QSARs – Presentation – Ceyda Oksel

Transcription

QSARs – Presentation – Ceyda Oksel
“Quantitative Structure‐Activity Relationship (QSAR) models”
Ceyda Oksel
Xue Z. Wang
University of Leeds
27.03.2013
A pan-European Infrastructure for Quality in Nanomaterials Safety Testing
NUID UCD | NHM | IOM | JRC | BFR | KIT | FUNDP | IST | UNIVLEEDS | NILU | HMGU | LMU | CIC | UU | ICN
DLO/RIKILT | WU | DGUV | TAU | VITO | SMU | TCD | FIOH | UOE | HW | CNRS | INERIS | UBRM | RIVM
Purpose of Today
•
•
•
•
•
Theory of QSAR
New Data Mining Methods
Nanomaterial Characterization
Nanotoxicity
Application of QSAR model in nanotoxicity modeling
QualityNano Training School 2
Part I
Theory of QSAR QualityNano Training School 3
QSAR History
1865
1893
Crum Brown and
Fraser[1] expressed the
idea that there
was a relationship between activity
and chemical
structure.
Richet[2] correlated toxicities of simple organic
molecules
with their solubility in water.
1900
Meyer[3] and Overton[4] found linear relationships
between the toxicity of organic compounds and their lipophilicity.
1969
Hansch[5] published a free‐energy related
model to correlate biological activities with
physicochemical
Properties.
The father of the concept of quantitative structure‐activity relationship (QSAR),
the quantitative correlation of the physicochemical properties of molecules with their biological activities[5]
Corwin Herman Hansch
(1918 – 2011)
[1] A. Crum‐Brown, T.R. Fraser, Trans. R. Soc. Edinb. 25 (1968–1969) 257. [2] M.C. Richet, Compt. Rend. Soc. Biol. 45 (1893) 775.
[3] H. Meyer, Arch. Exp. Pathol. Pharmakol. 42 (1899) 109. [4] E. Overton, Studien
u¨ber dir Narkose, Gustav Fischer, Jena, 1901
QualityNano Training School [5] Hansch, C. (1969) A quantitative approach to biochemical structure‐activity relationships Accounts of Chemical Research, 2, 232‐239.
4
What is QSAR ?
QSAR (Quantitative Structure‐Activity Relationship)
The Variance In
Molecular Structure
Physicochemical
or Biological Property
A QSAR is
that relates
descriptors
compound
activity.
a statistical model
a set of structural
of a chemical
to its biological
Mathematical Dependencies
Toxicity/Ecotoxicity Prediction
Bulk Form Nano‐particles
(widely used) (not adopted)
The presence of particular characteristics increase
the propability that the compund is toxic
QualityNano Training School [1] T. Puzyn et al. (eds.), Recent Advances in QSAR Studies, 103–125. DOI 10.1007/978‐1‐4020‐9783‐6_4, C Springer Science+Business Media B.V. 2010
5
nano‐QSAR
QSAR (Quantitative Structure‐Activity Relationship)
Successful QSAR studies
What about Nano‐size particles?
QSAR has been widely used to predict the toxicity of substances in bulk
form (especially drug‐like compunds) but, up to date, QSAR studies for the
prediction of nanoparticle toxicity have been rarely reported.
The European system for new chemical management (REACH) promotes
QSAR methods as an alternative way of toxicity testing.
QSAR models are very useful in case of the classic chemicals but the
concept of nano‐QSAR is still under development.
QualityNano Training School 6
Why do we need QSAR models?
Sources of information[1]
All chemical substances need to be tested in terms of their toxicological
and environmental properties before their use
There are several reasons to use QSAR Models : very fast, often free, reduce the number of animals used in experiments predict
predict
validate
IN SILICO
(computational)
Alternative, fast
developing applications
IN VITRO IN VIVO
(in container)
(on living organism)
The most important sources of information
Reduction in the time, the most
QSAR
cost and animal testing promising quantitative structure‐activity [1]
relationship
[1] A.. Gajewicz, et al., Advancing risk assessment of engineered nanomaterials: Application of computational approaches,
QualityNano Training School Adv. Drug Deliv. Rev. (2012), doi:10.1016/j.addr.2012.05.014[1]
7
Purpose of QSAR
To predict biological activity and physico‐chemical properties by rational means
To comprehend and rationalize the mechanisms of action within a series of chemicals
Savings in the cost of product development
Predictions could reduce the requirement for lengthy and expensive animal tests.
Reduction of animal tests, thus reducing animal use and obviously pain and discomfort to animals
Other areas of promoting green and greener chemistry to
increase efficiency and eliminate waste
QualityNano Training School 8
Requirements for QSAR
1.Set of molecules
Molecular
Modelling
2. Set of molecular
descriptors
3. Measured biological
activity or property
DESCRIPTORS
QSARs
Physcochemical,
biological, structural
e.g. Neural Networks
PLS, Expert Systems
How many compounds are required to develop a QSAR?
There is no direct and simple answer
“as many as possible”
To provide some guide, it is widely accepted that between five and ten compounds are required for every descriptor in a QSAR [1]
4.Statistical
methods
Toxicity
Endpoints
Daphnia magna
EC50
Cancinogenicity
Mutagenicity
Rat oral LD50
Mouse inhalation
LC50
Skin sensitisation
Eye irritancy
QualityNano Training School [1] Aptula AO, Roberts DW(2006) Mechanistic applicability domains for non‐animal based prediction
of toxicological end points: General principles and application to reactive toxicity. Chem Res Tox. 19:1097–1105
9
Applications of QSAR
There are a large number of applications of these models within
industry, academia and governmental (regulatory) agencies.

The estimation of physico‐chemical properties, biological
activities and understanding the physicochemical features
behind a biological response in drug desing.

The rational design of numerous other products such as surface‐active agents, perfumes, dyes, and fine chemicals.

The prediction of a variety of physico‐chemical properties of molecules.

The prediction of fate of molecules which are released
into the environment.

