One challenging aspect of modeling scRNA-seq data is data sparsity, that is, scRNA-seq measurements typically suffer from large fractions of observed zeros (i.e. dropouts), where a given gene in a given cell has no unique molecular identifiers or reads mapping to it [1].

What would be an appropriate dataset to benchmark supervised learning algorithms that operate on single-cell RNA-seq data? In particular, I am looking for challenging datasets (preferably with labels, i.e. cell type classification) where existing methods struggle due to the sparsity of the data. Additionally, does it ever happen that the dropouts are confounders of the labels? (i.e. for a given cell type A the dropout probability of a gene X is higher than for another cell type B).

Any pointers are welcome.

  • $\begingroup$ Can you say what is the purpose of your supervised learning method? to annotate cell types? $\endgroup$ – Chris_Rands Mar 24 at 22:11
  • $\begingroup$ Yes, that's right, cell type annotation is the first thing that comes to mind, but I am also willing to consider other tasks (e.g. classification of cancer cells). Even regression tasks would be fine too. Essentially, I have a method in mind that might be useful for modeling non-linear+sparse data and I am looking for an appropriate benchmark dataset. $\endgroup$ – rvinas Mar 24 at 22:40

You could try the PBMC 3k data from the Satija lab:


For well-annotated data, there's the Single-cell proteo-genomic reference map:


Data for that are available here.

If these datasets are too large (i.e. too good), you can use the existing counts as weighting for a random removal of reads. Genes with high expression are likely to be present regardless of sequencing depth, hence the count-adjusted removal.

FWIW, I don't think that supervised learning is a good fit for dealing with data sparsity in cDNA read counts. There are plenty of existing attempts to solve the issues by adjusting the assumed models, including research by the authors of commonly-used bulk differential expression tools (e.g. here).

  • $\begingroup$ Thanks! I like the second reference, it is fresh from the oven. If the data is "too good" (i.e. traditional models can easily distinguish the different cell types), you would then introduce random dropouts to increase the sparsity? $\endgroup$ – rvinas Mar 25 at 11:07
  • $\begingroup$ Random dropouts at the read level; I wouldn't randomly drop out genes. $\endgroup$ – gringer Mar 25 at 20:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.