Insitutype: likelihood-based cell typing for single cell spatial transcriptomics
Patrick J. DanaheriD, Edward J. Zhao, Zhi Gang YangiD, David RossiD, Mark GregoryiD, Zach Reitz, Tae Kyoung KimiD, Sarah K. BaxteriD, Shaun W. JacksoniD, Shanshan HeiD, Dave Henderson, Joseph BeechemiD
Abstract
Accurate cell typing is fundamental to analysis of spatial single-cell transcriptomics, but legacy scRNA-seq algorithms can underperform in this new type of data. We have developed a cell typing algorithm, Insitutype, designed for statistical and computational efficiency in spatial transcriptomics data. Insitutype is based on a likelihood model that weighs the evidence from every expression value, extracting all the information available in each cells expression profile. This likelihood model underlies a Bayes classifier for supervised cell typing, and an Expectation-Maximization algorithm for unsupervised and semi-supervised clustering. Insitutype also leverages alternative data types collected in spatial studies, such as cell images and spatial context, by using them to inform prior probabilities of cell type calls. We demonstrate rapid clustering of millions of cells and accurate fine-grained cell typing of kidney and non-small cell lung cancer samples.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 66
- Organism
- Assay
- Research Focus
- Approach
Abstract summary
Claude Opus 5.5 · xhigh effort · 18 Oct 2026 · from the abstract
Insitutype types cells in spatial transcriptomics with a likelihood model over every expression value, admits image and neighbourhood priors, and is reported to cluster millions of cells quickly.
Cell typing tools carried over from scRNA-seq are said to lose accuracy on single-cell spatial data. Insitutype instead builds a likelihood over every measured expression value and lets evidence outside the expression profile, such as cell images and spatial context, set the prior on each call.
Key results
ClaimedOne likelihood model over all expression values supports three modes of use: a Bayes classifier for supervised typing, and an expectation-maximization procedure for unsupervised and semi-supervised clustering.
Described as the design of the method; the abstract reports no likelihood formulation details, convergence behaviour or comparison with alternative models. · Abstract
ClaimedData collected alongside the transcripts, specifically cell images and spatial context, are folded in as prior probabilities on cell type calls rather than being discarded.
Stated as a capability of the method; no measurement of how much the priors change calls is given in the abstract. · Abstract
SuggestedThe authors report fast clustering at a scale of millions of cells and fine-grained typing in kidney and non-small cell lung cancer samples.
Demonstrations on kidney and non-small cell lung cancer tissue; the abstract gives no runtime, hardware, accuracy metric or baseline method for comparison. · Abstract
Methods, data and reproducibility
A likelihood model that scores every expression value in a cell's profile, used both as a Bayes classifier against reference types and inside an expectation-maximization loop for unsupervised and semi-supervised clustering, with cell images and spatial context entering as priors; demonstrated on kidney and non-small cell lung cancer spatial transcriptomics data.
- Kidney single-cell spatial transcriptomics samples
- Non-small cell lung cancer single-cell spatial transcriptomics samples
- Clustering demonstration at large scale · n = millions of cells
- Methods & tools
- Insitutype
Limitations (2)
- Accuracy is asserted rather than benchmarked: the abstract names no competing method, no accuracy or concordance metric and no reference annotation against which the calls were judged. · noted in this summary
- The speed claim comes without runtimes, memory figures or hardware, and the evaluation covers only two tissue settings, kidney and non-small cell lung cancer. · noted in this summary
Who should read:Read this if you assign cell types in single-cell spatial transcriptomics and want a likelihood framework that also runs unsupervised and can take image or neighbourhood priors.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Spatial mapping of cell types by integration of transcriptomics data
Alma Andersson, Joseph Bergenstråhle, … Joakim Lundeberg·bioRxiv·Dec 13, 2019
- Reference-based cell type matching of spatial transcriptomics data
Yun Zhang, Jeremy A. Miller, … Richard H. Scheuermann·bioRxiv·Mar 28, 2022
- Cell Type Assignments for Spatial Transcriptomics Data
Haotian Teng, Ye Yuan, Ziv Bar‐Joseph·bioRxiv·Feb 25, 2021
- SPICEMIX: Integrative single-cell spatial modeling of cell identity
Benjamin Chidester, Tianming Zhou, … Jian Ma·bioRxiv·Nov 29, 2020
- Robust decomposition of cell type mixtures in spatial transcriptomics
Dylan M. Cable, Evan M. Murray, … Rafael A. Irizarry·Nature BiotechnologyPeer-reviewed·Feb 18, 2021
- SpatioCell: Deep Integration of Histology and Spatial Transcriptomics for Profiling the Cellular Microenvironment at Single-Cell Level
Naiqiao Hou, Yue Yu, … Jia Song·中国科学通报：英文版Peer-reviewed·Jun 3, 2026