ABSTRACT
TO, KIMBERLY THOAIPHUONG. Predictive Modeling for Developmental Toxicity of Engineered Nanomaterials. (Under the direction of Dr. David Reif).
Nanomaterials are defined as particles with at least one dimension on the nanoscale (1-100nm).
Although their toxicity is not yet fully understood, nanomaterials present themselves in a wide
variety of consumer products, from sunscreens to laser printers. Nanomaterials have been
proposed as drug delivery systems and as improvements to in vivo molecular imaging
techniques. Despite the promising applications of nanomaterials, understanding of their health
impact remains elusive. Further, because nanomaterials are already present in consumer
products, there is a sense of urgency in understanding nanomaterial toxicity as a preventative
measure. Alternative testing strategies, such as quantitative predictive modeling can be
developed to accelerate nanomaterial risk assessment. The development of a predictive model for
engineered nanomaterials is difficult due to inconsistencies in nanomaterial characterization and
the selection of endpoints within existing data. The models that have been tested use a
combination of measured physicochemical characteristics and computationally derived
descriptors. Informed design of these models can identify the characteristics that underlie the
molecular mechanisms of nanomaterial toxicity. Because of the difficulties in data aggregation,
imputation of characteristics and multi-source endpoint summarization would provide a full-rank
matrix for which predictive modeling can be performed. Of course, the quality of data imputation
is dependent on an abundance of high quality data, a known drawback of existing nanomaterial
toxicity data. By identifying which characteristics are most influential to nanomaterial toxic
behavior, we can decrease the minimum information required in datasets for data curation.
Chapter 1 is a review of available nanotoxicity data and predictive modeling efforts through
imputation methods for high throughput toxicological data and subsequent hazard ranking.
Chapter 3 maps nanomaterial physicochemical features to multiple-endpoint developmental
© Copyright 2019 by Kimberly T. To
Predictive Modeling for Developmental Toxicity of Engineered Nanomaterials
by
Kimberly Thoaiphuong To
A dissertation submitted to the Graduate Faculty of North Carolina State University
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy
Bioinformatics
Raleigh, North Carolina 2019
APPROVED BY:
_______________________________ _______________________________ David Reif James Bonner
Committee Chair
DEDICATION
BIOGRAPHY
Kimberly T. To was born and raised in Raleigh, NC where she graduated from Wakefield High
School in 2010. She received her Bachelors of Science in Statistics from North Carolina State
University in 2014. Afterwards, she was coerced into attending graduate school by her
undergraduate mentor. She then began her doctoral studies in the Bioinformatics Program at NC
ACKNOWLEDGMENTS
I would first like to acknowledge my advisor Dr. David Reif for his guidance and support. Thank
you for encouraging me to never stop never stopping. I would like to thank the Bioinformatics
program director, Dr. Spencer Muse for always believing in me and pushing me when I was too
afraid to push myself. I would also like to thank my committee members Dr. Jamie Bonner, Dr.
Denis Fourches, and Dr. Brian Reich for their contributions to my research. Thank you to Dr.
Robert Tanguay and Dr. Lisa Truong at Oregon State University for making this work possible. I
would also like to thank members of the Reif Lab: Dr. Skylar Marvel, Dr. Guozhu “Dale” Zhang,
Dr. Michele Balik-Meisner, Dr. Kyle Roell, Marissa Kosnik, and Dylan Wallis.
Thank you to my parents for always being there. I would also like to thank my sister and
brother-in-law for their constant love and support. To the friends I’ve made along the way, whose
commiseration made graduate school all the more better: Desiree Unselt, Saddy Wisotsky, Kevin
Gillespie, Jeremy Ash, and Alex Eyre. Thank you to my second family Juni Cuevas and Chase
Benson, for always happily telling me what I don’t want to hear. Thank you to my soul sister,
Ravid Gur, who never hesitated to take me to the Hall of Gratuitous Praise. Finally, I would like
to thank my best friend and partner, JD Pittman, who has been my life support throughout this
TABLE OF CONTENTS
LIST OF TABLES ... vii
LIST OF FIGURES ... viii
CHAPTER 1: A Review of Predictive Modeling of Engineered Nanomaterial Toxicity ... 1
1.1. Introduction ... 2
1.2. Risk Prediction in Modern Toxicology ... 4
1.3. Nanotoxicology Data ... 6
1.3.1. Data Quality ... 6
1.3.2. Existing Sources ... 10
1.3.3. Nanomaterial Data Curation Initiative ... 13
1.4. Predictive Modeling ... 13
1.4.1. Grouping ... 16
1.4.2. Read-Across ... 19
1.4.3. NanoQSARs ... 22
1.4.3.1. Linear Regression Methods ... 22
1.4.3.2. Support Vector Machines ... 25
1.4.3.3. Neural Networks ... 26
1.4.3.4. Decision Trees ... 30
1.4.3.5. Comparison of Methods ... 31
1.5. Conclusion ... 31
1.6. References ... 33
CHAPTER 2: Characterizing the Effects of Missing Data and Evaluating Imputation Methods for Chemical Prioritization Applications using ToxPi ... 45
2.1. Background ... 46
2.2. Methods... 49
2.2.1. Data ... 49
2.2.2. Data Simulation ... 50
2.2.3. Data Imputation ... 50
2.2.4. ToxPi Calculation ... 51
2.2.5. ToxPi Score and Rank Evaluation ... 52
2.2.6. Statistical Software ... 54
2.3. Results ... 54
2.3.1. Missing Data ... 54
2.3.2. Wilcoxon signed-rank tests show most imputation methods result in significant changes to score ... 54
2.3.3. Minimum value imputation serves as a baseline for comparison ... 54
2.3.4. kNN causes unstable shifts in chemical ranks ... 57
2.3.5. Maximum value imputation causes uniform shifts in score and no shift in rank ... 57
2.3.6. Mean, binomial, and LLS imputation show similar effects ... 57
2.3.7. Variance of scores is stable in all methods except SVD ... 60
2.3.8. Overall ToxPi score variance differs significantly ... 60
2.4. Discussion ... 60
2.5. Conclusion ... 63
CHAPTER 3: Multivariate Modeling of Engineered Nanomaterial Features Associated
with Developmental Toxicity ... 66
3.1. Introduction ... 67
3.2. Materials and Methods ... 69
3.2.1. ENM Characterization ... 69
3.2.2. Experimental Design ... 70
3.2.2.1. ENM Exposure Solutions ... 70
3.2.2.2. Zebrafish ... 70
3.2.3. Modeling ... 71
3.3. Results ... 72
3.3.1. Physicochemical Characteristic Analysis ... 72
3.3.2. The full model identifies key physicochemical characteristics ... 75
3.3.3. Reduced bagged decision tree shows similar performance to full model ... 77
3.3.4. Underprediction of more toxic ENMs improves with the inclusion of PCC ... 79
3.3.5. Variable wAggE responses across PCC reveal intricate characteristic interactions for toxicity response ... 79
3.4. Discussion ... 81
3.5. Conclusion ... 84
3.6. References ... 85
CHAPTER 4: Review, Current Work, and Future Directions ... 90
4.1. Introduction ... 90
4.2. On the Development of Predictive Models for Engineered Nanomaterial Toxicity ... 91
4.2.1. Quantitative Predictive Models for Nanotoxicity ... 91
4.2.2. Differences and Similarities in Nanotoxicity models ... 93
