Argument `sparse = TRUE` does not work with `brms::lf`

Interfaces brms
1

Cross-posted from GitHub.

I’m writing an R package where a user passes an R formula specifying the experimental design of an RNAseq experiment, and then this formula is used to construct a Stan model (using brms) that is ultimately passed to rstan.

An example of this would be:

~ cell_type + (1 | donor:cell_type)

This internally is modified to include an interaction allowing for multiple genes to be modelled at once, producing:

~ (cell_type:feat_idx) + (1 | (donor:cell_type):feat_idx)

This formula is then included in the actual brmsformula I pass to brms:

# Construct formula for brms
formula <- paste0(
    "counts ~ intercept + design + log(sf) + log(",
    n_batches,
    ")"
)
formula <- brms::bf(formula, nl = TRUE) +
    brms::lf(design_formula, sparse = TRUE) +
    brms::lf(paste0("sf ~ 0 + ", batch_name)) +
    brms::lf(paste0("intercept ~ 0 + feat_idx"), sparse = TRUE)

For the current example, this final brmsformula looks like:

counts ~ intercept + design + log(sf) + log(6)
design ~ 0 + (cell_type:feat_idx) + (1 | (donor:cell_type):feat_idx)
sf ~ 0 + batch
intercept ~ 0 + feat_idx

My issue however is that although I specify sparse = TRUE, the resulting fixed/population effects model matrices are not stored as sparse model matrices:

model_data <- brms::standata(
    formula,
    data = model_data,
    family = "negbinomial",
    prior = priors,
    threads = brms::threading(n_threads)
)

> model_data$X_design |> class()
[1] "matrix" "array"

> model_data$X_intercept |> class()
[1] "matrix" "array"

These matrices are extremely large and contain many 0s, so I really would like to store them as sparse matrices.

What am I doing wrong? How can I make the sparse = TRUE argument work?

Supplementary Information

May or may not be helpful but I’ll include here:

  • The priors are constructed like so:
# Construct priors
priors <- brms::prior_string(
    paste0("dirichlet(rep_vector(1, ", n_batches, "))"),
    nlpar = "sf"
)

# Include optional priors
if (any(!design$random)) {
    # Induce sparsity in fixed effects
    priors <- priors +
        brms::prior(
            "horseshoe(main = TRUE, scale_slab = 10000)",
            class = "b",
            nlpar = "design"
        )
}
if (any(design$random)) {
    # Induce sparsity in random effects
    priors <- priors +
        brms::prior(
            "horseshoe(1, scale_slab = 10000)",
            class = "sd",
            nlpar = "design"
        )
}

And for the current example produce the following:

> priors
                                      prior class coef group resp dpar  nlpar
                dirichlet(rep_vector(1, 6))     b                          sf
 horseshoe(main = TRUE, scale_slab = 10000)     b                      design
           horseshoe(1, scale_slab = 10000)    sd                      design
   lb   ub source
 <NA> <NA>   user
 <NA> <NA>   user
 <NA> <NA>   user