a qsar-like analysis of the adsorption of endocrine disrupting compounds, pharmaceuticals, and...
TRANSCRIPT
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 1
Avai lab le a t www.sc iencedi rec t .com
journa l homepage : www.e lsev ie r . com/ loca te /wat res
A QSAR-like analysis of the adsorption of endocrinedisrupting compounds, pharmaceuticals, and personalcare products on modified activated carbons
Adam M. Reddinga,*, Fred S. Cannona, Shane A. Snyderb, Brett J. Vanderfordb
aDepartment of Civil and Environmental Engineering, The Pennsylvania State University, 212 Sackett Bldg, University Park, PA 16802, USAbSouthern Nevada Water Authority, Las Vegas, NV 89153, USA
a r t i c l e i n f o
Article history:
Received 10 October 2008
Received in revised form
6 May 2009
Accepted 16 May 2009
Published online 27 May 2009
Keywords:
Activated carbon
Endocrine disrupting compounds
Pharmaceuticals
QSAR
Chi index
* Corresponding author. Tel.: þ1 814 880 337E-mail address: [email protected]
0043-1354/$ – see front matter ª 2009 Elsevidoi:10.1016/j.watres.2009.05.026
a b s t r a c t
Rapid small-scale column tests (RSSCTs) examined the removal of 29 endocrine disrupting
compounds (EDCs) and pharmaceutical/personal care products (PPCPs). The RSSCTs
employed three lignite variants: HYDRODARCO 4000 (HD4000), steam-modified HD4000,
and methane/steam-modified HD4000. RSSCTs used native Lake Mead, NV water spiked
with 100–200 ppt each of 29 EDCs/PPCPs. For the steam and methane/steam variants,
breakthrough occurred at 14,000–92,000 bed volumes (BV); and this was 3–4 times more bed
volumes than for HD4000. Most EDC/PPCP bed life data were describable by a normalized
quantitative structure–activity relationship (i.e. QSAR-like model) of the form:
BVp ¼�ðTPV� rmcÞðe0:2812�pHs Þ
CV� Co
��0:2758� 8cp þ 0:0011� FOSA
�
where TPV is the pore volume, rmc is the apparent density, CV is the molecular volume, Co
is the concentration, 8cp depicts the molecule’s compactness, and FOSA is the molecule’s
hydrophobic surface area.
ª 2009 Elsevier Ltd. All rights reserved.
1. Introduction, overview, and objectives preventing, controlling, and curing diseases; then through the
Environmental engineers, scientists, and consumers have
considerable interest in predicting the extent to which envi-
ronmentally important compounds can be removed from
potable water sources by means of treatment unit operations.
This pursuit is particularly keen for endocrine disrupting
compounds (EDCs) and pharmaceutical and personal care
products (PPCPs). Analysts are finding these EDCs/PPCPs in
a wide array of fresh water sources at the parts per trillion
(ppt) level (Snyder, 2003; Snyder et al., 2003). These are often
the same chemicals that consumers ingest as a means of
1; fax: þ1 814 863 7304.(A.M. Redding).
er Ltd. All rights reserved
natural processes of human digestion and wastewater pro-
cessing, these EDCs/PPCPs find their way to down-stream
water sources.
Research continues to clarify the toxicological significance
of these trace EDCs/PPCPs in drinking water. The concerns of
consumers have caused increased regulatory focus on this
issue even though the EDCs/PPCPs appear at reportedly low
levels (Snyder, 2003; Snyder et al., 2003). Additionally, at
elevated concentrations, some EDCs/PPCPs appear to have
alarming effects on fish (van Aerle et al., 2001). Improved
analytical methods make possible the accurate and precise
.
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 13850
detection of these compounds at these ppt levels (Snyder
et al., 1999, 2000, 2001a,b, 2004; Trenholm et al., 2006; Van-
derford et al., 2003; Yoon et al., 2003). Given that EDCs and
PPCPs occur at ppt levels, removal of these compounds can be
challenging (Snyder, 2003). However, activated carbon has
proven well-suited to the task of removing low-level organic
contaminants from water.
The authors herein had earlier developed modified acti-
vated carbons that removed organic compounds for prolonged
time (Nowack et al., 2004; Rangel-Mendez and Cannon, 2005).
Specifically, this work demonstrated that by thermally treat-
ing lignite carbon with steam, or with methane plus steam,
the physicochemical properties of the commercial lignite-
based granular activated carbon (GAC) were altered to where
they removed 2-methylisoborneol (MIB) for 6–8 times longer
than their commercial lignite or bituminous counterparts,
when MIB was spiked at ppt levels. For these carbons, higher
pore volume and higher slurry pH corresponded to higher
removal of MIB, natural organic matter, and other
compounds. Because the odorant MIB possesses properties
similar to many EDCs/PPCPs (e.g. similar molecular weight,
solubility, and partitioning), the authors hypothesized that
these thermally modified lignite carbons would likewise
provide increased removal of EDCs and PPCPs.
To appraise this, the Penn State activated carbon team
collaborated with the Southern Nevada Water Authority
analytical team to process EDC/PPCP-spiked Lake Mead water
through conventional and modified activated carbons. For
these trials, the Lake Mead water was concurrently spiked at
ppt levels with 29 EDC/PPCP compounds that were monitored
by HPLC–MS, plus another 31 EDC/PPCP compounds that could
be monitored by GC–MS. The analysis of the GC–MS
compounds was not pursued due to technical issues. The
HPLC–MS compounds exhibited typically shaped GAC break-
through curves, with no breakthrough initially, then gradually
increasing breakthrough. These RSSCT results pertaining to
the lignite-based HYDRODARCO 4000 (HD4000), as compared
to the steam-modified lignite (STEAM), have appeared in
Snyder et al. (2007), with bed lives to 5%, 10%, and 20%
breakthrough reported for 28 EDCs/PPCPs. For the work
herein, the authors have added the results of the methane–
steam tailored lignite variant to that of earlier data set.
The regularity of these RSSCT breakthrough curves indi-
cated that this data set offered a valuable resource for devising
an initial QSAR-like predictive model for EDC/PPCP removal
through activated carbon beds. Clearly, any such compre-
hensive model that is applied to all molecules and all acti-
vated carbons in all native waters would be well beyond the
scope of any single journal paper. The authors, however,
sought to utilize this novel data set as a starting point that
other subsequent research could build upon.
Specifically, the objectives of the work herein were to (a) test
the hypothesis that thermally modified lignite variants would
offer prolonged bed life for removing EDCs/PPCPs at ppt levels
and (b) test the hypothesis that this breakthrough behavior
could be well-described by a QSAR-like model that included
relevant properties of the activated carbon, relevant proper-
ties of the molecules, and a few operational parameters.
To the authors’ knowledge, the RSSCT breakthrough of as
many ppt-level compounds concurrently in a native water like
this had not previously been studied by others; however, the
concurrent adsorption of EDCs/PPCPs at ppt levels in batch
tests has been demonstrated (Westerhoff et al., 2005). Inter-
estingly, as recently as 10–20 years ago, the prospect of
monitoring ppt levels of these compounds would have been
daunting. Because of this, virtually all of the prior molecular-
structure modeling in the refereed literature that has
appraised multiple compound sorption behavior has done so
at the ppb or even ppm levels (e.g. Blum et al., 1994; Brasquet
et al., 1997). Often, these models have been used for isotherm
conditions in deionized water, which can pose distinctions
from flow-through column conditions where background
NOM is present. The concentration disparities meant that the
parameters that affected ppb or ppm sorption performance
could be distinct from those that influenced the ppt sorption.
For example, at the ppt level, the compounds’ interactions
with the activated carbon surface could be considerably more
important than the compounds’ interactions with other
adsorbate molecules; this further meant that there could be
less relevance to the idea of pore-filling, where molecules
collect to form their own liquid-like sorbed phase (e.g. Dubinin
and Stoeckli, 1980; Manes and Hofer, 1969).
