I'm not sure if I know the answer to what is the "best" way to identify loss-of-function variants. A pre-mature stop codon towards the beginning of the gene is probably valid (unless the gene can and does undergo alternative splicing), but I think there are some recommendations here: https://github.com/konradjk/loftee

In my opinion, I think having access to specialized information for previous disease associations is the right solution if you are trying to analyze your own data, but I think "ClinVar" is currently the best thing that I can think of for that.

In terms of getting ideas from current specialized databases, I think these may be some examples to consider:

https://brcaexchange.org/

https://www.cftr2.org/

I think 0.01 or 0.05 would be common frequencies to define rare variants. However, I usually have multiple variant frequency programs. If you see very different frequencies with different reference sets, I would guess the pre-processing and/or variant calling could be a factor.

In general, for discovery, I think you will probably see a higher frequencies of false positives. As long as you have a way to identify a few possibly important mutations per sample, visualization of the alignment is very important.

Otherwise, if you have your own set of cases and controls (collected and processed the same way), you can test for differences for all variants and then check enrichment of variant categories (like loss-of-function). I've also seen people summarize gene counts (for a particular method of annotating variants) and then compare frequencies at the gene level (between their own cases and controls). In that situation, if you knew the gene involved, you could try various methods to see that results in high ranking for your gene (for your particular disease / project).