4.3. Implications of Data Imputation on Toxicity Profiling ... 93
4.3.1. Methods resulting in similar hazard score changes have different effects on mean rank change ... 93
4.3.2. Application to Nanotoxicity Data ... 95
4.3.3. Imputation Options for ENM Characteristics ... 96
4.4. Mapping Redundant Features to Toxicity Endpoint ... 97
4.4.1. Bagged decision tree identifies important variables and redundancy in feature space ... 97 95 4.4.2. Principal components analysis shows inherent groupings of PCC ... 97
4.4.3. Grid predictions show wAggE sensitivity to extreme PCC values ... 99
4.5. Current Work and Future Directions ... 101
4.5.1. Mapping Diameter, Circularity, and PDI to Behavioral Outcomes ... 101
4.5.2. Improving Published Experimental data for Predictive Modeling ... 103
4.5.3. Comparative Analysis of Imputation Methods for PCC ... 103
4.5.4. Comparative Analysis of Predictive Modeling Methods ... 104
4.6. Conclusions ... 104
LIST OF TABLES
Table 1.1 Summary of Nanomaterial Registry records for biological interactions, adapted from nanomaterialregistry.org ... 11
Table 1.2 Summary of eNanoMapper records for in vitro toxicity data ... 12
Table 2.1 Values for the range of RMSE and range of Rank Change for each of 7 imputation methods, separated by the number of assays per slice ... 59
Table 3.1 The fourteen engineered nanomaterials tested ... 69
Table 3.2 Numeric distributions for each physicochemical characteristic. The minimum, mean (average), and maximum values for the ENM set are presented ... 73
Table 3.3 Variable Importance. Variables are sorted by calculated variable importance. Percent increase in MSE measures how model accuracy decreases when a single variable’s values are permuted ... 76
LIST OF FIGURES
Figure 2.1 Conceptual overview of the simulation process and experimental design. Assays were randomly sampled from the original data based on a desired number of assays and assay sources (slices) so that the simulated datasets contained a subset of assays with arbitrarily assigned sources and all of 426 chemicals present in the original dataset. Simulated datasets were imputed and ToxPi profiles were calculated, with an overall summed ToxPi score given for analysis .. 55
Figure 2.2 Comparison of Imputation Methods Using ToxPi priority ranks. Mean ToxPi Rank Change between Imputed Simulated Data and Imputed Raw Data. Rank change was calculated by using the magnitude of difference between individual chemical ranks in the imputed raw data and the chemical ranks from each simulated dataset. Binomial, LLS, Maximum, and Mean show small magnitudes of change in rank. kNN shows a wider variation in rank change and therefore represents less stability in the method. Minimum value imputation and SVD present wider ranges in rank change, although the magnitude of change is smaller than kNN ... 56
Figure 2.3 Comparison of Imputation Methods by Toxpi Score. (a) Root Mean Square Error between Imputed Simulated Data ToxPi Scores and Imputed Raw Data Chemical Scores. After imputation and ToxPi calculation, scores were compared to the ToxPi scores using the standard “0” method. RMSE density distributions are separated by imputation method. The distribution of kNN is centered at the lowest RMSE compared to the other methods. Binomial, LLS, and Mean imputation are heavily overlapped. SVD is centered similarly, but shows a wider spread. Maximum imputation has the largest RMSE. (b) ToxPi Score Variance of Imputed Simulated Data ToxPi Scores. Amongst 1000 replicate simulations, the variance for each of 426 chemicals was calculated. Compared to SVD, all other methods present relatively low variability from chemical to chemical. SVD has an extremely wide range of ToxPi Score variability ... 58
Figure 3.1 Correlation plot of physicochemical characteristics (PCC). Plots along the diagonal show the distributions for each characteristic. Correlations are shown in the lower triangle, with colors corresponding to the direction and magnitude of the correlation ... 74
Figure 3.3 Over and underprediction by ENM for each model. For each model (column) and ENM (row), the predicted wAggE is measured along the vertical axis and the observed wAggE is measured on the x-axis. Reported MSE are calculated as departures in the predicted wAggE from the observed wAggE. Black points indicate ENMs with underpredicted wAggE and blue points indicate ENMs with overpredicted wAggE. The diagonal line separates under and over prediction and the dotted line is at the χ20.5 = 2.37 threshold ... 78 Figure 3.4 Weighted Aggregate Entropy (wAggE) Measurements for ENM across selected
PCC. The left plot shows wAggE values for each ENM across the concentration curve. Darker shades of blue indicate higher wAggE values. The right plot shows PCC measurements for each ENM. Yellow shades are scaled for each descriptor so that darker shades indicate higher values of any one descriptor and lighter shades indicate lower values ... 82
Figure 4.1 Multiple linear regression diagnostic plot. Horizontal axis shows fitted values, vertical axis shows residuals. Horizontal line indicates residuals = 0 and the red line is a smoothed trend line ... 92
Figure 4.2 Mean chemical hazard ranking changes. The horizontal axis shows the number of assays per slice and the vertical axis measures mean rank change. Plots are split by imputation method and number of slices ... 94
Figure 4.3 RMSE between ToxPi scores from each imputed dataset and true scores. The horizontal axis shows the number of assays per slice and the vertical axis measures RMSE. Plots are split by imputation method and number of slices ... 94
Figure 4.4 Scree plot from principal components analysis with 11 PCC. The horizontal axis shows each principal component. The vertical axis describes the proportion of variance explained by the principal components ... 98
Figure 4.5 Biplot between the first two principal components. The horizontal and vertical axes plot the first and second principal components, respectively. PCC loadings are mapped with lines and colored by contribution to the principal components. ENM scores are plotted and colored by wAggE values ... 98
Figure 4.6 Concentration vs. diameter prediction plot. The left shows a surface plot of predictions across a grid of possible values. The plot on the right shows an overhead view, where the horizontal axis shows concentration and they vertical axis shows diameter and is colored by wAggE values ... 100
Figure 4.8 Concentration vs. PDI prediction plot. The left shows a surface plot of predictions across a grid of possible values. The plot on the right shows an overhead view, where the horizontal axis shows concentration and they vertical axis shows PDI and is colored by wAggE values ... 100
CHAPTER 1
A Review of Predictive Modeling of Engineered Nanomaterial Toxicity
Chapter 1 is a review of available nanotoxicity data and predictive modeling efforts through
grouping, read across, and nano-QSAR. Chapter 2 describes a simulation study assessing data
imputation methods for high throughput toxicological data and subsequent hazard ranking.
Chapter 3 maps nanomaterial physicochemical features to multiple-endpoint developmental
toxicity. Chapter 4 reviews the previous work and discusses future directions.
