Global assessment of scaffold hopping potential for current pharmaceutical targets

Ye Hu; Jürgen Bajorath

doi:10.1039/C0MD00156B

View PDF VersionPrevious ArticleNext Article

DOI: 10.1039/C0MD00156B (Concise Article) Med. Chem. Commun., 2010, 1, 339-344

Show Chemical TermsShow Biomedical Terms

Global assessment of scaffold hopping potential for current pharmaceutical targets

Ye Hu and Jürgen Bajorath *
Department of Life Science Informatics, B-IT, LIMES Program Unit Chemical Biology and Medicinal Chemistry, Rheinische Friedrich-Wilhelms-Universität, Dahlmannstr. 2, D-53113 Bonn, Germany. E-mail: bajorath@bit.uni-bonn.de; Fax: +49-228-2699-341; Tel: +49-228-2699-306

Received 9th September 2010 , Accepted 22nd September 2010

First published on 13th October 2010

Abstract

Scaffold hopping is an intensely investigated topic, both in the context of computational method evaluation and practical compound screening applications. Scaffold hopping refers to the identification of different compound classes having similar biological activity and is typically explored on a case-by-case basis. However, how frequently scaffold hops occur across different targets is presently not well understood. We have investigated global scaffold hopping potential by systematically analyzing topologically distinct scaffolds in currently available bioactive compounds with defined target and activity annotations. The analysis reveals that for the majority of target proteins, active compounds representing between five and 49 topologically distinct scaffolds are available. Moreover, for 70 targets, between 50 and more than 300 distinct scaffolds are found. Thus, scaffold hops occur with rather high frequency among active compounds.

In medicinal chemistry, the search for different structural classes (chemotypes) having similar activity is generally of high interest,^1,2 for example, to support chemical optimization efforts or secure intellectual property positions. Moreover, the demonstration of scaffold hopping potential has become the “holy grail” of computational screening methods.^3–9 The “value” of any virtual screening approach is essentially judged upon its ability to identify different chemotypes having similar activity, mostly in benchmark calculations. Beyond the often rather artificial scenario provided by typical benchmark studies, in prospective applications, a virtual screen is generally claimed to be a “success” if at least one or a few novel compounds with different core structures (scaffolds) and desired biological activity have been identified. Unfortunately, the assessment of scaffold hopping potential often suffers from the lack of clear scaffold definitions and inconsistent analysis of scaffold hops.⁹ Moreover, it is currently unclear how “difficult” scaffold hopping really might be. No studies are available at present that provide a general assessment of scaffold hopping potential across different targets, although such insights would be of general interest, both for the evaluation of computational screening methods and practical medicinal chemistry applications.

General scaffold hopping potential might be estimated by systematically analyzing, on a per-target basis, how many well-defined scaffold hops are “encoded” by currently available bioactive compounds. Accordingly, we have carried out a large-scale analysis of scaffold hops among publicly available active compounds. All calculations reported herein were carried out with in-house Perl and Scientific Vector Language (SVL)¹⁰ programs and Pipeline Pilot¹¹ tools.

From two major public repositories of bioactive compounds, CHEMBLdb (CDB)¹² and BindingDB (BDB),¹³ 31,158 and 17,745 molecules with activity annotations (K_i or IC₅₀ values) against human targets were selected, respectively. These compounds were organized in 586 and 433 individual target sets and 12,047 and 6,291 atomic property-based Bemis & Murcko scaffolds¹⁴ were extracted from them, respectively. CDB and BDB currently show limited compound overlap¹⁵ and we therefore merged the CDB and BDB compound and scaffold sets, yielding a total of 795 individual target sets containing 45,263 compounds and 16,873 unique scaffolds.

