Maybe, just maybe, we will be able to find some of the epistatic genetic effects through clever search algos. https://arxiv.org/abs/1610.05108 pic.twitter.com/KlPgz0fIOh
-
Example: https://www.biorxiv.org/content/early/2017/12/31/241414 … Single trait w/no auxiliary data, very small n/p, shallow CNN, no hyperparameter optimization or data augmentation, but still beats rr-BLUP.
-
-
The scaling with 'p' isn't great and something like a 500k SNP chip might be too hard without devoting a whole cluster, but I figure maybe it can be pruned back to 50k or so based on a first-pass. Probably not too many relevant SNPs with exactly 0 additive effects...
-
Maybe do a rough filtering with lasso/p value, and then do NN on the chosen snps.
-
I think lasso's behavior in sparsifying is probably the exact opposite of what one would want. The other end of the spectrum, ridge, would be more useful b/c it ranks SNPs while shrinking, and small bets still let you take the top Nk SNPs (whatever you can feasibly fit the CNN).
-
Well, one will have to try a variety of approaches. Hsu's team currently prune the snp list by ranking them by p value from singular regression first, then taking the top 50k or whatever snps and feeding these to the lasso.
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.