Journal of Chromatography A, 1396 (2015) 34–44

Contents lists available at ScienceDirect

Journal of Chromatography A
journal homepage: www.elsevier.com/locate/chroma

Artificial neural network modelling of pharmaceutical residue

retention times in wastewater extracts using gradient liquid
chromatography-high resolution mass spectrometry data
Kelly Munro a , Thomas H. Miller a , Claudia P.B. Martins b , Anthony M. Edge c ,
David A. Cowan a , Leon P. Barron a,∗
a
Analytical & Environmental Sciences Division, King’s College London, 150 Stamford Street, SE1 9NH London, United Kingdom
b
Thermo Fisher Scientific, 355 River Oaks Parkway, San Jose, CA 95134, USA
c
Thermo Fisher Scientific, Tudor Road, Runcorn, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: The modelling and prediction of reversed-phase chromatographic retention time (tR ) under gradient
Received 10 November 2014 elution conditions for 166 pharmaceuticals in wastewater extracts is presented using artificial neural
Received in revised form 27 February 2015 networks for the first time. Radial basis function, multilayer perceptron and generalised regression neu-
Accepted 23 March 2015
ral networks were investigated and a comparison of their predictive ability for model solutions discussed.
Available online 30 March 2015
For real world application, the effect of matrix complexity on tR measurements is presented. Measured
tR for some compounds in influent wastewater varied by >1 min in comparison to tR in model solutions.
Keywords:
Similarly, matrix impact on artificial neural network predictive ability was addressed towards develop-
Artificial neural networks
Retention time prediction
ing a more robust approach for routine screening applications. Overall, the best neural network had a
Semi-targeted detection predictive accuracy of <1.3 min at the 75th percentile of all measured tR data in wastewater samples
Reversed-phase liquid chromatography (<10% of the total runtime). Coefficients of determination for 30 blind test compounds in wastewater
High resolution mass spectrometry matrices lay at or above R2 = 0.92. Finally, the model was evaluated for application to the semi-targeted
identification of pharmaceutical residues during a weeklong wastewater sampling campaign. The model
successfully identified native compounds at a rate of 83 ± 4% and 73 ± 5% in influent and effluent extracts,
respectively. The use of an HRMS database and the optimised ANN model was also applied to shortlist-
ing of 37 additional compounds in wastewater. Ultimately, this research will potentially enable faster
identification of emerging contaminants in the environment through more efficient post-acquisition data
mining.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction of PPCPs have used online [6] or offline [7,8] solid phase extraction
(SPE) [9] for clean-up and concentration of wastewater samples
Characterising the breadth of occurrence of pharmaceuticals followed by analysis with gas chromatography (GC) [10,11] or
and personal care products (PPCPs) in the aquatic environment liquid chromatography (LC) coupled to various modes of mass spec-
has been on-going for the past two decades [1]. In most developed trometry (MS) [8,12–15]. However, as >3000 licensed PPCPs exist,
countries, wastewater treatment plant effluents often act as identi- targeted approaches only allow characterisation of a small portion
fiable sources, and, given that they often serve large communities, of PPCP occurrence. Methods for such large numbers of compounds
the potential exists for larger numbers of PPCPs to enter receiv- are often limited by unfeasibly long cycle times for tandem MS
ing waters via effluent as a result. Several multi-residue analytical experiments or full m/z range scanning using lower resolution
methods have been applied to PPCP determination in environmen- instruments.
tal matrices [2–5] and mostly employ a targeted approach where Relatively recent efforts have focused on hyphenation to high
compounds are pre-selected. Recent methods for larger numbers resolution mass spectrometry (HRMS) for multi-residue screen-
ing [16–19]. The combination of full-scan m/z data acquisition at
or below ppm m/z accuracy and at high resolution has enabled
∗ Corresponding author. Tel.: +44 20 7848 3842; fax: +44 20 7848 4980. datasets to be interpreted in a non-targeted or semi-targeted man-
E-mail address: [email protected] (L.P. Barron). ner post-analysis [20–22]. By using full-scan HRMS instruments for

http://dx.doi.org/10.1016/j.chroma.2015.03.063
0021-9673/© 2015 Elsevier B.V. All rights reserved.
 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44 35

