Transcript Comparisons of Methods in Predictive Toxicology

Development of the Fathead Minnow Narcosis
Toxicity Data Base
Larry Brooke1, Gilman Veith2, Daniel Call3, Dianne
Geiger1, and Christine Russom4
1University
of Wisconsin-Superior, 2QSAR foundation, 3University of
Dubuque, and 4U.S. EPA Mid-Continent Ecology Laboratory
Log10 96-hr LC50 (mol/L)
Log Water Solubility (mol/L)
Bilinear Relationship Model
for Narcosis I MOA
(from Veith et al. 1983)
0
-2
-4
Log LC50 = -1.09 log P + 1.09 log (0.000068P + 1) - 0.79
R2 = 0.9986; n = 10
-6
-8
-2
0
2
4
Log10 P
6
8
Narcosis I Chemicals
(Acute Toxicity with Fathead Minnow)
Log10 LC50 (moles/L)
0
-2
-4
Y = -1.6417 - 0.7724X
r2 = 0.8944; n = 291
Where: Y = Log10 LC50 and X = Log10 P
-6
-8
-4
-2
0
2
Log10 P
4
6
8
Acute Toxicity to Fathead Minnow
with Narcosis I & II Chemicals
Log10 LC50 (moles/L)
0
Narcosis I (non-polar)
Y = -1.6417 - 0.7724X
r2 = 0.8944; n = 291
-2
-4
Narcosis II (Polar)
Y = -2.3244 - 0.6140X
r2 = 0.5599; n = 36
-6
Where: Y = Log10 LC50 (moles/L) and X = Log10 P
-8
-4
-2
0
2
Log10 P
4
6
8
Toxicity to Fathead Minnow
of Narcosis I, II, and III Chemicals
(From the U.S. EPA Data Base)
Log10 LC50 (moles/L)
0
-2
-4
Y = -1.7741 - 0.7513X
r2 = 0.8559; n = 351
-6
Where: Y = Log10 LC50 (moles/L) and X = Log10 P
-8
-4
-2
0
2
Log10 P
4
6
8
1
Nonpolar Narcotic Chemicals
(from Schultz et al. 1998)
Log10 LC50 (moles/L)
0
-1
Tetrahymena pyriformis
Y = -1.1728 - 0.7336X; r2 = 0.9442
n = 148
-2
-3
-4
Pimephales promelas
Y = -1.2140 - 0.8741X; r2 = 0.9569
n = 51
-5
-6
-2
-1
0
1
2
Log P
3
4
5
6
Nonpolar Narcotic Chemicals
(from Schultz et al. 1998 and U.S. EPA)
0
Log10 LC50 (moles/L)
-1
-2
-3
-4
-5
Fathead minnow
Tetrahymena pyriformis
-6
-7
-4
-2
0
2
Log10 P
4
6
8
Fathead Minnow Acute and Chronic Toxicity
with Narcosis Chemicals
Log10 LC50 or MATC (moles/L)
0
Acute Toxicity
Y = -1.6417 - 0.7724X
r2 = 0.8944; n = 291
-2
-4
-6
Chronic MATC
Y = -3.1562 - 6375X
r2 = 0.7576; n = 30
Where: Y= Log10 LC50 (moles/L) and X = Log10 P
-8
-2
0
2
4
Log10 P
6
8
Applying Predictive Data Mining to
Predictive Toxicology
From Narcosis to McKim Conference
Chihae Yang
28th June, 2006
Acknowledgment
•
•
•
•
Gilman Veith, International QSAR Foundation
J.F. Rathman, The Ohio State University
Leadscope team
Ohio Technology Action Fund
From Meyer-Overtone to McKim Conference
• Narcosis
– …”toxicity of neutral organics is related to their ability
to partition between water and a lipophilic biphase
where molecules exert their activity…”
• Model system for partition: olive oil/water.
• Evolution
Narcosis
Non-polar and polar narcosis
Reactivity
……
Paradigm shift
• How do we strategically leverage?
In silico
In vitro
In vivo
Omics
• How do we read across the species, endpoints,
structural classes, different knowledge domains?
Predictive data mining strategies
structural descriptions
analogs
chemical
stressor
profile
biological/environmental fate
Yang, C.; Richard, A.M., Cross, K.P. Current Computer-Aided Drug Design, 2006, 2, 1-19.
Steps in predictive data mining
Visualization
Analysis
Structure, data, graphs, models
SAR & QSAR
Profiling
Grouping
Searching
Hypothesis driven queries
Analog searching
Read across
Platform
Chemistry Biology integration
Knowledge addition
Relational database
Data mining analysis methods
Compound grouping
Analysis
Focused Data Sets
Prediction
QSAR
Classification
Classification
Rule Extraction
Pattern Recognition
Profiling
Clustering
Expert Grouping
Large diverse Data Sets
Applying to predictive tox
• Profiling “chem-bio” domain
– Cut across different knowledge domains
– Find hidden signals and relationships from data
• Qualify/quantify read-across
• Complementary to (Q)SAR
– Build hypothesis driven models
– Go beyond Yes/No question and answer
Predictive data mining examples
• Biological profile
– Relationships between fish narcosis and toxicological
findings in rat inhalation studies?