As illustrated in Fig. 1, property-based Bemis & Murcko scaffolds consist of core ring structures and linkers between them.¹⁴ Scaffolds only distinguished by heteroatom substitutions and bond orders display the same topology, as reflected by carbon skeletons (CSKs; i.e. scaffolds with all atom types set to carbon and all bond orders to one), as also illustrated in Fig. 1. We deliberately focused our analysis on topologically distinct scaffolds that are more relevant for scaffold hopping than scaffolds that are only distinguished by minor heteroatom substitutions or bond order alterations. Therefore, for each target set, we determined all Bemis & Murcko scaffolds yielding the same CSKs. In each of these cases, we only retained the scaffold that was represented by the largest number of compounds or, if several scaffolds had the same number of compounds, the scaffold represented by the largest number of compounds with highest median potency. An individual scaffold was retained instead of the CSK because compounds representing the scaffold were required for score calculations, as described below. Importantly, by retaining one Bemis & Murcko scaffold per CSK, all scaffolds selected for a target set at this stage were topologically distinct. This selection scheme yielded 10,989 topologically distinct scaffolds corresponding to 35,004 compounds. In order to further streamline the collection of target sets for meaningful scaffold hopping analysis, we only retained target sets containing at least five compounds with at least 1 μM potency (i.e., pKi or pIC₅₀ > = 6) and at least two scaffolds. Accordingly, our analysis was ultimately based on 8,693 topologically distinct scaffolds represented by 26,664 compounds organized into 502 different target sets. For the assignment of targets to families, we followed the CDB classification scheme and combined targets available in CDB and BDB. Table 1 reports the 19 target families considered in our analysis that contained between three and 137 individual targets.


	Fig. 1 Topologically distinct scaffolds. Nine representative scaffolds extracted from phosphodiesterase 5A inhibitors are shown. For each scaffold, the corresponding carbon skeleton (CSK) is shown and the number of compounds each scaffold represents is reported. Scaffolds 1 to 3 yield distinct CSKs, whereas scaffolds 4 to 9 share the same CSK. Scaffold 9 is selected for further analysis because it represents the largest number of compounds (i.e., 27), and the other five scaffolds are not further considered. This selection scheme ensures that only topologically distinct scaffolds are analyzed.

Table 1 Target families and scaffold distribution.^a

FamilyID	Target Family	# Targets
		Source			# Scaffolds
		BDB	CDB	Total	< 5	[5, 50)	[50, 100)	> = 100
a Nineteen target families are listed following the CHEMBLdb classification scheme. For each family, the numbers of targets taken from CHEMBLdb, BindingDB, and the total number of targets are reported (taking target overlap between these databases into account). In addition, for each family, the number of targets is reported whose compound sets contain different numbers of scaffolds. Target family abbreviations: GPCR, G-Protein Coupled Receptor; Others, all none classified targets.
1	Tyr protein kinases	30	32	38	4	28	2	4
2	Ser_Thr protein kinases	37	38	49	6	37	4	2
3	Ser_Thr_Tyr kinases	9	7	13	4	9	0	0
4	Phosphadiesterases	8	7	9	0	9	0	0
5	Protein phosphatases	1	3	3	0	3	0	0
6	Aspartic proteases	4	7	7	2	3	2	0
7	Cysteine proteases	11	12	14	1	9	2	2
8	Matrix metalloproteases	14	17	19	2	11	5	1
9	Serine proteases	18	21	25	2	20	0	3
10	Carbonic anhydrases	12	10	12	0	9	2	1
11	Histone deacetylases	8	5	8	0	7	1	0
12	CytochromeP450 enzymes	8	9	13	1	12	0	0
13	Transferases	4	4	8	1	7	0	0
14	Ion channels	4	20	22	8	14	0	0
15	GPCRs	45	129	137	13	92	19	13
16	Cytosolic-others	9	7	14	8	6	0	0
17	Electrochemical transporters	6	15	15	5	8	2	0
18	Nuclear receptors	15	16	20	5	13	2	0
19	Others	44	57	76	16	57	1	2

We first determined the number of distinct scaffolds present in each target set. The results are reported in Table 1 on a target family basis. Surprisingly, the majority of target sets were found to contain significant numbers of distinct scaffolds. A total of 354 target sets contained between five and 49 scaffolds, 42 target sets between 50 and 99, and 28 sets at least 100 scaffolds. Thus, the range of five to 49 scaffolds represents “average” scaffold diversity across current targets corresponding to average scaffold hopping potential. This is further illustrated by monitoring the scaffold distributions within target families (Fig. 2a). Many of these scaffolds were represented by compounds with in part very large potency differences (Fig. 2b).


	Fig. 2 Target family statistics. (a) Scaffold distribution and (b) target set median potency; presented as box plots. Target family IDs are according to Table 1. The box plots report the smallest value (bottom line), lower quartile (lower boundary of the box), median (thick horizontal line), upper quartile (upper boundary of the box), and the largest value (top line).

