makeTechTrend() throws error : 'to' must be a finite number

[marionilab]

makeTechTrend() throws error : 'to' must be a finite number

[LTLA]

The immediate error is caused by the fact that you have zero size factors - positive=TRUE is a misnomer as it only ensures that the size factors are non-negative. As a result, you end up with undefined log-transformed values in your log-expression matrix, NA subsequently complains because the internal calculations are not robust against NAs (and nor should they be; the NAs are the canary in the coal mine).

The real problem is with the fact that you need positive=TRUE at all. If you read ?computeSumFactors, you will see that I don't think particularly highly of positive=TRUE. The underlying issue is that you're observing cells for which the median count - even after pooling! - is zero. I would argue that there is no way to normalize these cells effectively. The preferred solution is to perform more stringent quality control, as these are typically damaged cells that sneaked through somewhere - see my comments in Section 4.

If you must keep these cells, you can replace their size factors with the scaled library size, see https://github.com/MarioniLab/EmptyDrops2017/blob/master/analysis/pbmc4k/analysis.Rmd as an example. This is not an accurate normalization and just papers over the problem, but it will allow you to proceed to downstream analyses. Keep in mind that log-fold changes involving these cells will not be accurately computed as library size normalization does not resolve composition biases.

[marcRDM]

More stringent QC did the job ! Thanks Aaron !