pubs.acs.org/acssensors Article A Sensor Array Based on Molecularly Imprinted Polymers and Machine Learning for the Analysis of Fluoroquinolone Antibiotics Mingyue Wang, Xavier Cetó, and Manel del Valle* Cite This: ACS Sens. 2022, 7, 3318−3325 Read Online ACCESS Metrics & More Article Recommendations * sı Supporting Information ABSTRACT: Fluoroquinolones (FQs) are one of the most important types of antibiotics in the clinical, poultry, and aquaculture industries, and their monitoring is required as the abuse has led to severe issues, such as antibiotic residues and antimicrobial resistance. In this study, we report a voltammetric electronic tongue (ET) for the simultaneous determination of ciprofloxacin, levofloxacin, and moxifloxacin in both pharmaceutical and biological samples. The ET comprises four sensors modified with three different customized molecularly imprinted polymers (MIPs) and a nonimprinted polymer integrated with Au nanoparticle-decorated multiwall carbon nanotubes (Au-fMWCNTs). MWCNTs were first functionalized to serve as a supporting substrate, while the anchored Au nanoparticles acted as a catalyst. Subsequently, MIP films were obtained by electropolymeriza- tion of pyrrole in the presence of the different target FQs. The sensors’ morphology was characterized by scanning electron microscopy and transmission electron microscopy, while the modification process was followed electrochemically step by step employing [Fe(CN)6]3−/4− as the redox probe. Under the optimal conditions, the MIP(FQs)@Au-fMWCNT sensors exhibited different responses, limits of detection of ca. 1 μM, and a wide detection range up to 300 μM for the three FQs. Lastly, the developed ET presents satisfactory agreement between the expected and obtained values when used for the simultaneous determination of mixtures of the three FQs (R2 ≥0.960, testing subset), which was also applied to the analysis of FQs in commercial pharmaceuticals and spiked human urine samples. KEYWORDS: molecularly imprinted polymers, electronic tongues, artificial neural networks, fluoroquinolones, antibiotics, pharmaceutical analysis F luoroquinolones (FQs), one of the most important families of antibiotics, are used individually or in combination against a broad spectrum of Gram-positive and To date, the detection of FQs using electrochemical sensors has been proposed as a cheaper and time- and labor-saving alternative to complex detection methods such as high- Gram-negative bacteria in clinical and animal husbandry.1,2 performance liquid chromatography or gas chromatography− Among them, ciprofloxacin (CFX), levofloxacin (LFX), and mass spectrometry.6 Nevertheless, most reported electro- moxifloxacin (MFX) are the most common in current clinical chemical sensors target only one specific FQ, and cannot practice. LFX and MFX are considered as respiratory drugs achieve the simultaneous multidetermination in practical that have good bactericidal activity against most of the applications or deal with the interferences that may arise respiratory pathogens, while CFX is used to treat the infections from certain complicated matrixes.7,8 This can be rather thorny of the urinary tract and intestine. Structurally, FQs are similar in the case of the FQ family given the common shared as they are all derived from nalidixic acid (NA), sharing the structure and similar natures. same quinolone ring, but with different substitutions. Although Compared with the conventional single sensor approach, by the usage of FQs is approved worldwide and their significant coupling a sensor array providing responses to various analytes curative effects have been demonstrated, consequences with appropriate data processing, electronic tongues (ETs) resulting from their misuse such as antibiotic residues (mainly allow the discrimination and simultaneous quantification of related to food safety and environmental contamination) or antimicrobial resistance (essentially requiring therapeutic drug Received: June 14, 2022 monitoring) cannot be neglected.3−5 Thus, FQs monitoring is Accepted: October 6, 2022 required in different stages such as in pharmaceutical factories Published: October 25, 2022 when produced, in hospitals during its administration to the patients, or in the environment for the treatment of wastewater. © 2022 The Authors. Published by American Chemical Society https://doi.org/10.1021/acssensors.2c01260 3318 ACS Sens. 2022, 7, 3318−3325 ACS Sensors pubs.acs.org/acssensors Article multiple analytes,9,10 while they might also be able to solve electrochemical characterization, while 100 mM PBS at pH 10.0 other problems related to sensors such as drifts, interferences, was used for the electrochemical extraction and cleaning. Then, 100 cross-sensitivity, and/or matrix effects. mM acetate buffer at pH 3.5 was used for the electropolymerization To further improve the performance of ETs, one of the and preparation of the stock solutions, while 100 mM acetate buffer at pH 4.5 was employed for the analysis of FQs. trending strategies is using a proper set of biorecognition Apparatus. An EVO MA10 (Zeiss, Oberkochen, Germany) and a elements such as enzymes, peptides, etc.11 Recently, the JEOL 1210 (Peabody, MA, USA) were used for scanning electron feasibility of incorporating molecularly imprinted polymers microscopy and transmission electron microscopy (SEM/TEM) (MIPs) into ETs has been investigated for the simultaneous analyses, respectively. An Autolab PGSTAT30 potentiostat (Ecoche- determination of structurally similar analytes in various mie, Utrecht, the Netherlands) with a multichannel configuration fields.12,13 As artificial customized receptors, MIPs not only controlled by GPES and FRA software packages was used for the can afford similar specificity to a bioreceptor but also can offer electrochemical measurements. higher physical and chemical stabilities than biological Fabrication of the MIP(FQs)@Au-fMWCNT Sensor. Hand- recognition elements.14,15 More importantly, the host-guest crafted GEC electrodes, which served as sensor platforms for the construction of the MIP-based sensor array, were prepared according recognition mechanism between MIPs and the target to established procedures in our laboratory.22 Briefly, the paste was compounds, defined by the size and structure of the template prepared by mixing 58% of graphite powder and 42% of epoxy resin. molecules, offers a simple and straightforward procedure for Next, this paste was placed into a PVC tube, in which a copper disc developing a series of MIP-based sensors to be combined into soldered to a connector had previously been inserted, and cured at 40 a sensor array for the development of ETs. However, despite °C for 2 days. Finally, the electrodes were polished and ready to use. MIPs’ many merits, there is still room for improvement on Pristine MWCNTs were pretreated with strong acids to obtain their integration into sensors, especially in regard to carboxylic-functionalized MWCNTs (fMWCNTs). First of all, sensitivity.16 Compared with other polymerization methods, MWCNTs were dispersed in an acid mixture of concentrated electropolymerization can generate controllable MIP thin films H2SO4/HNO3 (3:1 v/v ) by ultrasonication for 6 h at 40 °C. directly onto transducer surfaces, representing a simpler and Then, they were washed with Milli-Q water until neutralized and dried in an oven for further use. highly reproducible approach for the integration of MIPs with The decoration of fMWCNTs with Au NPs was performed electrochemical sensors.17,18 In addition, the lower sensitivity according to the previously reported procedure.23 Specifically, 500 μL derived from MIPs’ intrinsically inferior conductivity and of 10 mM HAuCl4, 500 μL of 10 mM trisodium citrate, and 18.4 mL electrocatalytic activity could be improved by introducing of Milli-Q water were mixed in a round-bottom flask, to which 20 mg nanomaterials,19 for example, carbon nanotubes (CNTs)20 or of fMWCNTs and 10 mL of anhydrous ethanol were added. After metallic nanoparticles (NPs).21 sonicating for 10 min, 600 μL of a 100 mM freshly prepared cold To this end, herein we explore