The experimental protocol employed herein aimed to offer
both practical engineering water treatment value and scien-
tific significance. Specifically, as a practical matter, the
authors chose organic species that represented common
EDCs/PPCPs. Moreover, the authors intentionally used a native
water for the spiked source, since in water treatment appli-
cations, there would always be an NOM background. With
regard to scientific significance, the authors have endeavored
to discern whether an equation like a quantitative structure–
activity relationship (QSAR) could be applied to this EDC/PPCP
breakthrough data. QSARs have related the activity of
a compound to chemical or physical characteristics that have
been obtained from structure or experimental data. In well-
characterized systems, QSARs have predicted activity (e.g.
partitioning, transport, adsorption) when experimental data
for the compound of interest are not available, including
Henry’s Law constants (Brennan et al., 1998), octanol–water
partitioning (Kier and Hall, 1986), pharmacological efficacy
(Pan et al., 2006), and activated carbon adsorption isotherms at
ppm levels (Blum et al., 1994; Brasquet et al., 1997; Qi et al.,
2000).
To this end, the authors found that such a QSAR-like
equation (Eq. (1)) could indeed predict the breakthrough of
most compounds, when the equation included just a few
QSAR descriptors of the organic species, plus a few parame-
ters that pertained to the activated carbon character.
Compound descriptors were selected through multi-variable
linear regression and then validated using an approach for
internal model validation, known as leave-one-out cross-
validation (LOOCV), with predictive error minimized accord-
ing to the prediction sum-of-squares (PRESS) statistic (Gra-
matica, 2007; Picard and Cook, 1984). LOOCV is particularly
well-suited to small data sets where additional data are either
difficult or impossible to obtain.
With regard to the activated carbon, the pertinent char-
acteristics in Eq. (1) are as follows: (1) the carbon’s internal
pore volume in the 4–1000 A width range (herein identified as
the total pore volume – TPV), (2) the activated carbon’s slurry
Table 1 – BET surface area, slurry pH, and pore volume ofGAC samples.
GAC sample HD4000 STEAM M/S
Slurry pH (pHs) 4.99 9.67 10.2
Mini-column apparent density
(rmc, g/mL)
0.33 0.33 0.33
Micropores (mL/g, 4.06–21.7 A) 0.15 0.25 0.31
Mesopores (mL/g, 21.7–504 A) 0.37 0.45 0.47
Total pore volume (TPV)
(mL/g, 4.06–1000 A)
0.54 0.73 0.1
BET surface area (m2/g) 607 848 1031
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 1 3851
pH, as an indicator of the carbon’s surface chemistry (pHs),
and (3) its apparent density in the column (rmc) (kept constant
herein). Prior work at Penn State (Nowack et al., 2004; Rangel-
Mendez and Cannon, 2005) and by many others (Leng and
Pinto, 1997; Moreno-Castilla, 2004) has shown that the pore
volume and slurry pH influence the extent of adsorption.
With regard to the EDC/PPCP compounds, Eq. (1) includes
three parameters, namely (1) the molecule’s ‘‘compactness’’,
as depicted by the compounds’ 8th-order simple-path Chi
index (8cp – unitless) (Kier and Hall, 1986). This parameter
depicts the number of unique 8-bond paths (non-hydrogen) in
a molecule; with more 8-bond paths, the value of the 8cp
descriptor increases, and the more compact the molecule
becomes relative to its molecular weight (see Support online
material for representative calculation). Also, the equation
includes (2) the molecule’s hydrophobic surface area (FOSA,
A2/molecule), as an indicator of how much of the molecular
surface can hydrophobically interact with the activated
carbon graphene layers, which are also generally hydrophobic
by nature. And finally, (3) the volume of Avogadro’s number of
molecules (CV, mL/mol). This CV is distinguished from the
molar volume (i.e. volume of one mole of pure compound),
which would effectively depict the inverse of the density of
a discrete organic liquid phase. The normalization to molec-
ular volume herein was included to account for the effect of
size exclusion during adsorption. For example, the size of the
adsorbate has been linked to both the access to the internal
pore structure and the available surface area for sorption
(Summers and Roberts, 1988; Kilduff et al., 1996a).
Eq. (1) also includes two parameters that pertain to the
RSSCT operations, namely (1) the molar concentration of the
EDC/PPCP compound (Co, nmol/L), and (2) the bed volumes
(BV) to characteristic initial breakthrough (L/L) for each of the
EDCs/PPCPs that were included in the spiked matrix.
BVp ¼�ðTPV� rmcÞðe0:2812�pHs Þ
CV� Co
��0:2758� 8cp þ 0:0011
� FOSA�
(1)
With an R2 of 0.861, the overall model (Eq. (1)) fits to the
performance of 23 out of the 29 EDCs/PPCPs that could be
analyzed by HPLC–MS. Of the remaining six, two exhibited no
breakthrough for one or several of the RSSCT runs, and four
had characteristics that caused them to be statistical out-
liersdsuch as an exceptionally large molecular volume (see
below).
2. Experimental
2.1. Pore volume distribution and BET surfacearea analysis
Pore volume distributions and BET surface areas (Table 1) of
the GAC samples were determined by interpretation of argon
adsorption isotherms following the protocol of Moore et al.
(2003). The adsorption of gaseous argon onto the GAC surface
was recorded through a pressure range of 10�6 to 0.993 atm
using ASAP 2000 Physisorption Instruments (Micromeritics
Instrument Co., Norcross, GA), with duplicate analyses. The
resulting argon adsorption isotherm data were processed
according to the density functional theory via the Micro-
meritics DFT Plus (V2.02) software to produce a pore volume
distribution for the sample.
2.2. Slurry pH
To monitor the net surface pH, a 0.010 M solution of potas-
sium chloride was prepared from nitrogen-sparged deionized
water. A suspension of carbon sample was prepared by
weighing 0.100 g of thoroughly washed 170 US Mesh� 200 US
Mesh GAC into a 20 mL glass scintillation vial and then adding
15 mL of 0.010 M KCl solution. After 2 h on a shaker table, the
pH of the supernatant was measured using a Mettler–Toledo
DG115-SC pH probe (Mettler–Toledo, Inc., Columbus, OH) that
was calibrated with reference standard buffers at pHs 4, 7, and
10 (VWR, Inc., West Chester, PA). The pH was measured again
after 24 h to verify that equilibrium had been reached. Values
listed below in Table 1 are the average of the three analyses,
with a standard deviation of 0.21. The authors noted that the
slurry pHs of these carbons did not change while the dry
carbon samples were stored in a vacuum desiccator for up to 1
year.
2.3. Activated carbon materials and preparation
The commercial lignite HD4000 used in this research origi-
nated from Norit Americas, Inc., Marshall, TX, USA. The
lignite variants were prepared utilizing a high-temperature
steam (STEAM) or methane/steam (MS) modification protocol
as described by Rangel-Mendez and Cannon (2005).
GAC samples for use in RSSCTs were prepared by grinding
the as-received commercial GAC and then wet-sieving it using
a 170-US mesh (0.090 mm) sieve combined with a 200-US
mesh (0.075 mm) sieve to obtain the 170� 200 fraction. In wet-
sieving, the surface of the sieve was rinsed with distilled water
to remove fine particles and prevent particles from sticking to
each other or to the surface of the mesh. Dry-sieving would
not have been adequate as electrostatics would have pre-
vented the correct particle size from passing the mesh. After
rinsing with about 0.5 L of additional distilled water per gram
of ground GAC, the samples were dried at 105 �C for 24 h
before storage. Samples were stored in a vacuum desiccator
for short periods (days) or under a slightly pressurized (2 psig)
nitrogen atmosphere for long-term storage (months). The
storage techniques aimed to minimize the exposure to
atmospheric oxygen and moisture.
