Robustness of MELD with small datasets

[Rumengol]

First of all, thank you for your work and this great package.

We have several datasets of scRNAseq case-control data with unfortunately no replicates.

For one of them, the results of the MELD algorithm are coherent with the observations, but for one other it is completely opposite: a cell type biologically observed to significantly decrease in the condition is deemed enriched by the algorithm. we thought it might be due to a bias introduced by either an extreme difference in the number of cells (2800 cells in the control dataset, 7700 in the case dataset), or the overall low cell counts.

Do you have any information about the possible limitations of the algorithm when ran on a small dataset or with a great difference between the conditions? We don't rule out the possibility of an experimental bias, but thought you might have an idea.

Thank you for your answer.