The identification of hazardous compounds at early stages of product development, the prediction of QualityNano Training School toxicity to humans
and environment.
10
What is required for a good QSAR Model ?
The QSAR model should meet the
requirements of the OECD principles [1]:
1
2
3
4
5
• a defined endpoint;
• an unambiguous algorithm;
• a defined domain of applicability ;
• appropriate measures of goodness‐of‐fit, robustness and predictivity;
• a mechanistic interpretation if possible.
Common QSAR Modelling
Errors
•Uninformative descriptors
•Poor descriptor selection and chance correlations
•Modelling complex, nonlinear structure property
relationships with linear models
•Incorrectly validating QSPR models
•Not understanding the domain of applicability of models
[1] OECD, Guidance Document on the Validation of (Quantitative) Structure– Activity Relationships Models, Organisation for Economic Co‐operation and Development QualityNano Training School 11
(2007).
Methods
•QSAR Modelling process consists of 5 main steps.
Main Steps
Molecular Structure
Representation
Molecular Descriptors
Predictive Model
Model Validation
Applicability Domain of QSAR
•Begins with the selection of molecules to be used
•Selection of descriptors; numerical
representer of molecular features (e.g. number
of carbon)
•Original descriptor pool must be reduced in size
•Model building
•The reliability of the model should be tested
Types of Molecular Descriptors
(topological, geometric, electronic and hybrid)
1D Descriptors
2D Descriptors 3D Descriptors
QualityNano Training School 12
Molecular Descriptors[1]
Molecular
Descriptors
ca
are numerical values that characterize properties of molecules.
calculated by
Mathematical
formulas
Experimental
measurements
Theoretical
Indexes
Computational
algortihms
Different Theories
(e.g. Quantum‐
chemistry, information theory, organic chemistry)
derived
from
More than 5000[2]
Molecular
Descriptors
processed
by
statistics
chemometrics
chemoinformatics
applied
in
Scientific Fields (e.g. QSAR,
medicinal chemistry, drug design, toxicology, analytical chemistry,
Environmetrics, virtual screening)
QualityNano Training School [1] T. Puzyn et al. (eds.), Recent Advances in QSAR Studies, 103–125. DOI 10.1007/978‐1‐4020‐9783‐6_4, C Springer Science+Business
Media B.V. 2010
[2] Todeschini R, Consonni V (2000) Handbook of molecular descriptors. Wiley‐VCH, Weinheim
13
Data Analysis Methods[1]
Different Statistical Methods have been used in QSAR for the extraction of useful information
from the data.
Regression Problems
Data Collection
•
•
•
•
Multiple linear regression
Partial least squares
Feed‐forward back propagation neural network
General regression neural network
Classification Problems
Molecular
Descriptor
Data Analysis
Model Validation
•
•
•
•
•
•
Linear Discriminant analysis
Logistic regression
Decision tree
K‐nearest neighbor
Probabilistic Neural Network
Support vector machine
Recently‐developed Methods
•
•
•
•
•
Kernel partial least square
Robust continuum regression
Local lazy regression
Fuzzy interval number k‐nearest neighbor
Fast projection plane classifier
QualityNano Training School [1] Ekins, S. (2007). Computational Toxicology: Risk Assessment for Pharmaceutical and Environmental Chemicals. New Jersey: Wiley-Interscience.
14
Data Mining Methods
•a relatively new field which share some techniques with traditional data analysis but also brings a lot of new visions and techniques. Value

Decision

Knowledge
Information
Data Data Data Data
Volume
‘DM is the discovery of useful Information
and Knowledge from a large collection of
Data, in order to help the Decision
making’
Data: records of numerical data, symbols, images,
documents
Knowledge: Rules: IF .. THEN ..
Cause‐effect relationships
Decision trees
Patterns: abnormal, normal operation
Predictive equations
QualityNano Training School 15
Data Mining Methods
•Data mining shares many techniques and tools with traditional data
analysis (such as the ones listed below). But it emphasises the analysis
of very large databases, and can deal with complicated cases which
cannot be handled by traditional methods.
Clustering
Summarisation
Classification
Regression
Conceptual Clustering
Case-based Learning
Inductive learning
•Data mining research has also generated some new techniques that
were not seen in traditional data analysis, such as Link Analysis
QualityNano Training School 16
New DM Technique Example 1: Link Analysis • Example : Ten data cases, three variables, x1, x2, and x3
• How can we automatically uncover the dependency relationship between them ?
WHICH RELATIONSHIP ?
DATA SET
x1
x2
x3
1
1
0
1
0
0
1
0
1
0
0
1
0
1
0
1
1
0
1
0
0
1
1
1
0
1
1
0
1
0
x1
x2
x3
x2
x1
x3
This relationship means, x2 is
dependent on x1, and x3 is dependent on x2, but x1 is independent of x2 and x3
This relationship means, x2 is
dependent on x1, x3 is also dependent on x1, but x2 and x3 are independent of each other
17
QualityNano Training School New DM Technique Example 2: Inductive Data Mining for Automatic Generation
of Decision Trees from Data
x1
x2
x3
x4
y
x1
x2
x3
x4
y
55.33
1.72
54
1.66219
92.19
71.31
3.44
55
1.60325
90.51
59.13
1.2
53
1.58399
92.74
72.3
4.02
55
1.66783
90.24
57.39
1.42
55
1.61731
91.88
68.81
6.88
55
1.69836
91.01
56.43
1.78
55
1.66228
92.8
66.61
2.31
52
1.77967
91.9
55.98
1.58
54
1.63195
92.56
63.66
2.99
52
1.81271
91.92
56.16
2.12
56
1.68034
92.61
63.85
0.24
50
1.81485
92.16
54.85
1.17
54
1.58206
92.33
67.25
0
53
1.72526
91.36
52.83
1.5
58
1.54998
92.22
67.19
0
52
1.86782
92.16
x1, x2, x3 are different measures of feed composition
x4 is the log of a combination of operating conditions
y – product quality measure y = ‘good’ if > 92, = ‘low’ if between 91‐92,
= ‘very low’ if < 91
QualityNano Training School Aim: to find which feed composition or operational condition variables have the most important impact on product quality
18
Automatically generated decision tree
Product quality is mainly determined by x1 and x4, almost has nothing to do with x2 and x3
•If x1 < 55.4, the probability that product quality is good (y>92) is high
•If x1 >55.4 but x4 >1.8, the probability that product quality is good is also high
•Other two regions, product quality is either low or very low
x1
Very low
> 55.4
< 55.4
Good
2.0
>1.8
<1.8
x1
< 70
91<y<92

