Evaluating molecular fingerprint-based models of drug side effects against a statistical control