A total of 70 target sets from twelve different target families (covering ∼14% of the current target spectrum) were characterized by what we considered high to very high scaffold diversity, each containing between 50 and more than 300 topologically distinct scaffolds. We next analyzed these sets in more detail. Table 2 shows the top 30 targets ranked by scaffold numbers. Well-known pharmaceutical targets appear high on the ranking. These targets, which are also popular for virtual compound screening studies, include, for example, different adenosine and dopamine receptor subtypes and other GPCRs, protein kinases, and various proteases. These targets are chemically well explored. We have recently shown that more than 80% of scaffolds from currently available bioactive compounds are topologically equivalent and/or display substructure relationships.¹⁶ Here we have exclusively focused on topologically distinct scaffolds, but we also determined substructure relationships between them, as reported in Table 2. For the target sets containing most scaffolds, at least approx. half of these scaffolds, but often more than 70% or 80% were found to be involved in substructure relationships (i.e. one scaffold is a substructure of another in the same set). From this point of view, it might not be very surprising that these targets have high scaffold hopping probability, also in benchmark calculations, and we would hence consider them “easy” virtual screening targets.

Table 2 Target sets ranked by scaffold number^a

Target Name	#Sc	FamilyID	%Sc-in-Sub
a The top 30 target sets ranked according to scaffold numbers are reported. For each set, the number of scaffolds (#Sc) and the percentage of these scaffolds (%Sc-in-Sub) that are involved in substructure relationships are reported.
Melanin-concentrating hormone receptor 1	318	15	57.2
Vascular endothelial growth factor receptor 1	302	1	66.2
Melanocortin receptor 4	207	15	63.3
Factor Xa	187	9	59.4
Cyclin-dependent kinase 2	180	2	70.6
Src tyrosinekinase	174	1	46.0
Thrombin	162	9	41.4
Adenosine receptor A3	160	15	73.1
Mu opioid receptor	157	15	86.6
Kappa opioid receptor	155	15	80.6
Delta opioid receptor	154	15	89.0
Cathepsin K	145	7	53.1
Serotonin receptor 5HT 1a	136	15	81.6
Acetylcholinesterase	136	19	53.7
Endothelial growth factorreceptor	134	1	76.1
Dopamine receptor D2	134	15	50.7
Adenosine receptor A1	129	15	74.4
Mitogen-activated proteinp38 alpha	129	2	64.3
Cathepsin S	129	7	55.8
Dipeptidyl peptidase 4	119	9	81.5
Adenosine receptor A2A	115	15	80.0
Serotonin transporter	110	15	43.6
Matrix metalloproteinase 3	108	8	61.1
Leukocyto-specific tyrosinekinase	106	1	69.8
Butyrylcholinesterase	104	19	51.9
Carbonic anhydrase II	101	10	55.4
Nociceptin receptor 1	100	15	72.0
Histamine H3 receptor	100	15	75.0
Protein kinase B Akt1	95	2	75.8
Matrix metalloproteinase 2	94	8	75.5

In order to assess scaffold hopping potential in quantitative terms, beyond scaffold numbers, we have also designed a function yielding a “hopping score” that incorporates compound potency information and is calculated over individual scaffold pairs in target sets. For a scaffold pair ij in target set T, all possible compound pairs C_ij are enumerated (i.e., compounds in a pair contain scaffold i and j, respectively). For each scaffold pair ij, a “raw” score is calculated as:

Here sim(i,j) reports the Tanimoto similarity¹⁷ of the two scaffolds in a pair calculated using MACCS structural keys¹⁸ and (1 − sim(i,j)) is a measure of their dissimilarity. Because similarity calculations are only carried out for topologically distinct scaffolds, a topologically insensitive molecular representation such as MACCS keys can be used here. P_Ci and P_Cj are the potency values of compound C_i and C_j and |C_ij| is the total number of all compound pairs representing the scaffold pair. Raw scores are transformed into conventional Z-scores by subtracting the sample mean and dividing standard deviation of the sample of all original raw scores. The Z-scores are then normalized with respect to a cumulative probability function in order to obtain final scores between 0 and 1.

It should be noted that the large-scale analysis of compound data inevitably involves at this stage the risk of comparing IC₅₀ and Ki values, which represents a potential error source. However, for compounds from a series representing an individual scaffold, as used for our raw score calculations, consistent potency measurements are usually reported. In addition, it should also be noted that IC₅₀ values are generally assay-dependent and hence often less reliable than Ki measurement. However, the potency weighting factor emphasizes large potency differences and the score is balanced by multiple pairwise contributions. Furthermore, the raw scores are converted into Z-scores. Taken together, these procedures make the scoring scheme fairly insensitive to limited fluctuations or inaccuracies of potency values.