Low
Good
x4
x4
y > 92
Low
Y > 92
> 70
1.5
Good
Y < 91
Very low
50
60
70

If x1 = [44.6, 55.4], then conditional probability of y = 3 is 0.9
QualityNano Training School If x4 = [1.8, 2.3], then the conditional probability of y
=3 is 1.0
x1
19
(1)Generation of a population of solutions
<s >s
<s >s
<s >s
<s >s
……
Automatic generation of decision tree from data
This is a new
.. .... .. .. .... .. .... .. .. .... ..
approoach
(2) Repeat steps (i) and (ii) until the stop criteria are based on a genetic programming method
satisfied:
(i) calculate the fitness function values for each solution candidate
(ii) perform crossover and mutation to generate the next generation
<s >s
<s >s
<s >s
.. .... .. .. .... .... .. .... ..
<s >s
.... ..
……
(3) the best solution in all generations: the final solution
QualityNano Training School 20
New DM Technique Example
Two Data Sets
Data Set 1:
Concentration lethal to 50% of the population, LC50,
1/Log(LC50), of vibrio fischeri, a biolumininescent bactorium
75 compounds
1069 molecular descriptors calculated by
molecular modelling software DRAGON
Data Set 2:
Concentration effecting 50% of the population,
EC50 of algae chlorella vulgaris, by causing fluorescein
diacetate to disappear
80 compounds
1150 descriptors
QualityNano Training School 21
Y‐Adapt for QSAR Toxicity Prediction: Data Set 1
A tree for the bacteria data using three classes.
A method to classify toxicity endpoints into classes
Sample tree from Y‐Adapt
For bacteria data:
of 1069 descriptors, only 5 descriptors are important ! Using the 5 descriptors, the tree model can predict with
the accuracy >96.7%
Polar surface area ≤ 40.368
No
Yes
Geary autocorrelation lag 4 weighted by atomic mass ≤ 0.007
Yes
Class 1 (21/21)
Geary autocorrelation lag 4 weighted by polarizability ≤ 1.181
No
H autocorrelation lag 0 weighted by Sanderson electro‐
negativities ≤ 2.111
Yes
Class 1 (8/8)
No
Yes
Class 3 (9/10)
No
Class 2 (9/9)
QualityNano Training School Broto‐Moreau autocorrelation lag 2 weighted by polarizability ≤ 4.085
Yes
Class 2 (5/6)
No
Class 3 (6/6)
22
Y‐Adapt for QSAR Toxicity Prediction : Data Set 1
Class 1:
Log(1/EC50) ≤ 4.08 29 cases
Class 2: 4.08 < Log(1/EC50) ≤ 4.51 16 cases
Class 3: 4.51 < Log(1/EC50) ≤ 4.99 9 cases
Class 4:
otherwise 6 cases
A decision tree from generation 46 for the bacteria data, finding four
classes from the continuous endpoint itself during induction.
Training accuracy: 93.3%
Test accuracy: 80.0%
Polar surface area ≤ 40.368
Yes
No
3D MoRSE signal 15 weighted 18 weighted by Sanderson electronegativites ≤ ‐0.644
H autocorrelation lag 0 weighted by Sanderson electro‐negativities ≤ 2.111
Yes
Class 1 (16/16)
Geary autocorrelation lag 4 weighted by atomic mass ≤ 0.007
Yes
Class 1 (13/13)
No
Yes
No
Class 2 (5/7)
No
Class 2 (9/9)
QualityNano Training School H autocorrelation lag 0 weighted by Sanderson electro‐negativities ≤ 2.394
Yes
Class 3 (7/9)
No
Class 4 (6/6)
23
Y‐Adapt for QSAR Toxicity Prediction : Data Set 2
Solvation connectivity index ≤ 2.949
No
Yes
Self‐returning walk count order 8 ≤ 3.798
Yes
Class 1
(16/16)
No
Class 2
(14/15)
2nd dataset ‐ algae data
Of the 1150 descriptors, only 5 descriptors are important !
Molecular multiple path count order 3 ≤ 92.813
Yes
No
Class 3
(6/8)
H autocorrelation of lag 2 weighted by Sanderson electro‐negativities ≤ 0.401
Yes
Class 4
(6/7)
No
Training: 92.2%
2nd component symmetry directional WHIM index
weighted by van der Waals volume ≤ 0.367
Test: 81.3% Yes
Class 3
(9/10)
QualityNano Training School No
Class 4
(8/8)
24
Integrated Process Data Mining Environment
Descriptor calculation
Data Pre‐processing
Scaling
Missing values
Outlier identification
Feature extraction
Data Mining Toolbox
Regression
 PCA and ICA
ART2 networks
Kohonen networks
K‐nearest neighbour
Fuzzy c‐means
Decision trees and rules
Feedforward neural networks (FFNN)
Summary statistics
Visualisation
Mixture QSAR models Discovery Validation
Statistical significance
Results for training and test sets
Data import
Excel
ASCII Files
Database
XML
Results Presentation
Graphs
Tables
ASCII files
QualityNano Training School Different data mining techniques can be combined into a single 25
system
QSAR Model Validation
Cross‐validation (Q2)
Model
Validation
Model Applicability
Model
Interpretation
•
•
K‐fold Cros Validation
Leave‐one‐out
Cross Validation
The statistical fit of model R2
Standart Error of Predicton (S)
•Model validation is required to ensure model reliability (quality of the model)
Ensures that the model is not due to chance factors!
•The statistical fit of model is usually performed
•Cross validation: dividing a data sample into subsets, performing the analysis on a single subset, using other subsets to confirm and validate the initial analysis
•There is no agreed method for the validation of QSAR Models
QualityNano Training School 26
QSAR Workflow[1]
[1] Tropsha, A., Gramatica, P., Gombar, V. (2003) The importance of being earnest:Validation is the Absolute Essential for Successful Application and Interpretation of QSPR QualityNano Training School 27
Models. Quant. Struct. Act. Relat. Comb. Sci. 22, 69‐77.
QSAR Model Applicability Domain
Descriptor 1
.
. . .
..
.
. . TRAINING SET
.
.
. . Descriptor 2
.. . . ..
. .
.. .
Inside Applicability
Domain
Outside Applicability
Domain
will be predicted
will not be predicted
•Model applicability allows us to decide whether the model will be useful for new data (correct use of the model).
•Finding the boundaries of the
model to see how well it will work
for other compounds.
•Not even a significant and validated QSAR model can be used for the entire universe of chemicals[1]!
[1] Gramatica P. (2007) Principles of QSAR models validation: internal and external. QSAR and Combinatorial Science; 26:694–701.
QualityNano Training School 28
QSAR Tools for Toxicity Prediction
1.
TOPKAT (by Komputer Assisted Technology ):
It uses QSARs to give numeric predictions, based on several descriptors
2. HazardExpert (by CompuDrug ):
HazardExpert, produced by CompuDrug, is a rule‐based system. The rules use toxicophores, or structural alerts derived from QSARs
3. CASE: It can derive QSAR model from user’s
6. DEREK:Deductive Estimation data.
of Risk from Existing Knowledge 4. ECOSAR: It uses QSARs to predict the aquatic (DEREK) is marketed by LHASA toxicity of chemicals for a variety of organisms
Ltd. It is an expert knowledge 5. OASIS: It uses hierarchical rules based on base system that predicts QSARs, which can be user derived. whether a chemical is toxic
•
•
Each of these QSAR tools have their own advantages and weaknesses for toxicity
predictions
There are also several available software packages for descriptor calculations (e.g. DRAGON)
QualityNano Training School 29
Challenges for QSAR [1]
1
• If there is a measurement error in the experimental data, it is very likely
that false correlations may arise
2
• If the training dataset is not large enough, the data collected may not reflect
the complete property space. Consequently, many QSAR results cannot be
used to confidently predict the most likely compounds of the best activity.
3
• There are many successful applications of this method but one cannot expect the QSAR approach to work well all the time. QSAR Models are easy to build but also very easy to get wrong.
QualityNano Training School 30
[1] Puzyn, T., Leszczynski, J.(2012) Towards Efficient Designing of Safe Nanomaterials: Innovative Merge of Computational
Approaches and Experimental Techniques, The Royal Society of Chemistry. Part II
Nano‐toxicity and
Nanomaterial Characterization
QualityNano Training School 31
Macro concerns with nano‐world
!
Macro concerns with nano‐world !
Uncertainities at the early stages of a new technology bring new concerns!
New Technology
New Benefits
New Concerns
[1]
•
•
Safety to human health and environment
Suitability of risk management strategies
More than 800 nano‐products
No standardized or validated methods for nanotoxicity testing
Available toxicity
data for NPs are
inconsistent and
confusing
Blocks the development of ENPs risk reduction strategies[2]
[1] Chaudhry, Q. et al. (2010). Knowns, Unknowns, and Unknown Unknowns . RSC Nanoscience & Nanotechnology (p. 201-218).
Royal Society of Chemistry.
[2]Tran, L., & Chaudhry, Q. (2010). Engineered Nanoparticles and Food: An Assessment of Exposure and Hazard. RSC Nanoscience &
QualityNano Training School 32
Nanotechnology (p 120-134). Royal Society of Chemistry.
[1]
Risk assesment
for
NMS Risk assesment for NMS [1]
 ENMs are already used in a number of commercial applications.
Potential safety issues of NMs should be addressed at the same time as the
tech. is developing
•What quantitative •What dangers to connections exist
human health may between the dose
basically arise from 2.Dose
and the extent of the
1.Hazard
the harmful agent ?
response
Identification
expected effect?
relationship
4.Risk •What is the nature and magnitude of the assessment
risk and how accurately can it be estimated?
3.Exposure
assessment
•To what extent is the relevant population group exposed to the harmful agent?
QualityNano Training School OECD (2003): Environment Directorate. Description of Selected
Key Generic Terms Used in Chemical / Hazard Assessment. OECD 33
Series on Testing and Assessment Number 44. ENV/JM/MONO(2003)15. Paris, 30-Oct-2003.
Dose‐response relationship
The evaluation of toxicity includes two steps:
Hazard Identification
Evaluation of toxicity
Dose‐response evaluation
involves determining the conditions or levels at which the presence of a contaminant may trigger a biological response from the body
Different amount of
CNT dust
At level of 0.1 mg/m3
No effects were
observed
At higher level of
0.5 mg/m3
Side effects were
detected in the
lung 34
QualityNano Training School Naidu, B. David (2009). Biotechnology and Nanotechnology - Regulation under Environmental, Health, and Safety Laws.. Oxford University Press.
Particle Characterization
•The only reasonable way to obtain toxicity information for the large number of nanoparticles without testing every single one is to relate the physicochemical characteristics of NPs with their toxicity in a SAR (Structure Activity Relationship) or QSAR model.
•Nanomaterial characterization is one of the first steps in toxicity test.
However, it is a very challenging issue!!!
Nanoparticle Structure Properties
Particle Size
Surface Area
Morphology
State of Dispersion,
…
Toxicity
QualityNano Training School 35
more research is required to reach a clear understanding of the
Possible toxicity‐related Physico‐chemical Properties
When a material is in a nano‐sized form, its properties become different from the
same material in a bulk form
 Fullerenes
 Carbon NTs
 Graphene
Carbon‐based
Nanomaterials Different and Exceptional
Properties
No certain information!!!
Recent studies showed that toxicity of NPs can be related to:

