search

gusevlab / fusion_twas

methods for functional summary-based imputation
http://gusevlab.org/projects/fusion/
GNU General Public License v3.0
75 stars 45 forks source link

Best model specified as enet, yet all values in wgt.matrix are reported as 0 #37

Closed travismize closed 3 years ago

travismize commented 3 years ago

Hello,

While parsing some of the weights obtained from http://gusevlab.org/projects/fusion/ (GTEx v7), I have noticed that the model with the highest R^2 is enet, yet the reported values in the wgt.matrix for enet are all 0.

Example files are: Brain_Caudate_basal_ganglia.ENSG00000162368.9.wgt.RDat, Brain_Caudate_basal_ganglia.ENSG00000230092.3.wgt.RDat

I am attempting to generate a list of SNPs from the best model of each weight, based on the model's specifications, but am unable to do so when all SNPs have a reported coefficient of 0.

problematic_weights.zip

Thank you

sashagusev commented 3 years ago

Hi, thanks for the delay and good catch. I think this is due to some stochasticity in the weights training. The R2 is computed by cross-validation, then a final set of weights is trained on all the data. It's possible that in this final weights a model that had significant R2 performance converges to all zero weights. I think the most likely interpretation is that this is a weak or unstable model and shouldn't be used for prediction. In principle you can revert to the top1 model instead if you need to test this gene.