Extracting SAR Rules from Compound Data

Extraction of SAR Rules from
Compound Data
Jürgen Bajorath
Department of Life Science Informatics
LIMES Program Chemical Biology and Medicinal Chemistry
University of Bonn
How to extract SAR information from
compound data sets?
Systematically
With the aid of graphical representations

Basic SAR Concepts
6 nM 6 nM
SAR continuity
distinct structures with
similar potency
SAR discontinuity
similar structures with
highly different potency
(“activity cliff”)
continuity
discontinuity
2.3 µM
Concept of Activity Landscapes
“Activity landscapes”: biological activity hypersurfaces
within chemical space;
visualized as a 2D projection of chemical space with
compound potency as the third dimension

Idealized Activity Landscapes and SARs
Continuous SAR Discontinuous SAR
gradual changes in
structure result in moderate
changes in activity
“rolling hills”
Basic SAR Concepts
small changes in
structure have dramatic
effects on activity
“activity cliffs”
6 nM 6 nM
Coexistence of
continuous and
discontinuous SAR
components:
SAR heterogeneity
corresponding to
variable activity landscapes
continuity
discontinuity
2.3 µM

y
Variable Activity Landscapes
Cathepsin S inhibitors
potency
- not idealized, but calculated -
“Coordinate-free” chemical
space (MACCS Tanimoto
coefficient distances)
2D projection through multidimensional
scaling
xy-plane: MACCS
Tanimoto similarity-based
projection
z-axis: interpolation of
potency values
color code: surface
elevation
Activity Landscapes and SARs
Cathepsin S inhibitors
x
2D projection of an activity
landscape
points represent molecules
color: potency
(red: high, green: low)
area shaded according to
interpolated potency

Activity Landscapes and SARs
Cathepsin S inhibitors
Cathepsin S inhibitors
2D vs. 3D landscape
representation
Activity Landscapes and SARs
0.1 nM
Activity cliff formed by highly
and weakly potent molecules
10 μM

Systematic SAR Analysis
What do we like to learn ?
Activity cliffs
SAR microenvironments
subsets of compounds
representing different local SARs
Graph Representations
Discontinuous SAR
components
lead optimization
Continuous SAR
components
QSAR, lead hopping
Basic data
A list of compounds
with potency values
pairwise comparison

Graph Representations
⎛ 1 ⎞
cont = weighted mean ⎜ ⎟
{ i , j i > j } ⎝ 1 + sim (i,j) ⎠
P i ⋅ P j
weight ij =
1 + P − P
i
j
disc =
1
SARI −
2
= ( cont − ( 1 disc ) )
Common features
edges determined by
2D structural similarity
potency used as
node annotation
SAR Index Scoring - Annotation
Numerical function to characterize SAR features
continuity score
emphasizes structurally
diverse compounds having
similar potency
balances
two parts
GLOBAL SCORE
all possible compound pairs
mean P i j
⎪⎧
⎪⎧
i > j , P Pi
Pj
1,
⎪⎫
i − −Pj
> 1,
⎪⎫
⎨ i , j
⎬
⎪⎩ ⎪⎩ sim( i , j ) > 0 . 65 65⎪⎭
⎪⎭
( − P ⋅ sim( i , j ) )
(P, potency; sim,
pairwise 2D similarity)
discontinuity score
emphasizes similar
compounds with large
potency differences

SAR Index Scoring
Numerical function to characterize SAR features
node size scaling
high compound score
low compound score
reflects the potency
deviation of a compound
from its structurally
similar neighbors
LOCAL SCORE
pairs formed by a given compound
disc
( )
i j
= mean P i − P
{ j sim( i , j ) > t , i ≠ j }
( ⋅ sim ( i , j ) )
discontinuity score
emphasizes similar
compounds with large
potency differences
Network-like Similarity Graph (NSG)
Exemplary graphical
SAR analysis method

Network-like Similarity Graph
Network-like Similarity Graph
NSG for a set of
squalene synthese
inhibitors
Annotated graph
representation of
similarity relationships
in compound data sets
Nodes: represent all
compounds in the data set
Edges: connect nodes with
high pairwise similarity
Clusters: Ward’s hierachical
clustering (gray background)
Layout: Fruchterman-Reingold
Annotated graph
representation of
similarity relationships
in compound data sets
Annotations:
node size
cluster scores
global scores

NSG – Score Annotations
node size compound compound compound compound
discontinuity discontinuity discontinuity discontinuity score score score score
cluster scores SAR SAR SAR SAR Index Index Index Index for for for for
compound compound compound compound clusters clusters clusters clusters
global scores SAR SAR SAR SAR Index Index Index Index for for for for the the the the entire entire entire entire
compound compound compound compound set set set set
highlights compounds that
introduce SAR discontinuity/
activity cliffs in a data set
indicates the level of
continuity/discontinuity in a
group of similar compounds
indicates the level of
continuity/discontinuity in the
data set
NSG Interpretation - Local SAR Features

