Baysor

Bayesian Segmentation of Spatial Transcriptomics Data

• Abstract
  • Method description
• Installation
  • Install as a Julia package
  • Build CLI application from source
  • Docker
• Run
  • Dataset preview
  • Full run
    • Normal run
    • Using a prior segmentation
    • Outputs
    • Choice of parameters
    • Multi-threading

Abstract

Spatial transcriptomics is an emerging stack of technologies, which adds spatial dimension to conventional single-cell RNA-sequencing. New protocols, based on in situ sequencing or multiplexed RNA fluorescent in situ hybridization register positions of single molecules in fixed tissue slices. Analysis of such data at the level of individual cells, however, requires accurate identification of cell boundaries. While many existing methods are able to approximate cell center positions using nuclei stains, current protocols do not report robust signal on the cell membranes, making accurate cell segmentation a key barrier for downstream analysis and interpretation of the data. To address this challenge, we developed a tool for Bayesian Segmentation of Spatial Transcriptomics Data (Baysor), which optimizes segmentation considering the likelihood of transcriptional composition, size and shape of the cell. The Bayesian approach can take into account nuclear or cytoplasm staining, however can also perform segmentation based on the detected transcripts alone. We show that Baysor segmentation can in some cases nearly double the number of the identified cells, while reducing contamination. Importantly, we demonstrate that Baysor performs well on data acquired using five different spatially-resolved protocols, making it a useful general tool for analysis of high-resolution spatial data.

Method description

See our pre-print for details on the method.

Installation

Install as a Julia package

To install Baysor as a Julia package run the following code:

using Pkg
Pkg.add([PackageSpec(name="UMAP", rev="master"), PackageSpec(url="https://github.com/kharchenkolab/Baysor.git")])
import Baysor

After that you can create Baysor executable by running

echo "#! /usr/bin/env julia\nimport Baysor: run_cli\nrun_cli()" >> baysor && chmod +x baysor

Build CLI application from source

On linux, you can build the command-line tool from source. To do that you need to clone this package and run the Makefile:

git clone https://github.com/kharchenkolab/Baysor.git
cd Baysor/bin
make

The command takes ~10 minutes to complete. It creates executable file ./Baysor/bin/baysor, which can be used for spatial data segmentation.

Docker

Alternatively, you can use Docker. It contains executable baysor to run Baysor from CLI, as well as IJulia installation to use Baysor with Jupyter. The repo also has images for older versions.

docker run -it --rm vpetukhov/baysor:latest

Build by hands:

docker pull julia:latest
cd Baysor/docker
docker build .

You can find more info about dockers at Docker Cheat Sheet.

Run

NOTE: The tool is in the beta-version now, so some parameters and functionality can be changed in the future. Please, report any problems and suggestions to Issues.

Dataset preview

As full run takes some time, it can be useful to run a quick preview to get meaning from the data and to get some guesses about parameters of the full run.

baysor preview [-x X_COL -y Y_COL --gene GENE_COL -c config.toml -o OUTPUT_PATH] MOLECULES_CSV

Here:

• MOLECULES_CSV must contain information about x and y positions and gene assignment for each transcript
• Parameters X_COL, Y_COL and GENE_COL must specify the corresponding column names. Default values are "x", "y" and "gene" correspondingly
• OUTPUT_PATH determines path to the output html file with the diagnostic plots

Full run

Normal run

To run the algorithm on your data, use

baysor run [-s SCALE -x X_COL -y Y_COL --gene GENE_COL] -c config.toml MOLECULES_CSV [PRIOR_SEGMENTATION]

Here:

• MOLECULES_CSV must contain information about x and y positions and gene assignment for each transcript
• PRIOR_SEGMENTATION is optional molecule segmentation obtrained from other method (see Using prior segmentation)
• Parameters X_COL, Y_COL and GENE_COL must specify the corresponding column names. Default values are "x", "y" and "gene" correspondingly.
• SCALE is a crutial parameter and must be approximately equal to the expected cell radius in the same units as "x" and "y". In general, it's around 5-10 micrometers, and the preview run can be helpful to determine it for a specific dataset (by eyes, for now).

To see full list of command-line options run

baysor run --help

For more info see examples (though probably are out-of-date).

For the description of all config parameters, see example_config.toml.

