Have you ever wondered which compound is the most popular in ChEMBL? And by popular I don't mean the one which cracks the best jokes at dinner parties; I mean the compound with the largest number of structural analogues or nearest neighbours (NNs). This number also gives an indication of the sparsity or density of the chemical space around a compound and is a useful concept during hit expansion and lead optimisation.
This number of course depends on the fingerprint, the hashing and folding parameters, the similarity coefficient and the threshold. So let's say 2048-bit RDKit Morgan fingerprints with a radius of 2 or 3 (equivalent to ECFP_4 or ECFP_6) and Tanimoto threshold of 0.5. Why so low threshold? For an explanation, see here and here.
To calculate this compound 'popularity', one would need to calculate the full similarity matrix of the 1.4M compounds in ChEMBL. This used to be prohibitively computationally expensive just a few years ago; nowadays, thanks to chemfp, it only takes 5 commands and a few hours to calculate the matrix and counts at a certain similarity threshold on a decent machine. Here's how it's done:
#install chemfp
pip install chemfp
#get the chembl sdf
#calculate rdkit morgan fps
rdkit2fps --morgan --radius 3 --id-tag "chembl_id" --errors report chembl_20.sdf.gz -o rdkit_chembl.fps
#calculate counts of neighbours at a given threshold, here 0.5
simsearch --threshold 0.5 --NxN -c rdkit_chembl.fps -o chembl_sim_matrix.txt
#sort by number of neighbours and find most 'popular' compound
sort -rn chembl_sim_matrix.txt | head -25
Here's a list of the most popular ones with the number of NNs:
1570 CHEMBL3037891
1540 CHEMBL2373012
1523 CHEMBL3086861
1461 CHEMBL2158309
1451 CHEMBL440810
1451 CHEMBL414360
1428 CHEMBL2401865
1425 CHEMBL408133
1425 CHEMBL354100
1425 CHEMBL344931
1425 CHEMBL310737
1425 CHEMBL303256
1425 CHEMBL302290
1425 CHEMBL27867
1425 CHEMBL137783
1403 CHEMBL1866472
1396 CHEMBL3102922
1392 CHEMBL3102921
1392 CHEMBL3102920
1387 CHEMBL91573
1387 CHEMBL70362
1387 CHEMBL601767
1387 CHEMBL533732
1387 CHEMBL503883
1387 CHEMBL477
Most of the most popular compounds are large, usually peptides, e.g. CHEMBL3037891. We'll leave it as an exercise for the reader why that is. But this not always the case, for example CHEMBL477 is adenosine and it genuinely has a lot of near neighbours in ChEMBL (including a lot of stereo variants).
And what about the least popular compounds, i.e. the ones with no neighbours above 0.5 Tanimoto? One would not expect a lot of them, since ChEMBL by design contains compounds from congeneric series as reported in the med. chem. literature. And yet, due to small size, symmetry (CHEMBL100050), possible mistakes, and peculiarities of fingerprint similarity, or due to being genuinely lonely (CHEMBL35416), they do exist.
Apart from popularity contests, a similarity/distance matrix of ChEMBL compounds is the first step for clustering and graph/community detection analysis.
George
PS: Thebacon has only 22 neighbours...
Apart from popularity contests, a similarity/distance matrix of ChEMBL compounds is the first step for clustering and graph/community detection analysis.
George
PS: Thebacon has only 22 neighbours...
Comments