Convergence Issues

[abadgerw]

@const-ae Thank you for a great tool! I wanted to generate a distance matrix (example data attached) and keep getting the following warning:

In dropout_curves(Y, model_matrix, Pred_reg, Pred_var_reg) :
Dropout curve estimation did not properly converge

Any insight into how I can address this?

Example Data.csv

[const-ae]

Thanks for providing the example data. The first thing I notice is that you have proteins in the columns and samples in the rows, and proDA expects it the other way around (I assume you are aware and simply transpose your data, but that could be an easy fix).

Otherwise, I would have to properly debug the function call and see where something seems to go wrong and unfortunately I currently don't have the capacity. You could try it yourself and call debugonce(dropout_curves) before calling proDA. This will allow you to step through the code and you might be able to figure out why the fitting goes wrong (could for example be a problematic initialization, very large data values, etc.).

[abadgerw]

Yes, I am transposing. Oddly enough, the convergence issue is not there when the format is wrong (ie not transposed). I tried calling debugonce and it really didn't guide me much though I am new to all of this. It just started at line 2 and didn't highlight where things are going wrong. Could it be because my data has positive and negative values because of the normalization I did?

[const-ae]

I tried calling debugonce and it really didn't guide me much though I am new to all of this

Yeah, debugging R code--especially if it's from someone else--is certainly an advanced skill. It is a super useful skill, but I can also see that it is a bit much to ask.

Could it be because my data has positive and negative values because of the normalization I did?

Hm, I did wonder about that myself. Most datasets have values, let's say, between 8 and 25, but the way the method is implemented, the data range shouldn't make a difference. You could check if you have some extreme outlier (e.g., a value above 100) but otherwise it is difficult to diagnose the problem without spending some serious time debugging this.