ABSTRACT
Engineered nanomaterials are man-made materials measuring with at least one dimension on the
nanoscale (1-100nm). Although their toxicity is not yet fully understood, nanomaterials present
themselves in a wide variety of consumer products, such as sunscreens or laser printers.
Nanomaterials have been proposed as drug delivery systems and as improvements to in vivo
molecular imaging techniques. Despite the promising applications of nanomaterials, their health
impact remains elusive. Further, because nanomaterials are already present in consumer
products, there is a sense of urgency in understanding nanomaterial toxicity as a preventative
measure. As such, the development of a predictive model for nanotoxicolgy will expedite
regulatory implementation for the safety of human and environmental health to nanomaterial
exposure. In this review, we discuss the limitations of existing nano toxicity data for the
purposes of predictive models. Implementations of grouping, read-across, and quantitative
nano-structure activity relationships (QNARs) for predicting toxicity are described. Reviewed QNAR
methods showed adequate prediction accuracy, but they used varying data sources with differing
experimental. For the purpose of developing a global predictive model, more uniform data are
Keywords: Nanomaterials, Nanotoxicity, Predictive Modeling, Hazard Assessment, Risk
prediction, Read-Across, QNAR
1.1. Introduction
Nanotechnology is a growing field dealing with materials on the nanoscale, whose size lends
them unique characteristics that support a high range of potential applications [1–3]. For
example, nanomaterials that are generally considered nontoxic are present in consumer products,
such as cosmetics and apparel [4]. Further, nanomedicine is a subfield dealing with applications
of nanotechnology in medicine, with growing interest in drug delivery systems owing to certain
nanomaterials’ having the ability to cross membrane barriers for more targeted drug therapy [5,
6]. Various studies have shown that the small size creates an advantage of greater surface area
exposure at equivalent concentrations to larger materials [7, 8]. Besides the small size, studies of
the effects caused by nanomaterial structural characteristics or in vivo molecular interactions show different behaviors from their well characterized bulk counterparts [9, 10]. Thus, further
toxicity screening and hazard risk assessment are needed for both novel and existing materials.
Recognition of toxicity concerns has stimulated the developing field of nanotoxicology. A major
challenge has been the standardization of definition, methodology, and data [11]. For several
years, the term “nanomaterial” has had no universally acknowledged definition, with different
regulatory agencies providing varying, albeit congruous, descriptions of what constitutes a
nanomaterial [12]. However, a formal and discrete definition is deemed necessary for future
research as well as consistent regulatory purposes. In 2011, the European Commission released a
recommendation for the definition, with guidelines on determining how to classify an item as a
size distribution for each unique nanomaterial, which, by design, will account for varying sizes
and aggregation or agglomeration that is common amongst nanomaterials. A flexible threshold
determines that a material will be classified as a nanomterial if at least 50% of the number size
distribution falls between the range of 1nm and 100nm [13]. The flexibility of this criterion
allows for future changes to threshold, should new information arise regarding size influences of
toxicity for a given nanomaterial.
Given this definition of the term “nanomaterial” one can consider the probabilistic nature of
nanomaterial measurements and descriptors as they relate to nanotoxicity. Common statistical
methods that have been explored use nanomaterial characteristics to categorize nanomaterials
with similar profiles and therefore potentially predict effects [14]. Amongst grouping and
categorization methods, nano-QSARs take advantage of an existing regression and
categorization framework that predicts biological outcome using material structural descriptors
as predictors. The overarching goal of these methods is to characterize nanomaterials so that
initial studies can benefit from accurate assumptions of toxic effects based on descriptors of
particular groups [15]. Grouping nanomaterials based on similar, yet naturally variable
characteristics underlies an intrinsically probabilistic view on predictive parameters. The
inconsistencies present among case-by-case chemical testing and screening studies highlight the
need for machine learning techniques that would effectively extract the maximum amount of
information from the available data and utilize this new information for newer substances with
similar parameter distributions. A handful of studies exist that show the accuracy of developed
networks for predicting toxicity and ranking chemicals for further testing. Implementation of
machine learning techniques for improvements to nano-QSAR predictions have been reported
Grouping and read across methods represent a preliminary step to the process of toxicity
prediction, where the primary theory is that similarly characterized materials will behave in a
similar manner in regards to toxic behavior. However, the addition of quantitative components
can allow for more precise predictions, both within groups and in a global dataset. In fact, some
described predictive models utilize data from various literature sources, with a generalized
toxicity endpoint accurately predicted for a variety of nanomaterials. [19, 20]. The success of
these types of models depends on common descriptors being present in each individual dataset.
Unfortunately, data aggregation can result in mismatched data, for which different descriptors
have been described and therefore hinders the development of a complex mathematical
relationship between descriptors and endpoint. Of course, data imputation can resolve these
issues by filling in gaps using a predictive process [21].
In this paper, we review the progress of predictive nanotoxicology and the methods that have
been implemented. We begin with an overview of traditional methods in toxicology and the
implications of current data concerns in nanotoxicology. In the next sections, we cover grouping
and read across for nanomaterials and describe the abundance of statistical methods applied to
nanotoxicity data.
1.2. Risk Prediction in Modern Toxicology
Currently, toxicity testing of new and existing chemicals is performed using the OECD
guidelines for the Testing of Chemicals [22]. These guidelines list a variety of relevant toxicity
endpoints that take into account both human and environmental health impacts. Having a
standardized reporting protocol for toxicity testing is beneficial so that data from different
in computational power and high throughput screening (HTS) limited toxicity testing to
case-by-case studies, making data aggregation and consequently chemical profiling quite tedious [23,
24]. Technological advancements have allowed for improved HTS methods for thousands of
chemicals, resulting in thousands of datasets that have been compiled and made publicly
available through the PubChem database [25].
Included in this database, the U.S. government consortia on Toxicity in the 21st Century (Tox21) and Toxicity Forecaster (ToxCast) have implemented in vitro HTS assays for quick and efficient toxicity testing of both well-studied chemicals and new or untested chemicals with an ultimate
goal of ranking and prioritizing chemicals to identify and further test those with highest toxicity
[26–28]. In vitro HTS is becoming more prevalent due to the growing number of insufficiently tested or untested chemicals and a greater emphasis on reducing animal testing, a recognized
goal of the EU's REACH program [29]. However, this type of testing typically results in large
amounts of data that need to be analyzed [24, 30]. As such, the developing a computational
method for predicting chemical toxicity is a necessary step, as both time and resources can be
saved by utilizing data at hand.
Moreover, calls for data collection and data gap filling in hazard assessment in the REACH
protocol are dependent on functional in silico methodology [31]. Typically, the developed methods, such as read-across and QSARs, take advantage of structural similarities amongst
materials to group them and predict chemical behaviors based on available data for tested
chemicals within the same groups [32]. The inclusion of molecular descriptors have been
recommended for better predictive power, as well as more complex methods such as neural nets
molecular descriptors that can predict NOEL using an artificial neural network with data from
the Japan Existing Chemical Database [36].