Fig. 2 – Volumetrically normalized adsorption (nL
compound per mL TPV) versus slurry pH for five
compounds: testosterone (-), ethynylestradiol (B),
diazepam (C), naproxen (6), and ibuprofen (:). These
compounds displayed the typical response to changing
slurry pH of the GAC. The value used for slope, as per Eq.
(3), was the average of all compounds’ slopes.
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 13852
2.4. Rapid small-scale column tests
The bench-scale evaluation consisted of RSSCTs as a measure
of GAC adsorption performance with EDCs/PPCPs. The RSSCTs
simulated a full-scale GAC bed of #12� 40 US mesh size grains
operating at a 7.6 min empty bed contact time, with protocol
as presented previously (Snyder et al., 2007). A 7.6-min empty
bed contact time translates to 70,000 bed volumes of water
processed per year. The EDC/PPCP tests herein employed
constant diffusivity similitude because of the compound’s low
molecular weight (average 257 Da) and relatively weak
polarity (refer to Crittenden et al., 1986, 1991).
These RSSCT experiments used the influent from a water
treatment plant at Lake Mead, NV, USA. This water had
a natural organic matter content of 3.0 mg/L TOC (Shimadzu
TOC-VCSN, Columbia, MD, USA). Water was stored at 4 �C in
stainless steel containers until testing. A total of 60 EDCs/
PPCPs were collectively dissolved in either methanol or
methylene chloride and dosed to reach concurrent concen-
trations of 100–200 ppt (0.24–1.49 nmol/L) for each compound
in the influent water (Snyder et al., 2007). Effluent concentra-
tions were measured for those 29 compounds that could be
monitored via high pressure liquid chromatography–mass
spectrometry, as per the analytical method presented in
Snyder et al. (2007). The STEAM and MS GACs slightly raised
the effluent pH for the first w900 bed volumes; and thereafter,
the effluent pH returned to the influent pH (8.05).
3. Results and discussion
3.1. Outline of model development steps
In light of the considerable complexity of the model development,
this text first presents a step-by-step outline of the approach
used herein. Details are included in the sections that follow.
Step 1: The authors conducted RSSCTs with natural Lake Mead
water that was concurrently spiked with many EDCs/PPCPs
Fig. 1 – Bed volumes treated at characteristic initial breakthroug
showed no breakthrough during the HD4000 test length. Proges
test length of the M/S carbon. Oxybenzone exhibited no breakth
arranged in order of M/S breakthrough.
and analyzed the effluents for these species. The authors also
characterized the three lignite GACs relative to pore volume
distribution, slurry pH, surface area, etc.
Step 2: The authors calculated a characteristic initial break-
through (BVobs, L/L) for each of these compounds from the
RSSCT data of each lignite variant (Fig. 1).
Step 3: The authors compiled and computed numerous (>40)
molecular descriptors for the EDCs/PPCPs (see Support online
material, Tables S1–S5).
Step 4: The authors plotted the GAC slurry pH (x-axis) versus
the volumetrically normalized breakthrough ( y-axis) (e.g.
Fig. 2) and characterized this plot as per Eq. (2) (where
parameters are as listed above and below):
h for the 28 EDCs/PPCPs exhibiting breakthrough. Triclosan
terone breakthrough was observed only during the longer
rough during any of the three test lengths. Compounds are
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 1 3853
BVobs ��
CV� Co
TPV� rmc
�¼ ez�pHs (2)
For each compound’s curve, the z-coefficient was calculated
and the z-values fell within a narrow range with an average of
z¼ 0.2812. This average z-value was used in the subsequent
model development.
Step 5: The authors calculated the ‘‘Relative Adsorbability’’
observed for each EDC/PPCP by plotting the observed charac-
teristic breakthrough (BVobs, y-axis) versus the volumetric
normalization (x-axis), now including the influence of slurry
pH (pHs) as per Step 4 above (e.g. Fig. 3). For each molecule’s
plot with the three GACs, the Relative Adsorbability was the
slope of this linear relationship (Eq. (3)).
BVobs ¼�ðTPV� rmcÞðe0:2812�pHs Þ
CV� Co
�� Relative Adsorbabilityobs
(3)
Step 6: The authors determined which molecular properties
could account for the variation in the compound’s Relative
Adsorbability by conducting a multi-variable regression
(Minitab V15, State College, PA) with an equation (Eq. (4)) of the
form (where ‘‘a’’ and ‘‘b’’ are best-fit coefficients):
Relative Adsorbability ¼ a�molecular descriptor1 þ b
�molecular descriptor2/ (4)
The authors exhaustively evaluated the molecular parameters
that produced the highest correlation between the predicted
and observed Relative Adsorbabilities. The best-fit coefficients
and the selected parameters ( p< a¼ 0.001) were then inserted
into the final equation (Eq. (1) shown again below for clarity),
Fig. 3 – Relative Adsorbability (RA) for several compounds
(as defined per Eq. (3)): androstenedione (C), caffeine (B),
diazepam (:), gemfibrozil (6), ibuprofen (-), and
sulfamethoxazole (,). Each dashed line shows the linear
regression and associated slope (RA) for each compound’s
data with the origin assigned as the intercept.
which thus offered a prediction of bed volumes (BVp) to
characteristic breakthrough (Fig. 4).
BVp ¼�ðTPV� rmcÞðe0:2812�pHs Þ
CV� Co
��0:2758� 8cp þ 0:0011
� FOSA�
(1)
Step 7: After determining the significant predictors (which
were 8cp and FOSA), the authors statistically evaluated if any
other parameters could be added that were significant in the
correlation between BVobs and BVp. Again, if p< a¼ 0.001, then
it was construed as significant. No other parameters were
found to be significant among more than 40 that were
examined.
Step 8: The authors tested this QSAR-like model by internal
validation per leave-one-out cross-validation (LOOCV), with
the final model form corroborated by a minimized predictive
sum-of-squares (PRESS-statistic) (Gramatica, 2007; Picard and
Cook, 1984).
Step 9: The authors examined the EDC/PPCP compounds that
were statistical outliers in this BVobs versus BVp analysis and
found that some of these also had physical–chemical outlier
characteristics. These characterizations could aid scientists in
extending the breadth of subsequent QSAR-like model
applicability.
3.2. GAC characteristics and RSSCT removalof EDCs/PPCPs
For the three lignite GAC variants, the pore volume, slurry pH,
surface area, and mini-column apparent density are pre-
sented in Table 1. The pore volume distributions appear in
Support online material, Fig. S1. Notably, for the three
carbons, the ratio of micropores-to-total pores was relatively
uniform. This meant that it was not possible to statistically
Fig. 4 – Predicted (Eq. (1), BVp) versus observed initial
breakthrough (Fig. 1, BVo) for 23 EDCs/PPCPs as described
by the QSAR-like model. Adsorption generally increased in
this order: M/S (,) > STEAM (6) > HD4000 (B). The
breakthrough of fluoxetine with steam (:) was a statistical
outlier (>3 standard deviations).
Fig. 6 – Estradiol breakthrough. Dashed lines indicate
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 13854
distinguish whether, for example, micropores or total pores
were more correlated to sorption performance (refer to
Nowack et al., 2004). Thus, the model herein used total pore
volume (width< 1000 A).
The RSSCT results have been plotted as the relative
concentration (C/Co) versus bed volumes of water processed
(Figs. 5 and 6; and Support online material, Figs. S2–S27).
STEAM and M/S removed EDCs/PPCPs 3–4 times longer (i.e.
14,000–92,000 BV for initial breakthrough) than did HD4000
(i.e. 2000–26,000 BV for initial breakthrough) (Fig. 1).