detection, larger numbers of compounds can be screened in a sin- compounds used for retention time modelling are listed in the
gle run and such methods have been applied to >400 compounds supplementary information (SI).
in wastewater [23]. Full scan and tandem MS/all-ion fragmenta-
tion experiments may also yield useful information to identify 2.2. Analytical instrumentation and conditions
new compounds, but a number of chromatographic peaks and/or
potential candidates with matching m/z may still exist before con- Unless otherwise stated, all instrumentation, equipment and
firmation becomes feasible. For this reason, there still exists a associated software was procured from Thermo Fisher Scientific
requirement to confirm compound identity using analyte retention (San Jose, CA, USA). An Accela ultra high performance liquid chro-
time from the chromatographic separation. Confirmation efforts matography system was used for all work presented herein. Sample
are further limited by the availability of PPCP reference materials, injection was performed using a HTS-A5 autosampler with sample
but also their metabolites and transformation products. Impor- storage compartment temperature control set at 10 ◦ C. Separa-
tantly, this may also require consideration of other isomeric and/or tions were performed on a 150 × 2.1 mm, 2.6 ␮m solid-core particle
isobaric compounds. Accucore C18 analytical column which was configured with a
The use of in silico retention time modelling tools may prove matching 10 × 2.1 mm, 2.6 ␮m particle size Accucore C18 guard.
useful in post-data acquisition analysis, at the very least, to Temperature was maintained at 24 ◦ C. The flow rate of mobile phase
help direct synthesis/acquisition of analytical reference materials was 0.4 mL/min and the injection volume was 20 ␮L in order to
for more rapid compound confirmation. Linear solvation-energy achieve acceptable sensitivity in wastewater extracts. A binary gra-
relationships (LSERs) have been used for modelling solute reten- dient elution profile of 90:10 to 20:80 10 mM ammonium acetate
tion behaviour in chromatography for some time [24–26]. As a in water:acetonitrile (mobile phase A and B, respectively; apparent
more flexible and practical alternative, researchers have also used pH = 5.2) was used over 27.5 min as follows: 0% B for 2.5 min; a lin-
artificial neural networks (ANNs) to model retention in several ear ramp to 30% B from 2.5 to 7.5 min; an isocratic stage of 30% B
chromatographic modes [27–29] and formats [30–32]; for method from 7.5 to 12.5 min; a linear ramp to 40% B from 12.5 to 15 min; a
optimisation purposes [33–37]; or to estimate unknown compound linear ramp to 100% B from 15.0 to 20.0 min; and then maintained
retention [38–40]. Fragkaki and colleagues recently compared the at 100% B from 20.0 to 27.5 min. Re-equilibration time was 7.5 min.
use of linear regression, partial least squares and ANNs for the Detection was performed using an Exactive HRMS instrument
prediction of relative retention time of 64 derivatised steroid com- fitted with a heated electrospray ionisation source (HESI-II). Sepa-
pounds in gradient GC-MS [41]. Recently, we developed and applied rate chromatographic runs of all samples and model solutions were
ANNs for the generation of sports doping compound retention time performed in either positive or negative ionisation mode. Nitrogen
in urine extracts, which were collected by LC-HRMS during the was used within the HESI-II source and the collision cell. Enhanced
London 2012 Olympic Games [42]. Using a short 8-min reversed- resolution mode was employed at 50,000 FWHM. Two MS experi-
phase gradient for 86 compounds, all predicted retention times (tRP ) ments occurred during each acquisition cycle by performing a full
were within 1.0 min of measured retention time (tRM ) and 12/15 scan without higher energy collisional dissociation (HCD) followed
blind test compounds were predicted to within 0.5 min of the true by a full scan with HCD enabled (collision energy: 20 eV). The cycle
value. However, the number of compounds used may be limited time was approximately 2 s, resulting in 1.2 scans per second due
for use as the basis for broad-scope application to in silico wastew- to the switching between full scan and HCD fragmentation exper-
ater analysis given that a potentially larger diversity and number iments. The automatic gain control target was set to balanced,
of compounds are likely to exist simultaneously in samples of such meaning that ∼106 ions were collected in the C-trap before being
high complexity. sent to the Orbitrap for acquisition. The sheath gas, auxiliary gas
The aim of this work was therefore to evaluate and confirm ANN and sweep gas flow rate settings were 50, 10 and 0 arbitrary units,
modelling as a useful tool for semi-targeted screening of residues in respectively. The capillary temperature was 350 ◦ C; the heater tem-
influent and effluent wastewater extracts. In particular, the objec- perature was 300 ◦ C; and the positive/negative spray voltages were
tives were to model a larger number of compound retention times +4.50 kV and −3.00 kV. The scan range was m/z 100–1000 for all
than included previously; to assess the effect of matrix on the relia- MS experiments. All acquisition data was processed using Xcalibur
bility of tRM data for ANN training; to assess ANN tRP performance and v 2.0 software. The HRMS instrument was calibrated prior to use
to apply ANNs to the identification of trace PPCPs in real samples by infusion of a mixture of caffeine, MRFA (met-arg-phe-ala) and
during a week-long wastewater sampling campaign. This repre- UltraMark 1621® (a mixture of fluorinated phosphazines) prepared
sents the first study of its kind to use ANN modelling tools with in a solvent of acetonitrile/methanol/acetic acid for positive mode
LC-HRMS datasets for semi-targeted PPCP screening purposes. ESI-HRMS. For negative ESI-HRMS mode calibration infusion of a
solution containing sodium dodecyl sulfate (SDS), sodium tauro-
cholate and Ultramark 1621® was performed in the same solvent.
2. Experimental
2.3. Sample preparation procedures
2.1. Reagents
Samples of wastewater influent (taken immediately after
All reagents were of analytical grade unless otherwise spec- the fine screen) and effluent were obtained from a major
ified. HPLC grade methanol, acetonitrile, dichloromethane and sewage treatment works in London (population equivalent = 3.4
dimethyldichlorosiloxane were purchased from Fisher Scien- million). Wastewater was collected as seven consecutive, sep-
tific (Loughborough, UK). Ammonium acetate and 37% (w/v) arate 24-h composite samples across a 1-week period from
hydrochloric acid solution were sourced from Sigma-Aldrich 11 to 17 March 2014 (using a set 60-min sampling interval).
(Gillingham, Dorset, UK). Ultrapure water was obtained from a Samples were refrigerated and transported to the labora-
Millipore Synergy-UV water purification system with a resistivity tory in Nalgene bottles and adjusted to pH 2 using 37% (w/v)
of 18.2 M cm (Millipore, Bedford, USA). Analytical grade reference hydrochloric acid solution and stored frozen at −20 ◦ C until
compounds were purchased from several different manufacturers analysis. Glassware was silanised to reduce analyte adsorp-
or donated by other laboratories. Stock standards were prepared tion to glass surfaces. This included a rinse with 50:50 (v/v)
at concentrations of 1 mg/mL in methanol and stored in silanised methanol:water before triplicate rinses with dichloromethane.
amber vials at 4 ◦ C in the dark when not in use. All 166 reference A 10:90 (v/v) dimethyldichlorosilane/dichloromethane
36 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44

