Transcript Document

Lowe et al., Molec. Pharmaceutics, 7, 1708 (2010)
Predicting Phospholipidosis Using
Machine Learning
•
•
•
•
•
Robert Lowe (Cambridge)
John Mitchell (St Andrews)
Robert Glen (Cambridge)
Hamse Mussa (Cambridge)
Florian Nigsch (Novartis)
1
John Mitchell; James McDonagh; Neetika Nath; Luna de Ferrari; Lazaros Mavridis; Rosanna Alderson
Rob Lowe; Richard Marchese Robinson
2
John Mitchell; James McDonagh; Neetika Nath; Luna de Ferrari; Lazaros Mavridis; Rosanna Alderson
Rob Lowe; Richard Marchese Robinson
3
Lowe et al., Molec. Pharmaceutics, 7, 1708 (2010)
Phospholipidosis
•
•
•
•
•
An adverse effect caused by drugs
Excess accumulation of phospholipids
Often by cationic amphiphilic drugs
Affects many cell types
Causes delay in the drug development
process
4
Lowe et al., Molec. Pharmaceutics, 7, 1708 (2010)
Phospholipidosis
• Causes delay in the drug development
process
• May or may not be related to human
pathologies such as Niemann-Pick
disease
5
Electron micrographs of alveolar macrophages (A and B) and peritoneal macrophages (C and D)
obtained from 3-month-old Lpla2+/+ and Lpla2-/- mice
Hiraoka, M. et al. 2006. Mol. Cell. Biol. 26(16):6139-6148
Tomizawa et al.,
Literature Mined Dataset
• Produced our own dataset of 185
compounds (from literature survey)
• 102 PPL+ and 83PPL• Each compound is an experimentally
confirmed positive or negative
R. Lowe, R.C. Glen, J.B.O. Mitchell Mol. Pharm. 2010 VOL. 7, NO. 5, 1708–1714
Some PPL+ molecules, from Reasor et al., Exp Biol Med, 226, 825 (2001)
10001101010011001101
10110101000011101101
10111101010001001100
10000001110011100111
10100101011101001110
10011111110001001010
Represent molecules using descriptors (we used E-Dragon & Circular Fingerprints)
Experimental Design
Split data into N folds, then train on
(N-2) of them, keeping one for
parameter optimisation and one for
unseen testing. Average results over
all runs (each molecule is predicted
once per N-fold validation).
We also repeat the whole process
several times with randomly
different assignments of which
molecules are in which folds.
Models are built using machine learning techniques such as Random Forest …
… or Support Vector Machine
Results
Average MCC Values:
RF
SVM
0.619
0.650
So we have built a good predictive model that can learn the
features that predispose a molecule to being PPL+, and can
make predictions from chemical structure.
This is useful – one could add it to a virtual screening protocol.
But can we understand anything new about how
phospholipidosis occurs?
Read up on gene expression studies related to phospholipidosis …
Sawada et al. listed genes which they found to be up- or down- regulated in phospholipidosis
As with all gene expression experiments, some of these will be highly relevant, others will
be noise. Can we help interpret these data?
Mechanism?
H. Sawada, K. Takami, S. Asahi Toxicological Sciences 2005 282-292
What expertise do we have available amongst our team, colleagues
& collaborators?
• Multiple target prediction
Florian Nigsch
• Maths
Hamse Mussa
• Programming
Rob Lowe
22
Predicting Targets using
ChEMBL: Application to the
Mechanism of
Phospholipidosis
23
• Multiple target prediction
Predicting off-target interactions of drugs. Not with the primary
pharmaceutical target, but with other targets relevant to side effects.
CHEMBL
Data mining and filtering
Filtered CHEMBL,
241145 compounds & 1923 targets
Random 99:1 split of the whole dataset, 10 repeats
10 models
Phospholipidosis dataset: 100 PPL+, 82 PPL- compounds
Predicted target associations
Target PS scores
ChEMBL Mining
• Mined the ChEMBL (03) database for
compounds and targets they interact with
• Target description included the word
"enzyme", "cytosolic", "receptor",
"agonist" or "ion channel"
• A high cut-off (weak binding) was used on
Ki/Kd/IC50 values (< 500μM) to define
activity
CHEMBL
Data mining and filtering
Filtered CHEMBL,
241145 compounds & 1923 targets
27
Method
• Number of Compounds : 241145
• Number of Targets : 1923
• Split the data into 10 different partitions
of training and validation
• Used circular fingerprints with SYBYL atom
types to define similarities between
molecules
28
Multi-class Classification
Algorithms:
• Parzen-Rosenblatt window
• Naive Bayes
Parzen-Rosenblatt window
• Rank likely targets using estimates of classcondition probabilities
1
p( xi |  ) 
N
 K x , x 
x j
i
j

using a Gaussian kernel
K(xi, xj) =
 ( x i  x j )T ( x i  x j ) 

exp 
2
d


2h
)