The most readily sorbed group of compounds were the
steroids: androstenedione, estradiol, estriol, estrone, ethy-
nylestradiol, progesterone, and testosterone (Fig. 1). On
observation, these steroids were found to be quite similar to
one another in their molecular volume, which averaged to
84 mL/mol.
regression through available data (C/Co < 0.50, circledpoints). The method of data interpretation is demonstrated
here (dotted line) where the M/S slope is used to define
breakthrough with STEAM. Influent concentration (Co)
equaled 0.66–0.85 nmol/L (179–232 ppt).
3.3. Computing the characteristic initial breakthrough
The authors sought to calculate the point of characteristic
initial breakthrough in a standardized manner while
including as much data as possible for each EDC/PPCP. To do
so, the authors fit a linear regression to all the data points with
0< C/Co� 0.50, plus the last C/Co¼ 0.00 data point before
breakthrough when possible. The authors defined the inter-
section of this regression line with the x-axis as the charac-
teristic initial breakthrough, as shown by an example in Fig. 5.
If only one data point or no points were available below
C/Co¼ 0.50, the first two available data points were used.
Support online material includes the breakthrough curves and
regressions for the additional test compounds not shown here
(Figs. S2–S27).
Occasionally, effluent sample selection offered insufficient
data to adequately define the initial breakthrough of an EDC/
PPCP with the STEAM carbon. This occurred with five
compounds where there was only one or two data points
above C/Co¼ 0. In those instances, the slope of the break-
through curve was assumed to be parallel to that observed for
the same EDC/PPCP compound with the M/S carbon (i.e. the
Fig. 5 – Meprobamate breakthrough. Circled points indicate
those used for interpretation of characteristic initial
breakthrough. Influent concentration (Co) equaled 0.86–
1.20 nmol/L (188–262 ppt).
carbon with the most similar pore volume distribution and
slurry pH) (see Fig. 6 for an example). It is noted that in more
than half of the cases where considerable data were available
for both STEAM and MS, the two breakthrough curves were
indeed parallel to one another. With this approach, the
authors employed more than w80% of the data points that
exceeded detection, so as to mathematically and narrowly
define the characteristic initial breakthroughs.
The authors note that other equation forms for charac-
teristic breakthrough could have been used (and indeed were
appraised), but the approach herein used the most data. With
the standard deviation inherent in analyzing these
compounds at the ppt level, the more comprehensive use of
the data set meant a more accurate and precise definition of
breakthrough. Also, as a practical matter, this ‘‘zero-break-
through’’ characterization reflects the consumer’s desire to
drink water without detectable EDCs/PPCPs.
3.4. Molecular descriptor calculations
The authors employed QikProp (Schrodinger LLC, New York,
NY, USA), a molecular analysis program to compute various
properties of the 29 EDCs/PPCPs herein. These properties
included molecular volume, various types of solvent-acces-
sible surface areas (1.4 A probe radius), dipole, polarizability,
and ionization potential. Two-dimensional ‘‘mol’’ files of the
compounds were acquired from the National Institute of
Standards and Technology WebBook (http://webbook.nist.
gov). The LigPrep computer program (Schrodinger LLC, New
York, NY, USA) produced optimized three-dimensional mol
files from the two-dimensional files.
The authors used the computer program Molconn-Z V4.10
(Edusoft, La Jolla, CA, USA) to calculate Chi connectivity
indices and kappa shape (topological) indices for the pertinent
EDCs/PPCPs. These descriptors have frequently appeared in
QSARs (Kier, 1986; Kier and Hall, 1986). The calculations
included the indices up to the 8th order, including simple-
Fig. 7 – Relative Adsorbability as calculated for all 22
compounds that exhibited breakthrough with all three
carbons.
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 1 3855
path, valence-path, valence-path-cluster, and valence-chain
types (Kier and Hall, 1986). For Chi indices above the 8th order,
many of the compounds produce ‘‘zero’’ values and the
molecule is thus too small to be described by indices that were
higher than the 8th order. A full list of these properties and
descriptors can be found in Support online material (Tables
S1–S5).
3.5. Accounting for the influence of the carbon’s slurrypH
To address the influence of GAC slurry pH (pHs), the authors
first plotted pHs versus the normalized adsorption for each
compound, as it appears in Fig. 2 and Support online material
(Figs. S28–S30). Because pH is a logarithmic parameter, the
authors characterized each curve by an exponential expres-
sion (see Section 3.1, Eq. (2)).
In Eq. (2), ‘‘z’’ reflected the influence of pHs on adsorption. It
is noted that since the slurry pHs were somewhat similar
between the STEAM lignite (9.67) and the MS lignite (10.2), the
two points corresponding to these did not vary greatly for any
given compound. Most importantly, the z-values showed
minimal variation (average¼ 0.2812, standard deviation¼0.033) and ranged from 0.230 to 0.357 (see Fig. 2 and Support
online material (Figs. S28–S30)). Thus, the selection of the
average z¼ 0.2812 was effectively based on 72 data points.
This average pre-exponential z was then used in Step 5 of the
QSAR-like model development, as below.
3.6. Relative Adsorbability
As in Step 5 of the QSAR-like model development, the authors
calculated a value herein termed the ‘‘Relative Adsorbability.’’
To find this value for each EDC/PPCP, the authors fit a line to
the plot of volumes of water processed at characteristic initial
breakthrough (L water/L carbon bed) versus the pore-to-
molecule ratio (x-axis) as per Eq. (3) (see Section 3.1). For the 22
compounds where this could be calculated, the pore-to-
molecule ratios ranged from 10,000 to 120,000 (mL of total
pores/L of GAC bed)/(nL of molecules/L of water). These plots
are shown for the 22 compounds in Fig. 3 and Support online
material (Figs. S25–S27).
For a given compound, the slope of the plot indicated the
Relative Adsorbability observed. These plots served as
a means of discerning a volume-to-volume based comparison
of adsorption behavior; and they posed some similarities to
the gravimetric-based isotherms. The difference is that for
these plots, the selection of GACs offered a range of total pore
volume; whereas in isotherms, the selection of GAC offers
a range of mass. Moreover, the Relative Adsorbability plots
employ column flow-through data at very low concentrations;
whereas isotherms employ batch data that must be collected
at inherently higher concentrations if true differences are to
exceed standard errors.
The Relative Adsorbabilities ranged from 0.13 to 1.24 (nL
molecules sorbed/mL pores available) (Fig. 7); this was an
order of magnitude span for these 22 EDCs/PPCPs (see Relative
Adsorbability (RA) in Table 2). The data were not sufficient to
calculate the Relative Adsorbability for progesterone or tri-
closan, as these compounds did not show breakthrough with
all three GACs. Using Fig. 7 and Table 2 data, one could
calculate the volume (nL) of all the compounds sorbed at
breakthrough, relative to the total pore volume. This
comparison showed that only w10�3 to 10�4 percent (%) of the
pores were occupied by these EDCs/PPCPs. Moreover, one
could compute that less than 0.2% of the pore surface would
have been covered with these EDCs/PPCPs.
3.7. Multi-variable linear regressions for identifyingsignificant molecular descriptors
As the sixth step, the authors compared the observed Relative
Adsorbabilities (Section 3.6) to the set of molecular properties
and descriptors in a stepwise linear regression, as per the
standard expression:
Relative Adsorbability ¼ a�molecular descriptor1 þ b
�molecular descriptor2/ (4)
The regression analysis identified 8cp and FOSA as the only
two statistically significant parameters. The best-fit coeffi-
cients for this regression against Relative Adsorbability were
as follows: a¼ 0.2758 and b¼ 0.001098. For this, the R2 corre-
lation was 0.851 when including 22 of the EDCs/PPCPs (see
Fig. 8).
These parameters were then included in the overall
equation to yield the final QSAR-like model (Eq. (1) and Fig. 4).
This equation described the characteristic breakthrough for 23
of the test compounds, with a coefficient of determination (R2)
of 0.861.