solution was then used to silanise the container, followed by networks were again ranked by correlation and output error. The
triplicate sets of rinses with each of dichloromethane and 50:50 maximum number of nodes tested in RBF hidden layers was 42
methanol:water. Before extraction, samples were thawed and fil- (software suggested value). Pruning was only applied to GRNNs
tered under vacuum using Whatman GF/F 0.7 ␮m glass microfibre and fan-out weight thresholds for input variables and hidden layer
filters. For quantification of mephedrone in influents, a pooled units were set to ≤0.05.
wastewater sample was used to generate a background-subtracted
matrix-matched calibration curve over 5–2500 ng/L. In addition to 2.5. Post-acquisition data mining and peak selection
this, n = 3 quality control 250 ng/L matrix-matched standards were
determined to ensure method inaccuracy remained below 20%. Post-acquisition automated peak selection was performed for
HyperSep Retain Polar Enhanced Polymer (PEP) SPE cartridges influent and effluent sample data using TraceFinderTM version 3.1
(200 mg, 6 mL barrel) were used for extraction of influent and software (Thermo); a separate peak selection process was per-
effluent wastewater samples. Cartridges were conditioned with formed for both the positive and negative ionisation mode runs
4 mL methanol followed by 4 mL ultrapure water. A 100 mL sample for each sample. Acquired HRMS data only was screened against
was loaded onto the cartridge followed by a wash step of 4 mL the commercially available ThermoFisher spectral library which
5:95 (v/v) methanol:water and dried under vacuum for 10 min. contained data for 1492 compounds comprising pesticides, her-
Bound solutes were subsequently eluted with 4 mL methanol into bicides, fungicides, pharmaceuticals, metabolites and illicit drugs
a pre-silanised tapered glass tube. Eluted extracts were evaporated (and inclusive of all 166 compounds selected for retention mod-
to dryness under N2 at 35 ◦ C and reconstituted in 100 ␮L of mobile elling herein). Neither empirical retention time nor peak width
phase A. For retention modelling purposes only, and to generate were used for database searching to simulate a semi-targeted
matrix-matched standards, dried extracts were reconstituted approach. Identification criteria were applied within the software
in 100 ␮L of a 250 ␮g/L standard of all compounds prepared in for automatic peak selection. With the exception of precursor ion
mobile phase A. The reconstituted samples were sonicated for peak area, the same identification criteria were applied in both the
10 min before being transferred to pre-silanised and crimped positive and negative ESI-HRMS screening methods. A mass toler-
septum-capped amber glass vials. Extracts were stored at 4 ◦ C and ance of 5 ppm was applied (i.e. precursors, fragments and isotope
in the dark and analysed within 24 h. ions). Where confirmation was required, and in addition to the
accurate mass of the precursor ion, the most abundant matching
2.4. Molecular description, ANN models and architectures fragment ion (with the HCD cell enabled) was used. Initial inten-
sity thresholds for automated screening of precursor ions were set
In line with previously published works [42,43], canonical sim- at 50:1 signal-noise ratio (S/N) and peak area ≥50,000 AU/peak
plified molecular line entry system strings (SMILES) were created for positively ionised precursor ions and ≥10,000 AU/peak for neg-
for all 166 compounds using ChemSpider freeware (Royal Soci- atively ionised. Lastly, an 80% fit threshold to theoretical isotope
ety of Chemistry, UK). Using Parameter Client freeware (Virtual profile was set, with an acceptable intensity threshold deviation
Computational Chemistry Laboratory, Munich, Germany), fifteen for each isotope ion set at 25% of the theoretical value. The isotopic
molecular descriptors were generated as inputs to ANN models fit threshold represents the deviation allowed between the mea-
which included the number of double and triple bonds (nDB or sured and theoretical isotopic patterns, based on the performed
nTB), the number of carbon and oxygen atoms (nC or nO), the pattern matching. The allowed intensity deviation considers the
number of 4–9 membered rings (nR04–nR09), unsaturation index intensity of the isotope ion relative to the monoisotopic ion. Follow-
(UI), hydrophilic factor [2], number of benzene rings (nBnz), and ing the automated screen of wastewater data, manual inspection
Moriguchi and Ghose-Crippen log P (MlogP and AlogP, respec- of all shortlisted peaks was performed to identify viable chro-
tively). Data for two additional molecular descriptors (log D and matographic peaks (see Section 3.4) for subsequent tR matching
pKa ) were generated using Percepta PhysChem Profiler (ACD Lab- with ANN and then confirmation with its reference standard tRM
oratories, ON, Canada). in the relevant matrix. Further criteria for acceptance included
Prediction of retention times was performed using the Trajan 6.0 peak profile and definition (Gaussian-like profile with n ≥ 10 dat-
neural network simulator (Trajan Software Ltd., Lincolnshire, UK) apoints/peak). A second manual check of the S/N ratio (≥3:1) was
and compared with measured retention time in matrix matched then applied to exclude erroneous peak identifications in the first
standards. Optimisation of each ANN type and architecture was automated screen.
performed in a number of stages. Firstly, all 17 input variables
listed were included in the design phase followed by a second stage 3. Results and discussion
which enabled subset selections to assess any improved correlation
or performance-limiting input variables. In these first two steps, 3.1. Modelling retention time in standard solutions using ANNs
case input variable data were randomised across training, verifi-
cation and blind test experiments. When a suitable architecture Initial assessment of the comparative performance of ANN tR
was identified, a third stage of optimisation involved randomising predictions was performed without including the influence of
all but the blind test cases to identify the best model. For appli- matrix. The choice of molecular descriptors was based on our pre-
cation, all case datasets were subsequently fixed. Networks were vious work which showed that these input variables contributed
set to balance their verification set errors against network diver- significantly to retention time modelling under reversed-phase
sity to cover as many architectures as possible across all model conditions. It was found that, for all ANN types, a 120:16:30 split
types. For multilayer perceptrons (MLPs) in particular, and for of cases across training, verification and blind test experiments
each set of design parameters, ANN architectures were tested over yielded the best results for consistent comparison purposes. A com-
30 min intervals and following this, the best 50 networks were parison between three network types and architectures is shown in
