dirichlet_multinomial noise model doesn't seem to function

[yifand64]

Hi, when I tried to change the noise model from the default beta-binomial to Dirichlet multinomial, it didn't seem to work. I wonder what could the solution be as I'm trying to benchmark sccomp against other methods.

res <- seurat_obj |>
sccomp_glm(
formula_composition = ~ type,
formula_variability = ~ type, #if you want to have variability analysis
.sample = sample,
.cell_group = cell_group,
#.count = count, only add this line for count object
bimodal_mean_variability_association = T,
noise_model = "dirichlet_multinomial",
#cores = 1
)

[stemangiola]

Hello, yes the method advanced so much that Dirichlet was a bit left behind in the new releases. Let me try to debug that.

Having said that, the Dirichlet was implemented only for comparative purposes, we suggest to use the sum-constrained Beta binomial.

[stemangiola]

I will get to this shortly.

[yifand64]

I will get to this shortly.

Hi Dr. Mangiola, thank you for addressing this. I'm currently doing a benchmark study on most of the compositional analysis methods. Currently, I use the brm model for the Dirichlet-Multinomial model but it is awfully slow, so I'm trying to use your beta-binomial as well as your implementation of the DM model. However, your DM model doesn't seem to select a reference type. I wonder if I'm interpreting this correctly.

[stemangiola]

Please disregard the dirichlet model at this stage, I will need to find the throughput to fix it (I just started a research group). I believe the one from brms will be fine.

In your benchmark, remember to consider the existence of outliers; they are very much present in a lot of real-world analyses!