Six compounds were not included in Fig. 4 regression; these
are the italicized compounds in Table 2. Two of the excluded
compounds exhibited no breakthrough for one or more
RSSCTs. Four of the excluded compounds were statistical
outliers; of these, some had chemical properties that could
have contributed to their outlier status, as discussed below.
As mentioned above, two descriptors, the 8th-order
simple-path Chi index (8cp) and the hydrophobic component
of the solvent-accessible surface area (FOSA, A2/molecule)
offered good statistical correlation. The 8cp and FOSA values
Table 2 – Influent concentrations and molecular properties of the 29 EDCs/PPCPs.
Compound RA(nL/mL)
Mol. weight(Da)
log Kow pKa log D(pH¼ 8.05)
Co (nmol/L) CV(mL/mol)
FOSA(A2)
8cp
HD4000 STEAM M/S
Acetaminophen 0.240 151.2 0.46 9.38 �0.89 0.82 1.13 0.71 40.25 93.6 0.136
Androstenedione 1.244 286.2 2.75 NA 2.75 0.71 0.75 0.89 83.35 389.8 2.717
Atrazine 0.628 215.7 2.61 1.7 �3.74 0.89 1.21 1.25 60.43 314.7 0.511
Caffeine 0.813 194.1 �0.07 10.4 �0.07 1.09 1.32 1.49 49.17 239.0 0.506
Carbamazepine 0.761 236.3 2.45 7.00 1.36 0.63 0.90 1.03 69.07 43.2 1.748
DEET 0.501 191.3 2.18 4.5 2.18 0.94 1.25 1.45 59.99 292.9 0.245
Diazepam 0.618 284.7 2.82 3.4 2.82 0.47 0.60 0.76 80.88 120.3 1.696
Diclofenac 0.244 294.0 0.7 4.51 �2.84 0.41 0.51 0.60 73.53 24.4 1.066
Dilantin 0.258 252.3 2.47 8.33 2.01 0.61 0.74 0.84 69.47 0.0 1.299
Erythromycin NA 734.5 3.06 8.88 3.00 0.02 0.02 0.01 189.2 802.5 3.239
Estradiol 1.192 272.4 4.01 10.4 4.01 0.66 0.81 0.85 80.02 304.1 2.693
Estriol 0.977 288.4 2.45 10.4 2.45 0.76 0.82 0.98 81.47 266.9 2.826
Estrone 1.098 270.4 3.13 10.4 3.13 0.78 0.91 0.98 79.08 300.2 2.693
Ethynylestradiol 0.947 296.4 3.67 10.4 3.67 0.64 0.77 0.84 87.34 306.2 2.978
Fluoxetine 0.683 309.1 4.05 8.7 3.31 0.27 0.48 0.46 81.04 156.9 1.218
Gemfibrozil 0.493 250.3 4.77 4.42 1.14 0.62 0.86 0.93 73.01 366.2 0.607
Hydrocodone 0.617 299.4 2.16 7.32 2.09 0.47 0.89 0.92 84.85 383.5 4.659
Ibuprofen 0.217 206.3 3.97 4.91 0.83 0.70 1.03 1.07 61.20 278.5 0.586
Lopromide NA 791.1 �2.05 10.2 �2.05 0.24 0.30 0.25 142.2 363.4 1.551
Meprobamate 0.272 218.3 0.7 10.9 0.07 0.86 0.95 1.20 53.73 234.4 0.167
Naproxen 0.430 230.3 3.18 4.15 �0.72 0.53 0.97 0.99 65.74 179.9 1.206
Oxybenzone NA 228.1 3.79 7.77 3.33 0.01 0.04 0.04 64.60 92.2 0.671
Pentoxifylline 0.624 278.3 0.29 0.97 0.29 0.61 0.66 0.83 72.46 366.2 1.207
Progesterone NA 314.5 3.87 NA 3.87 0.57 0.57 0.72 92.44 430.0 3.085
Sulfamethoxazole 0.130 253.3 0.89 5.5 0.89 0.58 0.71 0.54 64.83 96.4 0.492
TCEP 0.329 285.5 1.44 NA 1.44 0.52 0.72 1.02 56.13 241.0 0.188
Testosterone 1.184 288.4 3.32 NA 3.32 0.68 0.71 0.84 84.29 393.9 2.717
Triclosan NA 287.5 4.76 7.9 4.38 0.29 0.37 0.58 67.24 0.0 0.890
Trimethoprim 0.893 290.3 0.91 7.12 0.86 0.55 0.82 0.88 84.22 295.5 1.244
RA¼ Relative Adsorbability, Co¼ influent concentration, CV¼molecular volume, FOSA¼hydrophobic portion of solvent-accessible surface
area, 8cp¼ 8th order simple-path Chi index, Kow¼ log octanol–water partition coefficient. Outliers (>3 standard deviations), relative to the
averages for these values, are italicized. Italicized compounds were outliers from the final model as listed in Table 4. DEET: N,N-diethyl-meta-
toluamide, TCEP: tris(2-chloroethyl) phosphate.
Fig. 8 – Predicted versus observed Relatively Adsorbability
for 22 EDCs/PPCPs. Predictions were made using the 8cp
index and FOSA. Two compounds were outliers (Tables 3
and 4) on the basis of: (1) influent concentration (caffeine
(B)) and (2) 8cp value (hydrocodone (,)).
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 13856
for these EDCs/PPCPs are listed in Table 2. The 8cp Chi index
encodes for the number of unique 8-bond paths present in
a molecule, excluding bonds to hydrogen (Kier and Hall, 1986).
As the number of paths increases, the value of the 8cp index
will increase, and the compactness, relative to the molecular
weight, increases. This in turn corresponds to faster adsorp-
tion kinetics and greater access into smaller pores. Thus, for
the 22 compounds (Fig. 8), the positive correlation of the 8cp
index with the initial breakthrough implied that an increased
degree of compaction benefited adsorption. Kilduff et al.
(1996b) observed a similar phenomenon with the adsorption
of humic acids onto activated carbon: increasing ionic
strength caused macromolecules to coil and reduce their
effective size such that adsorption improved via increased
access to smaller pores and enhanced diffusivity.
The regression analysis selected the simple index (8cp) as
statistically significant, as opposed to the valence index (8cpv);
this result posed important implications. The simple index is
purely a structural descriptor and considers only the
compactness of the molecule, regardless of its heteroatom
substituents. Thus it encodes for only the s electrons. This
contrasts to the valence index which also accounts for
p-electrons and lone pair electrons additionally (Kier and Hall,
1986).
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 1 3857
The hydrophobic surface area (FOSA) represents the
aliphatic portion of the solvent-accessible surface area that is
composed of saturated carbons and their attached hydrogen.
It is noted that the FOSA value does not include any molecular
surface area due to p-bonds (PISA) (see discussion below). The
significance of the FOSA descriptor is logical, given that
hydrophobic interactions are known to strongly enhance
adsorption extent and kinetics. As a corollary to Traube’s rule
(Ward, 1946), the adsorbability of aliphatic acids increased
with increasing hydrocarbon length and their corresponding
increase in hydrophobicity (Moreno-Castilla, 2004). Likewise,
this relationship was manifested in our QSAR model herein:
adsorption increased with increasing hydrophobic surface
area, as would be expected.
Implied support for the selection of the 8cp and FOSA
variables can be found elsewhere in the literature. In their
application of the solvophobic theory to the adsorption of
homologous series of alcohols and ketones, the work of Belfort
et al. (1984) in isotherm tests demonstrated that for a set of six
hexanol isomers (2,3-dimethyl 2-butanol through 1-hexanol)
adsorption increased with increased branching, not with
decreased molecular size (i.e. molecular volume). Though not
explicitly calculated by Belfort et al. for this group of isomers,
the adsorbability bears some correlation to cavity surface area
(A2, R2¼ 0.82). Interestingly, when the cavity surface area (i.e.
molecular surface area) is normalized by the molecular
volume (A2/A3), then the correlation with adsorbability
increases noticeably (R2¼ 0.86).
