Professional Documents
Culture Documents
Dean and Mitchell 2020 P.A. Dose-Response Model
Dean and Mitchell 2020 P.A. Dose-Response Model
PII: S2352-3522(20)30021-9
DOI: https://doi.org/10.1016/j.mran.2020.100115
Reference: MRAN 100115
Please cite this article as: Kara Dean Graduate Student , Jade Mitchell Ph.D. Professor , A dose re-
sponse model for the inhalation route of exposure to P. aeruginosa, Microbial Risk Analysis (2020),
doi: https://doi.org/10.1016/j.mran.2020.100115
This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.
P. aeruginosa poses a potential human health risk via the inhalation exposure route
The multi-hit dose response model significantly fit available data
Single-hit theory may not capture P. aeruginosa behavior in immunocompetent hosts
Title: A dose response model for the inhalation route of exposure to P.
aeruginosa
Authors: Kara Dean1, Jade Mitchell2, Ph.D.
1
Graduate Student, Department of Biosystems and Agricultural Engineering, Michigan State University,
524 S. Shaw Ln, East Lansing, MI 48823. Email: deankara@msu.edu
2
Professor, Department of Biosystems and Agricultural Engineering, Michigan State University, 524 S.
Shaw Ln, East Lansing, MI 48823 (corresponding author). Email: jade@msu.edu
ABSTRACT
This study develops a dose response model for Pseudomonas aeruginosa for the
capable of causing community and hospital-acquired lung infections. As such, a dose response
model for this route of exposure is needed to assess risks posed by the inhalation of aerosols
from showers, humidifiers, or hot tubs contaminated with P. aeruginosa. Single-hit theory
models traditionally used for dose response modeling did not provide significant fits to the
limited available data. The multi-hit dose response model operates under the cooperativity theory
and did provide a significant fit, suggesting that a single P. aeruginosa bacterium is not
analysis was used to benchmark this model against additional dose response experiments and
results further suggested that the multi-hit model may better represent the dose response behavior
for this exposure route. The best fitting model has an LD50 of 2,588,047 Colony Forming Units.
This model can be used to quantify risk in inhalation exposure scenarios, however, due to the
limited amount of primary data, it is especially important for any future risk assessment to
analyze the impact that using different dose response models may have on the final risk estimates
Funding Source: Partial support for research was provided under Assistance Agreement No.
R836890 awarded by the U.S. Environmental Protection Agency to Purdue and Michigan State
University. This work has not been formally reviewed by the EPA. The views expressed in this
document are solely those of the authors and do not necessarily reflect those of the Agency.
INTRODUCTION
pneumonia, and immunocompromised hosts and patients with cystic fibrosis are at a higher risk
of infection (Driscoll et al. 2007; Sadikot et al. 2005). Although less common, P. aeruginosa can
also cause lung infections in immunocompetent hosts, and previous infections in healthy
individuals have reported a mortality rate of 33% (Hatchette et al. 2000; Sadikot et al. 2005). In
the United States, 51,000 healthcare-associated infections and about 440 deaths are caused by P.
aeruginosa annually (Azam & Khan 2019; CDC 2013; Moradeli et al. 2017). P. aeruginosa is
recognized worldwide as a public health risk and the bacterium is becoming more difficult to
treat, as it is naturally resistant to a number of antibiotics (Moradeli et al. 2017). In the U.S.
P. aeruginosa can exist and thrive in a range of natural and built environments, and water
has been implicated directly or indirectly in most of the investigated outbreaks (Bedard et al.
2016). Water-related sites like taps and showers and moist, humid environments like respiratory
therapy equipment are the most likely to be colonized in healthcare settings (Kerr & Snelling
2009). P. aeruginosa develops biofilms that aid in its production of virulence factors and
persistence in the environment, and biofilms in engineered water systems can become an ideal
long-term habitat for this opportunistic pathogen (Sadikot et al. 2005; Wingender & Flemming
2011). P. aeruginosa colonizes premise plumbing systems and point-of-use devices like
showerheads and faucets, and the threat this poses to the user is uncertain. It is possible that
during a showering event the pathogen may be aerosolized and inhaled. In order to be able to
1
understand the risk of a respiratory infection from P. aeruginosa in such an exposure scenario, it
is first necessary to have an understanding of the dose response relationship. Dose response
models characterize the relationship between the dose of a pathogen and the probability of an
adverse health outcome in the exposed population, such as infection, illness, or death (Haas et al.
2014). A dose response model for the inhalation route of exposure of P. aeruginosa has not
previously been developed but would be valuable in facilitating future risk assessments and to
support studies that determine if there is a need for remediation of premise plumbing.
Mycobacterium avium are able to persist and grow within premise plumbing and distribution
systems and all three have been associated with infections from exposure through premise
plumbing (Falkinham et al. 2015). Unlike L. pneumophila and M. avium, P. aeruginosa is not
currently on the Environmental Protection Agency’s Contaminant Candidate Lists (CCL) (U.S.
EPA 1998, 2005, 2009, 2016). Both L. pneumophila and M. avium pose inhalation risks and to
facilitate risk assessments that can quantify the risk they pose to human health, these pathogens
have had inhalation dose response models previously fit (Hamilton et al. 2017; Armstrong &
Haas 2007). An inhalation model for P. aeruginosa is a noted gap and this study aims to fit a
dose response model to pre-existing data. Such a model could be used to provide a greater
METHODS
2
A review of the literature was conducted to find a dose response study that simulated an
inhalation exposure, ran three or more dosing groups, and documented the positive and negative
responses from each dose. A study conducted by Ojielo et al. (2003) evaluated the risk of
pulmonary infection after bone marrow transplantation in mice. The researchers first evaluated
10 wild-type, specific pathogen-free B6D2F1/J mice with seven graded doses of P. aeruginosa.