selected based on correlation and output error. Three- and four- Fig. 1. A MLP network consists of a number of nodes housed within
layer MLPs were included and within each hidden layer the number discrete layers. Input nodes are fully interconnected to these layers
of nodes tested was ≤14 (software suggested value). For radial basis via non-linear activation functions. Inputs are mapped onto a set
functions (RBFs), probabilistic neural networks (PNNs) and gener- of required output values (tRM ) and supervised learning is iterative
alised regression neural networks (GRNNs), 108 architectures were and based on back-propagation. RBFs employ real-value functions
investigated in each optimisation step. Fifty of the best RBF/GRNN which depend only on proximity from the origin or a single discrete
 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44 37

Fig. 1. The best correlation of predicted versus measured retention times achieved for MLPs, RBFs and GRNNs (left) and associated predicted retention time errors for each
(on the right).
38 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44

point. Lastly, GRNN-type ANNs combine an RBF layer with a linear

activation function layer [9]. This network yielded the best perfor-
mance overall (Fig. 1) and, along with RBFs, required a fraction of the
time to optimise, train, verify and test (<10 s for 108 models) in com-
parison to MLPs. GRNNs also displayed excellent consistency across
different architectures and case allocations. The optimised GRNN
displayed correlation data of y = 0.8684x + 1.3313 (R2 = 0.9327) for
the training set; y = 0.7907x + 1.8897 (R2 = 0.8738) for the verifica-
tion set; and y = 0.8075x + 1.8936 (R2 = 0.8858) for blind test cases.
In particular, the average absolute tRP error and standard deviation
for the blind test dataset was 1.39 ± 1.29 min. It was clear that ANN
tR accuracy thresholds were required by considering a portion of
the case data which lay within a reasonable percentile of tRP error
and accepting that some false negatives would be encountered as a
result. Moreover, as the GRNN-type network offered markedly bet-
ter results than either the RBF or MLP-type networks at this stage
in the process, it was not likely, and indeed found, that this would
change with inclusion of a matrix effect.

3.2. Effect of wastewater matrix on measured retention time

Most retention modelling works have similarly focussed on

data derived from standards prepared in purified solvents [30,44].
Certainly for targeted methods, it is reasonable to expect that
an optimised extraction procedure would eliminate or suitably
minimise matrix effects. However, the value of semi-targeted
data screening lies in characterisation of extra matrix compo-
nents, which may later become analytes. Therefore, co-extraction
of a matrix component could be considered advantageous in
some respects. The SPE sorbent used here consisted of a
hydrophilic–lipophilic balanced (HLB) type chemistry which has
been widely adopted along with a similar protocol for this appli-
cation [45,46]. However, a disparity exists in SPE concentration
factors across the literature, which range from 100 to 2500-fold
to achieve acceptable sensitivity and mostly for target analytes
[47–53]. Several studies also used rapid gradient separations prior Fig. 2. The effect of matrix on (a) relative retention time in wastewater to a standard
to mass spectrometry and reported significant MS signal sup- in ultra-pure water and (b) tRM standard deviation for standard solutions and spiked
wastewater samples taken across a 1-week period.
pression (often >50%) showing that matrix exists at high enough
concentrations, even after application of SPE clean-up, to inter-
fere with validated analytical methods [47,48,50]. Matrix effects on consecutively. Past this initial 7-min elution window, the differ-
measured retention time are rarely reported, but can be overcome ences between influent and effluent matrix effects on tRM were less
by using matrix-matched standards instead. Herein, an intermedi- obvious. Thirdly, it was observed that inter-day tRM measurement
ate concentration factor of 1000-fold followed by a moderately long variance on the whole was again largest for influent, followed by
separation of 27.5 min was used in an attempt to balance this matrix treated effluent and then model solutions over the same 7-min elu-
effect with adequate sensitivity for broad, semi-targeted qualitative tion window (Fig. 2(b)). However, this variance was still well within
screening at the ng/L concentration level. acceptable method reproducibility limits (average tRM ± 15 s [54]).
Seven individual 24-h composite samples of both influent and Examination of total ion chromatograms revealed a broad matrix
effluent wastewaters across a week were used to prepare matrix- elution profile over this timeframe (maximum intensity = 4 × 108 ),
matched standards in order to compile average tRM data in each which may have accounted for retention shifts and inter-sample
matrix type separately. This was preferred over replicate data from variance. Even after SPE, influent and effluent wastewater extracts
a single composite sample of any one wastewater type (or combi- were coloured showing that such a matrix effect with such a widely
nation thereof) to show how daily composition may affect tRM . As adopted methodology was still possible. Most tRM variance after this
can be observed in Fig. 2(a), a mean tRM shift for all compounds in initial 7-min window for all sample types was <1% relative standard
wastewater extracts was evident in comparison to model solutions. deviation (RSD). The largest standard deviation of mean retention
As a result, several important considerations for tR predictions time was 10 s for MDMA in influent wastewater (tRM = 5.12 min).
became apparent. Firstly, under these conditions, the majority of The mean of the measured tRM standard deviations for all com-
compounds displayed an overall increase in tRM irrespective of sam- pounds was 2.5 ± 1.9 s and 2.1 ± 1.2 s for influent and effluent,
ple type meaning that use of non-matrix matched standards was respectively. The range of peak widths (at baseline) measured for
generally considered inappropriate for comparison with wastew- all compounds lay within the range of ∼0.2–0.5 min.
ater data for confirmation purposes. Secondly, compounds eluting
in the first 7 min of the gradient separation suffered the largest 3.3. Performance of GRNNs for spiked wastewater extracts
change in tRM in comparison to model solutions. More pronounced
matrix effects over this window were observed for compounds Direct replacement of tRM data in the GRNN for those measured in
in influent wastewater in particular with examples of the worst either matrix resulted in a poorer correlation overall. In an attempt
retention shifts being +0.99, +1.00 and +1.03 min for pindolol, to understand the robustness of GRNN networks, n = 6 identical
phenmetrazine and MDMA, respectively, which all eluted model architectures were created for each sample type and trained
 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44 39

