Using Cloud Computing To Improve the Accuracy and Probability of

Using Cloud Computing
To Improve the Accuracy
and Probability of Success
Of Drug Discovery
Dr. Ed Addison
Bio-IT World 2015
© 2015 Cloud Pharmaceuticals, Inc.
Today drug development faces big challenges
Traditional
discovery
High-throughput
screening
Costs and
timelines
Previous
in silico
Low probability
of success
Does not generate
novel molecules
Too high and
must come down
Not accurate
enough
DRUG
DISCOVERY
PRECLINICAL
compounds
compounds
10,000
© 2015 Cloud Pharmaceuticals, Inc.
250
CLINICAL
TRIALS
FDA
APPROVAL
compounds
DRUG
5
1
A Novel Approach: Quantum Molecular Design

Quantum chemistry design of drugs
from Duke University

Search of novel molecular space with
smart algorithm

Molecules have very strong binding
and freedom to operate

Elimination of poor candidates with
biophysical modeling

Stage gate by rapid synthesis,
optimization, and early assays

High preclinical success rate

Target to clinic possible in <18 months
© 2015 Cloud Pharmaceuticals, Inc.
Prediction of binding strength

Using QM/MM methods
– QM/MM is a multi-scale/multi-resolution approach
– Ligand is described by QM (no ligand parameterization
necessary) and the protein/solvent is described with
MM
– QM done with semi-empirical AM1 Hamiltonian,
allowing for faster computation
– MM is OPLS all-atom force field
– Ligand solvation term is calculated with implicit
solvation (GBSA) and solvent accessible surface area
(SASA)

Flexible protein, flexible ligand, explicit water
© 2015 Cloud Pharmaceuticals, Inc.
Binding activity parameterization



A set of ligands with
experimental results are
inserted in the binding
pocket
QM/MM-GBSA
calculations are
performed to determine
binding activity
Linear regression is used
to fit to the experimental
data
© 2015 Cloud Pharmaceuticals, Inc.
QM/MM-GBSA
Ebind
Casting the net into molecular space
Extract binding pattern
from X-ray structure
Search synthesizable
universe of scaffolds
Extract
roots
Explodes into
functional groups
Results in
10M-1B
molecules
(based on depth
Of study)
© 2015 Cloud Pharmaceuticals, Inc.
Quantum Molecular Design search algorithm

Search chemical space for a molecule with a certain desirable
property i.e. IC50
– Chemical space is very large ~1060 molecules, has a very complicated
landscape and is a discrete space, so continuous search methodologies
don’t work


Discrete chemical space is transformed into a vectorial space,
that allows us to search using integer programming methods
Originated from the Inverse Design algorithm (Duke U)
Property
Chemical Space
C
B
A
© 2015 Cloud Pharmaceuticals, Inc.
Searching with Quantum Molecular Design


AI SEARCH
Model molecular space
as a continuous space
AI search for optimum
in direction of increasing
scores
Results in diverse hit list
Better maximum
BINDING AFFINITY

Local maximum
CHEMICAL SPACE, MULTIPLE DIMENSION
© 2015 Cloud Pharmaceuticals, Inc.
Property filtering


Automated biophysical property filters
Current filters from external sources
– Probability of toxic side-effects, ADMET
– Molecular properties such as solubility
– Patent and literature searches

In-house developed filters, state of the art
– Blood-brain barrier permeability
– Synthesizability

Further research needed for
– Better toxicity predictions
© 2015 Cloud Pharmaceuticals, Inc.
Quantum Molecular Design workflow
Step 1:
Step 2:
Step 3:
Step 4:
Target Selection
Pre-Processing
Inverse Design
in the Cloud
Preclinical
Optimization
Parameterization
Training set
Protein 3D
Binding affinity
equation
Mol 1
Mol 2
Mol 3
Chemical space
generator
Pre-process
property filters
(ADME, etc.)
Designing leads
Synthesis and
cell-based assays
Mol 4
Mol 5
…
Post-process
property filters
Mol x
PROPERTY FILTERS
© 2015 Cloud Pharmaceuticals, Inc.
CLOUD
DATA ANALYSIS
Quantum Molecular Design is validated
Experimentally measured log (IC50)
© 2015 Cloud Pharmaceuticals, Inc.
>80% correlation to
actual lab data
Docking calculated score
QM/MM calculated score
Experimentally proven
binding affinity accuracy
Typically 500,000 CPU hours for de novo design
AMPK
Jak3
Aurora-A
JNK3
BASE1
Cloud Computing
MetAP2
bCR
has been deployed:
Metnase
DHFR
Amazon, Profit Brix,
PDC/PDK
eiF4E
GLP-1
HDAC8
hsp90
IGF
© 2015 Cloud Pharmaceuticals, Inc.
Cloud & Heat, Azure
And an emerging life science
data center in Iceland
QMD has been applied to many targets
PKCe
PKR
PLA2
ROCK2
Serpins
Methodology for GPCR targets
G-protein-coupled receptors are also known as seven transmembrane receptors.
 GPCRs represent 50-60% of current drug targets and 25% of
the top 200 best selling drugs are estimated to target
 GPCRs are hard targets for drug design because:

– Problems with obtaining 3D structures
– Several mechanisms of action
– “old” computational methods are not accurate enough to differentiate
between agonists, antagonists etc.

We employ homology modeling coupled with QM/MM to
obtain high accuracy of structure and binding prediction
© 2015 Cloud Pharmaceuticals, Inc.
Preclinical development
Outsourcing experimental
data collections
 Property filtering leads to
greater success probability
 Partner with CROs for
synthesis, expression
profiling, and assays

© 2015 Cloud Pharmaceuticals, Inc.
For further information
Contact
Ed Addison, CEO
Cloud Pharmaceuticals Inc.
6 Davis Dr.
Research Triangle Park, NC 27709
(910) 398-1200
[email protected]
© 2015 Cloud Pharmaceuticals, Inc.
Dr. Shahar Keinan
(919) 357-5319
[email protected]
Research Triangle Park, NC
© 2015 Cloud Pharmaceuticals, Inc.