P. aeruginosa PAO1 frozen stock was grown in 10 mL of tryptic soy broth at 37°C. The trachea
was exposed in a sterile fashion and a 26-gauge needle was used to administer the inoculum
intratracheally. The positive endpoint response was death (Ojielo et al. 2003). The seven dosing
Table 1: Dose Response Data for P. aeruginosa from Ojielo et al. (2003)
400,000 0 10
900,000 0 10
2,000,000 2 5
3,000,000 6 4
4,500,000 10 0
8,000,000 10 0
Ojielo et al. (2003) was the only data set found in the literature that met the
aforementioned criteria, however 13 other studies (16 experiments) were identified that
administered one or two doses of P. aeruginosa through the intranasal or intratracheal route to
mimic an inhalation exposure. The animal species, pathogen strain, dosages, and response
percentages are outlined for each study in Table 2. All of the studies in Table 2 had death as the
positive endpoint response. For the remainder of the paper, the Ojielo et al. (2003) data set in
3
Table 1 is referred to as Experiment 1, and the 16 additional experiments are numbered from 2-
17, as shown in Table 2. From these 13 studies, there were 13 experiments that recorded a
response for a single dose and three experiments that recorded responses for two doses (shaded
in grey in Table 2). Although these estimates could not be modeled like the data in Table 1, the
data were evaluated in conjunction with the Table 1 data through other methods to benchmark
and analyze the results of fitting the data in Table 1, given the limited data available for
modelers.
Exposure Positive
Experiment Animal Strain Dose (CFU) Resource
Route Response (%)
C3 +/+ (C57BL/6 Mueller et al.
2 PA103 Intranasal 100,000
mice) 0.222 2004
Factor B +/+ Mueller et al.
3 PA103 Intranasal 100,000
(C57BL/6 mice) 0.125 2004
C4 +/+ (C57BL/6 Mueller et al.
4 PA103 Intranasal 100,000
mice) 0.222 2004
Wolbeling et al.
5 C57BL/6JZtm mice TBCF10839 Intratracheal 600,000
0 2011
Kenawy et al.
6 MASP-2 mice KR420 Intranasal 1,000,000
0.182 2011
C3 +/+ (C57BL/6 Mueller et al.
7 PA243 Intranasal 3,000,000
mice) 0.273 2004
Morello et al.
8 Balb/c mice CHA Intranasal 3,000,000
1 2011
Factor B +/+ Mueller et al.
9 PA243 Intranasal 3,000,000
(C57BL/6 mice) 0.273 2004
Balb/c mice PAK Intranasal 5,000,000 Debarbieux et
10 0
al. 2010
Balb/c mice PAK Intranasal 15,000,000 1
PAK/PA01/ Hassan et al.
11 Balb/c mice Intranasal 10,000,000
DM125/DM126 2017
1
Balloy et al.
12 C57/BL6 mice PAK Intratracheal 15,000,000 2007; Ramphal
0.5 et al. 2005
13 C57BL6/J CHA Intranasal 20,000,000 0.78 Faure et al. 2014
4
Congenic B6.129P2-
PA M57-15 Intranasal 1,000,000,000
Cftrtm1Unc 1
infection pathogenisis in which Mueller et al. (2004) intranasally inoculated mice with the
PA103 and PA243 strains. Wolbeling et al. (2011) intratracheally inoculated C57BL/6J mice
with the TCVF10839 strain and did not obersve any mortality (Experiment 5) and Kenawy et al.
(2012) intranasally inoculated MASP +/+ (wild-type) mice with 1x106 CFU of P. aeruginosa KR
420 and observed a 18.2% response (Experiment 6). In Experiment 8, Morello et al. (2011)
intranasally inoculated a multidrug resistant mucoid strain of P. aeruginosa and observed 100%
mortality in Balb/c mice. DeBarbieux et al. (2010) used the bioluminescent PAK strain of P.
aeruginosa to also study Balb/c mice (Experiment 10). Within a study conducted by Hassan et
al. (2017), control Balb/c mice were intranasally inoculated with 1.0x107 CFU of the PAO1,
PAK, DM125, and DM126 strains of P. aeruginosa, and 100% mortality occurred in all cases, as
Balloy et al. (2007) used the PAK strain to challenge C57/Bl6 mice intratracheally and
calculated the dose where a 50% response was observed to be 1.5x107 CFU (Experiment 12).
Ramphal et al. (2005) also determined that a 50% response for the same mice species, strain, and
exposure route was 1.5x107 CFU. Faure et al. (2014) inoculated C57Bl/6J mice intranasally with
a mucoid strain CHA and saw a 78% mortality response (Experiment 13). George et al. (1991)
also studied CD-1 mice but with doses of the AC869 strain (Experiment 14). Studying potential
vaccine systems, de Souza Morais et al. (2018) intranasally inoculated Swiss Webster mice with
the P. aeruginosa strain PA14. The mortality response of the controls from this study are
5
El-Aziz et al. (2019) infected Balb/c mice with the PA9 strain of P. aeruginosa intranasally and
treated one group with PBS (Experiment 16). Finally, Van Heeckern et al. (2006) studied the
aeruginosa strain PA M57-15. The mortality of the wild-type mice are recorded in Table 2 as
Experiment 17.
Model Fitting
Using the statistical programming language, “R”, and a previously developed code that
uses maximum likelihood estimation (MLE) methods as outlined in Haas et al. (2014), the
exponential and beta-Poisson models were fit to the dose response data in Table 1 (R Core Team,
2019; Weir et al. 2017). The exponential model (Equation 1) determines the probability of a
response, P(d), based on the dose, d, and the parameter, k, which represents the likelihood that a
( ) Eq. 1
The approximate form of the beta-Poisson model (Equation 2) determines the probability
of response based on the α parameter, which dictates scale, and the median infective dose, N50.
The endpoint of response for the dose response data being considered is death, and as such the
⁄
( ) ( ) ( ) Eq. 2
Both the exponential and exact beta-Poisson models are “single-hit” models; they operate
under the assumption that just a single organism, kmin equal to 1, is needed to initiate infection
(Haas et al. 2014). The cooperativity theory assumes that for some pathogens a kmin greater than
6
1 may be necessary to initiate infection (Haas et al. 2014). This difference in underlying
assumptions results in the multi-hit dose response model (Equation 3). The multi-hit dose
response model is represented by the incomplete gamma function (Haas et al. 2014). As it
follows the gamma probability distribution, the multi-hit model can be coded in R using the
pgamma() function from the stats package. It is coded as pgamma(x, a), where x is the dose, d,
multiplied by the probability that the pathogen survives to initiate infection, k. The kmin value is
a. Note, when kmin is equal to one, the probabilities output by the multi-hit model are equivalent
( ) ( ) Eq. 3
The multi-hit model was fit iteratively, with the kmin parameter fixed at values ranging
from 1-187 and the k parameter determined using the MLE methods as described above (Haas et
al. 2014). Both parameters could not be solved for using MLE methods simultaneously because
they are inherently correlated and the kmin value should be an integer, as it represents a number of
pathogens.
