1) The document discusses methods for setting up similarity-driven virtual screening using various molecular similarity metrics and descriptor spaces. 2) It finds that traditional dogmas like only using Tanimoto similarity above 0.85 can be inaccurate, and recommends calibrating similarity cutoffs specifically for each target, query, and chemical space. 3) Tversky similarity with an alpha value of 0.7-0.9, which more heavily penalizes the query missing features of actives, is found to often give excellent results. The best approach is to test multiple options and calibrate for each individual virtual screening project.

1. Dealing with ‘exotic’ similarity
metrics
How to set up a (ChemAxon-powered)
Similarity-driven Virtual Screening server…
Dragos Horvath, dhorvath@unistra.fr
UMR 7140 CNRS – Université de Strasbourg

2. Introduction & Definitions
• Similarity-based Virtual Screening (SVS):
– Search, in a database of candidates m for similar analogues
of a query compound M of wanted properties, hoping that
the « similarity principle » magic will operate.
• Molecular Similarity S(M,m):
– distance (metric) between the two Descriptor Space (DS)
points 𝐷 𝑀 , 𝐷 𝑚 - let us call these 𝐷, 𝑑, for simplicity.
• Similarity Radius s defines « how similar is similar »
– Delimits a sphere in descriptor space around M, thought to
contain a minimum of inactive, but a maximum of active m.
• Virtual Hits – aka True & False « Positives » (TP,FP):
– Compounds m with S(M,m)<s

3. Compound Sets
• For server calibration:
– Candidate database: 165 ChEMBL ligand sets with >50
molecules of reported pKi values with respect to the 165
associated receptors & enzymes (targets T).
– Queries of T: MT
1, MT
2 … MT
i, i=1..QT is composed of the top
1/5 (max 100) actives on T, plus 1/5 (max 100) of binders of
medium potency, can be classified by pharmacophore
complexity (Nr. of populated FPT1 triplets)
– 10,000 randomly picked commercial molecules from ZINC,
assumed to be inactive “decoys”.
• Operational database:
– 1.5 M commercial compounds, from various sources
– Above « reference » molecules, for annotation purposes

4. Compound Sets

5. Compound Sets

6. Descriptor Spaces

7. Descriptor Spaces

8. All are Feature Counts
Di(M) = integer (positive
or null) population level
of « feature » i (a
substructure or a
pharmacophore triplet)
in molecule M
Descriptor Spaces

9. Dissimilarity Scores…
• Based on the comparison of descriptor vectors 𝐷, 𝑑
𝑁 𝑀 𝑁𝑂𝑅𝑀(𝑀) = 𝐷𝑖
2
𝑁(𝑀)
𝑖=1
𝑁𝐴𝑁𝐷 𝑚, 𝑀 𝐴𝑁𝐷(𝑚, 𝑀) = 𝐷𝑖 × 𝑑𝑖
𝑁 𝑂𝑅(𝑚,𝑀)
𝑖=1
𝑁𝐸𝑋𝐶 𝑀, 𝑚 𝐸𝑋𝐶(𝑀, 𝑚) = 𝐷𝑖
2
𝑖|𝑑 𝑖=0

10. Euclidean & Related…
𝐸 𝑚, 𝑀 = 𝐷𝑖 − 𝑑𝑖
2
𝑁 𝑂𝑅(𝑚,𝑀)
𝑖=1
𝑅 𝑚, 𝑀 =
𝐷𝑖 − 𝑑𝑖
2𝑁 𝑂𝑅(𝑚,𝑀)
𝑖=1
𝑁 𝑂𝑅(𝑚, 𝑀)
𝐴 𝑚, 𝑀 =
𝐷𝑖 − 𝑑𝑖
𝑁 𝑂𝑅(𝑚,𝑀)
𝑖=1
𝑁 𝑂𝑅(𝑚, 𝑀)
𝑅𝑊 𝑚, 𝑀 = 𝑅 𝑚, 𝑀
𝑁 𝐸𝑋𝐶(𝑚, 𝑀) + 𝑁𝐸𝑋𝐶(𝑀, 𝑚)
𝑁 𝑂𝑅(𝑚, 𝑀)
𝐴𝑊 𝑚, 𝑀 = 𝐴 𝑚, 𝑀
𝑁 𝐸𝑋𝐶(𝑚, 𝑀) + 𝑁𝐸𝑋𝐶(𝑀, 𝑚)
𝑁 𝑂𝑅(𝑚, 𝑀)

11. (A)Symmetric Correlation Scores –
Tanimoto & Tversky
𝑇 𝑀, 𝑚 = 1 −
𝐴𝑁𝐷(𝑀, 𝑚)
𝑁𝑂𝑅𝑀(𝑀) + 𝑁𝑂𝑅𝑀(𝑚) − 𝐴𝑁𝐷(𝑀, 𝑚)
𝑇𝑣 𝑀, 𝑚, 𝛼 = 1 −
𝐴𝑁𝐷(𝑀, 𝑚)
𝛼𝐸𝑋𝐶(𝑀, 𝑚) + 1 − 𝛼 𝐸𝑋𝐶(𝑚, 𝑀) + 𝐴𝑁𝐷(𝑀, 𝑚)
Situations where:
(a) candidate m misses a feature seen in active
M, and
(b) it contains some novel feature not seen in M
may be distinguished! At a>0.5, cases (a) will be
relatively more penalized than the symmetric
situation (b).
A raw guess of a should suffice! Three
implementations of Tv are considered:
• Tv+ (a=0.9)
• Tv (a=0.7)
• Tv- (a=0.3)

12. 2. Fine, but « how similar is similar »?
• You may be a believer of the dogma « Tanimoto>0.85 »
(T<0.15)
– But the Bible mentions not the other metrics, less subjected
to religious fervor.
• Alternatively, try to infer reasonable choices of
similarity radii for each Chemical Space (CS – the
combination of Descriptor Space & Similarity score)
– For each query, on every target, compute s* corresponding
to the « optimal » SVS scenario.
– This also allows to measure & benchmark SVS success with
respect to its Operational Premises (CS, nature of Target,
degree of complexity of the query, etc).

