4 ms·
No code in paper, correct. Very easy to implement via R or python. One of the authors (Gustavo de los Campos) has alot of code that could probably be coped/past
by kickout 9y ago
No code in paper, correct. Very easy to implement via R or python. One of the authors (Gustavo de los Campos) has alot of code that could probably be coped/pasted with little editing from his website.
Algorithm is actually painfully simple. Just relying on massive data (500K individuals)
- gwern 9y agoCoordinate descent is simple but the resource requirements may complicate your implementation running over n=500k with tens of thousands of markers each; as it was described to me, they were using a workstation with like 128GB RAM... (Plus of course you need access to the UKBB dataset in the first place.)
- moyix 9y agoThat's to replicate the study; presumably if the authors released their trained model it would be much less computationally expensive to run it on your own data?
- gwern 9y agoIf by trained model you mean PGS, yes, it would be trivial. While the fitting of the optimal weights may be computationally intensive, it's still just a simple little linear model: look up the SNP value, add or subtract a small coefficient for each of the 30k or whatever 'activated' SNPs, and get the net total.
- nextos 9y agoSeems code is derived from this Julia package written by the authors: https://github.com/gdlc/BGLR.jl https://github.com/gdlc/BGLR.jl