Model Selection
Goodness of fit was determined by comparing the optimized deviance of the model to the
χ2 distribution with the degrees of freedom equal to the number of parameters of the model
subtracted from the number of doses (Haas et al. 2014). The null hypothesis, that the model
provides an acceptable fit, is rejected if the deviance value exceeds the critical χ2 value. The best
fitting model was determined by comparing the difference in deviances between the models to
the critical χ2 value at one degree of freedom. Confidence bands were determined by
7
Benchmarking
The previously outlined methods for model fitting and selection were applied to the data
in Table 1. Since the data sets with one or two point estimates could not be modeled directly or
pooled within the classical framework, a hierarchical Bayesian analysis of the data in Tables 1
and 2 was completed. The Bayesian framework allows for the incorporation of dose response
data sets that are otherwise uninformative within the classical framework (Mitchell-Blackwood
et al. 2011). This model was used to benchmark the model predictions from the frequentist or
classical statistic approach. Under the Bayesian framework, inferences about a parameter, θ,
given data, , are drawn from the posterior distribution, ( | ), which is calculated with the
2014). The prior distribution allows for the incorporation of prior knowledge of the parameter
behavior. Using broad and non-informed prior distributions minimizes the impact of the prior
( ) ( | )
( | ) Eq. 4
∫ ( ) ( | )
Another benefit of a Bayesian analysis is the ability to look at data hierarchically. This is
important for dose response modeling because ultimately the goal is to be able to apply the
model and interpret the results across different animal species and pathogen strains. Under this
framework, it is assumed that dose response relationship from one experiment to the next is
related via a joint probability model for the parameters (Gelman et al. 2014). For example, Table
1 and Table 2 together yield 17 individual “experiments.” Although each experiment may have
its own best fitting dose response parameter(s), a hierarchical Bayesian analysis views each of
these best fitting parameters as a sample from a common population distribution. This common
8
distribution is described by hyperparameters (Gelman et al. 2014). The hyperdistibution can be
analyzed to draw conclusions about unobserved dose response data and to produce a
generalizable model. The relationship between the hyperdistribution and the individual model
Figure 1: The relationship between the individual k parameters as described by the hyperdistribution
Both the multi-hit (Equation 3) and exponential model (Equation 1) were fit to the 17
experiments using this framework and the hyperdistributions for the k parameter in each model
were compared. Analyzing the distributions described by the hyperparameters gave further
insight into the uncertainty of the dose response relationship for P. aeruginosa between
experiments, strains, and hosts. The hierarchical Bayesian analysis was completed using R and
Markov chain Monte Carlo methods. A random walk Metropolis-Hastings algorithm (Gelman et
al. 2014) was written for each model, with uninformative, lognormal priors on the mean of the k
parameter (µlnk) and the standard deviation of the k parameter (σlnk). These relationships are
9
Starting values and priors were modified to understand their influence on the posterior
distribution inferences. The algorithm was run sequentially with the final values for the
parameters from the first run becoming the starting values for the second, etc. to evaluate
convergence. The algorithm was also run in parallel with three different sets of starting values to
evaluate convergence. Trace plots and density plots were used to visually assess chain mixing
and convergence. Estimated potential scale reduction factors (R) were also calculated using the
gelman.diag() function in R to confirm that 10,000 runs was a sufficient number of iterations
(R<1.1) (Gelman et al. 2014; Plummer et al. 2006). Additionally, the Bayesian estimate of the
model fit to the data in Table 1 was compared to the corresponding MLE estimate as method of
validation.
RESULTS
Model Fitting
The exponential model fit the dose response data with a deviance of 16.88, which is
greater than the χ2 value at 5 degrees of freedom (11.1). The beta-Poisson model had a deviance
of 16.88 compared to the critical χ2 value of 9.49 for 4 degrees of freedom. The exponential
model was a better fit than the beta-Poisson, however neither model was a statistically good fit.
10
The multi-hit dose response model fit the data with a deviance of 1.09, which is below
the critical χ2 value of 11.1. The parameters for the best fitting model were a k value of 4.12x10-6
and a kmin of 11. The fit statistics are shown in Table 4. The multi-hit model is depicted in Figure
2 with 95% and 99% confidence bands. Figure 3 is a histogram of the k parameter estimates after
bootstrapping. The model was fit for kmin values ranging from 1 to 187 to ensure the optimal fit
statistics were identified. After a kmin of 187, the numerical search algorithm was no longer able
to find an optimum k value. Figure 4 illustrates the minimum deviance value for each iteration
with the lowest deviance and respective kmin value identified with a star.
Figure 2: The multi-hit dose response model with 95 and 99% confidence bands; Figure 3: Histogram of the k
11
Figure 4: The kmin value and minimum deviance from each iteration of fitting the multi-hit model, illustrating the
Benchmarking
A hierarchical Bayesian analysis was conducted for the 17 experiments in Tables 1 and 2
for the best fitting multi-hit model (Equation 3) with a set kmin of 11 and for the more
traditionally used exponential model (Equation 1). The exponential model was included in this
analysis for comparison purposes, with the goal of understanding if the multi-hit model was
more representative of the P. aeruginosa dose response relationship generally or only for the
Ojielo et al. (2003) data set. Figure 5 shows the doses and responses for the data in Tables 1 and
2. The triangles are the seven points from the Ojielo et al. (2003) data and the circles are the
additional estimates from Table 2. Figure 5 clearly demonstrates that the P. aeruginosa animal
data for mortality responses to inhalation exposures vary across different experiments, host
Figure 6 shows the 17 individual multi-hit models fit to the data in Figure 5. The black
dashed and dotted curves represent the 2.5%, 97.5%, and median estimates of the µlnk
hyperparameter for the multi-hit model. The seventeen k parameters varied from median values
as low as 7.9x10-8 to as high as 8.1x10-5. The k parameter for the Ojielo et al. (2003) experiment
had a median value of 4.1x10-6, similar to the MLE estimate (Table 4). The median value for
µlnk was 3.1 x10-6. All of the data points are captured by one or more of the 17 curves and the
95% confidence band for µlnk (1.0x10-6 to 1.0x10-5) captures six of the 17 curves, indicating
higher confidence for those parameter estimates, and approximately 18 of the 26 data points. The
trace plots for the hyperparameters showed high variation and no obvious trends, indicating
convergence was quickly achieved. The density plots demonstrated consistent parameter
estimates and all estimated R point values were less than 1.003, also indicating convergence.
