Russian Call Girl Brookfield - 7001305949 Escorts Service 50% Off with Cash O...
Discovery PBPK: Efficiently using machine learning & PBPK modeling to drive lead selection & optimization
1. Discovery PBPK:
Efficiently using machine learning & PBPK modeling
to drive lead selection & optimization
Viera Lukacova
Director – Simulation Sciences
viera@simulations-plus.com
1
2. • About 60-80% of animal studies conducted during Lead ID/Opt stages
Current Drug Discovery & Development Process
Thomas et al, ATLA 2010, 38, Supplement 1, 81–85
Lead ID and Lead Opt together contribute ~32-35% towards the total cost
2
3. Say “I do” to the Machine Learning/PBPK marriage!
Permeability,
solubility vs. pH,
pKa(s),
logD vs. pH,
Fup,
blood:plasma ratio,
tissue Kps,
CLint,
CLfilt
3
4. 4
The Machine Learning/PBPK Marriage:
Proof of concept
Using in silico machine learning & PBPK, Pfizer was interested in predicting Cmax, AUC, and the PK profile in
human for their PF-03084014 compound
Hosea et al., Mol Pharm. 2013 Apr 1;10(4):1207-15
4
5. Hosea et al., Mol Pharm (2013)
• So it worked on 1 compound – how about others?
The Machine Learning/PBPK Marriage:
Proof of concept (cont.)
5
6. Novartis Computational Chemistry Group – Using
‘Discovery PBPK’ to Guide Lead Optimization
Daga et al. (2018) Mol. Pharmaceutics 15(3):821
Daga et al. (2018) Mol. Pharmaceutics 15(3):831
6
7. • Trouble with clearance can be overcome by building a NCA CL model using a small number of compounds (CLloc)
• 49 Compounds: Single congeneric series reported by Merck in various papers
– RAT in vivo data: %F, CLp
Case Study #1: Dipeptidyl Peptidase-4 Inhibitors
CLglbExp CLp
CLloc
Daga et al. (2018) Mol. Pharmaceutics 15(3):821
Daga et al. (2018) Mol. Pharmaceutics 15(3):831
7
8. • Hepatocyte CL provides accurate estimate of CL and hence %F
• 81 Compounds: Single congeneric series reported by AstraZeneca in 4 papers
– RAT in vivo data: %F, CLp
– in vitro data: CLint(hep)
Case Study #2: 11-HSD1 Inhibitors
CLglbExp Hep CLintExp CLp
CLloc
Daga et al. (2018) Mol. Pharmaceutics 15(3):821
Daga et al. (2018) Mol. Pharmaceutics 15(3):831
8
9. • 61 compounds : Single congeneric series with experimental data
– In silico & in vitro physicochemical data: (solubility, Caco-2 permeability, plasma protein binding, CLint)
– Rat PK data (%F, AUC, Cmax, Tmax, CLplasma, Vss)
Case Study #3: Internal Kinase-“X” Inhibitor series
in silico prop & CLglbExp input prop & CLintExp CLp in silico prop & CLloc
Daga et al. (2018) Mol. Pharmaceutics 15(3):821
Daga et al. (2018) Mol. Pharmaceutics 15(3):831
9
10. Increasing training data size, improved performance
Local Models OK w/ only ~15 Rat Data Points
DPP-4
48 cpd
HSD1
81 cpd
Kinase-X
63 cpd
Global Local ~35Local ~15 Local ~50 Local ~70
>75% of compound predicted within 3-fold
from chemical structure Daga et al. (2018) Mol. Pharmaceutics 15(3):821
Daga et al. (2018) Mol. Pharmaceutics 15(3):831
10
12. %Fa
%Fb
Vd
CLsys
* Modified from van de Waterbeemd, H, and Gifford, E. Nat.
