Skip to contents

Fit a latent Dirichlet allocation model

run_lda.Rd

Topic model over a documents x terms count matrix, fitted by variational Bayes. The intended input is a binary matrix: this is the model behind cisTopic, where binarised cells x regions scATAC becomes cells x topics for clustering and topics x regions for region set discovery. Nothing here is ATAC-specific, so binarised regulon activity from binarise_regulon_activity() works the same way, as does any count matrix.

Usage

run_lda(x, k, lda_params = params_lda(), seed = 42L, .verbose = TRUE)

Arguments

x

Documents x terms matrix. Either a dgCMatrix, a dgRMatrix, or a dense numeric or logical matrix. Column names are required, row names are generated if absent.

k

Integer. Number of topics.

lda_params

List, see params_lda().

seed

Integer. Seed for the variational initialisation.

.verbose

Boolean or integer. Verbosity.

Value

An LdaResult object.

Details

Binarising first is what makes a topic model the right tool. LDA on raw single cell counts is dominated by library size, which is precisely the effect cisTopic sidesteps.

Use lda_k_sweep() if you do not already know k.

References

Hoffman, Blei and Bach, NIPS, 2010; Bravo Gonzalez-Blas, et al., Nat Methods, 2019

Examples

# two planted term blocks recovered as two topics
set.seed(42L)
corpus <- matrix(rbinom(200L * 40L, 1L, 0.05), nrow = 200L, ncol = 40L)
corpus[1:100, 1:10] <- rbinom(1000L, 1L, 0.6)
corpus[101:200, 11:20] <- rbinom(1000L, 1L, 0.6)
colnames(corpus) <- sprintf("term_%02d", 1:40)
lda_res <- run_lda(corpus > 0, k = 2L, .verbose = FALSE)
lda_res
#> LdaResult (latent Dirichlet allocation)
#>   Documents:        200
#>   Terms:            40
#>   Topics:           2
#>   Bound (ELBO):     -5137.97
#>   Perplexity:       31.6666
#>   Iterations:       60
#>