3.8. Regression statistics and leave-one-outcross-validation
The authors used a stepwise linear regression to identify the
significant predictors. With each successive step of the
regression, predictors and outlier observations were identified
(Table 3: starting with Model A and ending with Model H). The
statistics generated for each step are also listed (e.g. best-fit
coefficient and predictor p-value). Because of the moderate
size of the dataset, the authors used an internal validation
method, namely leave-one-out cross-validation (LOOCV) with
the predicted sum-of-squares (PRESS-statistic). This LOOCV
approach was used to verify the choice of model. In LOOCV,
one observation at a time is removed from the data set, and
the remaining observations are used successively to predict
the deleted observation (Gramatica, 2007; Picard and Cook,
1984). The PRESS-statistic is then the sum of the error for
predicting each deleted value using LOOCV; the model with
the lowest relative PRESS is thus the preferred model. Values
of the PRESS-statistic are also listed in Table 3. Additionally,
the statistical significance (i.e. p-value) is listed for each of
these other molecular parameters when added as a third
predictor in the model (Model H) where 8cp and FOSA were
already identified as predictors (Support online material,
Tables S1–S5).
3.9. Significance of other molecular descriptors
The authors compared many molecular descriptors to the
observed activity (i.e. adsorption) because each descriptor
encoded unique structural information. The large number of
descriptors required a rigorous critical significance level ( p< a
¼ 0.001) when accepting or rejecting predictors. Specifically,
Tables S1–S5 in Support online material list the statistical
significance of each molecular descriptor when added as an
additional predictor (i.e. in addition to FOSA and 8cp) for
Relative Adsorbability. As indicated, none of these QSAR
descriptors were significant ( p< a¼ 0.001); thus these others
were not included in the final equation. Indeed, the next
smallest p-value amongst these other 40 descriptors was 0.202
(Support online material, Table S4), which was far above 0.001.
Several of the particular parameters that posed no statis-
tical significance at these low ppt levels included Kow, log Kow,
log D, PISA, the valence index (8cpv), log S, or log S (at pH 8.05). It
is notable that the log of the octanol–water partition coeffi-
cient (log Kow) was not predictive of the observed break-
through. This compares to previous work by others that
showed that log Kow was important at the ppm concentration
level (Abe et al., 1988; Hu et al., 1998; Walters and Luthy, 1984).
Indeed, substituting log Kow for the ‘‘affinity’’ term (FOSA) in
the regression greatly reduced the quality of the correlation
(R2¼ 0.705). Further, if log Kow were to be added to the model
as a predictor alongside 8cp and FOSA, the resulting p-value of
0.878 (Support online material, Table S3 and Table 3: Model I)
identifies it as offering no more significant correlation. During
the research herein, the very low aqueous concentrations of
the adsorbates (ppt) appeared to greatly reduce the relevancy
of log Kow and log D (Table 3: Model J), and log S (Support
online material, Table S3). A similar irrelevancy of log Kow at
the ppt level has been observed by others (McDonough et al.,
2008; Yu et al., 2008) who conducted ppt single solute batch
experiments at ppt levels.
The authors are not aware of other studies that have
appraised concurrent sorption in flow-through tests of
numerous EDCs/PPCPs at the ppt level; the distinctions pose
important phenomenological differences. For example,
because Kow values are derived from the interaction of
a molecule immersed in liquid phases, the polar chemistry of
the entire molecule can influence partitioning. In contrast, for
these ppt-level RSSCTs, the 60 sorbed EDC/PPCP molecules
covered less than 0.2% of the GAC surface; thus interactions
with the GAC surface may have been more important than
interactions with other EDC/PPCP molecules.
Additionally, the authors observed that p surface area
(PISA) did not correlate with breakthrough (Table 3: Model K)
although 15 of the 22 compounds contained aromatic rings.
Others have noted that p–p dispersion interactions with these
rings were dominant in aromatic adsorption at much higher,
part-per-million concentrations (Radovic et al., 1997). But, for
this set of compounds, the PISA was not a significant
descriptor of the observed adsorption at the ppt level.
3.10. Breadth of QSAR-like model applicability
The QSAR-like model herein (Eq. (1)) utilized relatively few
parameters for characterizing the adsorption of numerous
EDCs/PPCPs. With regard to the activated carbon parameters,
this QSAR-like model appraised three lignite variants. With
regard to molecular parameters, the authors observed that the
model offered the greatest accuracy for those compounds
where influent concentration (Co) averaged 0.24–1.25 nmol/L,
Table 3 – Statistics produced from the linear stepwise regression for selecting predictors, using leave-one-out cross-validation (LOOCV) with the predicted sum-of-squares(PRESS-statistic) to verify model selection.
Model A B C D E F G H I J KSample size (n) 24 23 22 24 23 22 22 22 22 22 22
Predictor 1 8cp8cp
8cp8cp
8cp8cp
8cp8cp
8cp8cp
8cp
Coefficient 0.914 0.2938 0.3112 0.4119 0.1546 0.2553 0.2730 0.27579 0.2787 0.2727 0.271
p-value (a¼ 0.001) <0.001 <0.001 <0.001 <0.001 0.003 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001
Predictor 2 – – – – FOSA FOSA FOSA FOSA FOSA FOSA FOSA
Coefficient 0.00102 0.00111 0.00106 0.00110 0.00116 0.00110 0.00102
p-statistic (a¼ 0.001) 0.034 0.001 0.001 <0.001 <0.001 <0.001 <0.001
Predictor 3 – – – – – – – – Log Kow Log D PISA
Coefficient �0.00387 0.00291 0.00006
p-statistic (a¼ 0.001) 0.878 0.880 0.794
Constant 0.355 0.250 0.207 – 0.171 0.04533 0.01627 – – – –
p-value (a¼ 0.001) <0.001 0.002 0.007 0.147 0.577 0.818
PRESS-statistic 2.28 1.02 0.81 1.06 2.06 0.67 0.49 0.45 0.53 0.52 0.50
R-squared (%) 44.2 67.9 74.4 – 55.2 80.8 86.3 – – – –
R-squared (%, Pred) 16.0 62.5 69.8 – 24.0 75.5 81.8 – – – –
F-statistic 17.4 44.4 58.2 221.9 12.9 42.2 59.7 302.4 191.8 191.8 192.3
Outlier statistics Standardized residual
Hydrocodone �3.06 R R R �3.47 R R R R R R
Caffeine 1.42 2.12 R R 1.38 2.41 R R R R R
R¼ removed from sample set as a statistical outlier.
Steps proceeded from Model A to Model H, where Model H was selected, based on a minimum PRESS value. Two outliers were identified in model selection, hydrocodone and caffeine; these are shown
along with their standardized residual values. Additional possible models (I–K) are included to show the statistics generated when including additional probable predictors (e.g. Log Kow).
wa
te
rr
es
ea
rc
h4
3(2
00
9)
38
49
–3
86
13
85
8
Table 4 – Compounds that were excluded as statistical outliers, the properties that identified them as outliers, and result ifbreakthrough is nonetheless predicted.