Using a prior segmentation

In some cases, you may want to use another segmentation as a prior for Baysor. The most popular case is having a segmentation based on DAPI/poly-A stainings: such information helps to understand where nucleis are positioned, but it's often quite imprecise. To take this segmentation into account you can pass it as the second positional arguments to Baysor:

baysor run [ARGS] MOLECULES_CSV [PRIOR_SEGMENTATION]

Here, PRIOR_SEGMENTATION can be a path to a binary image with segmentation mask, an image with integer cell segmentation labels or a column name in the MOLECULES_CSV with integer cell assignment per molecule. In the later case, column name must have : prefix, e.g. for column cell you must use baysor run [ARGS] molecules.csv :cell. In case the image is too big to be stored in the tiff format, Baysor supports MATLAB '.mat' format: it should contain a single field with an integer matrix for either a binary mask or segmentation labels.

To specify expected quality of the prior segmentation you may use prior-segmentation-confidence parameter. The value 0.0 makes the algorithm ignore the prior, while the value 1.0 restricts the algorithm from contradicting the prior. Prior segmentation is mainly needed for the cases where gene expression signal is not enough, e.g. with very sparse protocols (such as ISS or DARTFISH). Another potential usecase is high-quality data with visible sub-cellular structure. In these situations, setting prior-segmentation-confidence > 0.7 is recommended. Otherwise, the default value 0.2 should work well.

If you have a non-segmented DAPI image, the simplest way to segment it would go through the following steps ImageJ:

1. Open the image (File -> Open)
2. Go to Image -> Type and pick "8-bit"
3. Run Process -> Filters -> Gaussian Blur, using Sigma = 1.0. The value can vary, depending on your DAPI, but 1.0 generally works fine.
4. Run Image -> Adjust -> Auto Threshold, using Method = Default. Different methods can give the best results for different cases. Often "Mean" also works well.
5. Run Process -> Binary -> Watershed
6. Save the resulting image in the .tif

Outputs

• segmentation_borders.html: visualization of cell borders for the dataset colored by local gene expression composition (first part) and molecule clusters (second part)
• segmentation_cell_stats.csv: diagnostic info about cells. The following parameters can be used to filter low-quality cells:
  • area: area of the convex hull around the cell molecules
  • avg_confidence: average confidence of the cell molecules
  • density: cell area divided by the number of molecules in cell
  • elongation: ration of the two eigenvalues of the cell covariance matrix
  • n_transcripts: number of transcripts per cell
• segmentation_config.toml: copy of the config to improve reproducibility
• segmentation_params.dump: aggregated parameters from the config and CLI
• segmentation.csv: segmentation info per molecule:
  • confidence: probability of a molecule to be real (i.e. not noise)
  • cell: id of the assigned cell. Value "0" correspond to noise.
  • cluster: id of molecule cluster
  • assignment_confidence: confidence that the molecule is assigned to a correct cell
  • is_noise: shows whether molecule was assigned to noise (it's equal "true" if and only if "cell" == 0)

Choice of parameters

Most important parameters:

• scale is the most sensitive parameter, which specifies expected radius of a cell. It doesn't have to be precise, but wrong set-up can lead to over- or under-segmentation. This parameter is inferred automatically if cell centers are provided.
• min-molecules-per-cell is the number of molecules, required for a cell to be considered as real. It really depends on a protocol. For instance, for ISS it's fine to set it to 3, while for MERFISH it can require hundreds of molecule.

Some other sensitive parameters (normally, shouldn't be changed):

• new-component-weight is proportional to probability of generating new cell for a molecule, instead of assigning it to one of existing cells. More precisely, probability to assign a molecule to a particular cell linearly depends on number of molecules, already assigned to this cell. And this parameter is used as number of molecules for a cell, which is just generated for this new molecule. The algorithm is robust to small changes in this parameter. And normally values in range 0.1-0.9 should work fine. Smaller values would lead to slower convergence of the algorithm, while larger values force emergence of large number of small cells on each iteration, which can produce noise in the result. In general, default value should work well.

Run parameters:

• num-cells-init expected number of cells in data. This parameter influence only convergence speed of the algorithm. It's better to set larger values than smaller ones.
• iters number of iterations for the algorithm. At the moment, no convergence criteria is implemented, so it will work exactly iters iterations. Thus, to small values would lead to non-convergence of the algorithm, while larger ones would just increase working time. Optimal values can be estimated by the convergence plots, produced among the results.
• n-frames determines maximum parallelization level. To allow multi-threaded processing of the data, all the space is splitted in n_frames parts, which contain approximately equal number of molecules. Results can be a bit inaccurate on the borders of the frames, so it's better to avoid very large values.

Multi-threading

Currently, running multi-threaded version is only available by splitting dataset to independent frames, which creates artifacts on the frame borders. If you're fine with that, please use the command

JULIA_NUM_THREADS=10; baysor run -n $JULIA_NUM_THREADS [-s SCALE -x X_COL -y Y_COL --gene GENE_COL] -c config.toml MOLECULES_CSV [CENTERS_CSV]