jgilis
commented
2 years ago Hi Chiara,
Yes, we indeed suggest using "raw" counts as input for satuRn, by which we simply mean non-normalized counts. If you import count using the tximport package, this basically means importing with the default setting for the countsFromAbundance argument, which is "no" (resulting in raw counts). Since setting other options for this argument like "scaledTPM" didnt have a pronounced effect on performance in our benchmark, we prefer working with the raw counts.
Such raw counts can be obtained from all quantifiers, alevin should be no exception.
Your reasoning is correct; this simply means that for the moment you are not able to pick up biological signal, which may have (at least) three reasons: (1) there isnt much biological signal altogether, (2) the biological signal is simply hidden by the technical noise or (3) the empirical correction by satuRn is too stringent.
If (1) is the case, there obviously is not much you can do.
To alleviate (2), you could consider a more stringent transcript-level filtering criterion. You could consider the edgeR::filterByExpression and DRIMSeq::dmFilter functions. To give you an idea, for our benchmarks I used
filterByExp(min.count = 1, min.total.count = 0, large.n = 0, min.prop = 0.5)
# n set to zero to force using min.prop
dmFilter(min_samps_feature_expr = ns/2,
min_feature_expr = 10,
min_samps_feature_prop=ns/2,
min_feature_prop=0.1,
min_samps_gene_expr=ns,
min_gene_expr=10)
# with ns the size of the smallest group in the comparisonas lenient and stringent filtering criteria, as described in the Methods of the paper.
Finally, issue 3. In all our benchmarks, empirical correction was vital to keep control of the type 1 error, and the correction did also a decent job as to not overcorrect. However, we only benchmarked this on three real single-cell datasets, so it may very well be that for your data something is going wrong. Therefore, I would really look at the two diagnostic plots in the testDTU function. For diagplot1, does the blue dashed curve nicely follow your test statistics, and are your test statistics nicely bell shaped? For diagplot2, does the dark green curve follow the bulk/mid portion of your test statistics? Note that more stringent filtering could alleviate such issues as well.
Kind regards,
Jeroen
ChiaraCaprioli
Hi,
I am using satuRn on a single-cell dataset to score DTU among groups of interest. Everything seems to work fine, but for the fact I obtain very few significant empirical FDR values in all contrasts I tried. I understand this might be due to both biological and technical reasons, especially zero-inflation and shallow sequencing, so I am checking potential ways to minimize this behavior.
Did I get correctly that in your model it is ok using raw count estimates as input data (i.e., not quantified by tools such as Alevin)? Should I expect any influence regarding this?
Thanks for your help