Compound Reason for outlier status Complete data set Effect
Caffeine High average Co (1.30 nM/L) Avg Co¼ 0.72, StDev¼ 0.29 Under predicted
Erythromycin Low average Co (0.018 nM/L) Avg Co¼ 0.72, StDev¼ 0.29 Over predicted
High molar
refractivity (189 mL/mol)
Avg CV¼ 77.7, StDev¼ 28.0
Hydrocodone High 8cp value (due to cross-linking, 8cp¼ 4.66) Avg 8cp¼ 1.55, StDev¼ 1.17 Over predicted
Lopromide High molar refractivity (142 mL/mol) Avg CV¼ 77.7, StDev¼ 28.0 Over predicted
Oxybenzone No breakthrough during tests Avg Co¼ 0.72, StDev¼ 0.29 N/A
Low average Co (0.029 nM/L)
Triclosan Simple index (8cp) cannot account for influence
of two chlorinated aromatic rings
Maximum of one chlorinated
aromatic ring per molecule
Under predicted
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 1 3859
molecular volume (CV) was 40–87 mL/mol, 8cp was 0.13–3.3,
and halogenation was limited to one aromatic ring. For four of
the compounds, at least one of these parameters was outside
this range (italicized in Table 2). The properties of these outliers
are listed in Table 4. For example, the molecular volumes of
erythromycin (189.2) and iopromide (142.2) are nearly twice as
high as for the other compounds. Hydrocodone has an 8cp that
is nearly twice as high as for other species. Also, triclosan has
three chlorine substituents attached to its ring structure; the
influence of this would not be seen in the 8cp index.
For three compounds, oxybenzone, progesterone, and tri-
closan, breakthrough was not observed in every case. No
breakthrough of oxybenzone was observed in any test,
apparently due to the low influent concentrations (Co¼ 0.01–
0.04 mmol/L); the model thus correctly predicts that break-
through would have occurred well beyond the sampling range
of the tests herein (e.g. HD4000 w 320,000 BV, STEAM w 405,000
BV, MS w 520,000 BV). Progesterone and triclosan did not
breakthrough during the HD4000 and STEAM tests, but did
breakthrough for the MS run. For progesterone, the model
predicts that breakthrough should have been observed with
HD4000 at w18,000 BV, but breakthrough was not observed.
Then, for STEAM the predicted progesterone breakthrough is
w92,000 BV, which was past the range of sampling herein. For
triclosan the model predicts breakthrough at w9000 BV for
HD4000; and at w36,000 BV for STEAM; but no breakthrough
was observed. As indicated below, triclosan has three chlorine
substituents on rings, which apparently render it an outlier.
In follow-up studies, which could be designed to include
a broader range of compounds, scientists could include
a spectrum of parameters that extend beyond the applicable
ranges above. For example, a set of compounds could be
intentionally selected that offered a range of chlorine heter-
oatomsdfrom none to four, so as to specifically address the
impact of such heteroatoms. As a practical first step, the
authors selected the compounds herein on the basis of their
environmental significance, rather than on the basis of
a particular span of some targeted parameter.
Further understanding the adsorption of EDCs/PPCPs from
a natural water requires consideration of the natural organic
matter (NOM) present. It stands to reason that the NOM in this
surface water (3.0 mg/L TOC) competed with the target
compounds by both occupying adsorption sites and blocking
access to pores. Because competition with NOM is strongly
size-dependent (Newcombe et al., 2002), the low variation in
molecular volume for the fitted compounds (average¼ 72 mL/
mol, standard deviation¼ 13 mL/mol) likely caused similar
levels of NOM competition for all EDCs/PPCPs. Because only
the level of NOM (as TOC) was measured herein, it was well
beyond the scope of this study to appraise the influence of
NOM character.
4. Conclusions
For 23 compounds, BVs to initial breakthough showed good
correlation with a relationship including GAC total pore
volume (TPV), GAC mini-column apparent density (rmc), GAC
slurry pH (pHs), volume of compound adsorbed (CV� Co), the
compound’s 8th-order simple Chi index (8cp), and the com-
pound’s hydrophobic surface area (FOSA).
Clearly, as a first attempt at devising a QSAR-like model for
characterizing ppt-level sorption in activated carbon columns
under real-world NOM conditions, the model will be limited in
its breadth of application. For example, the authors herein
employed three lignite-based activated carbons and these all
had similar proportions of micropores-to-total pores (where
total pores were micropores plus mesopores). Also, the molar
concentrations of the compounds were all similar; and the
model may not well-characterize behavior outside this
concentration range. Moreover, the conforming EDCs/PPCPs
hosted somewhat similar substituent groups. Nonetheless,
this QSAR-like model offers an important first step in pre-
dicting EDC/PPCP breakthrough behavior that other scientists,
engineers, and water quality specialists can build on, as yet
more analytical data are collected.
Through the process of developing this QSAR-like model,
the authors identified more than 40 molecular descriptors that
lacked statistical significance. For example, the octanol–water
partition coefficient (Kow) was not significant, even though
log Kow has been deemed important when modeling ppm-
level sorption.
As a practical conclusion regarding the overall bed life for
removing these EDCs/PPCPs, the modified lignite carbons
(STEAM and M/S) offered bed lives of a half a year to more than
a year for many of these compounds, when they were present
at intentionally high ppt levels (i.e. higher than they often
appear in native water supplies). The conventional activated
carbon would offer a shorter bed life in comparison. This is the
case when the activated carbon performed with a 7.6 min
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 13860
empty bed contact time, where compound removal was
simulated by constant diffusivity.
Acknowledgements
The authors thank Norit Americas for providing the
commercial carbon sample. This work was funded through
the National Science Foundation (grant #0202177), the
Consortium for Premium Carbon Products from Coal (CPCPC),
the City of Cincinnati, OH, USA and the American Water
Works Research Foundation (AwwaRF).
Appendix
Supplementary data associated with this article can be found
in the online version, at 10.1016/j.watres.2009.05.026.
r e f e r e n c e s
Abe, I., Kamaya, H., Ueda, I., 1988. Activated carbon as a biologicalmodel: comparison between activated carbon adsorption andoil–water partition coefficient for drug activity correlation.Journal of Pharmaceutical Sciences 77 (2), 166–168.
Belfort, G., Altshuler, G.L., Thallam, K.K., Feerick Jr., C.P.,Woodfield, K.L., 1984. Selective adsorption of organichomologues onto activated carbon from dilute aqueoussolutions: solvophobic interaction approach: part IV. Effect ofsimple structural modifications with aliphatics. AIChE Journal30 (2), 197–207.
Blum, D.J.W., Suffet, I.H., Duguet, J.P., 1994. Quantitativestructure–activity relationship using molecular connectivityfor the activated carbon adsorption of organic chemicals inwater. Water Research 28 (2), 687–699.
Brasquet, C., Subrenat, E., LeCloirec, P., 1997. Selective adsorptionon fibrous activated carbon of organics from aqueous solution:correlation between adsorption and molecular structure.Water Science and Technology 35 (7), 251–259.
Brennan, R.A., Nirmalakhandan, N., Speece, R.E., 1998.Comparison of predictive methods for Henry’s law coefficientsof organic chemicals. Water Research 32 (6), 1901–1911.
Crittenden, J.C., Berrigan, J.K., Hand, D.W., 1986. Design of rapidsmall-scale adsorption tests for a constant diffusivity. Journalof the Water Pollution Control Federation 58 (4), 312–319.
Crittenden, J.C., Reddy, P.S., Arora, H., Trynoski, J., Hand, D.W.,Perram, D.L., Summers, R.S., 1991. Predicting GACperformance with rapid small-scale column tests. Journal ofthe American Water Works Association 83 (1), 77–87.
Dubinin, M.M., Stoeckli, H.F., 1980. Homogeneous andheterogeneous micropore structures in carbonaceousadsorbents. Journal of Colloid and Interface Science 75 (1),34–42.
Gramatica, P., 2007. Principles of QSAR models validation:internal and external. QSAR and Combinatorial Science 26 (5),694–701.
Hu, J.Y., Aizawa, T., Ookubo, Y., Morita, T., Magara, Y., 1998.Adsorptive characteristics of ionogenic aromatic pesticides inwater on powdered activated carbon. Water Research 32 (9),2593–2600.
Kier, L.B., Hall, L.H., 1986. Molecular Connectivity in Structure–Activity Analysis. Research Studies Press, Letchworth,Hertfordshire, England.
Kier, L.B., 1986. Shape indexes of orders one and three frommolecular graphs. Quantitative Structure–ActivityRelationships 5 (1), 1–7.