The OECD has acknowledged the need to develop new testing strategies or harmonize existing
testing strategies for applicability to nanomaterial safety assessment. Because nanomaterials and
larger materials have been shown to differ in both size and biological behavior, the existing
methods may inadvertently miss significant behaviors of nanomaterials that typically would not
need to be considered in standard toxicology [37]. Nanotoxicology itself has adapted standard
QSARs for use with nanomaterials into nano-QSARs [38]. nQSARs or QNARs mirror the
methodology of standard QSARs while incorporating nano-specific descriptors such as
zeta-potential, surface descriptors, or molecular shape [39]. For example, Mu et. al used existing data
from 16 metal oxides to develop an improved nano-QSAR model that included two new
parameters: enthalpy of formation of a gaseous cation and polarization force [15].
Recommendations for continuing improvements to nano-QSARs suggest the inclusion of more
complex methods, such as Fjodorova's nano-QSAR built using artificial neural networks using
descriptors such as enthalpy of formation of a gaseous cation and those relating back to
placement on the periodic table [18]. It is pertinent that these differing characteristics be
considered for nanotoxicity predictions because nanomaterials have been shown to behave
differently than their bulk counterparts [40].
1.3. Nanotoxicology Data 1.3.1. Data Quality
As the field of nanotoxicology grows, the amount of data will continue to increase. However, we
acquiring data can be completed by (1) finding it in an existing repository (surveyed below), (2)
downloading the submitted data file from the corresponding publication, or (3) extracting the
data from the corresponding publication, when no explicit data submission exists. The methods
for acquiring data are fairly straightforward; extraction from a publication can be tedious, but
software have been developed for these purposes [41].
Once data are acquired, we need to assess data quality. Ideally, we would like for the data to be
within a repository and directly comparable to other datasets in the repository. Because
nanotoxicology is new and quickly developing, the research interests within the field are quite
diverse. Depending on specific research interests, different datasets may report different
physicochemical characteristics (PCC), making data comparability more difficult. This can be
prevented by the development of a globally accepted minimum information standard, similar to
those developed in other fields. Although this has not yet been established in nanotoxicology, the
Nanomaterial Registry (RTI) has developed their own Minimum Information about
Nanomaterials (MIAN) [42] that contains guidelines for reporting PCC, along with the methods
used to measure them. This MIAN is applied within the database itself, allowing for more
detailed filtering as well as more streamlined identification of similar nanomaterials [43].
However, if the goal is to use currently available data for risk prediction, an alternative solution
is with the use of Nano-Quantiative Structure-Property Relationships for predicting nanomaterial
properties. Nano-QSPRs have been successfully developed for zeta potential [44, 45], and could
therefore present a preliminary data processing step for predicting properties known to affect
toxicity, but may be difficult to measure or not available in a given dataset.
Another issue affecting data quality is the lack of a standardized naming convention. For
acquisition. By developing a standard naming convention, these different names can be linked
together, and data can be extracted more efficiently [46]. The Nanoparticle Ontology (NPO) was
developed for the purposes of sharing diverse data for nanomaterials in cancer research [47]. The
structured ontology allows for simple descriptions of complex structures, as well as the methods
used. Although NPO was developed for cancer research, it is designed for growth and has been
adapted for use in nanomaterial annotation. In this implementation, the NPO is converted to a
string expression that can be computationally processed for assessing similarities and differences
between nanomaterial structures [48]. For the purposes of data aggregation, the NPO has been
used to collect nanoparticle properties from cancer treatment literature with high precision [49].
In addition, the NPO has been adapted for use in the eNanoMapper database, discussed below.
A common data file format is also necessary for successful data curation because it allows for
dataset comparability. ISA-TAB-Nano is a file format using NPO, developed by the NCIP Nano
WG [50]. ISA-TAB-Nano is an extension of the existing file format ISA-TAB with adaptations
for nano-specific descriptors. ISA-TAB-Nano uses 4 labeled spreadsheets that together describe
reference information, study protocols, sample endpoint measurements, and descriptions of the
material. The information provided and ease of use makes ISA-TAB-Nano an important tool in
the goal of nanotoxicology data curation. ISA-TAB's predecessor MAGE-TAB included its own
set of minimum information standards for microarrays [51], so the use of standardized language
and an implementation of minimum information standards would make ISA-TAB-Nano the gold
standard for nanotoxicology data keeping.
The issues discussed thus far pertain to data keeping, with a seemingly simple solution of
standardization that can be easily adopted by researchers of nanosciences. A more difficult issue
in-house measurements [52], and therefore need to be validated. Sources of variation in these
descriptions are not limited to manufacturer error, but can arise from transport and handling [53].
Further, the Nanotechnology Characterization Laboratory (NCL) has published a list of common
pitfalls uncovered in their efforts to characterize nanoparticles [54]. A common theme is the
mis-attribution of toxic potential to nanoparticles in a sample, either due to contamination or
nanomaterial composition [55]. This has become increasingly problematic as evidence for
interactive and even synergistic effects of nanomaterials and co-existing contaminants has arisen
[56]. For example, a co-exposure of silver nanoparticles with cadmium have been shown to
decrease cell viability in HepG2 cells when compared to silver nanoparticles or cadmium as the
sole exposure [57]. Similarly, increased bioaccumulation of various heavy metals was shown in
C. Elegans after co-exposure with titanium dioxide nanoparticles [58]. At first glance, a solution to this issue is to test every sample in house for contamination and to perform experiments on
more complex nanoparticle formulations to ensure that an ingredient of any nanoformulation is
not the main driver of response. However, this can be a significant cost burden to individual labs.
A proposed alternative is to estimate probability distributions for nanoparticle characteristics to
allow for estimates of true particle characteristics [59]. Because we are dealing with disparate
datasets, statistical methods will be an important tool for imputing the data gaps inherent to data
curation. At present, the development of a robust predictive model for nanotoxicity is limited by
the amount of available high quality data. Imputing nanomaterial physicochemical characteristics
with manufacturer labels has been shown to accurately assess toxicity when used with a
physicochemical characteristic weighting scheme [60]. However, the known discrepancy
development using statistical imputation methods may improve the data gap issue while
providing the most accurate estimate of the missing physicochemical characteristic.
1.3.2. Existing Data Sources
Despite the known data quality challenges, there have been several efforts to aggregate
nano-related data. The Nanotechnology Consumer Products Inventory lists 1814 consumer products
containing nanomaterials and categorizes them based on verification that the product contains
nanomaterials [61, 62]. However, it does not contain scientific data that would be useful for risk
assessment. Similarly, the Nanowerk Nanomaterial Database lists nanomaterial suppliers for
4000 different nanomaterials along with the manufacturer product descriptions [63].
The DaNa 2.0 Knowledgebase was developed by an interdisciplinary team of researchers
providing input on 26 market-relevant nanomaterials and their effects on humans [64]. The
knowledgebase allows the user to search by application or by material and provides bountiful
information on potential exposure routes and behavior in both humans and the environment. The
team developed a “Literature Criteria Checklist” that assesses the quality of a literature source
for use in the knowledgebase. With a wealth of scientifically-based qualitative data, DaNa2.0 is
targeted to both non-scientific users, such as consumers or stakeholders, and to scientific users.