On the basis of this scoring scheme, scaffold pairs will be prioritized (and obtain scores close to 1) that consist of scaffolds with low similarity yielding comparably potent compounds; identifying such scaffolds is a primary goal of scaffold hopping analysis.⁹ By contrast, it is a priori not desired to facilitate scaffold transitions from highly potent to only weakly potent molecules. Therefore, not only target annotations, but also compound potency should be taken into consideration when assessing scaffold hopping potential on a large scale. For a target set T, the hopping score is then calculated as the median of all normalized scaffold pair scores:

score(T) = median{score_norm(i,j)|i,j ∈ T;i < j}

This score was calculated for the 70 target sets that were then ranked on the basis of decreasing scores, as reported in Table 3. This ranking differed from the one in Table 2 and highest scores were in this case obtained for carbonic anhydrases. Most of the target sets with significant scaffold hopping potential reported in Table 3 contained fewer than 100 scaffolds. Matrix metalloproteases and various GPCRs were also highly ranked. The rankings in Tables 2 and 3 were also combined on the basis of rank fusion. Table 4 shows the top 30 targets organized by increasing sum of ranks. These targets include many popular GPCRs, kinases, and proteases. Hence, on the basis of currently available compound data, these targets have highest scaffold hopping potential.

Table 3 Target sets ranked by scaffold score.^a

Target Name	#Sc	FamilyID	MedianPot	PotRange	Score
a The top 30 target sets ranked according to scaffold hopping scores are reported. For each set, the number of scaffolds (#Sc), median compound potency (MedianPot), potency range (PotRange), and hopping score are reported.
Carbonic anhydrase II	101	10	7.7	3.7	0.849
Carbonic anhydrase IX	84	10	7.4	3.8	0.839
Carbonic anhydrase I	67	10	7.2	3.2	0.744
Matrix metalloproteinase 8	53	8	8.0	4.0	0.741
Cannabinoid receptor 1	84	15	7.6	3.8	0.719
Matrix metalloproteinase 13	76	8	7.9	4.8	0.705
Neurokinin receptor 1	70	15	8.9	4.7	0.698
Estrogen receptor alpha	59	18	7.4	3.6	0.693
Histone deacetylase 1	65	11	7.2	3.0	0.689
Matrix metalloproteinase 2	94	8	7.9	4.0	0.665
Matrix metalloproteinase 9	79	8	7.7	3.6	0.663
Cannabinoid receptor 2	74	15	7.4	3.9	0.660
Estrogen receptor beta	57	18	7.7	3.8	0.659
Matrix metalloproteinase 3	108	8	7.3	3.6	0.628
Norepinephrine transporter	51	17	7.1	3.2	0.584
Matrix metalloproteinase 6	68	15	7.8	3.4	0.568
Acetylcholinesterase	136	19	7.3	5.1	0.567
Dopamine transporter	66	17	7.1	3.4	0.550
Cyclin-dependent kinase 1	80	2	6.9	4.0	0.546
Vascular endothelial growth factor receptor 2	302	1	7.3	3.3	0.545
Histamine H3 receptor	100	15	7.9	4.1	0.538
Beta-secretase 1	89	6	7.4	3.5	0.519
Protein kinase B Akt1	95	2	7.5	3.8	0.517
Alpha-1a adrenergic receptor	73	15	8.3	4.3	0.516
Poly (ADP-ribose) polymerase-1	75	19	7.5	3.0	0.513
Adenosine receptor A3	160	15	7.6	3.9	0.501
Matrix metalloproteinase 1	90	8	7.3	4.0	0.490
Checkpoint kinase	62	2	7.7	3.9	0.486
Cyclin-dependent kinase 2	180	2	7.2	3.5	0.483
Serotonin transporter	110	15	7.9	4.4	0.477

