Danielle,

While this sort of question often provokes something of a religious war, I think it is fair to say that a consistently well performing fingerprint method is the circular/radial fingerprint (aka ECFP/FCFP in Pipeline Pilot). The circular fingerprint approach has been implemented in a variety of tools, including the OpenEye GraphSim toolkit (full disclosure - I work at OpenEye), Canvas from Schrodinger and elsewhere. Honoured more by time are path fingerprints, first popularised by Daylight. A paper describing the ECFP idea is attached.

I notice that you focus on "structural fingerprints" and exemplify these with MACCS keys and the PubChem keys. These are fundamentally different in philosophy to the circular/tree/path fingerprint; MACCS keys and the like set a bit in the fingerprint on when a given substructure is present in the molecule (say a phenyl ring or a chlorine), so there is some ability to reverse the process and see which features are driving similarity. Such a thing is MUCH more difficult in so-called hashed fingerprints like circular/tree/path, where the bits set on the fingerprint do not correspond to the same structural feature from molecule to molecule.

You do not define what "better" would mean for you, so I assume virtual screening (VS) is your goal. 

As to VS performance, have a look at papers from Peter Willet's group over the years, as well as some recent ones on fingerprints in virtual screening.

doi:10.1021/ci0504723

doi:10.1016/j.jmgm.2010.05.008

Hi Paul,

Thank you very much for your answer! It is very useful and thank you for the references

Basically i am using similarity measures calculated from the fingerprints as molecular descriptors to carry out QSAR analysis. So i wanted fingerprints that are more interpretable i.e relate to a specific function group etc however i know that the hashed FPs are better - its a balance between interpretability and predictability!. I was interested to know the consensus relating to structural key FPs and i am not sure which one is better MACCS or pubchem. With MACCS only containing 166 bits i was worried i might loose some information. But from the literature it appears they have similar ability when measured objectively.

Thanks again for your answer it give me some more things to think about!

Danielle

Not doing QSAR I am only guessing, but I think a structural key fingerprint like MACCS is really the only appropriate fingerprint for QSAR. The hashed fingerprints won't work. Also if you use even a 166 bit f'print like MACCS you will need 167 compounds with measured activity or your model will be underdetermined. Don't be fooled by feature selection methods that promise to pick the features the correlate best with activity and use only those; you will still have a 166 component model, just some of the components will have zero weight.

Hi Danielle, check out FP4 fingerprints in open babel, they are a set of 'chemically intuative' fragments. I have used them quite a bit in models that lead to some chemical interpretability, though predictive power is a bit lower than the more sophisticated ones. You need to do bit of work to convert the open babel output into a matrix suitable for analysis, a few lines of R or Python will do the trick. Good luck! mark

Hello,

does anybody know which fingerprint is used in Pubchem? So I have one compound and I search for all its analogs within 80% similarity? So, 0.8 TC I guess, but for which fingerprint, Daylight, EFCP4 or so.