However, the availability of quantitative data is minimal, limiting the application of DaNa2.0 for
predictive modeling.
For quantitative toxicity data, two sources of nanomaterial data are the Nanomaterial Registry
and eNanoMapper. The Nanomaterial Registry was created in 2013 through a partnership with
the National Institute of Environmental Health Sciences (NIEHS), National Institute of
2014, the registry contains 2031 records for engineered nanomaterials, with biological data from
608 different studies. Data are sourced from databases, data partnerships, and manufacturer data
sheets (Table 1). Before entry into the registry, data are scrutinized for adherence to the
registry’s Minimum Information about Nanomaterials (MIAN) standards. The MIAN includes an
“Instance of Characterization” that indicates when the nanomaterial was characterized, which
can alleviate discrepencies in characteristic measurements.
Table 1. Summary of Nanomaterial Registry records for biological interactions, adapted from
nanomaterialregistry.org.
Data Source Host
Organization
Type Focus Source
Records Registry Record Ref Cancer Nanotechnology Laboratory (caNanoLab) National Cancer Institute (NCI)
Database Cancer research using
nanomaterials
2336 in vitro assays; 100 in vivo assays 348 unique nanomaterials from 266 studies [123, 124] Nanomaterial-Biological Interactions Knowledgebase (NBI) Oregon State University Data Partner (raw data) High throughput studies on nanomaterial toxicology using an embryonic zebra fish model 165 records for 31 unique nano cores
82 records for 22 unique cores
[125] NanoComposix, Inc. Data Partner (Product specification sheets and publications) Manufactured metal nanoparticles and the studies using them 35 unique nanomaterials [126, 127]
The eNanoMapper prototype database was published in 2015 from the European Commission’s
Seventh Framework Programme (FP7) with a distinct goal of facilitating the management of
literature source describing the protein corona formed due to nanoparticle-biological interactions
[67, 68]. Since its initial release, the eNanoMapper prototype database has grown to include
user-submitted data from European Commission funded projects. Among these data sources, 3
include in vitro toxicity measurements (Table 2).
Table 2. Summary of eNanoMapper records for in vitro toxicity data.
Nanomaterial Contributor Name
Materials with Toxicity Data
Focus Ref
FP7 MARINA 6 Substances Develop and validate the Risk Management Methods for Nanomaterials
[128]
MODENA 41 Substances Develop Quantiative Nanostructure Toxicity Relationship tools [129]
Protein Corona Paper
121 Substances Characterize protein corona and effect on net cell association [68]
In addition to this comprehensive database, the eNanoMapper project has released Jaqpot
Quattro, a web-based platform for generating nano-QSARs and read across models using data
provided by the eNanoMapper database [69]. Jaqpot supplies a number of algorithms for both
regression and classification with options for selecting predictor variables and an endpoint of
interest. Documentation for this utility is extensive, allowing for self-guided analysis of new or
existing nanotoxicity data.
The eNanoMapper project has developed its own ontology as an extension of the NanoParticle
Ontology (NPO), which gathers parts of existing ontologies to meet the overall ontological needs
for determining nanosafety from a wide array of sources. Further supporting this overarching
goal, the eNanoMapper database supports the input of different file formats and processes them
for inclusion in the database. So, through the developed ontology and ability to process a diverse
1.3.3. The Nanomaterial Data Curation Initiative
The Nanomaterial Data Curation Initiative (NDCI) from the NCIP NanoWG has released
publications detailing the challenges in a large scale data curation effort. Recommendations are
provided for standardization of a minimal information checklist, vocabulary to be used, and
evaluation of data quality and completeness [70]. These needs are apparent due to the differing
foci from different stakeholders. For example, individual laboratory groups are interested in
molecular and cellular aspects, while other groups are more concerned with consumer products.
These differing foci have different data requirements and the groups interested in them would
therefore have different ideas of minimal information required and data completeness and
quality. The workflow described by the NDCI includes communication with authors in attempts
to fill data gaps, as well as the development of a model describing the relationships between
physicochemical characterizations [71, 72].
The development of this model has potential for resolving the different informational
requirements from different stakeholders and highlights the modeling needs at the foundational
level of nanotoxicology. High quality, complete characterization data would be useful in the
understanding of nanomaterial structure and the development of nano-QSARs, which can inform
on the toxic potential of nanomaterials and determine relationships between structure and
behavior that would allow for predictive modeling of nanotoxicology.
1.4. Predictive Modeling
Legislation from the European Union has promoted decreased usage of non-human vertebrate
animal models, in favor of developing alternative testing strategies for chemical risk assessment
growing number of modifications altering their behavior, these alternative testing strategies can
improve time and money that would otherwise be spent on single case studies.
While high throughput screening (HTS) permits rapid testing of chemicals, the immense volume
of data generated introduces the need for new computational analysis methods to make sense of
the results [74–76]. Likewise, the use of high throughput methods for exploring nanotoxicity and
procuring data for the development of predictive models has been suggested due to coverage of a
wide variety of nanomaterials [77, 78]. Applications of high throughput screening methods for
nanomaterials have been described in various in vivo model organisms, such as the observation of hatching in zebrafish embryos [79] or behavioral alterations in C. elegans [80]. Similarly, in vitro methods have been described, with an emphasis on understanding molecular mechanisms for nanomaterial toxicity [81, 82]. Though the use of predictive models for these types of data
have been scarcely described, there is evidence for its success. For example, results from a
high-throughput screen using a zebrafish embryo toxicity test were used to develop a self-organizing
map that definitively separates materials and doses [83]. Therefore, we observe that these HTS
methods can preclude the development of a predictive model by heightened understanding of the
mechanism for nanomaterial toxicity.
A major potential use for these types of predictive models are for predicting the toxicity of
untested chemicals within a model’s applicability domain, commonly identified using leverage
values [84]. However, as previously mentioned, without a globally accepted minimum
information criteria for nanomaterials, the aggregation of data presents missing data for
measurements that may have predictive influence on nanomaterial toxicity [60]. It has been
shown in standard toxicological settings that this missingness can affect end-stage applications,
aggregating this heterogeneous data is a two-stage experimental design. The first stage uses
multi-source experimental data and applies geospatial stochastic kriging to identify data gaps to
be mediated in the second stage. Alternatively, similarities in ENM properties may allow for
model-based imputation [86]. The recent application of Quantiative Structure-Property
Relationships to nanomaterials uses quasi-SMILES, the coupling of SMILES structural
descriptors and intrinsic size measurements, to predict size-dependent properties such as Young's
modulus, zeta potential, dispersion quality, and thermal conductivity [87]. In fact, using only
quasi-SMILES to derive descriptors for multiple linear regression, one group has developed a
nano-QSPR for zeta potential that boasts a validation R2 ranging from 0.67 to 0.82. Of course it is described within these papers, that nanomaterials with less deviating characteristics require
further improvements to this nano-QSPR [45]. In general, these types of models aim to improve
the development and performance of predictive nanotoxicology models by providing structurally
and mathematically based methods of data imputation.
In this section, we first discuss the grouping of nanomaterials based on similar characteristics
and the necessity of high quality grouping paradigms for successful prediction in read-across.
