Transcript Drug Design Methods II: SVM

CZ3253: Computer Aided Drug design
Lecture 7: Drug Design Methods II: SVM
Prof. Chen Yu Zong
Tel: 6874-6877
Email: [email protected]
http://xin.cz3.nus.edu.sg
Room 07-24, level 7, SOC1,
National University of Singapore
Classification of Drugs by SVM
•
A drug is classified as either belong (+) or not belong (-) to a class
Examples of drug class: inhibitor of a protein, BBB penetrating, genotoxic
Examples of protein class: enzyme EC3.4 family, DNA-binding
•
By screening against all classes, the property of a drug or the function of a
protein can be identified
Drug
Class-1
SVM
-
Class-2
SVM
-
Class-3
SVM
+
-
Drug
belongs to
Family-3
2
Classification of Drugs or Proteins by SVM
What is SVM?
• Support vector machines, a machine learning method, learning by
examples, statistical learning, classify objects into one of the two
classes.
Advantages of SVM:
• Diversity of class members (no racial discrimination).
• Use of structure-derived physico-chemical features as basis for drug
classification (no structure-similarity required in the algorithm).
3
SVM References
•
C. Burges, "A tutorial on support vector machines for pattern recognition",
Data Mining and Knowledge Discovery, Kluwer Academic Publishers,1998
(on-line).
•
R. Duda, P. Hart, and D. Stork, Pattern Classification, John-Wiley, 2nd
edition, 2001 (section 5.11, hard-copy).
•
S. Gong et al. Dynamic Vision: From Images to Face Recognition, Imperial
College Pres, 2001 (sections 3.6.2, 3.7.2, hard copy).
•
Online lecture notes
(http://www.cs.unr.edu/~bebis/MathMethods/SVM/lecture.pdf )
•
Publications of SVM drug prediction:
– J. Chem. Inf. Comput. Sci. 44,1630 (2004)
– J. Chem. Inf. Comput. Sci. 44, 1497 (2004)
– Toxicol. Sci. 79,170 (2004).
4
Machine Learning Method
Inductive learning:
Example-based learning
Descriptor
Positive
examples
Negative
examples
5
Machine Learning Method
Feature vectors:
A=(1, 1, 1)
B=(0, 1, 1)
C=(1, 1, 1)
D=(0, 1, 1)
E=(0, 0, 0)
F=(1, 0, 1)
Descriptor
Feature vector
Positive
examples
Negative
examples
6
SVM Method
Feature vectors in input space:
Z
Input space
Feature vector
A=(1, 1, 1)
B=(0, 1, 1)
C=(1, 1, 1)
D=(0, 1, 1)
E=(0, 0, 0)
F=(1, 0, 1)
F
E A
B
Y
X
7
SVM Method
Protein family
members
Border
New border
Protein family
members
Nonmembers
Nonmembers
Project to a higher dimensional space
8
SVM method
New border
Support vector
Support vector
Protein family
members
Nonmembers
9
SVM Method
Support vector
Protein family
members
Nonmembers
New border
Support vector
10
Best Linear Separator?
11
Best Linear Separator?
12
Find Closest Points in Convex
Hulls
d
c
13
Plane Bisect Closest Points
x wb
w  d c
d
c
14
Find using quadratic program
min
1
2
c    i xi
i1
s.t.

i1
i
1
i  0
cd
d 
2
 x
i1

i1
i
i
i
1
i  1,..., 
Many existing and new solvers.
15
Best Linear Separator:
Supporting Plane Method
Maximize distance
Between two paral
supporting planes
x w  b 1
x w  b 1
Distance
= “Margin”
= 2
|| w ||
16
Best Linear Separator?
17
SVM Method
Border line is nonlinear
18
SVM method
Non-linear transformation: use of kernel function
19
SVM method
Non-linear transformation
20
SVM Method
21
SVM Method
22
SVM Method
23
SVM Method
24
SVM for Classification of Drugs
How to represent a drug?
•
Each structure represented by specific feature vector assembled from
structural, physico-chemical properties:
– Simple molecular properties (molecular weight, no. of rotatable bonds
etc. 18 in total)
– Molecular Connectivity and shape (28 in total)
– Electro-topological state polarity (84 in total)
– Quantum chemical properties (electric charge, polaritability etc. 13 in
total)
– Geometrical properties (molecular size vector, van der Waals volume,
molecular surface etc. 16 in total)
J. Chem. Inf. Comput. Sci. 44,1630 (2004)
J. Chem. Inf. Comput. Sci. 44, 1497 (2004)
Toxicol. Sci. 79,170 (2004).
25
SVM Feature Selection
CACO2 - 718 descriptors
Average of 10 Models
Predicted RT (min)
-3
-4
-5
Q2 is MSE scaled by
variance:
-6
-7
= (mean square error) /
Test Q2 = .7073 (true variance)
-8
-8
-7
-6
-5
-4
Observed RT (min)
-3
26
Feature Selection
Using subset of descriptors might greatly
improve results.
• Do feature selection using
Linear SVM with 1-norm regularization
1-norm
2-norm
27
Feature Selection via
Sparse SVM/LP
• Construct linear -SVM using 1-norm LP:
C
*
min *
z