 Size and shape

 Size distribution

 Agglomeration state

 Porosity

 Structure‐dependent elect. 
 Electronic properties

 Characterisation: 
In what medium?

Surface area
Surface chemistry
Surface charge
Crystal structure
Composition
Configuration
Coating
Aggregation state
Metal content
POSSIBLE factors[1]
inducing toxicity of chemicals
more research is required to reach a clear understanding
QualityNano Training School [1]
Puzyn, T., Leszczynska, D. & Leszczynski, J. Toward the development of ‘Nano-QSARs’: advances and challenges. Small 5, 2494–2509 (2009).
36
Nanomaterial Characterization
Several different methods and instruments are used to determine the properties on NMs
Shape and aggregation state
High Resolution Microscopy
Composition, purity and surface chemistry
Spectroscopy and Chromatography Methods
Surface charge
Zeta Potential Analysis
Crystal structure
X‐Ray Diffraction
Particle number and size distribution
Several Techniques (e.g. SEM,TEM,DLS)
Surface Area
Brunauer‐Emmett‐Teller Adsorption
There is no single instrument that is right tool for every test. There are in fact more than 400 QualityNano Training School 37
different techniques
for particle counting and characterization
What haven't we considered yet ?
Some NPs Have a tendency to form clusters (e.g. aluminium oxide)
in an agglomeration state,some NPs behave like larger particles in the media
Toxicity may also depends on the size of agglomerate
NOT on original particle size itself !!!
If the NPs aggregate
How to characterize aggregation ?
it is important to identify how Texture analysis ?
these changes affect toxicity
38
Hussain, S. e. (2008). Can silver nanoparticles be useful as potential biological labels? Nanotechnology 19 .
Challenges
for nano‐QSAR ?
Challenges for QSAR
Table shows the results of 3 different toxicity studies for silver NMs:
NP Size
type (nm)
Doses
Cell
(mg/ml)
Ag
7‐20
0.8‐50
Human fibrosorcoma and skin carcinoma Ag
<15
0.7
Human hepatoma
Ag
15 and
100
10‐250
Rat liver deliverd cell
line
Differences in NMs used
(size)
[1]
Differences in
experimental
conditions
Method
Result
XTT assay
Non‐toxic
MTT Alamar blue Oxidative stres and LDH assay
mediated toxicity
Kim et al. (2009)[2]
MTT, LDH, GSH
level, ROS and MMP
Hussain et al. (2005)[3]
Toxic
Araro et al. (2008)[1]
strongly affects
the results
Arora S, Jain J, Rajwade JM, Paknikar KM (2008). Cellular responses induced by silver nanoparticles: in vitro studies. Toxicol Lett 179:93–100
Kim S, Choi JE, Choi J, Chung KH, Park K, Yi J, Ryu DY (2009). Oxidative stress-dependent toxicity of silver nanoparticles in human
hepatoma cells. Toxicol In Vitro 23:1076–1084
[3] Hussain SM, Javorina AK, Schrand AM, Duhart HM, Ali SF,
QualityNano Training School 39
Schlager JJ (2006). The interaction of manganese nanoparticles with
PC-12 cells induces dopamine depletion. Toxicol Sci 92:456–463
[2]
Challenges for nano‐QSAR [1]
NMs are problematic for QSAR modellers because of the several reasons:
1
• Lack of high‐quality experimental data
2
• Lack of conceptual frameworks for grouping NPs according to mode of physicochemical properties and toxic action
3
• Lack of sufficient molecular descriptors able to express specificity of “nano” structure
4
• Lack of rational modeling procedures to screen large numbers of
structurally diversified nanoparticles.
5
• Limited knowledge on the interactions between NPs and biological systems
QualityNano Training School [1] T. Puzyn et al. (eds.), Recent Advances in QSAR Studies, 103–125. DOI 10.1007/978‐1‐4020‐9783‐6_4, C Springer Science+Business Media B.V. 2010
40
Part III
Case Study
despite all limitations…
QualityNano Training School 41
Nanotxicology 2012, 4‐7 September, Beijing
QSAR Toxicity (in vitro) Analysis for a Panel of Eighteen
NPs
The aim of this study was to test the hypothesis that nanoparticle
toxicity is a function of some measurable physical characteristics and
that Structure Activity Relationship (SAR) might ultimately be a useful
tool for nanotoxicology. A panel of 18 nanoparticles was chosen to
include carbon‐based materials and metal oxides. Comparative toxicity
of the nanoparticles in the panel was assessed, using a number of
different measures of apoptosis and necrosis; haemolytic effects and
the impact, on cell morphology were also assessed. Structural and
composition properties were measured including size distribution,
surface area, morphological parameters, metal concentration,
reactivity, free radical generation and zeta potential. Toxicity end
points were processed using principal component analysis and
clustering algorithms.
QualityNano Training School 42
The Set of Eighteen NPs
Nanooxicology 2012, 4‐7 September, Beijing
Particles
Qualifier
Carbon black
Diesel exhaust particles
Nanotubes
Fullerene
Printex 90
EPA
Japanese
Number
N1
N2
N3
N4
Unmodified
N5
Polystyrene latex
Amine
N6
Carboxylated
N7
7nm
N8
Aluminium oxide
50nm
N9
300nm
N10
Cerium oxide
N11
Nickel oxide
N12
Silicon oxide
N13
Zinc oxide
N14
Rutile
N15
Titanium dioxide
Anatase
N16
N17
QualityNano Training School Silver
* AHT: Allied High Tech, ** NAM: Nanostructure and Amorphous Material Inc
N18