Table 4 Combined target set ranking^a

Target Name	#Sc	FamilyID	MedianPot	PotRange	Score	Rank
Target Name	#Sc	FamilyID	MedianPot	PotRange	Score	Scaffold	Score	Sum
a Target sets are ranked according to the sum of the scaffold number- and scaffold score-based rankings. The top 30 targets are listed. For each set, the number of scaffolds (#Sc), median compound potency (MedianPot), potency range (PotRange), scaffold hopping score, and individual ranks (Scaffold and Score) and sum (SUM) are given.
Vascular endothelial growth factor receptor 2	302	1	7.3	3.3	0.545	2	20	22
Carbonic anhydrase II	101	10	7.7	3.7	0.849	26	1	27
Acetylcholinesterase	136	19	7.3	5.1	0.567	13	17	30
Adenosine receptor A3	160	15	7.6	3.9	0.501	8	26	34
Cyclin-dependent kinase 2	180	2	7.2	3.5	0.483	5	29	34
Matrix metalloproteinase 3	108	8	7.3	3.6	0.628	23	14	37
Carbonic anhydrase IX	84	10	7.4	3.8	0.839	37	2	39
Matrix metalloproteinase 2	94	8	7.9	4.0	0.665	30	10	40
Cannabinoid receptor 1	84	15	7.6	3.8	0.719	37	5	42
Cathepsin K	145	7	7.6	5.5	0.464	12	32	44
Histamine H3 receptor	100	15	7.9	4.1	0.538	27	21	48
Cathepsin S	129	7	7.4	3.9	0.467	17	31	48
Src tyrosinekinase	174	1	7.3	3.8	0.406	6	42	48
Melanin-concentrating hormone receptor 1	318	15	7.6	4.0	0.397	1	48	49
Matrix metalloproteinase 13	76	8	7.9	4.8	0.705	44	6	50
Thrombin	162	9	7.1	6.0	0.404	7	44	51
Protein kinase B Akt1	95	2	7.5	3.8	0.517	29	23	52
Serotonin transporter	110	15	7.9	4.4	0.477	22	30	52
Mitogen-activated proteinp38 alpha	129	2	7.6	4.3	0.436	17	36	53
Factor Xa	187	9	7.9	5.3	0.395	4	49	53
Matrix metalloproteinase 9	79	8	7.7	3.6	0.663	43	11	54
Endothelial growth factorreceptor	134	1	7.3	5.5	0.435	15	39	54
Neurokinin receptor 1	70	15	8.9	4.7	0.698	49	7	56
Carbonic anhydrase I	67	10	7.2	3.2	0.744	54	3	57
Beta-Secretase 1	89	6	7.4	3.5	0.519	35	22	57
Cannabinoid receptor 2	74	15	7.4	3.9	0.660	46	12	58
Leukocyto-specific tyrosinekinase	106	1	7.9	5.0	0.436	24	36	60
Cyclin-dependent kinase 1	80	2	6.9	4.0	0.546	42	19	61
Matrix metalloproteinase 1	90	8	7.3	4.0	0.490	34	27	61
Dipeptidyl peptidase 4	119	9	7.5	4.0	0.408	20	41	61

Vascular endothelial growth factor receptor-2 is the top-ranked target in Table 4 followed by carbonic anhydrase II. Fig. 3 shows scaffold pairs for these targets that yield high or low hopping scores. The top-scoring scaffold pairs display an astonishing degree of structural diversity, whereas low-scoring pairs are involved in close structural relationships. These observations are representative for many target sets that were found to contain a spectrum of topologically distinct scaffolds, ranging from closely related to virtually unrelated structures.


	Fig. 3 Highly ranked target sets. Scaffold pairs with high (Top) and low (Bottom) hopping scores are shown for two top-ranked target sets; (a) vascular endothelial growth factor receptor 2 ligands and (b) carbonic anhydrase II inhibitors. For each set, three high scoring and two low scoring scaffold pairs are shown. For each scaffold, the median potency of the compounds it represents is reported. For each scaffold pair, the hopping score and MACCS Tanimoto similarity are reported. For example, 1/0.17 means that the scaffold pair has score of 1 and their Tanimoto similarity is 0.17. For low-scoring scaffold pairs, structural differences are highlighted.

Finally, we also determined scaffold overlap between different target sets. The results are reported in Fig. 4 as a scaffold-based target network (drawn with Cytoscape¹⁹). Sixty of the 70 target sets shared one or more scaffolds with others. A total of 142 pair-wise target set relationships were detected among the 70 target sets; 106 of these relationships were formed exclusively within target families and 36 across different families. Substantial scaffold overlap between target sets was observed within the GPCR, kinase, and matrix metalloprotease target families. By contrast, inter-target family scaffold overlap was rather limited. These 142 relationships involved 1,298 scaffolds of a total of 5,232 scaffolds contained in the 70 target sets, i.e. ∼25%. Hence, scaffold overlap was generally limited and the majority of scaffolds belonged to individual target sets.