Fig. 3. Predicted versus measured retention times in (a) influent and (b) effluent wastewaters using the optimised 17-120-16-30 GRNN. The predicted retention time error
for each compound across the gradient run-time is shown as (c) for influent and as (d) for effluent.

ab initio using the mean of tRM for each compound in each matrix.
The predictive ability of the GRNN model recovered and corre-
lation coefficients even slightly improved. For the best network
applied to influent wastewater (Fig. 3(a) and (c)), the average and
standard deviation of absolute tRP error for the entire 166 cases
and the 30 blind test cases in isolation were 0.98 ± 1.16 min and
0.79 ± 1.26 min, respectively. All but one of the absolute blind test
case tRP errors lay within a window of 1.81 min. Mean and standard
deviation of tRP errors for the full dataset and blind test cases for
effluent (Fig. 3(b) and (d)) were 0.99 ± 1.17 min and 1.06 ± 1.15 min,
respectively and three of the absolute blind test case tRP errors lay
outside 1.81 min by comparison.
Examination of replicate GRNN models showed that each input
variable was used to similar degrees across matrices (Fig. 4). Error
ratios >1 represented a worsening of network performance with
removal of that descriptor. The top five molecular descriptors for
tR prediction in both wastewater matrix types were nDB, logD,
nO, pKa and nBnz in descending order. The consistency of input
variable dependency across all 18 networks tested showed that
this GRNN model could potentially be used robustly in bracketed-
type analyses where matrix-matched standards could be analysed
before and after each batch of wastewater samples (a) for rela- Fig. 4. Replicate network sensitivity to removal of individual molecular descrip-
tor input variables. The error ratio represents the additional error in tRP where that
tive comparison purposes to rule out within-batch column fouling
variable was omitted from the network to that of tRP error where all network vari-
events and (b) for better application to retention modelling by ables were included. Error bars represent standard deviation of the results from n = 6
accounting for the operational conditions at the time of analy- replicate networks for each of standards in ultrapure water and spiked wastewater
sis. extracts.
40 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44

mode). Despite the application of filters (as in Section 2.5), some

shortlisted peaks were not deemed satisfactory and were easily
excluded (e.g. they were not discrete/obvious chromatographic
peaks, or were integrated regions of noise). Manual inspection
reduced this number to ∼700 viable peaks in both influent and
effluent positive ionisation data, with a lower number observed
for negative ionisation (∼500 influent/∼300 effluent). As multiple
peaks often existed for each suspect compound, this correlated to
∼350 positively ionised compounds in both influent and effluent,
again with a lower number observed for negative ionisation (∼250
influent/∼200 effluent).
In the first instance, data for the 166 target compounds were
extracted for confirmation and ANN tRP evaluation. Retention time
confirmation was based on US EPA Method 1694 guidelines for
PPCP identification whereby chromatographic peaks must elute
within ±15 s of the matrix-matched standard tRM [54]. Here, this
was taken as the average of all seven spiked influent or effluent
retention times recorded across the week (inter-day tRM ). For ANN
tRP matching, the window was set as before at ±1.30 min. Out of
the full 166 compounds used in this work, post-acquisition LC-
HRMS data-mining yielded matching m/z and isotope profiles for
between 70 and 82 compounds per day in wastewater (positive
mode ESI-HRMS), with a lower number (19–67) observed using
negative mode ESI-HRMS (Table 1 and SI). Individual compound
occurrence across all samples, as well as comparison of ANN sus-
pect shortlisting and matrix-matched standard confirmation are
shown in the SI. On average, similar numbers of positively ionised
compounds were confirmed using a matrix-matched standard in
Fig. 5. Plots of absolute tRP error versus percentile of each training, verification and
influent and effluent wastewaters (30–36 compounds/day). When
blind test dataset for (a) influent wastewater and (b) effluent wastewater.
comparing to tRP data, this corresponded to an overall mean % ANN
success rate of 83 ± 4% and 73 ± 5% for PPCPs at trace concentra-
As can be seen in Fig. 5, plots of absolute tRP error in either matrix tions in influent and effluent samples, respectively (p = 0.007). The
relative to the percentile of data from each of the training, verifica- ANN correctly excluded the presence of between 32 and 44% of
tion and blind test datasets revealed a moderate increase in error up TraceFinder-shortlisted compounds in wastewater each day.
to ∼75–85% of the data population where it subsequently increased Mephedrone represented an interesting example of an illicit
markedly due to the contribution of outliers. Therefore, taking the compound which occurred in both influent and effluent wastew-
75th percentile of all datasets across both sample types, this corre- ater on all days. Taking influent on its own, it also represented a
sponded to a tRP error of 1.30 min (i.e. an elution window of 2.60 min blind test compound which could be used to reliably evaluate the
in total and <10% of total runtime). Whilst this threshold repre- semi-targeted approach using LC-HRMS data and ANN together. As
sented a potential 25% false negative rate, extension or reduction can be observed from Fig. 6(a), an excellent retention time pre-
of this threshold resulted in either an impractically wide or narrow diction match was achieved with the optimised GRNN at realistic
retention window relative to peak width for this chromatographic concentration levels. Retention time was then confirmed using a
system. As discussed later, the false negative rate was shown to be matrix-matched standard along with matching isotopic m/z pro-
sample dependent in practice. Therefore, this limit enabled prac- files (Fig. 6(b)). Following this, the concentration of this compound
tical and rapid shortlisting of larger numbers of viable suspect was estimated in influent wastewater across the week at con-
compounds in a poorly understood sample. Obviously, thresholds centrations ranging between 42 and 160 ng/L with the highest
may be adjusted to extend the scope of the semi-targeted approach concentration measured on a Saturday (R2 = 0.9980; method per-
in the future if needed by including more compounds during ANN formance given in the SI).
training or by using a different set of chromatographic conditions. For the majority of ions with m/z within the ±5 ppm match win-
The application of ANNs to semi-targeted screening, however, is not dow, more than one viable chromatographic peak was detected
intended for confirmatory identification and should only be used to (two to four viable peaks were often identified across the run-
reduce the time required to source appropriate standard reference time). These mostly constituted a number of well separated peaks
materials for this purpose. (e.g. codeine, betaxolol, efaproxiral, cathinone, mephedrone and
bunolol) or, in some cases, a number of closely eluting/unresolved
3.4. Trace screening of wastewater samples and evaluation of peaks (e.g. isometheptene and metipranolol). In a few cases, larger
GRNN retention time predictions numbers of viable peaks were detected even when using such a
narrow m/z match window. Table 2 shows that up to 15 viable
Unspiked aliquots of the same 24-h composite samples of chromatographic peaks were identified for suspected tramadol
wastewater were analysed to characterise target PPCP occurrence occurrence (included as a blind test case) in wastewater. Prelim-
across the week. Peak selection was performed using a much inary identification of this compound without a reference standard
larger database than the 166 target compounds used to simulate would be very challenging due to the presence of other isobaric
a more realistic data mining exercise. In several cases, matching species. Application of the ANN reduced this number to three
m/z data existed within this wider database along with target suspect peaks in both influent and effluent data on all days. Exam-
compound data. Automated peak selection yielded between 2500 ination of CID data yielded no identifiable ion signatures within
and 3000 peaks for influent and between 1200 and 1900 peaks this m/z scan range, making this an especially difficult case for
in effluent samples depending on the day (irrespective of the ESI occurrence confirmation. For this application using this type of
 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44 41