13
Figure 6: Plot of the 17 individually fit multi-hit models and the multi-hit model fit with the
median and 95% confidence band of µlnk
Figure 7 shows the 17 different best fitting exponential models fit to the data in Figure 4.
The black dashed and dotted curves represent the 2.5%, 97.5% and median estimate of µlnk
hyperparameter for the exponential model. The 17 k parameters varied from median values as
low as 4.3x10-9 to as high as 1.9x10-6. The median k parameter for the Ojielo et al. (2003)
experiment was 3.2x10-7, similar to the 3.22x10-7 from the MLE fit (Table 3). The median
estimate of µlnk was 1.4x10-7. The 95% confidence band for the µlnk hyperparameter (5.1x10-8 to
1.4x10-7) captured six of the 17 curves, indicating a higher level of confidence for those
parameter estimates, and approximately 12 of the 26 data points. The trace plots, density plots
14
Figure 7: Plot of the 17 individually fit exponential models and exponential model fit with the
median and 95% confidence bands of µlnk
The summary statistics for both the multi-hit and exponential hierarchical Bayesian
analyses are shown in Table 5. Both analyses achieved convergence, however the k parameter
estimates for the 13 experiments with only a single data point were less stable than those for the
experiments with two or seven data points, as expected. The estimates for the k parameters for
the experiments with only a 0% or 100% response were the least stable. The 17 multi-hit curves
in Figure 6 capture all of the 24 observed data points and the 17 exponential models in Figure 7
clearly exclude at least three of the data points. The 95% confidence band of the µlnk
hyperparameter for the multi-hit model also captures more of the observed data points than the
exponential model.
15
Parameter 2.50% Median Mean 97.50% 2.50% Median Mean 97.50%
k1 3.4x10-6 4.1x10-6 4.1x10-6 5.0 x10-6 2.1x10-7 3.2x10-7 3.2x10-7 4.7x10-7
k2 5.6x10-5 8.1x10-5 8.0x10-5 1.1 x10-4 1.9x10-7 1.4x10-6 1.3x10-6 5.4x10-6
k3 4.1x10-5 6.8x10-5 6.8x10-5 1.0 x10-4 3.5x10-8 5.6x10-7 4.9x10-7 3.5x10-6
k4 5.5x10-5 8.1x10-5 7.9x10-5 1.1 x10-4 1.9x10-7 1.4x10-6 1.3x10-6 5.3x10-6
k5 2.4x10-8 1.4x10-6 1.1x10-6 1.0 x10-5 1.1x10-9 3.5x10-8 2.8x10-8 2.8x10-7
k6 6.2x10-6 7.9x10-6 7.8x10-6 9.5 x10-6 5.5x10-8 1.8x10-7 1.7x10-7 4.3x10-7
k7 2.2x10-6 2.9x10-6 2.9x10-6 3.7 x10-6 2.7x10-8 9.9x10-8 9.4x10-8 2.5x10-7
k9 5.4x10-6 2.0x10-5 2.8x10-5 8.3 x10-4 5.4x10-7 1.9x10-6 2.4x10-6 3.9x10-5
k10 2.2x10-6 2.9x10-6 2.9x10-6 3.7 x10-6 2.5x10-8 9.9x10-8 9.4x10-8 2.6x10-7
k11 9.0x10-7 1.2x10-6 1.2x10-6 1.7 x10-6 4.5x10-8 9.0x10-8 8.9x10-8 1.7x10-7
k12 1.7x10-6 1.1x10-5 1.5x10-5 1.0 x10-3 1.9x10-7 7.4x10-7 9.7x10-7 1.9x10-5
k13 5.6x10-7 7.0x10-7 7.1x10-7 8.9 x10-7 1.7x10-8 4.6x10-8 4.4x10-8 1.0x10-7
k14 5.2x10-7 6.7x10-7 6.7x10-7 8.8 x10-7 3.2x10-8 7.6x10-8 7.4x10-8 1.5x10-7
k15 3.1x10-7 3.9x10-7 3.9x10-7 5.0 x10-7 9.6x10-9 2.6x10-8 2.6x10-8 5.7x10-8
k15 1.2x10-7 1.6x10-7 1.6x10-7 2.1 x10-7 6.3x10-9 1.7x10-8 1.6x10-8 3.9x10-8
k16 2.3x10-7 4.2x10-6 5.1x10-6 4.4 x10-4 2.9x10-8 2.6x10-7 3.3x10-7 1.4x10-5
k17 4.9x10-8 7.9x10-8 7.6x10-8 1.1 x10-7 1.5x10-9 4.3x10-9 4.2x10-9 1.1x10-8
DISCUSSION
This study fit the multi-hit dose response equation to dose response data for P.
aeruginosa after neither the exponential or beta-Poisson model, models which represent the
majority of microbial dose response relationships, provided a statistically significant fit using
MLE methods. The optimized multi-hit model was further confirmed with a hierarchical
Bayesian analysis of additional dose response estimates used to benchmark the MLE analysis.