We then discuss the use of nanomaterial read-across and the inclusion of quantitative approaches
for improved prediction. Next, we cover nano-QSARs and the underlying statistical methods
quantifying the relationships between nanomaterial descriptors and toxicity endpoints. We begin
with an overview of multiple linear regression, a relatively simple modeling technique, and then
discuss more complex methods and the use of Bayesian inference to improve model performance
1.4.1. Grouping
Grouping methods are used to categorize chemicals based on similar physicochemical
characteristics or toxicological profiles. Chemicals that are grouped together can then be
assumed to exhibit similar behaviors. Grouping facilitates the use of chemical read-across,
wherein the prediction of a chemical's toxicity is based on its group assignment. Therefore,
grouping methods have been recommended by the OECD and ECHA as a means to reduce
unnecessary testing by allowing for in silico toxicity screening [31, 88].
The ECHA has published its own recommendations for performing grouping and read-across for
nanomaterials. In this stepwise procedure, nanomaterials are characterized by their
physico-chemical characteristics and subsequently grouped by these attributes as well as known
behaviors and reactivity. The results from chemical read-across using these methods is then
assessed and used to improve grouping criteria [89]. A case study of 6 nanoform types of TiO2
implemented a simplified version of the ECHA grouping framework to group an aggregated
dataset consisting of nanomaterial attributes and comet assay results. The working hypothesis for
this study was that TiO2 surface coatings or surface impurities decrease genotoxicity of the
material. Hierarchical clustering and principal components analysis revealed two distinct groups,
genotoxic and non-genotoxic. The relevancy of surface coatings and surface impurities was
reaffirmed using random forests variable selection and predictions of genotoxicity based on these
characteristic-based groupings consistency with existing comet assay results [90]. A similar
grouping method has been applied using pulmonary inflammation data from nanomaterial
exposed rodents. In this study, hierarchical clustering was applied to the toxic endpoint, in this
case potency calculated from benchmark doses, to generate 4 potency groups. A predictive
subsequently used for predicting the potency of newer materials [91]. This study differs from the
previously described TiO2 study in that it utilizes toxicity information as the grouping criteria,
highlighting the utility of non-structurally based descriptors for grouping and model
development.
Although grouping and read across based on structural similarities is quite standard, it is noted
that the poorly understood structure of nanomaterials requires use of extraneous attributes such
as mechanistic based descriptors or toxicological endpoints [89]. However, an early
implementation of nanomaterial grouping bypasses this notion of nano-specific mechanisms and
instead uses 3 general attributes that allude to mode of action. The first category consists of
chemicals exhibiting chemically mediated toxicity (i.e. release of metal ions); the second
category consists of fibrous nanomaterials; and the third category consists of respirable granular
and biopersistent, but not fibrous nanoparticles [92]. This grouping scheme has not been widely
utilized, however, the use of mode of action as a characteristic attribute underscores a growing
emphasis on the use of a broader perspective for defining grouping criteria.
Arts et.al reviewed grouping methods for nanomaterials and highlighted the need for
nanomaterial life-cycle based grouping, arguing that the proposed grouping methods focus on a
single point in the nanomaterial life-cycle, whereas an entire life-cycle based approach would
capture a more informed toxicological profile [93]. This proposed grouping mechanism was
subsequently implemented in the DF4nanoGrouping framework by the European Centre for
Ecotoxicology and Toxicology of Chemicals (ECETOC) Nano Task Force [14]. Nanomaterials
can be assigned to one of four groups: (1) soluble nanomaterials, (2) biopersistent high aspect
ratio nanomaterials (HAR NMs), (3) passive nanomaterials, and (4) active nanomaterials. These
properties, Tier 2 assesses system-dependent properties, and Tier 3 confirms passive vs. active
nanomaterials using in vivo screening. The utility of DF4nanoGrouping was discussed in a case
study, which exemplifies the reduction of animal-testing needs by using different hazard
assessments for each group [94].
This robust framework has been coupled with specific testing methods and analytical procedures
that allow for life-cycle based grouping. In one such study, data from assays satisfying the Tier 3
initiative were collected and processed through classification trees using nanomaterial
descriptors. In this way, specific toxicological endpoints can be linked to the most relevant
descriptors for which between-group diversity can be detected [95]. From this study, we observe
the applicability of the DF4 framework to existing data for grouping nanomaterials as passive or
active. This tiered methodology can also be implemented outside of the confines of the
predetermined grouping. In a 2018 study by Scott-Fordsmand et.al, 8 different nanomaterials are
partial order ranked within the scope of each of the three tiers, as described by Bruggeman and
Patil [96]. The inclusion of a threshold in the chemical matrix allows for rank comparison of the
nanomaterials to the threshold rank. This naturally yields three groups: (1) Nanomaterials that
rank lower than the threshold, (2) Nanomaterials that rank higher than the threshold, and (3)
Nanomaterials that rank equally to the threshold [97]. Though this method can be useful for
determining relative toxicity amongst a subset of nanomaterials, it does not support a global
grouping framework, as group assignments are dependent on the selected threshold parameters
and the nanomaterials in question. An accurate grouping scheme is important for read across, as
it precedes the prediction of toxicity; that is, recognizing the nanomaterial parameters which
influence within-group variability and between-group variability will improve grouping and lead
1.4.2. Read Across
Chemical read across is a successful non-testing strategy in toxicology that has been
recommended for filling data gaps in nanotoxicology data [89]. The general principle of read
across is that two physically or structurally similar chemicals should have similarly measured
activity for any given endpoint. As a result, we expect that the activity of one or more chemicals
can be predicted by the activity of one or more similar chemicals.
Successful read-across is dependent on the available chemical descriptors and an accurate
grouping mechanism, based on either these descriptors or known toxicity measurements, as
described above. The European Union FP7 MARINA Risk Assessment Strategy is broken into
two phases. The first phase focuses on problem framing, where relevant exposure scenarios are
considered for a target material. Data are then collected and the chemicals within this aggregated
data are grouped based on a determined grouping mechanism. So in this risk assessment strategy,
we see a high emphasis on adequate grouping and ample, high quality data. The second phase
focuses on risk assessment. Based on data collected in Phase 1, a hypothesis is created that
emphasizes similarities between the target material and relevant source material. Specific
endpoints can then be predicted for the target material, if enough data exists. Otherwise, further
testing is needed [98].
To date, the described nano-read-across approaches have been predominantly performed on sets
of metal oxide nanoparticle data, an important subgroup of nanomaterials that are prevalent in
consumer products [99]. Gajewicz et. al performed two case studies on data from metal oxide
exposure to E. Coli and to HaCaT cells. For E. Coli, grouping was performed on 10
same oxidation state as that in the metal oxide structure (ΔHMe+). Cytotoxicity for both the
training set of 10 nanomaterials and the validation set of 7 nanomaterials was 100% accurate. For
the HaCaT cell line, 10 nanomaterials were grouped using a single descriptor, Mullikan's
electronegativty. The training set featured 80% accuracy, while the validation set of 8
nanomaterials featured accuracy of 87.5% [100]. From this example, we see the importance of
deriving an accurate grouping scheme prior to read-across, as it allows for data-supported
toxicity prediction for untested materials. It is also noted that the read-across framework is best
suited for a dataset of this size, compared to QSAR methods which require a larger number of
observations.
