Seed gene discovery on Binary phenotypes

[madhushanka]

Dear team,

I am currently trying to run the pipeline (more specifically seed gene discovery) for some binary phenotypes and having some problems in that regard:

envs/deeprvat/lib/python3.9/site-packages/numpy/linalg/linalg.py", line 89, in _raise_linalgerror_singular raise LinAlgError("Singular matrix") numpy.linalg.LinAlgError: Singular matrix

Do you think this is due to how the binary phenotypes are represented in the phenotype file? In my case control: 0 and case :1

I am stuck here. Any insights or suggestions would be highly appreciated!
PS: Do you think its fine to run the training and association testing(including seed gene discovery) for a small sample size about 4000?

Best,
Madhu

[ThibaultBechtler]

Hi!

Thank you for raising this issue.

We will look into the error on our end as soon as possible. Could you maybe provide us with a minmal workable example of your phenotype file that raises the error?
So far we haven't tried to train a model on binary phenotypes. This could lead to unforeseen numerical instabilities.

Regarding your question on sample size:
Is there a specific reason that you want to retrain a model from scratch, instead of using the pretrained model? 4000 samples to train a model seem insufficient to me. The gene impairment scores produced by the pretrained model work for testing on binary phenotypes as well.

Best,
Thibault

[madhushanka]

Dear Thibault,

Thanks for getting back on this. I have attached here a minimal phenotype file(excluding the sample column/index column).

Regarding the sample size, I first tried using the pretrained models and it ran without any problem but got no significant results. I got reasonably significant results even with only using Regenie burden tests. The whole idea of using DeepRVAT was to enhance those and see if I could improve it. With pretrained models, I could also not see if I can run deepRVAT for different burden masks. In order to include some gene burden masks(i.e plof, missense) I thought of using the training and testing pipeline(that includes seed gene discovery) where it lead me to the above error.

Hope I explained the situation properly.
Looking forward to your thoughts!

best,
Madhu

pheno_minimal.txt