	Fig. 4 Scaffold-based target network. Scaffold overlaps between target sets are viewed in a network representation. Nodes represent target sets that are connected by an edge if they share one or more scaffolds. The width of edges is scaled by scaffold numbers. Nodes are colored to reflect target family membership and their size is scaled by median scaffold hopping scores.

In summary, in order to better understand how frequently scaffold hops occur in compounds active against different targets, we have systematically derived topologically distinct scaffolds for sets of compounds representing 502 targets belonging to 19 target families. The occurrence of different scaffolds in target sets provides an estimate for the likelihood that scaffold hops can be identified for given targets. In 354 of our target sets, between five and 49 distinct scaffolds were detected, providing a range for average scaffold hopping frequency. In 70 target sets, between 50 and 318 different scaffolds were found. A subset of these scaffolds was structurally highly diverse but yielded similarly potent compounds, thus meeting “ideal” scaffold hopping criteria. However, many other scaffolds (on average ∼60% of all scaffolds in a target set) displayed well-defined substructure relationships. Thus, in these cases, it is not surprising that similarity-based virtual screening methods often display scaffold hopping potential (although scaffold hopping ability is usually considered the ultimate “proof” that a computational screening method is useful). By contrast, identifying scaffolds that are truly distinct is much more difficult, given the observed distributions of structurally related and unrelated scaffolds. However, on the basis of our analysis, we conclude that there is considerable scaffold hopping potential across the spectrum of currently available targets. Thus, searching for structurally diverse active compounds should be promising in many cases.

References

J. Brown and E. Jacoby, Mini-Rev. Med. Chem., 2006, 6, 1217–1229 CrossRef.
H. Zhao, Drug Discovery Today, 2007, 12, 149–155 CrossRef CAS.
S. Renner and G. Schneider, ChemMedChem, 2006, 1, 181–185 CrossRef CAS.
E. J. Barker, D. Buttar, D. A. Cosgrove, E. J. Gardiner, V. J. Gillet, P. Kitts and P. Willett, J. Chem. Inf. Model., 2006, 46, 503–511 CrossRef CAS.
K. Tsunoyama, A. Amini, M. J. E. Sternberg and S. H. Muggleton, J. Chem. Inf. Model., 2008, 48, 949–957 CrossRef CAS.
N. Wale, I. A. Watson and G. Karypis, J. Chem. Inf. Model., 2008, 48, 730–741 CrossRef CAS.
S. Senger, J. Chem. Inf. Model., 2009, 49, 1514–1524 CrossRef CAS.
M. Vogt, D. Stumpfe, H. Geppert and J. Bajorath, J. Med. Chem., 2010, 53, 5707–5715 CrossRef CAS.
H. Geppert, M. Vogt and J. Bajorath, J. Chem. Inf. Model., 2010, 50, 205–216 CrossRef CAS.
MOE (Molecular Operating Environment); Chemical Computing Group Inc.: Montreal, Quebec, Canada, 2007 Search PubMed.
Scitegic Pipeline Pilot, Student Edition, Version 6.1; Accelrys, Inc.: San Diego, CA, 2007 Search PubMed.
CHEMBLdb. http://www.ebi.ac.uk/chembl/(accessed May 11, 2010).
T. Liu, Y. Lin, X. Wen, R. N. Jorissen and M. K. Gilson, Nucleic Acids Res., 2007, 35, D198–D201 CrossRef CAS.
G. W. Bemis and M. A. Murcko, J. Med. Chem., 1996, 39, 2887–2893 CrossRef CAS.
Y. Hu, A. M. Wassermann, E. Lounkine and J. Bajorath, J. Med. Chem., 2010, 53, 752–758 CrossRef CAS.
Y. Hu and J. Bajorath, ChemMedChem, 2010, 5, 1681–1685 CrossRef.
P. Willett, J. M. Barnard and G. M. Downs, J. Chem. Inf. Comput. Sci., 1998, 38, 983–996 CrossRef CAS.
MACCS Structural Keys; Symyx Software: San Ramon, CA, 2005 Search PubMed.
P. Shannon, A. Markiel, O. Ozier, N. S. Baliga, J. T. Wang, D. Ramage, N. Amin, B. Schwikowski and T. Ideker, Genome Res., 2003, 13, 2498–2504 CrossRef CAS.