With STRING version 8, we began to add prior correction to the scores. That means that we consider how close scores are to the background rate of about 6% for protein–protein interactions, and we remove the random expectation from the scores before we combine them. (At the end, we put the random contribution back in to make it consistent with the case that only a single evidence channel contributes to the score.)
For some applications, you might want to remove certain evidence types from the STRING download files. To compute a new combined score, you can now use the same prior correction we've been using internally. We have started a new BitBucket repository, and added a script to discard channels from the full STRING download files.