The multi-hit model has not been traditionally applied to microbial data but has been used in
toxic chemical risk assessment (Janardan 1986). The slope of the multi-hit model is greater than
that of the exponential at the median infectious dose and most microbial experimental data show
16
slopes equal to or less than the exponential model (Haas et al. 2014). One-hit models are based
on the hypothesis of independent action; a hypothesis that states that pathogenic individuals
behave independently of one another and each have an independent probability of causing
infection or death (Cornforth et al. 2015; Druett 1952). The statement of independent action
equates to a kmin value equal to 1. Single-hit models are not threshold models and exhibit low
dose linearity. The multi-hit model assumes a kmin >1, does not have low dose linearity, and is a
simple threshold model. Historically there has been stronger evidence for the biological
plausibility of the one-hit model than for the multi-hit (Haas et al. 2014).
several cooperative behaviors within pathogens help facilitate the evasion of host defenses and
initiate infection. There is concern that single-hit theory models cannot capture these behaviors
and host-pathogen interactions, making the models overly conservative (Coleman et al. 2017;
Coleman et al. 2018). It was demonstrated that the independent action theory failed in Bacillus
thuringiensis, an insect pathogen, because of the cooperative nature of the bacterial toxins. Cell
proliferation (Cornforth et al. 2015). A study of Bacillus anthracis demonstrated that disruption
of the pathogen’s quorum sensing inhibited growth and virulence gene expression in vitro,
suggesting that a single spore would not result in growth, disease progression, or mortality
(Coleman et al. 2008; Jones et al. 2005). Coleman et al. (2008) presented evidence supporting
the concept of threshold regions for resistance to inhalation anthrax and suggested that in
response to the restrictions overestimated low dose risk place on risk management, it may be
necessary to analyze both threshold and non-threshold dose response models for B. anthracis.
Though the previous examples are spore-forming pathogens, limitations of single-hit theory
17
models have also been acknowledged for non-spore forming bacteria. Rahman et al. (2018)
developed a mechanistic dose response model for Listeria monocytogenes that quantifies the
interaction between the pathogen and the human gut environment with parameters such as
bacterial killing rate in the stomach, dispersal rate, and growth rate, and links infection to the
pathogen strain, the digestion process, and the host immune status. This model when compared
to the exponential model, predicted similar risks at high doses but predicted lower responses to
low doses and demonstrated that strain virulence and host immune status considerably affected
These examples highlight a few demonstrated cases where the independent action
hypothesis fails and adds credence to the application of a simple threshold model to the P.
aeruginosa dose response data. Similar to the previously mentioned pathogens, there are
cooperative action indicators in the pathology of P. aeruginosa. P. aeruginosa rarely infects the
lungs of an immunocompetent host, despite its ubiquitous nature. Some of the bacterial factors
involved in the pathogenesis of P. aeruginosa lung infections include pili and flagella, a Type III
secretion system, and quorum sensing (Sadikot et al. 2005). Data suggests that P. aeruginosa
needs to express several virulence factors to initiate a pulmonary infection and the outcome of
the infection is dependent on the host response, which in the lung environment can be quite
robust (Tang et al. 1996; Sadikot et al. 2005). Quorum sensing is a form of cell-to-cell
communication that controls the expression of more than 300 genes in P. aeruginosa (Azam &
Khan 2017; Schuster et al. 2003; Schuster & Greenberg 2006). P. aeruginosa uses quorum
sensing to regulate the expression of extracellular virulence factors and to overcome innate
immune responses in the host (Pearson et al. 2000; Smith & Iglewski 2003; Kariminik et al.
2017). This communication also promotes biofilm formation which enables the pathogen to
18
evade host defenses and persist in the environment, as is common for P. aeruginosa (Shrout et
al. 2011; Sadikot et al. 2005). Targeting the quorum sensing system of P. aeruginosa is a
possible alternative to antibiotic treatment and previous studies demonstrated increased survival
of hosts when P. aeruginosa quorum sensing was targeted or quenched (Hraiech et al. 2014; Wu
lung environment suggests that a single P. aeruginosa bacterium may not be enough to initiate
response modeling of microbes may not be limited to only the single-hit theory models
traditionally used. The multi-hit model attempts to capture some of the biological factors
involved in pathogen-host dynamics for the immunocompetent population with the kmin
parameter; for the immunocompetent population specifically, the results of the model fitting in
this study suggest that 11 P. aeruginosa bacteria are needed to overcome immune system
defenses and initiate a pulmonary infection. The LD50 of the simple threshold model fit in this
study for an inhalation exposure to P. aeruginosa is 2,588,047 CFU. The multi-hit model
indicates survivability below a certain point, and then a rapid increase in mortality with dose.
Compared to the exponential model, the multi-hit model estimates much lower probabilities of
response at low doses, as did the mechanistic model developed for L. monocytogenes (Rahman et
al. 2018).
Although more conservative estimates have often been preferred, it has more recently
been suggested that QMRA models often overestimate risk, likely due to overly conservative
dose response models (Coleman et al., 2017; Snary et al., 2016). However, it is important to note
with its application that the multi-hit model estimates de minimis risk at low doses. This model
19
attempts to describe the dose response relationship for the immunocompetent population only
and it is likely that the dose response curve is shaped differently for immunocompromised hosts.
As such, possible methods to utilize the same dose response model for more susceptible hosts,
such as using the 95th percentile parameter estimates, will likely greatly underestimate risk in the
case of P. aeruginosa.. Additionally, this model was fit to data of an intratracheal exposure, in
which the bacteria bypasses the additional defenses provided by the nasal passages (Ojielo et al.
2003). In a natural inhalation exposure, the additional nasal defense mechanisms may increase
A major limitation of this study is that the MLE fitting was only done with a single dose
response data set. Fitting dose response models to several different data sets to evaluate
consistency across experiments or possible pooling would be much preferred. The absence of
additional data sets with more than three doses hindered the validation of the model and the
evaluation of its applicability. An attempt at validation was made, however, through the
hierarchical Bayesian analysis of single and double dose response estimates found in the
literature. A hierarchical Bayesian analysis was conducted for both the multi-hit and exponential
models in an effort to better understand if the multi-hit model was merely a feature of the Ojielo
et al. (2003) circumstances, or if it may be more consistently applicable for P. aeruginosa than
It is clear from Figures 5, 6, and 7 that there is over three orders of magnitude of
uncertainty between experiments for the P. aeruginosa dose response relationship for an
inhalation exposure. The ranges of σlnk for the hyperdistributions of both the multi-hit and
exponential k parameter suggest that it may be unlikely that a single set of dose response
parameters would be applicable across host species and pathogen strains for P. aeruginosa.
20
Although future dose response studies may narrow this uncertainty, the data suggests that the
dose response relationship may just be highly variable for P. aeruginosa. For this pathogen and
exposure route, it may be necessary to select a model based on the specific pathogen strain or to
In addition to quantifying some of the uncertainty associated with this dose response
relationship, the benchmarking analysis also gave further credence to the multi-hit behavior of P.
aeruginosa. The 95% confidence interval for µlnk of the multi-hit model includes 17 of the
observed dose response data points, six more than the confidence interval for the exponential
µlnk. All of the data were captured by the 17 multi-hit curves in the analysis and the exponential
curves excluded three observations. Interestingly, one of the points excluded by the 17
exponential models fit in this analysis was the 1.0x108 CFU dose from the van Heeckern et al.
