Wolfram Language Paclet Repository
Community-contributed installable additions to the Wolfram Language
Substructure Screening | |
Molecule fingerprints can be used for fast substructure searching due to the fact that if a particular substructure appears in a molecule, then all bits set in the substructure's fingerprint will also be set in the molecule's fingerprints. The
Prepare a dataset of fingerprints
Prepare a dataset of fingerprints
First import one hundred thousand molecules randomly selected from the ChEMBL database.
In[22]:=
mols=Import[PacletObject["WolframChemistry/MoleculeFingerprints"]["AssetLocation","SMILES strings for 100K molecules from ChEMBL"]];
It is important to end the expression with a semicolon here, suppressing the output so that the system does not try to format each molecule.
Now precompute the fingerprints for these molecules, using the "BitVector" output format:
In[140]:=
fprints=
[mols,"BitVector"];
 | PatternFingerprint |
Search the dataset for a query molecule
Search the dataset for a query molecule
Now that we have precomputed the pattern fingerprints we can search through them for a query substructure. This works because all bits set in the query fingerprint will also be set in the fingerprint for a molecule containing the query.
In[47]:=
query=Molecule["caffeine"];MoleculePlot[query]
Out[48]=
Compute the query fingerprint:
In[49]:=
queryFP=
[query,"BitVector"]
 | PatternFingerprint |
Out[49]=
DataStructure
 |
|
In[50]:=
prescreened=Pick[mols,queryFP["Copy"]["BitAnd",#]===queryFP&/@fprints];//AbsoluteTiming
Out[50]=
{0.309261,Null}
It is possible that molecules that don't contain the query could still have the same bits set, due to bit collision:
In[51]:=
CountsBy[prescreened,MoleculeContainsQ[query,IncludeHydrogensFalse]]
Out[51]=
True167,False26
By using the fingerprints to quickly screen out the majority of the molecules without doing a full substructure search.
In[52]:=
CountsBy[mols,MoleculeContainsQ[query,IncludeHydrogensFalse]]//AbsoluteTiming
Out[52]=
{21.0927,False99833,True167}
The speedup by using fingerprints to prescreen is quite large:
In[53]:=
%〚1〛/%%%〚1〛
Out[53]=
68.2037
Of course computing the fingerprints all at once is an expensive operation, and the time savings will only fully be realized when searching for many substructures in a large set of molecules.
""