z
  i i   C  || w ||1
w , b , , z , z
s.t
i 1
 xi  w  b  yi   zi  
*
x

w

b

y

z
 i
i
i 
zi , zi* ,   0 i  1,..,
• Pick best C, for SVM
• Keep descriptors
| wi |
with nonzero coefficients
0
28
Bagged Feature Selection
Partition Training Data
Training Set
Validation Set
Linear SVM Algorithm
For Feature Selection
Random Variable - r
A Linear Regression Model
Repeat B times
Bag B Models and Obtain Subset of Features
Make 20 models of the form
w  x - b  w1  x 1  w2  x 2  ...  w718  x 718  wr r  b
with only a few wi  0
Keep attributes with w i  w r
29
Bagged SVM (RBF)
CACO2 - 31 Descriptors
-3
Predicted RT (min)
-4
-5
-6
-7
Test Q2 = .134
-8
-8
-7
-6
-5
-4
-3
Observed RT (min)
30
Starplot Caco2 - 31 Descriptors
ABSDRN6
a.don
KB54
SMR.VSA2
BNP8
DRNB10
DRNB00
KB11
PEOE.VSA.4
PEOE.VSA.FPPOS
ANGLEB45
PIPB53
SlogP.VSA6
apol
ABSFUKMIN
PIPB04
PEOE.VSA.FPOL
PIPMAX
PEOE.VSA.FHYD
PEOE.VSA.PPOS
EP2
PEOE.VSA.FNEG
SlogP.VSA0
BNPB31
FUKB14
BNPB50
SlogP.VSA9
pmiZ
BNPB21
ABSKMIN
SIKIA
31
Chemistry In/Out Modeling
Data
+Descriptors
Feature Selection
Test Data
Visualize Features
Assess Chemistry
Construct SVM
Nonlinear model
SVM Model
Chemistry
Interpretation
Predict bioactivities
32
Bagged SVM (RBF)
CACO2 - 15 Descriptors
-3
Predicted RT (min)
-4
-5
-6
-7
Test Q2 = .166
-8
-8
-7
-6
-5
-4
-3
Observed RT (min)
33
CACO2 – 15 Variables
a.don
DRNB10
PEOE.VSA.FNEG
BNPB31
KB54
ABSDRN6
ABSKMIN
FUKB14
SMR.VSA2
PEOE.VSA.FPPOS
SIKIA
SlogP.VSA0
ANGLEB45
DRNB00
pmiZ
34
Chemical Insights
•
•
Hydrophobicity - a.don
SIZE and Shape
ABSDRN6, SMR.VSA2, ANGLEB45, PmiZ
Large is bad. Flat is bad. Globular is good.
• Polarity –
PEOE.VSA.FPPOS, PEOE.VSA.FNEG:
negative partial charge good.
Correspond to conventional wisdom – rule of 5.
35
Hybrid TAE/SHAPE
•
Shape important overall factor
– DRNB10, DRNB00: del rho dot N
– BNP31: bare nuclear potential
– KB54: kinetic energy descriptors
very large lipophilic molecules don’t work
– FUKB14: Fukui Surface
•
•
Interpretations difficult
Point to chemistry challenges/hypotheses
36
Final SVM Approach
• Construct large set of descriptors.
• Perform feature selection:
– Sensitivity Analysis or SVM-LP
• Construct many SVM models
– Optimize using QP or LP
– Evaluate by Validation Set or Leave-one-out
– Select best models by grid or pattern search
• Bag best k models to create final function
37
Drug Discovery Results (LOO)
Data
#
Sampl
e
# Var.
Full
# Var.
FS (Avg)
Q2
Full
Q2
FS
Caco2
27
713
41
0.33
0.29
Barrier
62
569
51
0.31
0.28
HIV
64
561
17
0.46
0.40
Cancer
46
362
34
0.50
0.16
LCCK
66
350
69
0.40
0.37
Aquasol
197
525
57
0.08
0.06
38
SVM-based drug design and property prediction software
Useful for inhibitor/activator/substrate prediction, drug safety and
pharmacokinetic prediction.
Drug
Chemical
Structure
Option 1
Chemical
Structure
Your drug
structure
Option 2
http://jing.cz3.nus.edu.sg/cgi-bin/svmprot.cgi
Which class your
drug belongs to?
Send structure to classifier
Input structure
through internet
Computer loaded
with SVMProt
Input structure
on local machine
Drug designed
or property
predicted
Support vector machines
classifier for every
Drug class
Identified
classes
J. Chem. Inf. Comput. Sci. 44,1630 (2004)
J. Chem. Inf. Comput. Sci. 44, 1497 (2004)
Toxicol. Sci. 79,170 (2004).
SVM Drug Prediction Results
Protein inhibitor/activator/substrate prediction:
•
•
86% of the 129 estrogen receptor activators and 84% of 101 non-activators correctly
predicted.
81% of 116 P-glycoprotein substrates and 79% of 85 non-substrates correctly
predicted
Drug Toxicity Prediction:
•
•
97% of 102 TdP+ and 84% of 243 TdP- agents correctly predicted
73% of 229 genotoxic and 93% of 631 non-genotoxic agents correctly predicted
Pharmacokinetics prediction:
•
•
95% of 276 BBB+ and 82% of 139 BBB- agents correctly predicted
90% of 131 human intestine absorption and 80% of 65 non-absoption agents
correctly predicted.
J. Chem. Inf. Comput. Sci. 44,1630 (2004)
J. Chem. Inf. Comput. Sci. 44, 1497 (2004)
Toxicol. Sci. 79,170 (2004).