Re: [Rdkit-discuss] Incorrect results for substructure search obtained with Tversky similarity.
Hi Axel, The RDKit's Morgan Fingerprint is not a substructure screening fingerprint. If you want to use a fingerprint for screening, your best bet is the Pattern fingerprint. As an aside, the RDKit has a function, DataStructs.AllProbeBitsMatch (http://www.rdkit.org/Python_Docs/rdkit.DataStructs.cDataStructs-module.html#AllProbeBitsMatch) which is a somewhat more efficient way of doing the check you're looking to do. -greg On Mon, Dec 12, 2016 at 7:06 PM +0100, "Axel Rudling" wrote: Hi Brian and thank you for your respons. Yes, so Tversky with alpha parameter set to 1.0 and a cutoff for the similarity at 1.0 (100 % of me in you) will equal substucture search, at least at a theoretical level. I guess my question is, does imperfections in the fp model likley to generate these kind of results? So I use ecfp4 with 2048 bits. Regards Axel On Dec 12, 2016 6:57 PM, "Brian Kelley" wrote: I'm not really sure what you mean by tversky searching in substructure mode. Fingerprinting methods do not guarantee the presence of an exact substructure. You can think of tversky asking what percentage of me is in you and that percentage doesn't have to be a substructure. However they are correlated in that a good screening fingerprint can throw out molecules that will never be a substructure match. You still have to check the substructure match however. Using a screen fingerprint to filter out true negatives, I generally go from 5-10k substructure matches/sec to around 500-600k/sec in real world searches. I'm happy to provide an example of this if you need it. I hope this helps. Brian Kelley On Dec 12, 2016, at 11:29 AM, Axel Rudling wrote: Hello all, Currently I'm doing a project with Tversky searching in substructure mode and use smiles for creating fingerprints. For most molecules I get the correct result but there are some molecules where I get an overflow of falsely predicted substructure molecules. In brief, I get a large amount of compounds as a result from the substructure search that are not actually substructures of the query compound. I'm not certain of why but it might have to do with the FP representation as these molecules have a very unusual curricular structure ex.: C1C[NH2+]CCC[NH2+]CCCNCCC[NH2+]C1 I use 2048-bit ECFP4 fingerprints. tverskySim = DataStructs.TverskySimilarity(ffp1,ffp2,1.0,0.0) Does anyone have an idea? best Axel -- Check out the vibrant tech community on one of the world's most engaging tech sites, SlashDot.org! http://sdm.link/slashdot___ Rdkit-discuss mailing list Rdkit-discuss@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/rdkit-discuss -- Check out the vibrant tech community on one of the world's most engaging tech sites, SlashDot.org! http://sdm.link/slashdot___ Rdkit-discuss mailing list Rdkit-discuss@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/rdkit-discuss
Re: [Rdkit-discuss] Incorrect results for substructure search obtained with Tversky similarity.
Hi Brian and thank you for your respons. Yes, so Tversky with alpha parameter set to 1.0 and a cutoff for the similarity at 1.0 (100 % of me in you) will equal substucture search, at least at a theoretical level. I guess my question is, does imperfections in the fp model likley to generate these kind of results? So I use ecfp4 with 2048 bits. Regards Axel On Dec 12, 2016 6:57 PM, "Brian Kelley" wrote: > I'm not really sure what you mean by tversky searching in substructure > mode. > > Fingerprinting methods do not guarantee the presence of an exact > substructure. You can think of tversky asking what percentage of me is in > you and that percentage doesn't have to be a substructure. However they > are correlated in that a good screening fingerprint can throw out molecules > that will never be a substructure match. You still have to check the > substructure match however. > > Using a screen fingerprint to filter out true negatives, I generally go > from 5-10k substructure matches/sec to around 500-600k/sec in real world > searches. I'm happy to provide an example of this if you need it. > > I hope this helps. > > > Brian Kelley > > On Dec 12, 2016, at 11:29 AM, Axel Rudling wrote: > > Hello all, > > Currently I'm doing a project with Tversky searching in substructure mode > and use smiles for creating fingerprints. > > For most molecules I get the correct result but there are some molecules > where I get an overflow of falsely predicted substructure molecules. In > brief, I get a large amount of compounds as a result from the substructure > search that are not actually substructures of the query compound. I'm not > certain of why but it might have to do with the FP representation as these > molecules have a very unusual curricular structure ex.: > > C1C[NH2+]CCC[NH2+]CCCNCCC[NH2+]C1 > > > I use 2048-bit ECFP4 fingerprints. > > tverskySim = DataStructs.TverskySimilarity(ffp1,ffp2,1.0,0.0) > > Does anyone have an idea? > > > best > > Axel > > > -- > Check out the vibrant tech community on one of the world's most > engaging tech sites, SlashDot.org! http://sdm.link/slashdot > > ___ > Rdkit-discuss mailing list > Rdkit-discuss@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/rdkit-discuss > > -- Check out the vibrant tech community on one of the world's most engaging tech sites, SlashDot.org! http://sdm.link/slashdot___ Rdkit-discuss mailing list Rdkit-discuss@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/rdkit-discuss
Re: [Rdkit-discuss] Incorrect results for substructure search obtained with Tversky similarity.
I'm not really sure what you mean by tversky searching in substructure mode. Fingerprinting methods do not guarantee the presence of an exact substructure. You can think of tversky asking what percentage of me is in you and that percentage doesn't have to be a substructure. However they are correlated in that a good screening fingerprint can throw out molecules that will never be a substructure match. You still have to check the substructure match however. Using a screen fingerprint to filter out true negatives, I generally go from 5-10k substructure matches/sec to around 500-600k/sec in real world searches. I'm happy to provide an example of this if you need it. I hope this helps. Brian Kelley > On Dec 12, 2016, at 11:29 AM, Axel Rudling wrote: > > Hello all, > > Currently I'm doing a project with Tversky searching in substructure mode and > use smiles for creating fingerprints. > > For most molecules I get the correct result but there are some molecules > where I get an overflow of falsely predicted substructure molecules. In > brief, I get a large amount of compounds as a result from the substructure > search that are not actually substructures of the query compound. I'm not > certain of why but it might have to do with the FP representation as these > molecules have a very unusual curricular structure ex.: > > C1C[NH2+]CCC[NH2+]CCCNCCC[NH2+]C1 > > > I use 2048-bit ECFP4 fingerprints. > > tverskySim = DataStructs.TverskySimilarity(ffp1,ffp2,1.0,0.0) > > Does anyone have an idea? > > > > best > > Axel > > > -- > Check out the vibrant tech community on one of the world's most > engaging tech sites, SlashDot.org! http://sdm.link/slashdot > ___ > Rdkit-discuss mailing list > Rdkit-discuss@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/rdkit-discuss -- Check out the vibrant tech community on one of the world's most engaging tech sites, SlashDot.org! http://sdm.link/slashdot___ Rdkit-discuss mailing list Rdkit-discuss@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/rdkit-discuss