Table 1
Application of LC-HRMS and ANN prediction models to determine the occurrence of 166 selected PPCPs in unspiked influent and effluent wastewaters over a week long
sampling campaign at a London-based WWTP (11th–17th March, 2014). See SI for individual compound occurrence and matching tRM data.

Influent wastewater Effluent wastewater

Sampling Day 1 2 3 4 5 6 7 1 2 3 4 5 6 7

Number of targeted compounds shortlisted 130 134 89 134 137 140 137 116 120 121 119 121 122 112
by TraceFinder software
Number of shortlisted compounds 59 61 42 69 67 62 64 49 53 60 56 58 56 53
excluded
Number of confirmed compounds using 39 37 28 36 41 36 38 28 29 31 29 35 34 31
ANN + matrix-matched standard
Number of extra confirmed compounds 6 10 6 6 7 11 7 13 17 9 12 11 10 10
using a matrix-matched standard only
(i.e. ANN false negative)
Number of extra suspected compounds 26 26 13 24 22 31 28 26 21 21 22 17 21 18
using ANN only (i.e. ANN false positive)
Total number compounds confirmed (with 45 47 34 42 48 47 45 41 46 40 41 46 44 41
standard only)
Success rate for ANN (%) 87 79 82 86 85 77 84 68 63 78 71 76 77 76

Mean % ANN success rate ± standard 83 ± 4 73 ± 5

deviation (n = 7)

Fig. 6. Extracted ion chromatograms confirming the presence of mephedrone in (a) an unspiked influent wastewater sample and (b) a 250 ␮g/L matrix matched standard.
The left panel in (a) and (b) both show the [M + H]+ ion of mephedrone (acquired in full scan mode) and its corresponding product ion (m/z 145.0886). Values listed for the
[M + H]+ and product ion represent the calculated m/z. The right panel shows the isotopic profile comparison between the measured [M + H]+ ion in wastewater relevant to
a software-generated theoretical isotopic profile.
42 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44

Table 2
Shortlist of viable chromatographic peak retention times recorded using LC-HRMS for suspected tramadol occurrence in unspiked influent and effluent wastewaters. Retention
times given in bold italics shading represent confirmatory identification of tramadol using HRMS data as well as ANN and from a matrix-matched standard. Retention data
given in italics shading represent only those additional peaks suspected to be tramadol using HRMS and ANN data alone.

Calculated tramadol [M+H]+ m/z: 264.1958

Influent wastewater
Analyte in influent: 7.44 min (Range: 7.19–7.69 min)a
Analyte ANN in influent: 7.56 min (Range: 6.26–8.86 min)b

Viable peak (min)

Sampling day tR 1 tR 2 tR 3 tR 4 tR 5 tR 6 tR 7 tR 8 tR 9 tR 10 tR 11 tR 12 tR 13 tR 14 tR 15
Day 1 2.57 5.78 6.54 7.41 8.53 11.74 16.30
Day 2 2.58 5.75 6.55 7.42 8.53 11.76 16.31
Day 3 2.57 5.73 6.52 7.39 8.52 11.77 16.30
Day 4 2.45 5.75 6.50 7.39 8.54 9.31 10.04 15.95 16.71 18.43 19.05 19.92
Day 5 2.46 5.73 6.48 7.35 8.49 9.20 11.74 15.93 16.28
Day 6 2.56 5.77 6.57 7.47 8.58 16.33
Day 7 2.59 5.76 6.56 7.43 8.54 16.30

Effluent wastewater
Analyte in effluent: 7.38 min (Range: 7.13–7.63 min)a
Analyte ANN in effluent: 7.49 min (Range: 6.19–8.79 min)b

Viable peak (min)

Sampling day tR 1 tR 2 tR 3 tR 4 tR 5 tR 6 tR 7 tR 8 tR 9 tR 10 tR 11 tR 12 tR 13 tR 14
Day 1 2.27 4.33 5.66 6.39 7.31 8.45 8.83 11.73 15.91 16.24 18.39 19.03 19.88
Day 2 2.28 4.36 5.69 6.41 7.31 8.46 8.83 11.74 15.94 16.27 18.39 19.03 19.88
Day 3 2.29 4.37 5.70 6.44 7.35 8.50 8.85 11.75 15.95 16.30 18.43 19.05 19.92
Day 4 2.29 4.40 5.70 6.44 7.35 8.48 8.86 9.31 11.79 15.98 16.32 18.42 19.05 19.92
Day 5 2.29 4.42 5.72 6.44 7.35 8.48 11.75 16.28 18.41 19.89
Day 6 2.30 4.42 5.71 6.44 7.34 8.48 8.83 11.78 15.96 16.28 18.42 19.04 19.89
Day 7 2.33 4.45 5.74 6.45 7.34 8.48 8.83 9.27 11.79 15.98 18.42 19.05 19.90
a
Range for confirmation of tramadol occurrence defined as ±15 s of in a spiked sample extract.
b
Range defined as ±1.30 min.

LC-HRMS instrumentation, it reinforces the requirement for good yielded multiple chromatographic peaks with matching m/z. How-
chromatographic separation of compounds during screening. After ever, whilst perhaps more regular occurrence was observed for
comparison with a matrix-matched standard, tramadol was iden- other peaks, one peak at ∼8.90 min was found to be a match with
tified chromatographically as one of the ANN-shortlisted peaks in ANN (which was subsequently confirmed with a matrix-matched
all samples across the week. Pethidine, as another example, was standard at 8.80 min). In contrast, effluent samples yielded signifi-
identified in much the same manner (Table 3). Influent samples cantly more viable peaks with matching m/z. Some of these peaks

Table 3
Shortlist of viable chromatographic peak retention times recorded using LC-HRMS for suspected pethidine occurrence in unspiked influent and effluent wastewaters. Retention
data given in bold italics shading represent confirmatory identification of pethidine using HRMS data as well as ANN and from a matrix-matched standard. Retention data
given in italics shading represent only those additional peaks suspected to be pethidine using HRMS and ANN data alone.

