Better start value for x0 when estimating dropout parameters

[duohongrui]

Hi,
Splatter is an ideal tool to simulate different kinds of single-cell profiling data. Recently I met a commo problem for most of my reference datasets using the function splatEstimate when Splatter was estimating the dropout parameters. The error said:

Error in nls(y ~ logistic(x, x0 = x0, k = k), data = df, start = list(x0 = 0,  : singular gradient
In addition: Warning message:
In value[[3L]](cond) :
  Fitting dropout using the Gauss-Newton method failed, using the Golub-Pereyra algorithm instead

It seems like fitting the reference data to the logistic model failed. After that, I looked for the solutions on the Internet and found one. https://github.com/broadinstitute/infercnv/blob/master/R/SplatterScrape.R. It adds x_approx_mid <- median(x[which(y>0.2& y < 0.8)]) step before fit <- nls(y ~ .logistic(x, x0 = x0, k = k), data = df, start = list(x0 = x_approx_mid, k = -1)) and considers x_approx_mid as the x0 starting point.
I accepted this alteration and rerun the estimating step. Finally, it works for all of my datasets. However, I am not good at statistics and cannot find out why it works. Here is my R environment.

R version 4.1.1 (2021-08-10)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 10 x64 (build 19044)

Matrix products: default

locale:
[1] LC_COLLATE=Chinese (Simplified)_China.936  LC_CTYPE=Chinese (Simplified)_China.936   
[3] LC_MONETARY=Chinese (Simplified)_China.936 LC_NUMERIC=C                              
[5] LC_TIME=Chinese (Simplified)_China.936    

attached base packages:
[1] parallel  stats4    stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
 [1] splatter_1.16.1             SingleCellExperiment_1.14.1 SummarizedExperiment_1.22.0 Biobase_2.52.0             
 [5] GenomicRanges_1.44.0        GenomeInfoDb_1.28.4         IRanges_2.26.0              S4Vectors_0.30.2           
 [9] BiocGenerics_0.38.0         MatrixGenerics_1.4.3        matrixStats_0.61.0          amap_0.8-18 

Could you check this problem and fix this bug in splat-estimate.R? Thanks very much!

[lazappi]

Hi @duohongrui

Thanks for the issue and the suggested solution! This seems like it is related to #133. After that issue I added another fallback method which might work in this case, you can try it using the development version of {splatter}.

This solution from {infercnv} is probably more robust though so I will look at adding it. I can definitely see how having a better starting estimate of x0 would improve the fitting.

[duohongrui]

I see. Thanks again!

[lazappi]

I have added this to the development version and it will be in the upcoming release