I don't think you could just subtract outright as each sample is going to have a very different total frequency of OTUs. However, I wonder if you would be able to scale the control's counts to a sample's total frequency. It won't be perfect as there'll still be sequences which the control didn't see, but could have, if it had happened to have more sequencing depth, but that's a problem with a qualitative filter anyhow.
@mortonjt does a lot of compositional statistics, he would have a better grasp than I of what could make sense here.