1
(h 2
(xi - xj)T(xi - xj) corresponds to the number of features in which xi and xj disagree
Partition No.
PRW Rank
NB Rank
1
17.049
74.104
2
16.343
76.251
3
18.424
79.078
4
16.212
73.539
5
17.339
73.535
6
18.630
77.244
7
20.694
78.560
8
18.870
74.464
9
16.584
76.235
10
18.200
78.077
Average
17.835
76.109
When we test the two methods, PRW ranks known targets better
than Naïve Bayes does. Hence we use PRW for our study.
Filtered CHEMBL,
241145 compounds & 1923 targets
Random 99:1 split of the whole dataset, 10 repeats
10 models
So we generate 10 separate validated models which we will use
to predict off-target interactions for our PPL+/PPL- set.
Assemble List of Targets Relevant to
Sawada’s Suggested Mechanisms
Mechanisms:
1. Inhibition of lysosomal phospholipase activity;
2. Inhibition of lysosomal enzyme transport;
3. Enhanced phospholipid biosynthesis;
4. Enhanced cholesterol biosynthesis.
Assemble List of Targets Relevant to
Sawada’s Suggested Mechanisms
Inhibition of
lysosomal
phospholipase
activity
Enhanced
phospholipid
biosynthesis
Enhanced
cholesterol
biosynthesis
Assigning Scores to Targets
• Use these 10 models of target interactions
• Predict targets for phospholipidosis dataset
• Score targets according to the likelihood of
involvement in phospholipidosis
• Use the top 100 predicted targets per
compound as we seek off-target interactions
N
PS   C p ( xi ) ( )
i 1
N
PS   C p ( xi ) ( )
i 1
• Score measures tendency of target to interact with
PPL+ rather than PPL- compounds.
10 models
Phospholipidosis dataset: 100 PPL+, 82 PPL- compounds
Predicted target associations
Target PS scores
M1 & M5 are involved in phospholipase C regulation & may be relevant; but not in Sawada’s list.
Our Scores for 8 of Sawada’s PPL-Relevant Targets
Mechanism Target
1
Sphingomyelin phosphodiesterase (SMPD) (h)
55
163=
90
152=
97
1203=
-10
610=
0
3-hydroxy-3-methylglutaryl-coenzyme A reductase (HMGCR) (h)
456=
10
Squalene monooxygenase (SQLE) (h)
437=
14
Lanosterol synthase (LSS) (h)
114=
134
Phospholipase A2 (PLA2) (h)
3
Elongation of very long chain fatty acids protein 6 (ELOVL6) (h)
Enhanced
phospholipid
biosynthesis Acyl-CoA desaturase (SCD) (m)
Enhanced
cholesterol
biosynthesis
PS
225
Inhibition of
lysosomal
Lysosomal Phospholipase A1 (LYPLA1) (r)
phospholipase
activity
4
Rank
41
We consider a PS score significant if the target is
predicted to interact with at least 50 more PPL+
compounds than PPL- compounds.
Our Scores for Sawada’s PPL-Relevant Targets
Mechanism Target
1
Sphingomyelin phosphodiesterase (SMPD) (h)
55
163=
90
152=
97
1203=
-10
610=
0
3-hydroxy-3-methylglutaryl-coenzyme A reductase (HMGCR) (h)
456=
10
Squalene monooxygenase (SQLE) (h)
437=
14
Lanosterol synthase (LSS) (h)
114=
134
Phospholipase A2 (PLA2) (h)
3
Elongation of very long chain fatty acids protein 6 (ELOVL6) (h)
Enhanced
phospholipid
biosynthesis Acyl-CoA desaturase (SCD) (m)
Enhanced
cholesterol
biosynthesis
PS
225
Inhibition of
lysosomal
Lysosomal Phospholipase A1 (LYPLA1) (r)
phospholipase
activity
4
Rank
Assemble List of Targets Relevant to
Sawada’s Suggested Mechanisms
Mechanisms:
1. Inhibition of lysosomal phospholipase activity;
2. Inhibition of lysosomal enzyme transport;
3. Enhanced phospholipid biosynthesis;
4. Enhanced cholesterol biosynthesis.
Sawada’s Suggested Mechanisms
Mechanism:
1. Inhibition of lysosomal phospholipase activity
We find evidence for this mechanism operating
through three target proteins:
Sphingomyelin phosphodiesterase (SMPD)
Lysosomal phospholipase A1 (LYPLA1)
Phospholipase A2 (PLA2)
Sawada’s Suggested Mechanisms
Mechanisms:
2. Inhibition of lysosomal enzyme transport;
There were no targets relevant to this mechanism
with sufficient data to test.
Sawada’s Suggested Mechanisms
Mechanisms:
3. Enhanced phospholipid biosynthesis
We were able to test two targets relevant to this
mechanism and found no evidence linking them
to phospholipidosis.
Sawada’s Suggested Mechanisms
Mechanisms:
4. Enhanced cholesterol biosynthesis
We find evidence for this mechanism operating
through one target protein:
Lanosterol synthase (LSS)
Sawada’s Suggested Mechanisms
• The mechanisms and targets suggested here
are insufficient to explain all the PPL+
compounds in our data set.
• We expect that other targets and possibly
mechanisms are important.
• Our method can’t test direct compound –
phospholipid binding.
50
Acknowledgements
•
•
•
Alexios Koutsoukas
Andreas Bender
Richard Marchese-Robinson