(2006) study. This study included a second observation of 100% mortality for a dose of 1.0x109
CFU (van Heeckern et al. 2006). Much like for the Ojielo et al. (2003) data, the exponential
model is not steep enough to be able to capture both observed responses. The steeper curve of the
multi-hit model is able to capture both points, as evidenced in Figure 5. The inclusion of a higher
percentage of the observed data points and this second instance of threshold-like behavior further
This study provides the best current estimate of the dose response relationship for P.
significant fit to the Ojielo et al. (2003) experiment and a more comprehensive look at dose
response estimates in the literature (Table 2) yielded a hyperdistribution with a similar median
value for k (3.11x10-6). Using the model in Figure 1 to describe all of the data in Tables 1 and 2
would provide an accurate or conservative estimate for 20 of the 26 dose response observations.
21
While the hierarchical Bayesian analysis may further validate the applicability of the multi-hit
model, it does suggest that an uncertainty factor or strain-specific model should be used in future
applications. It is important to note that this hierarchical analysis was only completed with a set
kmin of 11. If kmin were instead treated as a distribution, with less bacteria being needed to initiate
infections in some cases and more in others, it is possible that this would lead to more narrow
confidence bands of both the resulting kmin and k hyperdistributions. Further work is needed to
explore the validity of applying a distribution to the kmin value. In addition, more research into
the biological and physiological factors that contribute to the kmin parameter is needed to expand
the application of the multi-hit model to other populations at risk and additional strains of the
pathogen, which has specific relevance to P. aeruginosa as multi-drug resistant strains have been
CONCLUSIONS
The dose response model created in this study is the first dose response model for the
inhalation route of exposure for P. aeruginosa. The multi-hit model provided a significant fit to
the data and indicates that perhaps microbial dose response modeling should not be limited to
single-hit theory models. This was further suggested by the hierarchical Bayesian analysis of the
k parameter for both the multi-hit and exponential models. A dose response model is needed for
this route of exposure to facilitate the completion of quantitative microbial risk assessments
(QMRA) to address exposure scenarios of concern. Possible exposure scenarios may include
inhaling aerosols in a showering event, using a humidifier, or in pools and hot tubs. There were
limited primary data sets in the literature that could be modeled for this pathogen and exposure
route. This is a common type of uncertainty in microbial dose-response modeling as most data
sets available for modeling are small, dated and collected for other purposes. Because of this
22
limitation, the use of this model in future QMRAs should be coupled with an analysis of the
impact that using other dose response models may have on the risk estimates. If and when new
data sets become available, an uncertainty analysis using future models fit to new data sets,
modified versions of this model using the results of the benchmarking analysis, or surrogate
models from similar pathogens should be evaluated. This model and future QMRA models will
aid risk managers and decision makers in elucidating the potential risk posed by P. aeruginosa in
water.
Author Statement
Kara Dean: Formal Analysis, Writing-Original Draft. Jade Mitchell: Supervision, Project
Administration, Writing-Review and Editing.
REFERENCES
Abd El-Aziz, A. M., Elgaml, A., & Ali, Y. M. (2019). Bacteriophage Therapy Increases
Complement-Mediated Lysis of Bacteria and Enhances Bacterial Clearance after Acute
Lung Infection with Multidrug-Resistant Pseudomonas aeruginosa. Journal of Infectious
Diseases, 219(9), 1439–1447. https://doi.org/10.1093/infdis/jiy678
Armstrong, T. W., & Haas, C. N. (2007). A quantitative microbial risk assessment model for
Legionnaires’ disease: Animal model selection and dose-response modeling. Risk Analysis,
27(6), 1581–1596. https://doi.org/10.1111/j.1539-6924.2007.00990.x
Azam, M. W., & Khan, A. U. (2019). Updates on the pathogenicity status of Pseudomonas
aeruginosa. Drug Discovery Today, 24(1), 350–359.
https://doi.org/10.1016/j.drudis.2018.07.003
Balloy, V., Verma, A., Kuravi, S., Si‐ Tahar, M., Chignard, M., & Ramphal, R. (2007). The
Role of Flagellin versus Motility in Acute Lung Disease Caused by Pseudomonas
aeruginosa. The Journal of Infectious Diseases, 196(2), 289–296.
https://doi.org/10.1086/518610
Bédard, E., Prévost, M., & Déziel, E. (2016). Pseudomonas aeruginosa in premise plumbing of
large buildings. MicrobiologyOpen, 5(6), 937–956. https://doi.org/10.1002/mbo3.391
23
Centers for Disease Control and Prevention (CDC). 2013. Antibiotic resistance threats in the
United States, 2013. Atlanta: CDC. Available from:
http://www.cdc.gov/drugresistance/threat-report-2013/pdf/ar-threats-2013-508.pdf
Coleman, M. E., Thran, B., Morse, S. S., Hugh-jones, M., & Massulik, S. (2008). Inhalation
Anthrax: Dose Response and Risk Analysis. Biosecurity and Bioterrorism: Biodefense
Strategy, Practice, and Science, 6(2), 147–160. https://doi.org/10.1089/bsp.2007.0066
Coleman, M. E., Marks, H. M., Hertzberg, R. C., & Stephenson, M. M. (2017). Mechanistic
modeling of salmonellosis: Update and future directions. Human and Ecological Risk
Assessment, 23(8), 1830–1856. https://doi.org/10.1080/10807039.2017.1356682
Coleman, M., Elkins, C., Gutting, B., Mongodin, E., Solano-Aguilar, G., & Walls, I. (2018).
Microbiota and Dose Response: Evolving Paradigm of Health Triangle. Risk Analysis,
38(10), 2013–2028. https://doi.org/10.1111/risa.13121
Cornforth, D. M., Matthews, A., Brown, S. P., & Raymond, B. (2015). Errors in Pathogen Risk
Assessment due to the Failure of the Independent Action Hypothesis. PLoS Pathogens,
11(4), e1004775. https://doi.org/10.1371/journal.ppat.1004775
Debarbieux, L., Leduc, D., Maura, D., Morello, E., Criscuolo, A., Grossi, O., … Touqui, L.