Supplier
Degussa
EPA
Vicki Stone
Sigma
Polysciences
Sigma
Polysciences
Krahn Chemie
AHT*
AHT*
NAM**
NAM**
NAM**
NAM**
NAM**
NAM**
NAM**
NAM**
Form
Dry
Dry
Dry
Dry
Sus.
Sus.
Sus.
Dry
Dry
Dry
Dry
Dry
Dry
Dry
Dry
Dry
Dry
43
Sus.
Characterisation of Physico‐chemical Properties
Nanotoxiclogy 2012, 4‐7 September, Beijing
•Particle shape was analysed using LEO 1530 Scanning Electron Microscope (SEM) or Philips
CM20 Transmission Electron Microscope (TEM) aspect ratio and mean size
•Surface area and porosity were measured using TriStar 3000 BET. Thirteen measurements,
including five surface areas based on different definitions, three pore volumes, three pore
sizes as well as the mean size and particle density.
•Particle size and size distribution were analysed using a Malvern MasterSizer 2000. Size
distribution, and seven other size properties (mass diameter, uniformity, specific surface
area, surface area mean diameter and three mass diameters)
•The free radical activities were measured by EPR (Electron paramagnetic resonance)
using the spin traps, DMPO and Tempone H separately. Tempone H has partial selectivity for
trapping superoxide anions, whilst DMPO traps mainly hydroxyl radicals
•Particle reactivity in solution, the dithiothreitol (DTT) consumption test as a reducing species
was carried out. Since the DTT consumption test can only be conducted on dry powders, only
fourteen of the panel of nanoparticles were assessed using this assay.
•Metal Concentration water soluble concentration of ten heavy metals (Ti, V, Cr, Mn, Fe, Co,
Ni, Cu, Zn and Cd) was measured
•Charge: z potential was measured using Malvern Instrument’s Zetasizer Nano instrument
QualityNano Training School 44
Total Number of Measurements
13 • BET measurements (for 14 dry samples)
2 • SEM/TEM measurements (for all samples)
7 • laser diffraction size statistical measurements (Master Sizer 2000) for all samples
84 • laser diffraction size distribution measurements
2 • EPR measurements (for all samples)
1 • reactivity measurement (for dry samples)
10 • metal content measurements
Total:119 Measurements
Particle
name
BET Measurements
M2000
Measurements
SEM Measurements
Zeta measurements
Other
Measurements
N1
A(1,1),A(1,2),…,A(1,12)
B(1,1),B(1,2)…
C(1,1),C(1,2)…
D(1,1),D(1,2)…
E(1,1),E(1,2)…
N2
A(2,1),A(2,2),…,A(2,12
B(2,1),B(2,2)…
C(2,1),C(2,2)…
D(2,1),D(2,2)…
E(2,1),E(2,2)…
…
N18
…
A(18,1),A(17,2),…
…
…
QualityNano Training School B(18,1),B(18,2)…
C(18,1),C(18,2)…
…
…
D(18,1),D(18,2)…
E(18,1),E(18,2)…
45
SEM and TEM images of the 18 nanoparticle samples
Nanotoxicolgy 2012, 4‐7 September, Beijing
QualityNano Training School 46
Table 2. The water soluble concentrations of various metals for the 18 nanoparticle samples
The heavy metal concentration of each NP
Nanotoxicology 2012, 4‐7 September, Beijing
Particle names
Metal Concentration (µg/g)
Ti
V
Cr
Mn
Fe
Co
Ni
Cu
Zn
Cd
Carbon Black N1
0.000
0.057
0.000
0.305
0.000
0.002
0.263
1.161
9.687
0.008
Diesel Exhaust N2
2.320
0.312
16.47
20.81
208.4
0.508
4.940
2.235
3181
0.387
Japanese Nanotubes N3
0.000
0.100
0.000
0.161
13.10
0.000
0.299
0.123
0.460
0.001
Fullerene N4
0.000
0.084
0.000
0.000
0.000
0.000
0.534
0.142
4.837
0.000
Polystyrene Latex Beads N5
0.000
1.172
0.000
0.000
0.000
0.000
0.000
0.625
19.56
0.000
Polystyrene Latex Beads N6
4.368
0.779
0.343
0.199
0.000
0.006
0.440
18.23
58.71
0.010
Polystyrene Latex Beads N7
0.000
1.018
0.000
0.000
0.000
0.000
0.000
2.588
22.323
0.000
Aluminuim Oxide N8
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.714
0.000
0.000
Aluminuim Oxide N9
0.000
0.031
0.000
0.000
0.000
0.000
0.056
0.099
4.987
0.000
Aluminuim Oxide N10
0.000
0.062
0.000
0.000
0.000
0.000
0.023
0.088
0.000
0.000
Cerium Oxide N11
0.000
0.004
1.058
0.128
0.000
0.000
0.616
0.191
112.9
0.007
Nickel Oxide N12
0.000
0.056
0.000
0.000
0.000
0.042
16963
0.286
0.000
0.038
Silicon Oxide N13
0.000
0.088
0.000
0.000
0.000
0.000
14.163
0.106
0.000
0.000
Zinc Oxide N14
0.000
0.096
0.000
0.000
0.000
0.000
1.336
0.047
8185
9.711
Titanium Dioxide Rutile N15
0.000
0.083
0.000
0.000
0.000
0.000
0.286
0.276
39.623
0.000
Titanium Dioxide Anatase N16
0.000
0.104
0.000
0.160
0.000
0.006
0.692
0.303
402.3
0.000
Silver N17
0.000
0.111
0.000
0.503
0.000
0.000
0.475
0.597
17.867
0.019
Silver N18
0.000
0.096
0.000
0.256
0.000
0.059
0.561
162.8
194.8
0.037
QualityNano Training School 47
Characterisation of Particle Size and Shape
Nanotoxicology 2012, 4‐7 September, Beijing
Carbon Black N1
LEO 1530 Gemini FEGSEM
Scanning Electron Microscopy
The information about Size, Shape can be obtained from SEM images
Mean size obtained by measuring the particles on the SEM/TEM image: Mean_Size_SEM
Aspect ratio obtained from SEM/TEM images: Aspect_Ratio
QualityNano Training School 48
BET Measurements for 14 dry powder samples
Nanotoxicology 2012, 4‐7 September, Beijing
Micromeritics TriStar 3000
Surface Area Technique: Gas Adsorption
Minimum Surface Area: 0.01m2/g
Support Gases: Nitrogen and Helium
Measurement Parameters: BET Measurement
Adsorption Isotherms
Desorption Isotherms
t‐plots
Pore Volume Analysis
Pore Size Analysis
Surface Area (m2/g)
Pore Volume (cm3/g)

