Service notice: Several of our team members are on vacation so service will be slow through at least July 13th, possibly longer depending on how much backlog accumulates during that time. This means that for a while it may take us more time than usual to answer your questions. Thank you for your patience.

VariantRecalibration, numBadVariants, and size of the data set

I'm somewhat struggling with the new negative training model in 2.7. Specifically, this paragraph in the FAQ causes me trouble:

Finally, please be advised that while the default recommendation for --numBadVariants is 1000, this value is geared for smaller datasets. This is the number of the worst scoring variants to use when building the model of bad variants. If you have a dataset that's on the large side, you may need to increase this value considerably, especially for SNPs.

And so I keep thinking about how to scale it with my dataset, and I keep wanting to just make it a percentage of the total variants - which is of course the behavior that was removed! In the Version History for 2.7, you say

Because of how relative amounts of good and bad variants tend to scale differently with call set size, we also realized it was a bad idea to have the selection of bad variants be based on a percentage (as it has been until now) and instead switched it to a hard number

Can you comment a little further about how it scales? I'm assuming it's non-linear, and my intuition would be that smaller sets have proportionally more bad variants. Is that what you've seen? Do you have any other observations that could help guide selection of that parameter?

Best Answer


  • Ahh, that would be cool. I'm just now moving my production pipelines to 2.7, maybe I'll hold off for a bit longer. I should be able to scare up a dataset or two I could test with, please do let me know when it appears in the nightlies. Thanks!

Sign In or Register to comment.