(2010). Bacteriophages Can Treat and Prevent Pseudomonas aeruginosa Lung Infections.
The Journal of Infectious Diseases, 201(7), 1096–1104. https://doi.org/10.1086/651135
de Souza Morais, S.M., Rodigues, N.F, da Silva, N.I.O… Coelho, L.F.L. (2018). Serum albumin
nanoparticles vaccine provides protection against a lethal Pseudomonas aeruginosa
challenge. Vaccine, 36(43), 6408–6415. https://doi.org/10.1016/j.vaccine.2018.08.070
Driscoll, J. A., Brody, S. L., & Kollef, M. H. (2007). The epidemiology, pathogenesis and
treatment of Pseudomonas aeruginosa infections. Drugs, 67(3), 351–368.
https://doi.org/10.2165/00003495-200767030-00003
Falkinham III, J. O., Hilborn, E. D., Arduino, M. J., Pruden, A., & Edwards, M. A. (2015).
Review Epidemiology and Ecology of Opportunistic Premise Plumbing Pathogens:
Legionella pneumophila, Mycobacterium avium, and Pseudomonas aeruginosa.
Environmental Health Perspectives, 123(8), 749–758.
Faure, E., Mear, J. B., Faure, K., Normand, S., Couturier-Maillard, A., Grandjean, T., … Kipnis,
E. (2014). Pseudomonas aeruginosa type-3 secretion system dampens host defense by
exploiting the NLRC4-coupled inflammasome. American Journal of Respiratory and
Critical Care Medicine, 189(7), 799–811. https://doi.org/10.1164/rccm.201307-1358OC
24
Gelman, A., & Hill, J. (2007). Data Analysis Using Regression and Multilevel/Hierarchical
Models. New York, New York: Cambridge University Press.
Gelman, A., Carlin, J.B., Stern, H.S., Dunson, D.B., Vehtari, A., & Rubin, D.B. (2014). Bayesian
Data Analysis (Third). Boca Raton, Florida: Taylor & Francis Group.
George, S. E., Kohan, M. J., Whitehouse, D. A., Creason, J. P., Kawanishi, C. Y., Sherwood, R.
L., & Claxton, L. D. (1991). Distribution, clearance, and mortality of environmental
pseudomonads in mice upon intranasal exposure. Applied and Environmental Microbiology,
57(8), 2420–2425.
Haas, C., Rose, J., & Gerba, C. (2014). Quantitative Microbial Risk Assessment (Second).
Hoboken, New Jersey: John Wiley & Sons, Inc.
Hassan, R., El-Naggar, W., Abd El-Aziz, A. M., Shaaban, M., Kenawy, H. I., & Ali, Y. M.
(2018). Immunization with outer membrane proteins (OprF and OprI) and flagellin B
protects mice from pulmonary infection with mucoid and nonmucoid Pseudomonas
aeruginosa. Journal of Microbiology, Immunology and Infection, 51(3), 312–320.
https://doi.org/10.1016/j.jmii.2016.08.014
Hatchette, T. F., Gupta, R., & Marrie, T. J. (2000). Pseudomonas aeruginosa Community‐
Acquired Pneumonia in Previously Healthy Adults: Case Report and Review of the
Literature. Clinical Infectious Diseases, 31(6), 1349–1356. https://doi.org/10.1086/317486
Hamilton, K. A., Weir, M. H., & Haas, C. N. (2017). Dose response models and a quantitative
microbial risk assessment framework for the Mycobacterium avium complex that account
for recent developments in molecular biology, taxonomy, and epidemiology. Water
Research, 109, 310–326. https://doi.org/10.1016/j.watres.2016.11.053
Hraiech, S., Hiblot, J., Lafleur, J., Lepidi, H., Papazian, L., Rolain, J. M., … Chabriere, E.
(2014). Inhaled lactonase reduces Pseudomonas aeruginosa quorum sensing and mortality
in rat pneumonia. PLoS ONE, 9(10), 1–8. https://doi.org/10.1371/journal.pone.0107125
Jones, M. B., Jani, R., Ren, D., Wood, T. K., & Blaser, M. J. (2005). Inhibition of Bacillus
anthracis Growth and Virulence-Gene Expression by Inhibitors of Quorum-Sensing. The
Journal of Infectious Diseases, 191, 1881–1888.
Kariminik, A., Baseri-Salehi, M., & Kheirkhah, B. (2017). Pseudomonas aeruginosa quorum
sensing modulates immune responses: An updated review article. Immunology Letters,
190(December 2016), 1–6. https://doi.org/10.1016/j.imlet.2017.07.002
Kenawy, H. I., Ali, Y. M., Rajakumar, K., Lynch, N. J., Kadioglu, A., Stover, C. M., &
Schwaeble, W. J. (2012). Absence of the lectin activation pathway of complement does not
25
increase susceptibility to Pseudomonas aeruginosa infections. Immunobiology, 217(2),
272–280. https://doi.org/10.1016/j.imbio.2011.10.001
Kerr, K. G., & Snelling, A. M. (2009). Pseudomonas aeruginosa: a formidable and ever-present
adversary. Journal of Hospital Infection, 73(4), 338–344.
https://doi.org/10.1016/j.jhin.2009.04.020
Mitchell-Blackwood, J., Gurian, P. L., Lee, R., & Thran, B. (2012). Variance in Bacillus
anthracis virulence assessed through Bayesian hierarchical dose-response modelling.
Journal of Applied Microbiology, 113(2), 265–275. https://doi.org/10.1111/j.1365-
2672.2012.05311.x
Moradali, M. F., Ghods, S., & Rehm, B. H. A. (2017). Pseudomonas aeruginosa Lifestyle: A
Paradigm for Adaptation, Survival, and Persistence. Frontiers in Cellular and Infection
Microbiology, 7, 1–29. https://doi.org/10.3389/fcimb.2017.00039
Morello, E., Saussereau, E., Maura, D., Huerre, M., Touqui, L., & Debarbieux, L. (2011).