Single point surface area at P/Po = 
0.197 (m²/g) – SPSA

BET Surface Area: BET_SA

Langmuir Surface Area: LM_SA

BJH Adsorption cumulative surface
area of pores between 17.000 Å and
3000.000
Å
diameter
:

BJH_AC_SAOP

BJH Desorption cumulative surface
area of pores between 17.000 Å and
3000.000
Å
diameter:
BJH_DC_SAOP

Single point adsorption total pore
volume of pores at P/Po = 0.98:
SP_TP_VOP
Pore Sizes (Å)

Adsorption average pore width:
BET_PW

BJH Adsorption average pore
diameter : BJH_A_PD
BJH Adsorption cumulative volume
of pores between 17.000 Å and

3000.000 Å diameter:
BJH_AC_VOP
BJH Desorption cumulative volume

of pores between 17.000 Å and
3000.000 Å diameter :
BJH_DC_VOP

QualityNano Training School BJH Desorption average pore
diameter: BJH_D_PD
Mean size calculated
from BET surface area:
Mean_Size_BET
Particle density:
Density
** 14 dry powder samples, no available for three suspensions (N5,N6, and N7) and the new Silver sample.
49
Characterisation of NPs‐Mastersizer
Nanotoxicology 2012, 4‐7 September, Beijing
10
9
P r o b a b ilit y ( % )
8
7
6
5
4
3
2
1
0
0.01
0.1
1
10
100
1000
10000
Size (micron)
Malvern Mastersizer 2000
Type of Analyser: Laser Diffraction
Sample Volume: 1000ml, 150ml (Hydro G and S)
Measurement Range: 20nm – 2000μm
QualityNano Training School The information about mean Size and Size distribution, can be obtained from this sizer. 50
Characterisation of NPs‐Mastersizer
Nanotoxicology 2012, 4‐7 September, Beijing
Mastersizer 2000 Measurements: 91**








