses that treat each cell as a node. Six public spatial transcriptomics datasets were compared: Visium HD mouse small intestine and Stereo-seq mouse brain against published ENACT and CellBin results, and Xenium mouse kidney and MERFISH mouse liver against published Baysor and Cellpose segmentation, with cell-type distributions consistent with spatial marker-gene expression.

Cell-type labels come from the agent CeLLMan, which selects a knowledge-graph module by species and organ and descends from broad cell classes to specific cell types. CeLLMan judges against a knowledge graph of cell types, marker genes and hierarchy curated from PubMed literature rather than relying on the language model's own memory; in each round three independent agents evaluate marker-gene evidence and a consensus decides whether to refine further. On the Tabula Sapiens v2.0 and Tabula Muris Senis single-cell atlases, CeLLMan's default annotations were compared cell by cell with published atlas annotations using marker-gene expression as the criterion; across all organs, marker-gene evidence supported CeLLMan for 45.0% and 49.1% of cells, versus 2.8% and 7.1% supporting the published annotations.

The segmentation-free cell-type map is fed back as a prior that constrains nucleus-seeded watershed segmentation, producing individual cells compatible with the existing analysis ecosystem. Conventional segmentation relies only on stains such as DAPI, Poly-A or H&E, which do not themselves provide boundaries between different cell types; SSAF-T lets cell-type boundaries limit where segmentation can go. On Xenium mouse kidney and MERFISH mouse liver, SSAF-T and the published segmentations shared most transcripts, with median transcript containment of 0.82 and 0.75 and median Jaccard indices of 0.45 and 0.21; SSAF-T assigned 95.1% of transcripts in Xenium and 81.9% in MERFISH versus 89.0% and 52.9% for the published segmentations, and completed both 2D and 3D segmentation faster than Baysor even on a CPU.

The demonstrated scope is public data with published annotations to compare against: Visium HD mouse small intestine, Stereo-seq mouse brain, Xenium mouse kidney, MERFISH mouse liver, and Xenium human placenta and human heart; two in-house Xenium datasets serve as case studies, wounded mouse skin at post-wounding day 15 and human lung adenocarcinoma. Segmentation uniformly uses a nucleus-seeded watershed constrained by the cell-type map; the authors note there is no universally accepted gold standard for cell segmentation and plan to integrate better nuclei segmentation or segmentation that needs no nuclei.

CeLLMan's performance depends on the completeness and accuracy of the knowledge graph, which the authors plan to update and expand regularly through an automated process with human-on-the-loop validation. In the human lung adenocarcinoma analysis all cells came from a single patient, so the P values reflect within-sample differences rather than interpatient variation; the split between tissue-resident and monocyte-derived macrophages rests on transcriptional annotation rather than lineage tracing, and the reference atlas lacks a distinct interstitial macrophage label. The detailed relation between Satb2, Wnt pathway mediators and hair follicle cell types in mouse skin was not quantified or experimentally validated.