Calculated pethidine td:paraenter [M+H]+ m/z: 248.1645

Influent wastewater
Analyte in influent: 8.80 min (Range: 8.55–9.05 min)a
Analyte ANN in influent: 7.86 min (Range: 6.56-9.16 min)b

Viable peak (min)

Sampling day tR 1 tR 2 tR 3 tR 4 tR 5 tR 6
Day 1 – 2.29 – 3.40 – –
Day 2 – 2.28 – 3.36 – –
Day 3 – 2.28 – – – –
Day 4 2.18 3.17 – 8.90 –
Day 5 2.18 3.13 – 8.89 –
Day 6 – 2.28 – – – –
Day 7 – 2.32 – – – 14.22

Effluent wastewater
Analyte in effluent: 8.78 min (Range: 8.53-9.03 min)a
Analyte ANN in effluent: 7.79 min (Range: 6.49-9.09 min)b

Viable peak (min)

Sampling day tR 1 tR 2 tR 3 tR 4 tR 5 tR 6 tR 7 tR 8 tR 9 tR 10
Day 1 – 2.90 – – 6.48 – 8.87 17.82 – –
Day 2 1.43 2.95 3.93 – 6.53 8.55 8.88 17.44 17.82 –
Day 3 1.45 2.98 – – – 8.57 8.90 17.84 – –
Day 4 – 2.96 – – 6.54 – 8.91 17.48 17.86 18.28
Day 5 1.44 2.98 – 6.30 6.52 – 8.89 17.84 – –
Day 6 1.44 3.02 – – 6.53 – 8.90 17.48 17.83 18.25
Day 7 1.46 3.06 – – 6.50 8.56 8.90 – –
a
Range for confirmation of pethidine occurrence defined as ±15 s of in a spiked sample extract.
b
Range defined as ±1.30 min.
 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44 43

occurred on most/all days. The ANN narrowed a total of ten viable Appendix A. Supplementary data
peaks to three in this case and confirmation was then achieved with
the matrix-matched standard. Obviously, and where the informa- Supplementary data associated with this article can be found, in
tion exists, HCD fragments may be used to eliminate more of these the online version, at http://dx.doi.org/10.1016/j.chroma.2015.03.
peaks, but these would not necessarily be known in a semi-targeted 063
application however.
Lastly, this semi-targeted approach was applied to additional References
pharmaceutical residue matches occurring in an influent wastew-
ater sample. Following HRMS data screening, 95 compounds were [1] C.G. Daughton, T.A. Ternes, Pharmaceuticals and personal care products in the
environment: agents of subtle change? Environ. Health Persp. 107 (1999) 907.
shortlisted, again with many showing multiple chromatographic
[2] J.-W. Kim, B.R. Ramaswamy, K.-H. Chang, T. Isobe, S. Tanabe, Multiresidue
peaks at isobaric m/z. Molecular descriptors were generated and analytical method for the determination of antimicrobials, preservatives, ben-
following the GRNN prediction, 37 compounds were found to zotriazole UV stabilizers, flame retardants and plasticizers in fish using ultra
high performance liquid chromatography coupled with tandem mass spec-
match measured sample retention times. All shortlisted com-
trometry, J. Chromatogr. A 1218 (2011) 3511.
pounds, along with their HRMS and tRM data are detailed in the SI. [3] J. Nurmi, J. Pellinen, Multiresidue method for the analysis of emerging
Interestingly, whilst none of these compounds were used in the contaminants in wastewater by ultra performance liquid chromatography-
development of ANN models, many have been previously detected time-of-flight mass spectrometry, J. Chromatogr. A 1218 (2011) 6712.
[4] E. Perez-Carrera, M. Hansen, V.M. Leon, E. Bjorklund, K.A. Krogh, B. Halling-
in wastewater systems. Such examples included clozapine, feno- Sorensen, E. Gonzalez-Mazo, Multiresidue method for the determination of 32
profen, ibuprofen, naproxen, sertraline and salbutamol [55–57]. It human and veterinary pharmaceuticals in soil and sediment by pressurized-
is important to note however, such SPE methods may not recover liquid extraction and LC-MS/MS, Anal. Bioanal. Chem. 398 (2010) 1173.
[5] D.R. Baker, L. Barron, B. Kasprzyk-Hordern, Illicit and pharmaceutical drug con-
all compound types and coupled with a relatively small database sumption estimated via wastewater analysis. Part A: chemical analysis and
in comparison to the number of pharmaceuticals available, in real- drug use estimates, Sci. Tot. Environ. 487 (2014) 629.
ity this is still likely to represent an under-estimation of total PPCP [6] S. Idder, L. Ley, P. Mazellier, H. Budzinski, Quantitative on-line
preconcentration-liquid chromatography coupled with tandem mass
occurrence. Moreover, it must be re-stated here that, despite appli- spectrometry method for the determination of pharmaceutical compounds in
cation of ANNs and an HRMS database, confirmatory identification water, Anal. Chim. Acta 805 (2013) 107.
of these compounds is still required using a reference standard, and [7] M. Laven, T. Alsberg, Y. Yu, M. Adolfsson-Erici, H. Sun, Serial mixed-
mode cation- and anion-exchange solid-phase extraction for separation
preferably one prepared in matrix.
of basic, neutral and acidic pharmaceuticals in wastewater and analysis
by high-performance liquid chromatography-quadrupole time-of-flight mass
spectrometry, J. Chromatogr. A 1216 (2009) 49.
[8] M.J. Martinez Bueno, A. Agueera, M. Dolores Hernando, M. Jose Gomez, Eval-
4. Conclusion uation of various liquid chromatography-quadrupole-linear ion trap-mass
spectrometry operation modes applied to the analysis of organic pollutants
in wastewaters, J. Chromatogr. A 1216 (2009) 5995.
Modelling of PPCP retention times from gradient reversed- [9] D.F. Specht, A general regression neural network, Neural Networks IEEE Trans.
phase chromatography of influent and effluent wastewater extracts 2 (1991) 568.
has been successfully demonstrated for the first time. By selecting [10] O.A.H. Jones, N. Voulvoulis, J.N. Lester, Analytical method development for
the simultaneous determination of five human pharmaceuticals in water
a range of molecular descriptors, predictions at the 75th percentile and wastewater samples by gas chromatography-mass spectrometry, Chro-
of all compound data were within 1.30 min of tRM . Discrimina- matographia 58 (2003) 471.
tion of structurally similar/isobaric compounds was also possible [11] N. Migowska, M. Caban, P. Stepnowski, J. Kumirska, Simultaneous analysis of
non-steroidal anti-inflammatory drugs and estrogenic hormones in water and
in most cases and demonstrated acceptable tRP accuracy. Appli-
wastewater samples using gas chromatography-mass spectrometry and gas
cation to real wastewater from a week-long sampling campaign chromatography with electron capture detection, Sci. Tot. Environ. 441 (2012)
showed that ANNs could be used successfully for trace detection 77.
[12] R. Grabic, J. Fick, R.H. Lindberg, G. Fedorova, M. Tysklind, Multi-residue method
with similar performance. Furthermore, the application of ANNs
for trace level determination of pharmaceuticals in environmental samples
for retention time prediction markedly reduced the number of using liquid chromatography coupled to triple quadrupole mass spectrometry,
viable chromatographic peaks for each compound and also accu- Talanta 100 (2012) 183.
rately excluded compounds from wastewater data. Ultimately, this [13] R. Loos, S. Tavazzi, B. Paracchini, E. Canuti, C. Weissteiner, Analysis of polar
organic contaminants in surface water of the northern Adriatic Sea by
approach will allow more rapid characterisation of PPCP occur- solid-phase extraction followed by ultrahigh-pressure liquid chromatography-
rence in wastewaters by directing the synthesis of likely candidate QTRAP(A (R)) MS using a hybrid triple-quadrupole linear ion trap instrument,
reference materials. Anal. Bioanal. Chem. 405 (2013) 5875.
[14] M. Petrovic, B. Skrbic, J. Zivancev, L. Ferrando-Climent, D. Barcelo, Determina-
tion of 81 pharmaceutical drugs by high performance liquid chromatography
coupled to mass spectrometry with hybrid triple quadrupole-linear ion trap in
different types of water in Serbia, Sci. Tot. Environ. 468 (2014) 415.
Acknowledgments [15] W. Peysson, E. Vulliet, Determination of 136 pharmaceuticals and hormones in
sewage sludge using quick, easy, cheap, effective, rugged and safe extraction
followed by analysis with liquid chromatography-time-of-flight-mass spec-
The authors gratefully acknowledge primary funding for this
trometry, J. Chromatogr. A 1290 (2013) 46.
work from the Environmental Sustainability Knowledge Trans- [16] C.L. Chitescu, E. Oosterink, J. de Jong, A.A.M. Stolker, Ultrasonic or accelerated
fer Network, Engineering and Physical Sciences Research Council solvent extraction followed by U-HPLC-high mass accuracy MS for screening
(EPSRC) and Thermo Fisher Scientific to support a CASE indus- of pharmaceuticals and fungicides in soil and plant samples, Talanta 88 (2012)
653.
trial scholarship for K. Munro (Reference: EP/J502029/1). Thanks [17] F. Wode, C. Reilich, P. van Baar, U. Duennbier, M. Jekel, T. Reemtsma,
are also extended to the Biotechnology and Biological Sciences Multiresidue analytical method for the simultaneous determination of 72
Research Council (BBSRC) and AstraZeneca (under the Global SHE micropollutants in aqueous samples with ultra high performance liquid
chromatography-high resolution mass spectrometry, J. Chromatogr. A 1270
research programme) for award of a CASE industrial scholar- (2012) 118.
ship for T. Miller who collaborated on this project (Reference: [18] L. Vergeynst, H. Van Langenhove, P. Joos, K. Demeestere, Suspect screening and
BB/K501177/1). Funding bodies played no role in the design of the target quantification of multi-class pharmaceuticals in surface water based on
large-volume injection liquid chromatography and time-of-flight mass spec-
study or decision to publish. The authors declare no financial con- trometry, Anal. Bioanal. Chem. 406 (2014) 2533.
flict of interest. Thanks are also extended to Alessandro Musenga [19] L. Bijlsma, E. Emke, F. Hernandez, P. de Voogt, Investigation of drugs of abuse
for his assistance with reference standards and to the staff within and relevant metabolites in Dutch sewage water by liquid chromatography
coupled to high resolution mass spectrometry, Chemosphere 89 (2012) 1399.
the Centre of Excellence in Mass Spectrometry at KCL for technical
[20] M. Schriks, J.A. van Leerdam, S.C. van der Linden, B. van der Burg, A.P. van
support. Wezel, P. de Voogt, High-resolution mass spectrometric identification and
44 K. Munro et al. / J. Chromatogr. A 1396 (2015) 34–44

