Baysor: cell segmentation in imaging based spatial transcriptomics

Based on in situ sequencing or multiplex sequencing RNA The single molecule space transcriptomics scheme of fluorescence hybridization can reveal the detailed tissue structure . However , Distinguishing the boundaries of individual cells in these data is challenging , And may hinder downstream analysis . Current methods usually use nuclear staining to approximate the location of cells . Based on this , A research team from the United States developed a segmentation method ：Baysor. It can be segmented using only molecular location data or combined with evidence of auxiliary staining , So as to improve the segmentation quality , Increase the number of cells and segmented molecules . Relevant research results have been reported in 《Nature Biotechnology》 publish .

Baysor What is it? ？

Many analyses in spatial transcriptomics can be expressed as label assignment problems . for example , Cell segmentation is the assignment of cell tags to observed molecules . The separation of intercellular background is a process that marks molecules as “ The signal ” And “ background ” The problem of . The salient features of these problems are , Labels often show strong spatial aggregation . From a mathematical point of view , This spatial clustering trend can use Markov random fields on simple subdivision graphs （MRF） Preset to snap , Different label problems can be solved by selecting appropriate label probability model and observable data .

Baysor It's based on MRF Segmentation algorithm , It takes into account the joint possibility of transcriptional composition and cell morphology , Optimized two-dimensional （2D） Or three-dimensional （3D） Cell boundaries . It not only considers the segmentation based on CO dyeing , It can also be segmented separately according to the detected transcripts .

Baysor It can be used to analyze data from various experimental schemes （ Pictured above ）, And cell segmentation can be carried out by using molecular location alone or by combining additional information . This method models each cell as a distribution , Combine the spatial location and genetic characteristics of each molecule . therefore , The entire data set is considered to be a mixture of the specific distribution of such cells . then ,Baysor Using Bayesian hybrid model (BMM) To separate the mixture . Optimization depends on MRF A priori prediction to ensure the spatial separability of cells and encode additional information about molecular spatial relationships

Baysor Performance evaluation of

To evaluate performance , The researchers expanded MERFISH, Immunostaining for incorporation into cell boundaries . Use this and other benchmarks , Studies have shown that Baysor In some cases, segmentation can nearly double the number of cells than existing tools , At the same time, reduce the artifacts of segmentation . meanwhile , The researchers also proved that Baysor It performs well on the data obtained using five different schemes , Make it a powerful and universal tool for analyzing imaging based spatial transcriptomics .

Baysor And other segmentation methods in generating data sets using five different schemes ： When checking summary statistics , Find out Baysor The reported cells contain the same number and area of molecules as originally published （" The paper "） The segmentation results are roughly the same ; Compared with other segmentation methods ,Baysor More cell numbers and higher molecular proportions were reported .

The researchers developed a pan cellular cell surface marker Na+/K+-ATPase Immunofluorescence of (IF) And MERFISH A combined solution , This protocol was used in the mouse small intestine to provide an additional baseline data set with defined cell boundaries . The model dataset will be Baysor Extending to dense and complex tissue types provides a favorable setting . The research team quantified Baysor Segmentation and other methods of segmentation and IF The degree of consistency of membrane signals , Overall speaking Baysor Better than other methods .

Baysor The core of depends on MRF The general method of , This method can be used to solve other labeling problems on spatial data , Such as background molecular separation or clustering . Even though Baysor The algorithm performs well in most published schemes , But still Some potential improvements can be introduced , Such as improving the modeling of cell shape . By extending the hierarchical Bayesian model to introduce the shape and composition characteristics of cell types , It is best to incorporate a clear model of cell type specific transcript separation structures , Further improvements can be made .

Auxiliary dyeing is very valuable in solving difficult cases . therefore , Optimal segmentation may depend on the combination of transcriptional component signals and information from auxiliary dyes . because Baysor You can use an uncertain a priori prediction , Probabilistic auxiliary image segmentation method will provide an advantage in this regard .

Baysor The software package can be through Github obtain ：https://github.com/kharchenkolab/Baysor

The code to reproduce the result is shown in the following link ：https://github.com/kharchenkolab/ BaysorAnalysis/

MERFISH Probe design and analysis software is available at the following link ：https://github.com/ZhuangLab/ MERFISH_analysis

First public number ： National Gene Bank big data platform  

reference

Petukhov V, Xu R J, Soldatov R A, et al. Cell segmentation in imaging-based spatial transcriptomics[J]. Nature Biotechnology, 2021: 1-10.

Image from Nature Biotechnology Official website and references , If there is infringement, please contact to delete .