Very interesting topic. I have faced these challenges while working with fingerprints and here are few observations from my end. By the way I agree that mathematically the best bet is ~ 13%.
1) The hashed FP (CDK) is good enough to separate patterns which are not common but on a large dataset (in my case 10000+ mols), the performance drops drastically. Top 1% hits were good but then rest of the started to loose specificity (esp when Tanimoto score was around 0.77).
2) First I thought it was an artifact of the Tanimoto score… but I wasn’t convinced spl. in cases where we had rings (close vs open). I ended up writing a new FP based on the pubchem patterns as coded in the CDK and added few more patterns to resize it to 1024 from 881. Well! It’s works like magic and I could find much more serialised hits than before. I think the extensions of the fingerprint which I made based on the patterns in my db also helped.
At the end of the day, I believe that all these searches are heuristic and hashed FP is faster to generate but prone to bit clashes where as SMARTS based FPs are slower to generate (as u spend time in MCS) in matching patterns but they are more sensitive and specific as you can trace the patterns (u get what u see) as the patterns and bitset relationship is know and static