quantification of glucocorticoid compounds in various wastewaters in The [39] K. Petritis, L.J. Kangas, B. Yan, M.E. Monroe, E.F. Strittmatter, W.-J. Qian,
Netherlands, Environ. Sci. Tech. 44 (2010) 4766. J.N. Adkins, R.J. Moore, Y. Xu, M.S. Lipton, D.G.C. Ii, R.D. Smith, Improved
[21] F. Hernandez, M. Ibanez, E. Gracia-Lor, J.V. Sancho, Retrospective LC-QTOF-MS peptide elution time prediction for reversed-phase liquid chromatography-
analysis searching for pharmaceutical metabolites in urban wastewater, J. Sep. MS by incorporating peptide sequence information, Anal. Chem. 78 (2006)
Sci. 34 (2011) 3517. 5026.
[22] M. Ibanez, C. Guerrero, J.V. Sancho, F. Hernandez, Screening of antibiotics in sur- [40] A. Malenovic, B. Jancic-Stojanovic, N. Kostic, D. Ivanovic, M. Medenica, Opti-
face and wastewater samples by ultra-high-pressure liquid chromatography mization of artificial neural networks for modeling of atorvastatin and its
coupled to hybrid quadrupole time-of-flight mass spectrometry, J. Chromatogr. impurities retention in micellar liquid chromatography, Chromatographia 73
A 1216 (2009) 2529. (2011) 993.
[23] J. Robles-Molina, F.J. Lara-Ortega, B. Gilbert-Lopez, J.F. Garcia-Reyes, A. Molina- [41] A.G. Fragkaki, E. Farmaki, N. Thomaidis, A. Tsantili-Kakoulidou, Y.S. Angelis,
Diaz, Multi-residue method for the determination of over 400 priority and M. Koupparis, C. Georgakopoulos, Comparison of multiple linear regression,
emerging pollutants in water and wastewater by solid-phase extraction and partial least squares and artificial neural networks for prediction of gas chro-
liquid chromatography-time-of-flight mass spectrometry, J. Chromatogr. A matographic relative retention times of trimethylsilylated anabolic androgenic
1350 (2014) 30. steroids, J. Chromatogr. A 1256 (2012) 232.
[24] J.W. Li, Prediction of internal standards in reversed-phase liquid chromatog- [42] T.H. Miller, A. Musenga, D.A. Cowan, L.P. Barron, Prediction of chromatographic
raphy IV. Correlation and prediction of retention in reversed-phase ion-pair retention time in high-resolution anti-doping screening data using artificial
chromatography based on linear solvation energy relationships, Anal. Chim. neural networks, Anal. Chem. 85 (2013) 10330.
Acta 522 (2004) 113. [43] L. Barron, J. Havel, M. Purcell, M. Szpak, B. Kelleher, B. Paull, Predicting sorption
[25] M. Tian, K.H. Row, Study of retention in micellar liquid chromatography on a of pharmaceuticals and personal care products onto soil and digested sludge
C(18) column on the basis of linear solvation energy relationships, B. Korean using artificial neural networks, Analyst 134 (2009) 663.
Chem. Soc. 29 (2008) 979. [44] D.R. Albaugh, L.M. Hall, D.W. Hill, T.M. Kertesz, M. Parham, L.H. Hall, D.F. Grant,
[26] M. Vitha, P.W. Carr, The chemical interpretation and practice of linear solva- Prediction of HPLC retention index using artificial neural networks and IGroup
tion energy relationships in chromatography, J. Chromatogr. A 1126 (2006) E-state indices, J. Chem. Inf. Model. 49 (2009) 788.
143. [45] N.V. Heuett, C.E. Ramirez, A. Fernandez, P.R. Gardinali, Analysis of drugs of abuse
[27] J.E. Madden, N. Avdalovic, P.R. Haddad, J. Havel, Prediction of retention times for by online SPE-LC high resolution mass spectrometry: communal assessment of
anions in linear gradient elution ion chromatography with hydroxide eluents consumption, Sci. Tot. Environ. 511 (2015) 319.
using artificial neural networks, J. Chromatogr. A 910 (2001) 173. [46] M. Petrovic, M. Farré, M.L. de Alda, S. Perez, C. Postigo, M. Köck, J. Radjenovic, M.
[28] N.S. Quiming, N.L. Denola, S.R. Bin Samsuri, Y. Saito, K. Jinno, Development of Gros, D. Barcelo, Recent trends in the liquid chromatography–mass spectrome-
retention prediction models for adrenoreceptor agonists and antagonists on try analysis of organic contaminants in environmental samples, J. Chromatogr.
a polyvinyl alcohol-bonded stationary phase in hydrophilic interaction chro- A 1217 (2010) 4004.
matography, J. Sep. Sci. 31 (2008) 1537. [47] M. Petrovic, M. Gros, D. Barcelo, Multi-residue analysis of pharmaceuticals in
[29] V.K. Gupta, H. Khani, B. Ahmadi-Roudi, S. Mirakhorli, E. Fereyduni, S. Agarwal, wastewater by ultra-performance liquid chromatography–quadrupole–time-
Prediction of capillary gas chromatographic retention times of fatty acid methyl of-flight mass spectrometry, J. Chromatogr. A 1124 (2006) 68.
esters in human blood using MLR, PLS and back-propagation artificial neural [48] E. Gracia-Lor, J.V. Sancho, F. Hernández, Simultaneous determination of acidic,
networks, Talanta 83 (2011) 1014. neutral and basic pharmaceuticals in urban wastewater by ultra high-pressure
[30] A.A. D’Archivio, A. Giannitto, M.A. Maggi, F. Ruggieri, Cross-column retention liquid chromatography-tandem mass spectrometry, J. Chromatogr. A 1217
prediction in reversed-phase high-performance liquid chromatography by arti- (2010) 622.
ficial neural network modelling, Anal. Chim. Acta 717 (2012) 52. [49] K.H. Langford, K.V. Thomas, Determination of pharmaceutical compounds in
[31] A.A. D’Archivio, A. Incani, F. Ruggieri, Cross-column prediction of gas- hospital effluents and their contribution to wastewater treatment works, Env-
chromatographic retention of polychlorinated biphenyls by artificial neural iron. Int. 35 (2009) 766.
networks, J. Chromatogr. A 1218 (2011) 8679. [50] A.L. Batt, M.S. Kostich, J.M. Lazorchak, Analysis of ecologically relevant phar-
[32] A.A. D’Archivio, M.A. Maggi, F. Ruggieri, Multiple-column RP-HPLC retention maceuticals in wastewater and surface water using selective solid-phase
modelling based on solvatochromic or theoretical solute descriptors, J. Sep. extraction and UPLC-MS/MS, Anal. Chem. 80 (2008) 5021.
Sci. 33 (2010) 155. [51] A.L.N. van Nuijs, B. Pecceu, L. Theunis, N. Dubois, C. Charlier, P.G. Jorens, L.
[33] A.A. D’Archivio, M.A. Maggi, F. Ruggieri, Multi-variable retention modelling in Bervoets, R. Blust, H. Neels, A. Covaci, Cocaine and metabolites in waste and
reversed-phase high-performance liquid chromatography based on the solva- surface water across Belgium, Environ. Pollut. 157 (2009) 123.
tion method: a comparison between curvilinear and artificial neural network [52] B. Shao, D. Chen, J. Zhang, Y. Wu, C. Sun, Determination of 76 pharmaceutical
regression, Anal. Chim. Acta 690 (2011) 35. drugs by liquid chromatography–tandem mass spectrometry in slaughter-
[34] A.A. D’Archivio, F. Ruggieri, P. Mazzeo, E. Tettamanti, Modelling of retention of house wastewater, J. Chromatogr. A 1216 (2009) 8312.
pesticides in reversed-phase high-performance liquid chromatography: quan- [53] B. Kasprzyk-Hordern, R.M. Dinsdale, A.J. Guwy, The effect of signal suppres-
titative structure-retention relationships based on solute quantum-chemical sion and mobile phase composition on the simultaneous analysis of multiple
descriptors and experimental (solvatochromic and spin-probe) mobile phase classes of acidic/neutral pharmaceuticals and personal care products in surface
descriptors, Anal. Chim. Acta 593 (2007) 140. water by solid-phase extraction and ultra performance liquid chromatography-
[35] R. Berges, V. Sanz-Nebot, J. Barbosa, Modelling retention in liquid chromatog- negative electrospray tandem mass spectrometry, Talanta 74 (2008) 1299.
raphy as a function of solvent composition and pH of the mobile phase, J. [54] USEPA, Method 1694: Pharmaceuticals and Personal Care Products in Water,
Chromatogr. A 869 (2000) 27. Soil, Sediment, and Biosolids by HPLC/MS/MS (2007).
[36] M. Pompe, M. Razinger, M. Novic, M. Veber, Modelling of gas chromatographic [55] S. Yuan, X. Jiang, X. Xia, H. Zhang, S. Zheng, Detection, occurrence and fate of 22
retention indices using counterpropagation neural networks, Anal. Chim. Acta psychiatric pharmaceuticals in psychiatric hospital and municipal wastewater
348 (1997) 215. treatment plants in Beijing, China, Chemosphere 90 (2013) 2520.
[37] N.S. Quiming, N.L. Denola, I. Ueta, Y. Saito, S. Tatematsu, K. Jinno, Reten- [56] L. Lishman, S.A. Smyth, K. Sarafin, S. Kleywegt, J. Toito, T. Peart, B. Lee, M. Servos,
tion prediction of adrenoreceptor agonists and antagonists on a diol column M. Beland, P. Seto, Occurrence and reductions of pharmaceuticals and per-
in hydrophilic interaction chromatography, Anal. Chim. Acta 598 (2007) sonal care products and estrogens by municipal wastewater treatment plants
41. in Ontario, Canada, Sci. Tot. Environ. 367 (2006) 544.
[38] K. Shinoda, M. Sugimoto, N. Yachie, N. Sugiyama, T. Masuda, M. Robert, T. Soga, [57] A. Lajeunesse, S.A. Smyth, K. Barclay, S. Sauvé, C. Gagnon, Distribution of antide-
M. Tomita, Prediction of liquid chromatographic retention times of peptides pressant residues in wastewater and biosolids following different treatment
generated by protease digestion of the Escherichia coli proteome using artificial processes by municipal wastewater treatment plants in Canada, Water Res. 46
neural networks, J. Proteome Res. 5 (2006) 3312. (2012) 5600.