D[4,3] – the Volume Weighted Mean or Mass Moment Mean Diameter, represent by D[4,3]
Uniformity: a measure of the absolute deviation from the median, represented by Uniformity
Specific surface area: the total area of the particles divided by the total weight, SSA_2000
D[3,2] – the Surface Weighted Mean or Surface Area Moment Mean Diameter, represented by D[3,2]
D(v,0.5) – the size in microns at which 50% of the sample is smaller and 50% is larger. This value is also known as the Mass Median Diameter or the median of the volume distribution, represented by D(0.5)
D(v,0.1) – the size of particle below which 10% of the sample lies, represented by D(0.1)
D(v,0.9) – the size of particle below which 90% of the sample lies, represented by D(0.9)
Size distribution: there are 100 values. Among them there are 16 values equal to zero for all particles. So 84 size distribution attributes are used for further analysis
** Note: available for all samples
QualityNano Training School 51
Characterisation of NPs‐Zeta Potential
Nanotoxicology 2012, 4‐7 September, Beijing
Malvern ZetaSizerNano
instrument
Size ranges: 0.6nm ‐ 6µm
Zeta Potential Suitable: suspensions
QualityNano Training School 52
Toxicity Results
Cytotoxicity by LDH Release
Viability
25.00
N6
80.00
40.00
100.00
20.00
Apoptosis
60.00
120.00
N12
N14
Viable
100.00
LDH Release
Apoptosis
80.00
15.00
60.00
10.00
N6
40.00
N14
20.00
20.00
5.00
N6
0.00
31.25
62.50
125
250
Dose (g/ml)
0.00
0.00
31.25
62.50
125
Pro‐inflammation effects
80
62.50
125
250
Dose (g/ml)
Dose (g/ml)
Necrosis
100
Haemolysis
90
Haemolysos at dose 500 g/ml
70
N6
60
Necrosis
31.25
250
50
40
N14
30
20
N3
10
80
70
60
50
40
30
20
10
0
31.25
62.50
125
Dose (g/ml)
250
0
QualityNano Training School
53
Principal Component Analysis of Cytotoxicity Data
LDH
Viability
Apoptosis
MTT
Total‐Toxicity
Necrosis
Haemolysis
QualityNano Training School 54
Principal Component Analysis of Cytotoxicity Data
‐ The major cluster contains low toxicity particle samples Zinc oxide
‐Aminated beads (N6), zinc oxide (N14), Japanese Nanotubes (N3) and nickel oxide (N12) are separate from the major cluster. ‐ Aminated beads (N6) has the highest toxicity values in nearly all assay results (LDH, apoptosis, Japanese nanotubes
necrosis, haemolytic, MTT and cell morphology assays).
Polystrene larex (Amine)
‐ Zinc oxide (N14) has high toxicity value in LDH, apoptosis, necrosis and inflammation assays. Nickel oxide
‐ Japanese nanotubes (N3) have high toxicity values in viability and MTT assays. Figure : Plot of the first and second PCs based on PCA analysis of toxicity data ‐Nickel oxide (N12) has shown highly toxic in LDH and haemolytic assays results..
The distinction of the four particles, N6, N14, N3 and N12, from other particles is clear.
QualityNano Training School 55
NP6‐Average Toxicity Comparison
NP14‐Average Toxicity Comparison
N6 has the highest toxicity values in nearly all assay
results (LDH, apoptosis, necrosis, haemolytic, MTT and cell morphology assays). N14 has high toxicity value in LDH, apoptosis,
necrosis and inflammation assays.
NP12‐Average Toxicity Comparison
NP3‐Average Toxicity Comparison
N12 has shown highly toxic in LDH and haemolytic
assays results N3 have high toxicity values in viability and MTT assays. The toxicity results of N6, N3, N12 and N14 which are likely to be toxic
QualityNano Training School 56
Next Step: Principal Component Analysis of Descriptors
The focus of the (Q)SAR Analysis is to identify the possible structural and
compositional properties which may contribute to the toxicity of zinc
oxide (N14), polystrene latex amine (N6), Japanese nanotubes (N3) and
nickel oxide (N12).
The theory is that there should be structural or/and compositional
distinctions for these nanoparticles from the rest of the non‐toxic
particles.
PCA and clustering were then applied to the physicochemical
descriptors with the hypothesis that the toxic nanoparticles might also
be discriminated from the non‐toxic particles based on the analysis of
the measured physicochemical descriptors.
If this can be proved, further analysis can be conducted to identify the
key physicochemical descriptors that lead to the observed toxicity.
QualityNano Training School 57
Principal Component Analysis of Descriptors
Principal component analysis of the structural and compositional descriptors for the panel of 18 nanoparticles (excluding BET and DTT measured descriptors since they are not available for the four wet samples). ‐The scaling operation was performed, 3PCs were then applied.
N5
N6
‐PC1‐PC2‐PC3 three dimensional plot shows N7
N3
that the 18 nanoparticles were clearly grouped N2
N1 2
into clusters.
N1 4
‐The largest cluster contains particles that showed low toxicity values.
‐Japanese Nanotubes (N3), nickel oxide (N12) and zinc oxide (N14) and diesel exhaust Figure : PCA analysis of the physicochemical
descriptors for the panel of 18 nanoparticles
(excluding BET and DTT measured descriptors
since they are not available for the four wet
samples).
particles (N2) are outside this low toxicity sample cluster.
‐There must be another reason that leads to high toxicity value for N6
QualityNano Training School 58
Results
•To examine the clustering results more clearly, PC1‐
PC2, PC1‐PC3 and PC2‐PC3 two dimensional plots are also examined .
•It is PC2 that separates N3(Japanese nanotubes), N12 (nickel oxide) and N14 (zinc oxide) from the rest samples, as can be seen from PC1‐PC2 and PC2‐PC3 plots.
QualityNano Training School 59
Results
•The contribution plot for, PC1, PC2 and
PC3 is shown in the Figure, we are mainly
interested in the contribution to PC2 (solid
red
colour),
since
it
is
PC2
that
discriminated Japanese nanotubes (N3)
nickel oxide (N12) and zinc oxide (N14)