• Fathead minnow EPA dataset
• Rat acute toxicity dataset from RTECS
• Thermodynamics consideration
Theoretical bases:
Vapor-liquid equilibrium
• Non-ideal Raoult’s law:
- The equilibrium distribution between liquid and vapor
phases for a chemical species i
 i xi piv  yi P  Pi  partial pressure
i : activity coefficient
xi : mole fraction of i in the liquid phase
piv : vapor pressure of pure liquid i at the same temperature T
yi : mole fraction in the vapor phase.
Study sources for rat and FHM correlations
- rat exposure time 2-8 hours
- narcosis
RTECS 2006
2341
• single dose
• inhalation
chamber
921
EPA FHM
617
• dose unit (mg/mL)
• defined LD50
76
179
LC50 at 96 hr
Profiling examples
Structures
Liver
O
Lung
UBL
GI
pLC50
Rat
FHM
present
present
absent
absent
0.489
-1.37
absent
present
present
absent
0.799
-0.729
absent
present
absent
absent
1.98
0.44
absent
absent
absent
present
2.54
1.49
OH
N
OH
Representing structures with
Leadscope molecular descriptors
O
Ak
Benzenes
N
Functional groups
O
O
N
Any
N
H
N
N
Heterocycles
N
Pharmacophores
PCC
PCC
HBA
O
Spacers
User defined features
NH2
Read-across using structural descriptors
Structural descriptors
profiles of rat organ lesions
LC50 FHM
Structural
descriptors
%
structures
liver
ubl
lung
GI
43.4
0
0
0.24
1,2-subst
13.2
0
0
1,3-subst
10.5
0
1,4-subst
18.4
pLC50
FHM
Rat
0.06
1.38
2.02
0.3
0
1.3
1.63
0
0.25
0
1.88
1.54
0
0
0.36
0.07
1.46
1.79
30.3
0.09
0.04
0.3
0.04
-0.2
1.4
alcohol, p-alkyl-
13.2
0.2
0.1
0.4
0
-1.02
1.21
alcohol, aryl-
13.2
0
0
0.3
0.1
1.01
1.61
aldehyde
6.6
0
0
0.2
0.2
1.31
0.67
amines
18.4
0
0
0.29
0
0.66
2.24
carbonyl
26.3
0.05
0
0.1
0.05
1.3
1.91
ether
13.2
0.2
0
0.1
0
0.56
0.94
13.2
0.2
0
0.1
0
0.56
0.94
18.4
0
0
0.29
0.07
1.91
2.02
13.2
0
0
0.4
0
1.54
2.3
5.3
0.25
0
0.25
0
-0.87
0.63
Benzenes
alcohol
ether, alkylhalide
halide, arylketone
23 structural descriptors were selected.
Quantitative read-across
Liver
kidney
ubl
Lung
GI
Pearson correlations
Liver – GI
Lung – Kidney
Liver – Lung
pLC50Rat – pLC50FHM
- 0.52
0.45
-0.31
0.55
pLC50FHM – Liver
- 0.72
pLC50FHM – Kidney
- 0.75
pLC50
Rat
pLC50
FHM
From a surface scientist point of view
• Passive diffusion through lipid bilayer
– Headgroup interaction
– Hydrophobic tail interaction
– Hydrophilic to lipophilic balance (HLB)
• Partition model of molecules in lipid layer :
 species i 
bulk
 activity i 
bulk


 species i 
 activity i 
lipid
lipid
at equilibrium
 ibulk x ibulk   ilipid x ilipid
partition coefficient: K x
x ilipid
 ibulk
 bulk  lipid
xi
i
 : activity coefficient
UNIFAC activity coefficient model
ln i  ln  iC  ln  iR
“combinatorial” term
molecular volume and
surface area effects
(size, shape, packing)
“residual” term
intermolecular energy effects
(interaction)
The properties of Gases & Liquids, 4th ed., R. Reid, J. Prausnitz, B. Poling, McGraw Hill, 1987
Advantages of UNIFAC model
• Group contribution method
– Molecular descriptors-based activity coefficients
• Flexibility to vary liquid phases compositions
–
–
–
–
–
octanol/water
octanol-water solution/water
hexadecane/water
lipid/water
etc.
Example: Lipid as a solvent phase
O
O
O
O
O
O
N
P O
O
O
O
O
O
O
O
O
O
O
O
Example of activity coefficients
in various environment
Solvent
Log10 
Water
5.23
Octanol
0.05
Lipid tail
-0.40
Lipid head
0.12
Hexadecane
0.73
O
O H
Activity coefficients at infinite dilution can be used to model
solubility in various phases.
measured
LogP
LogP
(ow/w)
Pearson correlations
against measured LogP
LogP(o/w)
0.93
LogP(ow/w)
0.92
LogP(h/w)
0.92
LogP(dppc/w)
0.90
LogP
(o/w)
LogP
(h/w)
LogP
(dppc/w)
Reflection
…We’re committed to nothing less than a point-forpoint transcript of everything there is. Only one
problem: the index is harder to use than the book.
We’ll live to see the day when retrieving from the
catalog becomes more difficult than extracting
from the world that catalog condenses….
“The gold bug variations, Richard Powers”, 2004
Distribution of LC50s for FHM and rats
pLC50 of FHM
Mean: 0.669
pLD50 of rats
Mean: 1.52
2
Narcosis I Chemicals
Acute Toxicity with Fathead Minnow
and Water Solubility of Chemical
0
Water Solubility (moles/L)
Log10 LC50 (moles/L)
1
-1
-2
-3
-4
LC50 vs Log P
Solubility vs Log P
-5
-6
-2
-1
0
1
2
Log10 P
3
4
5
6