Activity Cliff Index
NSGs provide interactive
graphical access to prominent
activity cliffs
Cliff Index (CI) enables
systematic mining and ranking
of activity cliffs
CI prioritizes pairs of similar
compounds having large
potency differences:
( ) i j P P j i ⋅
+
2
1 sim( , )
j i −
= CI( , )
CI = 15.2
SAR Pathways from NSGs
Pathways are annotated with compound
discontinuity scores to emphasize compounds
forming activtiy cliffs
7 μM
CI = 13.4
activity cliff marker
1 μM
0.015 nM
A sequence of pairwise
similar compounds with
balanced chemical and
activity similarity
potency
increases from start to end
node
potency gradient
smooth gradients are preferred

SAR Pathways
Cytochrome P450
2C19 PubChem
screening data set
Preferred pathways with:
pairwise similar compounds
scaffold hop
Pathway SAR Model
small increase in potency per compound
large potency difference between start- and endpoint
smooth potency gradient
many compounds
deviation deviation from a linear
potency increase
number number number of of compounds compounds in
the pathway
Pathways are based on
a predefined SAR model
potency potency
potency
difference
difference
between
start- and
endpoint

SAR Trees
Cytochrome P450
2C19 PubChem
screening data set
SAR Trees
SAR Trees provide a
structural context for
individual pathways
Activity cliff pathways
can be monitored
A set of pathways
organized in a tree
root
all pathways begin (or lead to)
the same compound
branches
identical pathway sections are
fused into one branch
leaves
endpoints of potency gradients
(highest/lowest potent
compounds)
activity cliff

Advanced Application: Studying
Multi-target SARs Using NSGs
NSGs can also be utilized
to compare SAR behavior
for multiple targets
Node color reflects
compound selectivity
instead of potency
From Activity Cliffs to Selectivity Cliffs
Multi-target SARs
Target-pair selectivity:
difference between
logarithmic potency
SA / B(
i)
= −SB
/ A = PA(
i)
− PB
( i)
Structure-selectivity
relationships (SSRs)
cathepsin L
cathepsin B
pIC50 = 9 pIC50 = 7
S L/B = 2

SAR and SSR Network Analysis
cathepsin L
0.48
0.05
Potency-based NSG
Potency:
10.4 3.0
Compound discontinuity score:
1
0
activity cliff markers
Cluster discontinuity score
SAR and SSR Network Analysis
cathepsin L
0.48
0.05
cathepsin B
1
0
“rough” SAR
“smooth” SAR
0.10
0.27

SAR and SSR Network Analysis
cathepsin L /
cathepsin B
0.73
0.72 1
Local SSR Environments
cathepsin L /
cathepsin B
0.73
0.72
Selectivity-based NSG
Selectivity:
3.2 (L) – 3.2 (B)
Compound discontinuity score:
1
0
selectivity cliff markers
Cluster discontinuity score
0
“rough” SSR
“smooth” SSR
discontinuous SSR

Activity Cliffs vs. Selectivity Cliffs
L B L/B
discontinuous SAR continuous SAR
L: 15 nM
B: 3.5 μM
L/B: 1.4
discontinuous SSR
activity cliff markers selectivity cliff markers
Local SSR Environments
cathepsin L /
cathepsin B
0.73
0.72
L: 10 μM
B: 170 nM
L/B: -1.8
discontinuous SSR

Activity Cliffs vs. Selectivity Cliffs
L B L/B
continuous SAR continuous SAR
L: 3.6 μM
B: 102 μM
L/B: 1.5
Selectivity Determinants
L/B
discontinuous SSR
selectivity cliff markers
L: 26 μM
B: 5.3 μM
L/B: -0.7
Molecules with different selectivity are found in the neighborhood of
selectivity cliff markers
Selectivity rules can be formulated
sel: -0.7
sel: 0.1
halogens with increasing
bulkiness and decreasing
electronegativity shift
selectivity toward cat L
sel: 1.5
sel: 2.0

Selectivity Determinants
sel: 2.3
Conclusions
K/L
bulkier substituents shift
selectivity towards cat L
sel: -0.6
sel: -0.8
sel: -0.8
Numerical and graphical analysis tools are developed for
mining of SAR information in compound data sets
Annotated similarity-based compound networks play an
important role for graphical SAR analysis
NSGs enable a systematic comparison of global and local
SAR features in compound data sets and the identification
of activity cliffs
SAR Trees are based on pre-defined SAR model
NSG enable a comparative analysis of multi-target SARs