from the rest particle samples.
•The largest contributions to PC2 are: aspect
ratio measured by SEM imaging, and volume
weighted mean, uniformity, and D(0.9), all
measured by Mastersizer, as well as Ni and
Zn metal contents
Figure: PC1, PC2 and PC3 Contribution plots based on PCA analysis of the structural and compositional descriptors for the panel of 18 nanoparticles
•Further analysis is required.
QualityNano Training School 60
Results
N3
•In order to determine exactly which
physicochemical descriptors lead to high toxicity values of N3, N12 ,N14
N14
and N6, several PC analysis were
N2
performed.
N 12
•For the 14 dry particles, PC analysis with BET and DTT measurements gave
similar result with the first PC analysis
in which BET and DTT data was
Figure : Principal component analysis of the structural excluded.
and compositional descriptors (all such descriptors including BET and DTT measurements) for the 14 dry nanoparticles.
QualityNano Training School 61
Results
Figure: PC1 and PC2 contribution plots for PCA analysis of the compositional descriptors
•To identify the relative
contribution of each metal
concentration to PC1 and
PC2, the contribution plot to
PC1 and PC2 is shown in the
Figure.
• Ni content made the most important contribution to PC1, suggesting that nickel content may be the main reason for the high toxicity of N12 (nickel oxide). • While Zn content made the most obvious contribution to PC2, suggesting that zinc content may be the reason for the high toxicity of N2 (diesel exhaust particles) and N14 (zinc oxide).
QualityNano Training School 62
Table 2. The water soluble concentrations of various metals for the 18 nanoparticle samples
The heavy metal concentration of each NP
Nanotoxicology 2012, 4‐7 September, Beijing
Particle names
Metal Concentration (µg/g)
Ti
V
Cr
Mn
Fe
Co
Ni
Cu
Zn
Cd
Carbon Black N1
0.000
0.057
0.000
0.305
0.000
0.002
0.263
1.161
9.687
0.008
Diesel Exhaust N2
2.320
0.312
16.47
20.81
208.4
0.508
4.940
2.235
3181
0.387
Japanese Nanotubes N3
0.000
0.100
0.000
0.161
13.10
0.000
0.299
0.123
0.460
0.001
Fullerene N4
0.000
0.084
0.000
0.000
0.000
0.000
0.534
0.142
4.837
0.000
Polystyrene Latex Beads N5
0.000
1.172
0.000
0.000
0.000
0.000
0.000
0.625
19.56
0.000
Polystyrene Latex Beads N6
4.368
0.779
0.343
0.199
0.000
0.006
0.440
18.23
58.71
0.010
Polystyrene Latex Beads N7
0.000
1.018
0.000
0.000
0.000
0.000
0.000
2.588
22.323
0.000
Aluminuim Oxide N8
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.714
0.000
0.000
Aluminuim Oxide N9
0.000
0.031
0.000
0.000
0.000
0.000
0.056
0.099
4.987
0.000
Aluminuim Oxide N10
0.000
0.062
0.000
0.000
0.000
0.000
0.023
0.088
0.000
0.000
Cerium Oxide N11
0.000
0.004
1.058
0.128
0.000
0.000
0.616
0.191
112.9
0.007
Nickel Oxide N12
0.000
0.056
0.000
0.000
0.000
0.042
16963
0.286
0.000
0.038
Silicon Oxide N13
0.000
0.088
0.000
0.000
0.000
0.000
14.163
0.106
0.000
0.000
Zinc Oxide N14
0.000
0.096
0.000
0.000
0.000
0.000
1.336
0.047
8185
9.711
Titanium Dioxide Rutile N15
0.000
0.083
0.000
0.000
0.000
0.000
0.286
0.276
39.623
0.000
Titanium Dioxide Anatase N16
0.000
0.104
0.000
0.160
0.000
0.006
0.692
0.303
402.3
0.000
Silver N17
0.000
0.111
0.000
0.503
0.000
0.000
0.475
0.597
17.867
0.019
Silver N18
0.000
0.096
0.000
0.256
0.000
0.059
0.561
162.8
194.8
0.037
QualityNano Training School 63
Conclusions
Identifying the factor that discriminates N6 (polystrene latex amine) from N5 (unmodified) and N7 (carboxylated) PCA analysis: PCA analysis of metal contents only, or PCA analysis of structural descriptors only, or PCA analysis of both structural and compositional descriptors together, could not discriminate N5, N6 and N7 from the largest clusters of non‐toxic particle samples. PCA Conclusion: neither structural descriptors nor metal contents made N6 different fro N5 and N7 in toxicity
It was at this stage of analysis that we suspected that there are other possible reasons for toxicity specific to N6: either it was due to the charge difference, or due to N6 being amine, or both. zeta potential for all three polystyrene latex samples.
N6 (amine) 37.8, 37.5, and 40.3 (mV), positive N5 (unmodified) ‐36.2, ‐38.8 and ‐36.8, negative
N7 (carboxylated)
‐54.9, ‐55.3, and ‐58.6.,
negative
According to Verma et al. any particle with positive charge is likely to interact electrostatically since most cell surfaces and other biological membranes are negatively charged under physiological conditions(Verma et al. 2008). Therefore, it is most likely that the large positive charge of N6 contributed its observed high toxicity, despite it structurally similar to N5 and N7.
QualityNano Training School 64
Conclusions
• The SAR analysis generates the following conclusions:
1. The most likely reason of high toxicity for Japanese nanotube (N3) is its shape (high aspect ratio). 2. The most likely cause of high toxicity for zinc oxide (N14) and nickel oxide (N12) is their high contents of zinc and nickel. 3. Although, Aminated beads (N6) has the highest toxicity values in nearly all assay results , it was always grouped closely with the other two polystyrene latex beads, unmodified beads (N5) and carboxylated beads (N7), which showed no toxicity. Therefore, the zeta potential of the three polystrene latex was measured. The result suggests that the large positive charge of N6 contributed to its observed high toxicity.
• The work has shown that SAR based on molecular descriptors is a useful
tool for describing factors influencing the toxicity of nanoparticles. • A panel of eighteen nanoparticles is still considered to be too small, therefore, a QSAR model that can be applied to predict the toxicity of new nanoparticles cannot be developed with this small set of data.
QualityNano Training School 65