13. s
W
1.0
)()(
)()(
)( E
FN
E
FP
FNFP
NN
NN
s


W
SS
S


A basic SVS Optimality Criterion: W
L(M,m) l L(M,m)> l
S(M,m)s
True
Positives
(TP)  
False
Positives
(FP) 
False (?)
Negatives
(FN) 
True
Negatives
(TN) 
)()(
)()(
)( E
FN
E
FP
FNFP
NN
NN
s


W
SS
S


s
Activity (profile) differences L(m,M)
Λ 𝑀, 𝑚 =
0 𝑖𝑓 𝑝𝐾𝑖(𝑀) − 𝑝𝐾𝑖(𝑚) < 0.5
1 𝑖𝑓 𝑝𝐾𝑖(𝑀) − 𝑝𝐾𝑖(𝑚) > 3.0
𝑝𝐾𝑖(𝑀) − 𝑝𝐾𝑖(𝑚) − 0.5
2.5
𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

14. s
W
1.0
)()(
)()(
)( E
FN
E
FP
FNFP
NN
NN
s


W
SS
S


A basic SVS Optimality Criterion: W
L(M,m) l L(M,m)> l
S(M,m)s
True
Positives
(TP)  
False
Positives
(FP) 
False (?)
Negatives
(FN) 
True
Negatives
(TN) 
)()(
)()(
)( E
FN
E
FP
FNFP
NN
NN
s


W
SS
S


s
Activity (profile) differences L(m,M)

15. The Ascertained Optimality Excess X
Compound Pairs selected at cutoff s
Random S
values
Meaningful
S values
Var(W)W
X
Fraction of Compound Pairs selected at cutoff s
     sVars randrand
WWWX
X

16. Workflow
ForEach Target T
Set database db=set of tested ligands (known pKi ) + decoy set (pKi=0);
ForEach Query M of T
ForEach DescriptorSpace D
ForEach SimilarityScore S
# Start Current SVS experiment defined by Target, Query, Descriptors & Similarity Score
ForEach m!=M in db
Calculate S(M,m)|D ;
EndLoop(m)
Scan over s → X(s) and return s* such that X(s*)=maximal;
Classify SVS(T,M,D,S) wrt X(s*) as « failed », « acceptable », « good » or « excellent » ;
EndLoop S;
EndLoop D;
EndLoop M;
EndLoop T;
Analyze Success Rates & s* distributions in terms of various Operational premises (nature of T,
complexity of M, choice of D, of S or of D-S combinations)

17. Insights: (1) – So much for dogmas!
0
5
10
15
20
25
30
35
40
0.04 0.08 0.12 0.16 0.2 0.24 0.28 0.32 0.36 0.4 0.44 0.48 0.52 0.56 0.6
%ofTanimoto-basedqueriesofGoodoptimalitylevelatd*
d*
FPT1
treeSY03
s*
Use distribution to
« teach » the web
server how to
rank prospective
SVS hits!
Top Hits (0)
Good
Hits (1)
Average
Hits (2)
Acceptable
Hits (3)
Are these Hits? (4) Ignore…
Top Hits (0)
Good
Hits (1)
Average
Hits (2)
Acceptable
Hits (3)
Are these Hits? (4) Ignore…

18. Insights: (2) – Tversky at a>0.5: an
excellent similarity scoring scheme.
0
2
4
6
8
10
12
14
16
Tv+ Tv T RW AW Tv- E A R
R:all-acceptable
R:all-good
R:all-excellent
Relative«marketshare»ofmetric:fractionofSVSrunsbasedonshown
metric,outofallSVSexperimentshavingreachedgivensuccesslevels

19. Tv+ may pick actives that are more
complex than queries (NK1 example)
T

20. Insights: (3) – Trends with respect to
target classes could be evidenced…
Relative«marketshare»ofmetric:fractionofSVSruns–withintargetclasses-based
onshownmetric,outofallSVSexperimentshavingreachedgivensuccesslevels
0
2
4
6
8
10
12
14
16
Tv+ Tv T RW AW Tv- E A R
R:all-good
R:Kinases-good
R:monoamineGPCR-good
R:otherGPCR-good

21. Insights: (4) – when the query compound is
complex, the metric matters less
Relative«marketshare»ofmetric:fractionofSVSruns–withinquerycomplexity
classes-basedonshownmetric,outofallSVShavingreachedgivensuccesslevels
0
2
4
6
8
10
12
14
16
Tv+ Tv T RW AW Tv- E A R
R:all-good
R:pharm-high-good
R:pharm-low-good

22. Some conclusions
• The study has highlighted many interesting aspects
– Intrinsic usefulness of Tversky scores biased towards of query feature loss
penalty: a=0.9…0.7 will do!
– Other target-, query complexity-, query activity-, descriptor space-dependent
trends of the SVS success
– Some inevitable sources of bias, showing that not even ChEMBL is not
large/diverse enough to cover it all…
• Main message: use this protocol – or related – to calibrate web
servers, rather than sticking to well-studied metrics and descriptors
for which « Universal » similarity cutoffs are believed to hold.
• Try infochim.u-strasbg.fr/webserv/VSEngine.html – to our
knowledge, the only public SVS server to support atypical, but
powerful metrics coupled to chemically relevant, pH-sensitive
descriptor spaces… all while exploiting the power of ChemAxon
tools!