Rev. Drug Disc. 2003, 2:192-204
HTPK Simulation Module
[HTPK = High Throughput PharmacoKinetic]
ACAT™ Model
Fraczkiewicz, AAPS 2018, Rapid Fire presentation
12
13. HTPK Simulation Module – Inputs
Input required for %Fa and %Fb Corresponding ADMET Predictor model
Diffusivity in water DiffCoef
Solubility in water S+Sw
Salt solubility factor SolFactor
Solubility in fasted intestinal fluid S+FaSSIF
Solubility in fed intestinal fluid S+FeSSIF
pKa S+pKa
Human jejunal permeability S+Peff
log P S+logP
Volume of distribution Vd
Percent unbound in plasma hum_fup%
Microsomal fraction unbound S+fumic
Metabolic clearance CYP_HLM_CLint
Blood to plasma ratio RBP
Precipitation time <user input>
Dose <user input>
Dose volume <user input>
Particle radius <user input>
Body weight <user input>
* Any predicted input can be
replaced by experimental values
Fraczkiewicz, AAPS 2018, Rapid Fire presentation
13
14. HTPK Simulation Module – Outputs
Outputs generated by the %Fa/%Fb model
Fraction absorbed
Fraction bioavailable
Cp(t) = plasma concentration vs. time profile
C_max = maximum plasma concentration
t_max = time of peak plasma concentration
AUC = area under the Cp(t) curve
Sensitivity to permeability and solubility
Mechanistic estimation of the volume of distribution
Outputs generated by the OptDose model
Optimal dose to reach desired plasma concentration
at steady state
Fraczkiewicz, AAPS 2018, Rapid Fire presentation
14
15. • Gut clearance is not considered
• Passive absorption kinetics only
• Clearance is assumed to follow linear kinetics
• Metabolism models only consider CYP enzymes
• Enterohepatic circulation and biliary excretion not considered
• First-order precipitation kinetics assumed
• Dosage form defined as IR tablet
• Physiology set to either fasted human (adult) or rat
HTPK Simulation Module – Limitations
Fraczkiewicz, AAPS 2018, Rapid Fire presentation
15
16. 90% predicted within 2-fold of
the observed value.
83% predicted within 1.5-fold
Performance Snapshot: %Fa
125 passively-absorbed compounds from Zhao [Zhao et al. J. Pharm. Sci., 2001, 90: 749.]
ReportedFa%%Fa(observed)
%Fa (predicted)
Some outliers are predicted to
be actively transported
Fraczkiewicz, AAPS 2018, Rapid Fire presentation
16
17. 81% predicted within 2-fold of
the observed value.
~70% predicted within 1.5-fold
Performance Snapshot: %Fb
62 CYP-metabolized compounds from Toshimoto data set [Toshimoto et al. Drug Metab. Dispos. 2014, 42:1811.]
ReportedFb%%Fb(observed)
%Fb (predicted)
Some outliers are predicted to
be UGT/esterase substrates
Fraczkiewicz, AAPS 2018, Rapid Fire presentation
17
18. HTPK Simulation Module – Performance
• Calculation of fraction absorbed (%Fa) and fraction bioavailable
(%Fb) in human after 24 h at three different dose levels: 1 mg,
10 mg, and 100 mg.
• 2284 diverse drugs extracted from World Drug Index
Platform (*)
Processing time for
2284 drugs
Laptop A 3.9 min (~0.1 s/drug)
Laptop B 2.5 min (~0.06 s/drug)
(*)
• A = DELL XPS with Intel® Core™ i7-
3537U CPU 2.5 GHz, 8 GB RAM, 64-
bit, running Windows 7.
• B = ASUS R.O.G. with Intel® Core™
i7-7700HQ CPU 2.8 GHz, 16 GB
RAM, 64-bit, running Windows 10.
Fraczkiewicz, AAPS 2018, Rapid Fire presentation
18
19. • Combination of accurate QSAR and PBPK models shows potential of
using in silico approaches in drug discovery settings
• Simplified models may be applied to screen large compound libraries
• The main limitations of the in silico approach:
– Quantitative clearance prediction
• Local models built on small sets of compounds were shown as possible solution to
improve the in vivo predictions
– Quantitative prediction of drug interactions with specific enzymes and
transporters
• Availability of experimental data is the limiting factor to building the in silico models
• There are still gaps in the physiological information (expression levels in different tissues)
Summary
19
20. Clearance Classification Models
El-Kattan AF, Varma MVS; Drug Metab Dispos 46: 729-739
• Three binary (Yes/No) models for
Metabolism, Renal and Hepatic Uptake
• 3-class ANNE model
• Risk model based on rules from El-
Kattan and Varma
20