However, the ability to aggregate data from different sources would increase the amount of data
available for nanomaterial toxicity and allow for a more comprehensive grouping analysis and
therefore, more accurate read-across. A multi-nano-read-across approach has been described for
metal oxide nanoparticles. In this study, data was obtained from 15 studies, each using one of 9
different cell types under different testing environments. Grouping was performed using
nanoparticle structural descriptors and normalized toxicity measurements in a self-organizing
map. Overall, species-dependent mode of action and variable experimental conditions were
found to affect similarities between prediction results from different species, although
inter-species similarities were found, highlighting the promise for developing an accurate read across
model based on data from multiple species and studies [101].
Although the grouping mechanism plays a vital role in read-across performance, understanding
the mathematical relationships between nanomaterial descriptors in eliciting toxicity would
improve the predictive power of the read-across model. A criticism of read-across is the
predictions [102]. Therefore, a quantitative predictive approach would alleviate this bias by
relying on numerical relationships for determining similarities between materials, improving the
reproducibility of nanomaterial read-across results. The Quantiative Read-Across Approach for
Nanomaterials (Nano-QRA) uses an algorithmic approach for filling in missing data about a
chemical’s toxicity endpoint by identifying linear trends within its assigned group based on one
to three nanomaterial descriptors. In this preliminary study, the accuracy of a quantitative
approach was demonstrated using a limited set of descriptors with the previously described metal
oxide nanoparticle dataset [103]. Although the authors note similarities in methodology between
Nano-QRA and kNN-QSAR, Nano-QRA is considered more appropriate for scarce datasets.
A streamlined framework for quantitative nano-read across was developed using the
protein-corona dataset [104]. This framework, nano-Lazar, is an extension of an existing read-across
framework, Lazar [105]. To explore extension options, three similarity protocols for grouping
were tested based on structural similarity, property similarity, or biological similarity. After
grouping, three local regression algorithms were applied to each group to model the toxicity end
point: net cell association. Results from this study indicate more accurate predictive outcomes
when using measured nanoparticle descriptors with the random forests model. Because
Nano-Lazar performs grouping and local prediction, it can be described as both a read-across model
and nano-QSAR model. In fact, Nano-Lazar was developed using a large, well-characterized
dataset, which is a known necessity for developing QSARs and nano-QSARs. So, although
read-across can give qualitative information about a nanomaterial’s potential toxicity, the inclusion of
quantitative prediction may alter the underlying advantages of qualitative read-across. We see
from these read-across examples that despite the advantages of read across, quantitative solutions
1.4.3. Nano-QSARs
A common quantitative method for predicting nanotoxicity is the application of QSAR-like
models to nanomaterials. Various names have been given to describe this use of mathematical
modeling for predicting nanotoxicity, e.g. Nano Quantitative Structure-Activity Relationships
(nano-QSAR), Quantitative NanoStructure Activity Relationships (QNAR), or Quantitative
Structure-Toxicity Relationship (QSTR). A common practice in applying QSAR to
nanomaterials is to extend the descriptor space past structural descriptors to include inherent
nanomaterial features as well as experimental descriptors.
1.4.3.1. Linear Regression Methods
Perhaps the most simple nano-QSAR technique is to use linear regression, wherein a specified
nanotoxicity endpoint is modeled through a linear relationship to the selected descriptors. A
common practice is to use a descriptor selection technique that will pick descriptors most
relevant to the endpoint prior to model development. Use of a subset selection method decreases
computation time needed to perform an exhaustive search over all possible subsets of descriptors
to identify the globally optimized model. In one study, Puzyn et. al began with 12 calculated
structural descriptors for a set of 17 metal oxide nanomaterials, ranging in size from 15 to 90nm
[106]. The genetic algorithm was implemented, where each row of features for a particular
nanomaterial is regarded as a parental chromosome. Evolution of these chromosomes is then
modeled through subsequent generations which represent the fittest combination of features
selected [107]. In the Puzyn study, only a single descriptor, ΔHMe+, was selected for predicting
cell viability in E. Coli. The reported model R2 of 0.85 and cross-validation regression
In one study, cytotoxicity data for these 17 metal oxide nanomaterials was used along with
descriptors derived from the material's molecular formula and data from the periodic table. In
this case, variable selection was performed using stepwise multiple linear regression, so that
inclusion of a particular descriptor is based on statistical significance in the linear regression
model [108] The final model included only a single descriptor, the charge of the metal cation
corresponding to a given oxide, χOX, with reported R2 and Q2 values of 0.84 and 0.81,
respectively. Interestingly, neither ΔHMe+ nor χOX are size-dependent descriptors, but rather are
mechanistic descriptors, informing on expected biochemical behavior of the materials in a given
medium. Of course, the final model is dependent on the subset selection method. In a subsequent
analysis, cytotoxicity measurements from Puzyn’s study and another metal oxide nanomaterial
study [109] were aggregated, forming a training dataset of 16 metal oxides and 26 calculated
chemical descriptors [15]. Variables were selected by first considering Pearson's correlation
coefficient, removing redundant variables to avoid autocorrelation. Clustering and principal
components analysis were performed to select four optimal descriptors. All potential models
derived from this set of four descriptors were considered, and a final model containing two
descriptors, polarization force (Z/r) and ΔHMe+, was selected based on the coefficient of
determination. R2 for the 16 metal oxides was 0.882, while Q2 for internal validation using various validation methods ranged from 0.846 and 0.910, indicating improved model
performance when including the additional Z/r parameter. It is therefore clear that the
performance of the model depends on the selection of accurate molecular descriptors specific to
a given endpoint.
An adapted version of multiple linear regression is k Nearest Neighbors (kNN) regression, where
predicted using a subset of data points most closely resembling attributes for the endpoint in
question. Of course, performance of kNN regression is dependent on the selection of k, which
can be determined by cross-validation or some optimization procedure. Another important factor
in the performance of the resulting model is the distance metric, with Euclidean distance being
the common method used. In a study of 34 synthesized gold nanoparticles ranging in size from
5-10nm and with differing surface ligands, Euclidean distance was used to compare nanoparticle
profiles, comprised of 29 descriptors [110]. The value of k was selected using leave-one-out
cross-validation and a range of possible k values and selecting the k that maximizes Q2. Variable selection is performed in a similar manner, with an initial model built using randomly selected
variables, updated iteratively and selecting for the model with the highest Q2 [111]. Using this kNN strategy, 4 final QNAR models were developed for each of 4 endpoints: cellular uptake in
human lung cells (A549), cellular uptake in human kidney cells (HEK293), oxidative stress
(measured by HO-1 protein levels in A549 cells), and hydrophobicity (measured by partition
coefficient). Reported R2 for these models ranged from 0.990 to 0.995, while external validation Q2 values ranged from 0.768 to 0.930. As with the multiple linear regression nano-QSARs, the variables selected for each of these four models give insight into the mechanism for toxic effect.
