CorSpatialDims: Find dims with high spatial autocorrelation

In jbergenstrahle/STUtility: Analysis and visualization of Spatial Transcriptomics data

Find dims with high spatial autocorrelation

Description

This function can be used to find dims with spatial structure in ST datasets. A more detailed decription of the algorithm is outlined in the Details section below.

Usage

CorSpatialDims(
  object,
  dims = NULL,
  reduction = "pca",
  nNeighbours = NULL,
  maxdist = NULL
)

Arguments

object	Seurat object
dims	dims to rank by spatial autocorrelation. If no dims are provided, the all dims will be used
reduction	dimensionality reduction object to pull data from
nNeighbours	Number of neighbours to find for each spot, For Visium data, this parameter is set to 6 because of the spots are arranged in a hexagonal pattern and should have maximum 6 neighbors.
maxdist	Maximum allowed distance to define neighbouring spots [default: 1.5]. If not provided, a maximum distance is automatically selected depending on the platform. For Visium data, this maximum distance is set to 150 microns.

Details

overview of method:

• Build a connection network from the array x,y coordinates for each sample. For a 'Visium' array, this would typically be 6 neighbours because of the hexagonal structure of spots.

• Combine connection networks from multiple samples

• Compute the lag vector for each feature

• Compute the correlation between the lag vector and the original vector

The connection network is build by defining edges between each spot and its 'nNeighborurs' closest neighbours that are within a maximum distance defined by 'maxdist'. This is to make sure that spots along the tissue edges or holes have the correct number of neighbours. A connection network is built for each section separately but they are then combined into one large connection network so that the autocorrelation can be computed for the whole dataset.

Now that we have a neighbour group defined for each spot, we can calculate the lag vector for each feature. The lag vector of a features is essentially the summed expression of that feature in the neighbour groups, computed for all spots and can be thought of as a "smoothing" estimate.

If we consider a spot A and its neighbours nbA, a feature with high spatial corelation should have similar expression levels in both groups. We can therefore compute the a correlation score between the lag vector and the "normal" expression vector to get an estimate of the spatial autocorrelation.

Value

data.frame with gene names and correlation scores

jbergenstrahle/STUtility documentation built on March 14, 2023, 7:15 a.m.