SpaTopic_inference {SpaTopic}

'SpaTopic': fast topic inference to identify tissue architecture in multiplexed images

Description

This is the main function of 'SpaTopic', implementing a Collapsed Gibbs Sampling algorithm to learn topics, which referred to different tissue microenvironments, across multiple multiplexed tissue images. The function takes cell labels and coordinates on tissue images as input, and returns the inferred topic labels for every cell, as well as topic contents, a distribution over celltypes. The function recovers spatial tissue architectures across images, as well as indicating cell-cell interactions in each domain.

Usage

SpaTopic_inference(
  tissue,
  ntopics,
  sigma = 50,
  region_radius = 400,
  kneigh = 5,
  npoints_selected = 1,
  ini_LDA = TRUE,
  ninit = 10,
  niter_init = 100,
  beta = 0.05,
  alpha = 0.01,
  trace = FALSE,
  seed = 123,
  thin = 20,
  burnin = 1000,
  niter = 200,
  display_progress = TRUE,
  do.parallel = FALSE,
  n.cores = 1,
  axis = "2D"
)

Arguments

tissue

(Required). A data frame or a list of data frames. One for each image. Each row represent a cell with its image ID, X, Y coordinates on the image, celltype, with column names (image, X, Y, type), respectively. You may add another column Y2 for 3D tissue image.

ntopics

(Required). Number of topics. Topics will be obtained as distributions of cell types.

sigma

Default is 50. The lengthscale of the Nearest-neighbor Exponential Kernel. Sigma controls the strength of decay of correlation with distance in the kernel function. Please check the paper for more information. Need to be adjusted based on the image resolution

region_radius

Default is 400. The radius for each grid square when sampling region centers for each image. Need to be adjusted based on the image resolution and pattern complexity.

kneigh

Default is 5. Only consider the top 5 closest region centers for each cell.

npoints_selected

Default is 1. Number of points sampled for each grid square when sampling region centers for each image. Used with region_radius.

ini_LDA

Default is TRUE. Use warm start strategy for initialization and choose the best one to continue. If 0, it simply uses the first initialization.

ninit

Default is 10. Number of initialization. Only retain the initialization with the highest log likelihood (perplexity).

niter_init

Default is 100. Warm start with 100 iterations in the Gibbs sampling during initialization.

beta

Default is 0.05. A hyperparameter to control the sparsity of topic content (topic-celltype) matrix Beta. A smaller value introduces more sparse in Beta.

alpha

Default is 0.01. A hyperparameter to control the sparsity of document (region) content (region-topic) matrix Theta. For our application, we keep it very small for the sparsity in Theta.

trace

Default is FALSE. Compute and save log likelihood, Ndk, Nwk for every posterior samples. Useful when you want to use DIC to select number of topics, but it is time consuming to compute the likelihood for every posterior samples.

seed

Default is 123. Random seed.

thin

Default is 20. Key parameter in Gibbs sampling. Collect a posterior sample for every thin=20 iterations.

burnin

Default is 1000. Key parameter in Gibbs sampling. Start to collect posterior samples after 1000 iterations. You may increase the number of iterations for burn-in for highly complex tissue images.

niter

Default is 200. Key parameter in Gibbs sampling. Number of posterior samples collected for model inference.

display_progress

Default is TRUE. Display the progress bar.

do.parallel

Default is FALSE. Use parallel computing through R package foreach.

n.cores

Default is 1. Number of cores used in parallel computing.

axis

Default is "2D". You may switch to "3D" for 3D tissue images. However, the model inference for 3D tissue is still under test.

Value

Return a gibbs.res-class object. A list of outputs from Gibbs sampling.

See Also

gibbs.res-class

Examples

## tissue is a data frame containing cellular information from one image or
## multiple data frames from multiple images.

data("lung5")
## NOT RUN, it takes about 90s
library(sf)
#gibbs.res<-SpaTopic_inference(lung5, ntopics = 7,
#                               sigma = 50, region_radius = 400)
                             
                              
## generate a fake image 2 and make an example for multiple images
## NOT RUN
#lung6<-lung5
#lung6$image<-"image2"  ## The image ID of two images should be different
#gibbs.res<-SpaTopic_inference(list(A = lung5, B = lung6), 
#                 ntopics = 7, sigma = 50, region_radius = 400)