For example, a common feature in models for cellular uptake were descriptors related to
hydrophobicity. Similarly, Fourches et.al modeled the cellular uptake of 109 superparamagnetic
nanoparticles with different surface modifiers in PaCa2 cells using k Nearest Neighbors
Regression [112]. The resulting internal validation R2 ranged from 0.65 to 0.80 and after removing materials outside of the applicability domain, 0.67 to 0.90. The external validation Q2 was 0.72 and after applying the applicability domain, 0.77. Among the descriptors chosen for
variable selection in deriving a mechanism based toxicity prediction. By using the k nearest
neighbors for prediction, kNN regression is inherently based on the same principle as grouping
and read-across; that is, a material's toxicity can be predicted by the toxicity of similar materials.
1.4.3.2. Support Vector Machines
In addition to kNN regression, Fourches et. al implemented Support Vector Machine (SVM)
classification by converting bioactivity to a binary endpoint indicating biological activity.
External validation of the developed model yields 73% accuracy, with 60% specificity. SVM
classification and regression use a kernel function to non-linearly discretize endpoints. Parameter
selection for these kernel functions can be performed using a grid-search coupled with
cross-validation. In one such case, EC50 data from 17 metal oxide nanoparticles was obtained from
literature, with six derived molecular descriptors Using a radial basis function, a support vector
machine regression model resulted in training and test set R2 values of 0.903 and 0.916, respectively [113]. When compared with the results from stepwise multiple linear regression
performed on the same dataset, R2 values indicate high agreement between the two models (MLR Training and Test R2 = 0.877 and 0.848, respectively). From this example, we find that non-linear SVM benefits from marginally higher accuracy in prediction, however model
interpretability is higher for multiple linear regression, where feature weights can be directly
associated with endpoint variability. Regardless, variable selection can be performed for
non-linear SVM using cross-validation to search through models of different sizes and selected
variables can then be individually analyzed against the endpoint to understand impacts to toxicity
as has recently been done with Puzyn's metal oxide nanoparticle EC50 data [114]. In this case,
37 variables were derived using minimum redundancy feature selection to remove highly
algorithm with support vector regression, a 3-variable model with the lowest RMSE identified
three relevant variables: atomic ratio of oxygen to metal (RAatom), enthalpy of melting ΔHm, and
cohesive energy (Ecoh). In the next step, a residual bootstrapping technique was implemented to
improve model predictions by refitting the model to endpoints updated with residual information.
Comparison of the standard support vector regression and the support vector regression with
residual boostrapping indicate lower R2 values for each test set using residual boostrapping, however this difference is not statistically significant and R2 values using the bootstrap method are less variable across test sets, indicating better model stability. We see from these results that
SVM regression is quite similar to the described linear methods above, wherein a dataset of
nanomaterial features can be mathematically modeled to the endpoint of interest. Using more
advanced methods for refining these models can result in robust predictive models for
nanotoxicity.
1.4.3.3. Neural Networks
When compared to the more complex neural network method, SVM shows varying performance,
dependent on the data to be analyzed and the linearity of the mathematical technique applied.
Neural networks represent another class of analytical methods derived from a simplified model
of neurological activity. The most basic neural network, the feedforward neural network, moves
in a single direction connecting an input layer, hidden layer, and output layer. Applied to QNAR,
the input layer will consist of nanomaterial descriptors that are fed to the hidden layer. The
hidden layer contains an activation function that transforms the descriptors using some
mathematical function. Each node within the hidden layer is then sent to the output layer via a
between layers are weighted, with these weights being continuously updated based on prediction
errors to tune the model for accuracy.
A comparative study quantifying lipid membrane disruption in rat lung epithelial cells and rat
lung alveolar macrophages implemented multiple linear regression, linear SVM, radial SVM,
radial basis function neural networks, and general regression neural networks [115]. Radial SVM
separates descriptor data in such a way that in higher dimensions, circular boundaries discretize
particular endpoints. Similarly, radial basis neural networks process descriptors through radial
basis functions and use their summation for prediction. General regression neural networks are a
type of radial basis function that assumes the training sample represents the mean of the
activation functions. This particular study found that the relative performance of each of these
techniques varied depending on the dataset used. For example, linear SVM (R2 = (0.77,0.85), Q2ext = (0.65, 0.96)) and RBFNN (R2 = (0.78,0.88), Q2ext = (0.36,0.91)) performed equivalently
to MLR (R2 = (0.78,0.85), Q2ext = (0.65,0.98)) when using aggregated results from exposure to
TiO2 and ZnO. Optimal models were selected using results from leave-one-out cross validation,
however R2 is reported here for comparability to other studies. When TiO2 was studied
separately, radial SVM presented the highest agreement between predicted and true values
(Radial SVM (R2 = (0.95,0.99), Q2ext = (0.77, 1.00)). When studying ZnO separately, radial
SVM (R2 = (0.94, 1.00), Q2ext = (0.49, 1.00)) and GRNN (R2 = (0.94, 0.98), Q2ext = (0.52, 1.00))
outperformed MLR (R2 = (0.88, 0.97), Q2ext = (0.41, 0.99)). From these results, we identify a
potential need for discrete models for subsets of materials. However, further comparative studies
are needed given that in one particular study, three RBFNN models for three groups of
nanomaterials were found to outperform MLR (MLR: R2 = (0.792,0.857); RBFNN: R2 =
the sample size and the descriptors selected for the input layer. These two studies differ in both,
with the Papa et. al study featuring a training set of 31 nanoparticles and the Luan et.al study
featuring 15 to 36 nanoparticles per group. Five descriptors were used in the comparative study,
describing concentration and nanoparticle size in varying media. On the other hand, Luan et.al
derived molecular descriptors from software using chemical structure, which could be the reason
for higher accuracy in these RBFNN models.
Besides radial basis functions, neural networks can vary by the model selected for the hidden
layer or by the overall organization of the network. In one study, an asymmetrical sigmoid
function was used in the hidden layer, essentially applying an S-shaped relationship between
inputs and outputs for each node. In this case, chitosan nanoparticles, paired with streptokinase
were applied to Mrc-5 cells and analyzed for cell viability [117]. The input variables, size,
chitosan concentration, solution pH, and stirring time, describe experimental parameters rather
than molecular descriptors. The relatively high agreement for predictions with the training set
(R2 = 0.90) and with the external test set (Q2 = 0.96) indicate the influence of experimental conditions in nanoparticle behavior. Typical neural networks, such as this, benefit from the
implementation of backpropagation, where errors from the output layer are used to update
weights between layers. Another type of neural network that has been applied to nanotoxicology
data is counter-propagation artificial neural networks (CP-ANN), a type of self-organizing map
(SOM). In CP-ANN both unsupervised and supervised methods are used to first discretize
inputs, separating dissimilar training samples, such as in a SOM, and then combining these nodes
into the output layer. In a study using data from Puzyn's study of 17 metal oxide nanoparticles, a
CP-ANN was developed for the cytotoxicity endpoint, as well as a classification endpoint