Kilduff, J.E., Karanfil, T., Chin, Y.P., Weber, W.J., 1996. Adsorptionof natural organic polyelectrolytes by activated carbon: a size-exclusion chromatography study. Environmental Science andTechnology 30 (4), 1336–1343.
Kilduff, J.E., Karanfil, T., Weber, W.J., 1996. Competitiveinteractions among components of humic acids in granularactivated carbon adsorption systems: effects of solutionchemistry. Environmental Science and Technology 30 (4),1344–1351.
Leng, C.C., Pinto, N.G., 1997. Effects of surface properties ofactivated carbons on adsorption behavior of selectedaromatics. Carbon 35 (9), 1375–1385.
Manes, M., Hofer, J.E., 1969. Application of the polanyi adsorptionpotential theory to adsorption from solution on activatedcarbon. Journal of Physical Chemistry 73 (3), 584–590.
McDonough, K.M., Fairey, J.L., Lowry, G.V., 2008. Adsorption ofpolychlorinated biphenyls to activated carbon: equilibriumisotherms and a preliminary assessment of the effect ofdissolved organic matter and biofilm loadings. Water Research42 (3), 575–584.
Moore, B.C., Cannon, F.S., Metz, D.H., DeMarco, J., 2003. GAC porestructure in Cincinnati – during full-scale treatment/reactivation. Journal of the American Water WorksAssociation 95 (2), 103–112.
Moreno-Castilla, C., 2004. Adsorption of organic molecules fromaqueous solutions on carbon materials. Carbon 42 (1), 83–94.
Newcombe, G., Morrison, J., Hepplewhite, C., Knappe, D.R.U.,2002. Simultaneous adsorption of MIB and NOM ontoactivated carbon – II. Competitive effects. Carbon 40 (12),2147–2156.
Nowack, K.O., Cannon, F.S., Mazyck, D.W., 2004. Enhancingactivated carbon adsorption of 2-methylisoborneol: methaneand steam treatments. Environmental Science andTechnology 38 (1), 276–284.
Pan, X.L., Tan, N.H., Zeng, G.Z., Han, H.J., Huang, H.Q., 2006. 3D-QSAR and docking studies of aldehyde inhibitors of humancathepsin K. Bioorganic and Medicinal Chemistry 14 (8),2771–2778.
Picard, R.R., Cook, R.D., 1984. Cross-validation of regressionmodels. Journal of the American Statistical Association 79(387), 575–583.
Qi, S.Y., Hay, K.J., Rood, M.J., Cal, M.P., 2000. Carbon fiberadsorption using quantitative structure–activity relationship.Journal of Environmental Engineering – ASCE 126 (9), 865–868.
Radovic, L.R., Silva, I.F., Ume, J.I., Menendez, J.A., Leon, C.A.L.Y.,Scaroni, A.W., 1997. An experimental and theoretical study ofthe adsorption of aromatics possessing electron-withdrawingand electron-donating functional groups by chemicallymodified activated carbons. Carbon 35 (9), 1339–1348.
Rangel-Mendez, J.R., Cannon, F.S., 2005. Improved activatedcarbon by thermal treatment in methane and steam:physicochemical influences on MIB sorption capacity. Carbon43 (3), 467–479.
Snyder, S.A., Keith, T.L., Verbrugge, D.A., Snyder, E.M., Gross, T.S.,Kannan, K., Giesy, J.P., 1999. Analytical methods for detectionof selected estrogenic compounds in aqueous mixtures.Environmental Science and Technology 33 (16), 2814–2820.
Snyder, S.A., Snyder, E., Villeneuve, D., Kurunthachalam, K.,Villalobos, A., Blankenship, A., Giesy, J., 2000. Instrumentaland bioanalytical measures of endocrine disruptors in water.In: Analysis of Environmental Endocrine Disruptors, vol. 747,pp. 73–95.
w a t e r r e s e a r c h 4 3 ( 2 0 0 9 ) 3 8 4 9 – 3 8 6 1 3861
Snyder, S.A., Kelly, K.L., Grange, A.H., Sovocool, G.W., Snyder, E.M.,Giesy, J.P., 2001. Pharmaceuticals and Personal Care Products inthe Environment: Scientific and Regulatory Issues. In:Daughton, C.G., Jones-Lepp, T.L. (Eds.). American ChemicalSociety, Washington, D.C., pp. 116–140.
Snyder, S.A., Villeneuve, D.L., Snyder, E.M., Giesy, J.P., 2001.Identification and quantification of estrogen receptor agonistsin wastewater effluents. Environmental Science andTechnology 35 (18), 3620–3625.
Snyder, S.A., Westerhoff, P., Yoon, Y., Sedlak, D.L., 2003.Pharmaceuticals, personal care products, and endocrinedisruptors in water: implications for the water industry.Environmental Engineering Science 20 (5), 449–469.
Snyder, S.A., Leising, J., Westerhoff, P., Yoon, Y., Mash, H.,Vanderford, B., 2004. Biological and physical attenuation ofendocrine disruptors and pharmaceuticals: implications forwater reuse. Ground Water Monitoring and Remediation 24(2), 108–118.
Snyder, S.A., Adham, S., Redding, A.M., Cannon, F.S.,DeCarolis, J., Oppenheimer, J., Wert, E.C., Yoon, Y., 2007. Roleof membranes and activated carbon in the removal ofendocrine disruptors and pharmaceuticals. Desalination 202(1–3), 156–181.
Snyder, S.A., 2003. Endocrine disruptors as water contaminants:toxicological implications for humans and wildlife. SouthwestHydrology 2 (6).
Summers, R.S., Roberts, P.V., 1988. Activated carbon adsorption ofhumic substances. 2. Size exclusion and electrostaticinteractions. Journal of Colloid and Interface Science 122 (2),382–397.
Trenholm, R.A., Vanderford, B.J., Holady, J.C., Rexing, D.J.,Snyder, S.A., 2006. Broad range analysis of endocrine
disruptors and pharmaceuticals using gas chromatographyand liquid chromatography tandem mass spectrometry.Chemosphere 65 (11), 1990–1998.
van Aerle, R., Nolan, M., Jobling, S., Christiansen, L.B., Sumpter, J.P.,Tyler, C.R., 2001. Sexual disruption in a second species of wildcyprinid fish in United Kingdom freshwaters. EnvironmentalToxicology and Chemistry 20 (12), 2841–2847.
Vanderford, B.J., Pearson, R.A., Rexing, D.J., Snyder, S.A., 2003.Analysis of endocrine disruptors, pharmaceuticals, andpersonal care products in water using liquid chromatography/tandem mass spectrometry. Analytical Chemistry 75 (22),6265–6274.
Walters, R.W., Luthy, R.G., 1984. Equilibrium adsorption ofpolycyclic aromatic-hydrocarbons from water ontoactivated carbon. Environmental Science and Technology 18(6), 395–403.
Ward, A.F.H., 1946. Thermodynamics of monolayers on solutions.1. The theoretical significance of Traube’s rule. Transactionsof the Faraday Society 42 (5), 399–407.
Westerhoff, P., Yoon, Y., Snyder, S., Wert, E., 2005. Fate ofendocrine-disruptor, pharmaceutical, and personal careproduct chemicals during simulated drinking water treatmentprocesses. Environmental Science and Technology 39 (17),6649–6663.
Yoon, Y.M., Westerhoff, P., Snyder, S.A., Esparza, M., 2003. HPLC-fluorescence detection and adsorption of bisphenol A, 17 beta-estradiol, and 17 alpha-ethynyl estradiol on powderedactivated carbon. Water Research 37 (14), 3530–3537.
Yu, Z., Peldszus, S., Huck, P.M., 2008. Adsorption characteristics ofselected pharmaceuticals and an endocrine disruptingcompound-Naproxen, carbamazepine and nonylphenol-onactivated carbon. Water Research 42 (12), 2873–2882.