Not quite - if there are a ton of zeros, the result will be quite biased already (changes in low abundance taxa could really throw off the balance calculations). So I wouldn't trust the coefficients / pvalues coming out of balances that were computed from many low abundance taxa.
That has been proposed before
And that is a totally fine approach - its just balance-taxonomy won't be available, since it won't work on a collapsed table. So if you go this route, you'll need to do your own processing offline via R / Python