Pulmonary bacteriophage therapy on Pseudomonas aeruginosa cystic fibrosis strains: First
steps towards treatment and prevention. PLoS ONE, 6(2).
https://doi.org/10.1371/journal.pone.0016963
Mueller-Ortiz, S. L., Drouin, S. M., & Wetsel, R. A. (2004). The Alternative Activation Pathway
and Complement Component C3 Are Critical for a Protective Immune Response against
Pseudomonas aeruginosa in a Murine Model of Pneumonia. Infection and Immunity, 72(5),
2899–2906. https://doi.org/10.1128/IAI.72.5.2899-2906.2004
Ojielo, C. I., Cooke, K., Mancuso, P., Standiford, T. J., Olkiewicz, K. M., Clouthier, S., …
Moore, B. B. (2003). Defective Phagocytosis and Clearance of Pseudomonas aeruginosa in
the Lung Following Bone Marrow Transplantation. The Journal of Immunology, 171(8),
4416–4424. https://doi.org/10.4049/jimmunol.171.8.4416
Pearson, J. P., Feldman, M., & Iglewski, B. H. (2000). Pseudomonas aeruginosa Cell-to-Cell
Signaling Is Required for Virulence in a Model of Acute Pulmonary Infection. Infection and
Immunity, 68(7), 4331–4334.
R Core Team. (2019). R: A language and environment for statistical computing. R Foundation
for Statistical Computing, Vienna Austria. https://R-project.org/
Plummer, M., Best, N., Cowles, K., & Vines, K. 2006. CODA: Convergence Diagnosis and
Output Analysis for MCMC. R News, 6:7-11.
R Core Team. (2019). R: A language and environment for statistical computing. R Foundation
for Statistical Computing, Vienna Austria. https://R-project.org/Rahman, A., Munther, D.,
Fazil, A., Smith, B., & Wu, J. (2018). Advancing risk assessment: mechanistic dose-
response modelling of Listeria monocytogenes infection in human populations. Royal
Society Open Science, 5: 180343. http://dx.doi.org/10.1098/rsos.180343
26
Ramphal, R., Balloy, V., Huerre, M., Si-Tahar, M., & Chignard, M. (2005). TLRs 2 and 4 Are
Not Involved in Hypersusceptibility to Acute Pseudomonas aeruginosa Lung Infections .
The Journal of Immunology, 175(6), 3927–3934.
https://doi.org/10.4049/jimmunol.175.6.3927
Sadikot, R. T., Blackwell, T. S., Christman, J. W., & Prince, A. S. (2005). Host Interactions in
Pseudomonas aeruginosa Pneumonia. American Journal of Respiratory and Critical Care
Medicine, 171, 1209–1223. https://doi.org/10.1164/rccm.200408-1044SO
Schuster, M., Lostroh, C. P., Ogi, T., & Greenberg, E. P. (2003). Identification, Timing and
Signal Specificity of Pseudomonas aeruginosa Quorum-Controlled Genes: a Transcriptome
Analysis. Journal of Bacteriology, 185(7), 2066–2079.
https://doi.org/10.1128/JB.185.7.2066
Shrout, J. D., Tolker-Nielsen, T., Givskov, M., & Parsek, M. R. (2011). The contribution of cell-
cell signaling and motility to bacterial biofilm formation. MRS Bulletin, 36(5), 367–373.
https://doi.org/10.1557/mrs.2011.67
Smith, R. S., & Iglewski, B. H. (2003). P. aeruginosa quorum-sensing systems and virulence.
Current Opinion in Microbiology, 6, 56–60. https://doi.org/10.1016/S1369-5274(03)00008-
0
Snary, E. L., Swart, A. N., Simons, R. R. L., Rita, A., Domingues, C., Vigre, H., … Hill, A. A.
(2016). A Quantitative Microbiological Risk Assessment for Salmonella in Pigs for the
European Union. Risk Analysis, 36(3), 437–449. https://doi.org/10.1111/risa.12586
Tang, H. B., Mango, E. D. I., Bryan, R., Gambello, M., Iglewski, B. H., Goldberg, J. B., &
Prince, A. (1996). Contribution of Specific Pseudomonas aeruginosa Virulence Factors to
Pathogenesis of Pneumonia in a Neonatal Mouse Model of Infection. Infection and
Immunity, 64(1), 37–43.
U.S. Environmental Protection Agency. (2005). Drinking Water Contaminant Candidate List 2;
Final Notice. Federal Register, 70(36), 9071–9077.
https://doi.org/10.1017/cbo9781107337442.008
U.S. Environmental Protection Agency. (2009). Drinking Water Contaminant Candidate List 3-
Final. Federal Register, 74(194), 51850–51862.
27
U.S. Environmental Protection Agency. (2016). Drinking Water Contaminant Candidate List 4-
Final. Federal Register, 81(222), 81099–81114.
van Heeckeren, A. M., Schluchter, M. D., Xue, W., & Davis, P. B. (2006). Response to acute
lung infection with mucoid Pseudomonas aeruginosa in cystic fibrosis mice. American
Journal of Respiratory and Critical Care Medicine, 173(3), 288–296.
https://doi.org/10.1164/rccm.200506-917OC
Weir, M. H., Mitchell, J., Flynn, W., & Pope, J. M. (2017). Development of a microbial dose
response visualization and modelling application for QMRA modelers and educators.
Environmental Modelling and Software, 88, 74–83.
https://doi.org/10.1016/j.envsoft.2016.11.011
Wingender, J., & Flemming, H. (2011). Biofilms in drinking water and their role as reservoir for
pathogens. International Journal of Hygiene and Environmental Health, 214(6), 417–423.
https://doi.org/10.1016/j.ijheh.2011.05.009
Wölbeling, F., Munder, A., Kerber-Momot, T., Neumann, D., Hennig, C., Hansen, G., …
Baumann, U. (2011). Lung function and inflammation during murine Pseudomonas
aeruginosa airway infection. Immunobiology, 216(8), 901–908.
https://doi.org/10.1016/j.imbio.2011.02.003
Wu, H., Song, Z., Hentzer, M., Andersen, J. B., Molin, S., Givskov, M., & Høiby, N. (2004).
Synthetic furanones inhibit quorum-sensing and enhance bacterial clearance in
Pseudomonas aeruginosa lung infection in mice. Journal of Antimicrobial Chemotherapy,
53(6), 1054–1061. https://doi.org/10.1093/jac/dkh223
28