• bayesVG
• Installation
• Usage
  • Libraries
  • HVG detection
  • SVG detection
• Contact information

bayesVG

Installation

You can install the most recent version of bayesVG using:

remotes::install_github("jr-leary7/bayesVG")

Usage

Libraries

library(Seurat)
library(bayesVG)

HVG detection

Data

First, we load the 10X Genomics pbmc3k dataset, which is composed of 2,700 peripheral blood mononuclear cells from a single healthy donor.

data("seu_pbmc")

Modeling

Now we’re able to model gene expression, summarize the posterior distribution of variance for each gene, and classify the top 3000 most-variable genes as HVGs. The findVariableFeaturesBayes() function can take as input either a Seurat or a SingleCellExperiment object.

seu_pbmc <- findVariableFeaturesBayes(seu_pbmc, 
                                      n.cells.subsample = 500L, 
                                      algorithm = "meanfield",
                                      save.model = TRUE) %>% 
            classifyHVGs(n.HVG = 3000L)

We can extract the summary table (which is sorted by default) and classify the top 3,000 genes as HVGs like so. These genes can then be used as the basis for downstream analyses such as PCA, clustering, UMAP visualization, etc.

summary_hvg <- getBayesianGeneStats(seu_pbmc)
top3k_hvgs <- summary_hvg$gene[1:3000]

SVG detection

Data

First, we load the 10X Genomics anterior mouse brain dataset.

data("seu_brain")

Before running bayesVG for SVG detection it’s necessary to normalize the expression data and identify a set of naive HVGs.

seu_brain <- NormalizeData(seu_brain, verbose = FALSE) %>% 
             FindVariableFeatures(nfeatures = 3000L, verbose = FALSE)

Modeling

Now we can model gene expression with an approximate multivariate hierarchical Gaussian process (GP), summarize the spatial component of variance for each gene, and classify the top 1000 most spatially variable genes as SVGs. The findSpatiallyVariableFeaturesBayes() function can take as input either a Seurat or a SpatialExperiment object.

seu_brain <- findSpatiallyVariableFeaturesBayes(seu_brain, 
                                                naive.hvgs = VariableFeatures(seu_brain), 
                                                kernel = "matern", 
                                                kernel.smoothness = 1.5, 
                                                algorithm = "meanfield", 
                                                n.cores = 4L, 
                                                save.model = TRUE) %>% 
             classifySVGs(n.SVG = 1000L)

We can extract the summary table (which, like the HVG summary table is sorted by default) and classify the top 1,000 genes as SVGs like so. These genes can then be used as the basis for downstream analyses such as PCA, clustering, UMAP visualization, etc.

summary_svg <- getBayesianGeneStats(seu_brain)
top1k_svgs <- summary_svg$gene[1:1000]

We can cluster the SVG set (using a Bayesian Gaussian mixture model) into spatial modules as shown below. The clustering function returns a PCA embedding of the SVGs, a table of the soft cluster assignment probabilities, and the log-likelihood and Bayesian information criterion (BIC) of the clustering.

svg_clusters <- clusterSVGsBayes(seu_brain, 
                                 svgs = top1k_svgs, 
                                 n.clust = 5L)

Lastly, we can score the SVG clusters using UCell under the hood. These scores can then be visualized using e.g., violin plots or UMAPs.

seu_brain <- scoreSpatialModules(seu_brain, svg.clusters = svg_clusters)

Contact information

This package is developed & maintained by Jack R. Leary. Feel free to reach out by opening an issue or by email (j.leary@ufl.edu) if more detailed assistance is needed.