Clarification needed on --flip argument

[RiccardoIannaco]

Hi,
I am not sure about the exact purpose of the --flip argument as described in the README:

G2PT/README.md

Line 28 in e4d2987

* _--flip_ argument will flip ref. and alt. allele (`--flip` argument, which make homozygous ref. as 2)

Is this intended only for merging datasets with different ref/alt conventions, or is there another reason to use it in this pipeline? Some more context or examples would be super helpful.

This doubt arose because, when training with my own data, the model (even as epochs increased) kept returning R² values close to zero, but never positive value. In the prediction files generated by predict_attention.py, the prediction column had the same (or almost the same) predicted phenotype for every individual, no significant variability at all. As epochs went up, that value seemed to approach the mean phenotype, but never showed significant differences between patients. This made me suspect a possible error with my input PLINK data.
Any ideas or suggestions? Could the use (or lack) of --flip cause this behavior?

Thanks again for your support and for sharing this tool.
Riccardo

[dlsrnsi]

Hi,

Would you please try this code:

from sgkit.io import plink

plink_data = plink.read_plink(path=path_to_your_data)
plink_data.call_genotype.as_numpy()

It will be a numpy array whose rows represents variants and columns represent samples.

Thank you for reading,

Best regards,
Ingoo Lee

[dlsrnsi]

Also, would you try all sex to be 1 not -9 in your case?

[RiccardoIannaco]

Hi Ingoo,

Thank you for your reply and your suggestions.

I tried using the code you provided:

I got the expected numpy array, where rows represent variants and columns represent samples, and the output looks like this (for the training set):

xarray.DataArray 'call_genotype' (variants: 321, samples: 273, ploidy: 2)>
array([[[1, 1],
        [1, 1],
        ...
       ],
       [[1, 1],
        [1, 0],
        ...
       ],
       ...
])

and for the validation set:

xarray.DataArray 'call_genotype' (variants: 321, samples: 69, ploidy: 2)>
array([[[1, 1],
        [1, 1],
        ...
       ],
       [[1, 1],
        [1, 1],
        ...
       ],
       ...
])

Just to clarify:

• The number of samples is low because I am working on a rare disease cohort.
• Similarly, the number of SNPs is also limited (321 variants), since they were filtered after GWAS using a p-value threshold of 1e-4.

However, I'm not sure what exactly I should check or do with this output. Could you please clarify what you expect me to verify or modify at this step?

As for your other suggestion, I also tried setting all sex values to 1 (instead of -9), but unfortunately this did not change the outcome: the model is still predicting almost the same phenotype value for every individual.

Thank you for your time and support!

Best,
Riccardo

[dlsrnsi]

Hi,

1. Would you check MAFs (average along rows)?
2. Are they in the same LD blocks?
3. Would you try to use a lower p-value threshold?
4. Also can you try to get a performance from PRS using PLINK?

Best regards,
Ingoo Lee

[RiccardoIannaco]

Hi,
Do you mean more stringent or less stringent p-value, in that with too stringent a p-value I would have lost additional snp, and being already only at 321 with a sparse population it seemed risky to me.

I am now calculating LD, PRS and MAF, to assess whether that might be casuating the problem. Thank you very much for the suggestion.

Best,
Riccardo

[dlsrnsi]

Hi, sorry for the late reply,

would you try less stringent p-value?

Best regards,
Ingoo Lee

[RiccardoIannaco]

Hi,

Unfortunately, , by making several attempts with different p-values, no difference can be seen and indeed at some point the single GPU (rtx 4090) becomes limiting in the amount of memory with the stock hyperparameters

Best,
Riccardo