the electropolymerization of NaBH4 solution was also added into the mixture. The reaction lasted pyrrole (Py) onto multiwall CNTs (MWCNTs) as a for 10 h under continuous stirring. The obtained product (denoted as supporting substrate that can lead to MIPs with a larger Au-fMWCNTs) was thoroughly washed with Milli-Q water and dried in an oven at 40 °C overnight. surface area, higher conductivity, and amplified signal. In For the modification of the sensors, 2.0 mg of the dried Au- addition, the incorporation of metallic gold NPs (Au NPs) can fMWCNT powder were dispersed in 1 mL of DMF into which 100 also act as a catalyst for the electrochemical reactions and μL of Nafion were added to improve the stability of Au-fMWCNTs further enhance the sensitivity of the obtained sensors. In this on the surface of the electrodes upon drop-casting. Finally, 4 μL of the manner, MIPs integrated with these nanomaterials were Au-fMWCNT dispersion in DMF were evenly dropped onto the synthesized toward CFX, LFX, and MFX. Next, the developed surface of the GEC and dried in the oven. sensors were characterized both morphologically and electro- The synthesis of the MIP(FQs) film on the surface of Au- chemically and finally combined into a sensor array, which was fMWCNTs was performed using the electropolymerization method. successfully employed to develop an MIP-based ET for the The solution containing 0.05 M Py as the functional monomer, 0.01 M FQ (either CFX, LFX, or MFX) as the template molecule, and 20 analysis of FQ antibiotics in commercial drugs and spiked mL of acetate buffer (pH 3.5) as the supporting electrolyte was human urine samples. purged with nitrogen for 10 min before being used for the synthesis of ■ EXPERIMENTAL SECTION Materials and Reagents. Graphite powder (particle size <50 μm, the MIP(FQ) film. Electropolymerization was conducted by cycling the potential from 0.0 to +1.3 V versus Ag/AgCl at a scan rate of 50 mV·s−1(Figure S1). Subsequently, extraction of the template obtained from BDH, Poole, UK) and an epoxy resin kit (supplied by molecules from the polymer film was carried out in a two-step Resineco, Barcelona, Spain) were used for fabricating the graphite− procedure: cycling the potential from 0.0 to +1.5 V versus Ag/AgCl at epoxy composite (GEC) electrodes. Pristine MWCNTs (outer a scan rate of 100 mV·s−1 in PBS (pH 10.0) (Figure S1), followed by diameter ∼ 10 to 30 nm) were obtained from SES Research Inc. the immersion of the electrodes into a mixture of EtOH/0.1 M (Houston, TX, USA). Py and anhydrous trisodium citrate were NaOH (1:1 v/v) for 30 min. purchased from Alfa Aesar (Ward Hill, MA, USA). Gold(III) chloride Similarly, the same procedures as abovementioned were used for solution in hydrochloric acid (30 wt % HAuCl4 in diluted HCl), N,N- the synthesis of the nonimprinted polymer on Au-fMWCNTs (NIP@ dimethylformamide (DMF), perfluorinated resin solution containing Au-fMWCNTs), except that no FQ was added during electro- Nafion, CFX, LFX, MFX, flumequine (FLQ), and NA were purchased polymerization. The purpose of the NIP is to be used as a control. from Sigma-Aldrich (St. Louis, MO, USA), while sodium borohy- Electrochemical Measurements. All electrochemical measure- dride, potassium hexacyanoferrate(II) trihydrate, and potassium ments were performed at room temperature employing a standard hexacyanoferrate(III) were purchased from Panreac (Barcelona, three-electrode cell made up of a Ag/AgCl (3 M KCl) reference Spain). Ethanol (EtOH, 96 vol %) and NaOH were bought from electrode, a Pt wire as the auxiliary electrode, and the MIP(FQs)@ Scharlau (Barcelona, Spain). Au-fMWCNT- or NIP@Au-fMWCNT-modified GECs as the work- All solutions employed in the experiments were prepared using ing electrode. Both cyclic voltammetry (CV) and electrochemical deionized water from a Milli-Q System (Millipore, Billerica, MA, impedance spectroscopy (EIS) for electrochemical characterization USA). Various buffer solutions were used in this study for different were conducted in a 5 mM [Fe(CN)6]3−/4− solution in PBS (pH 7.0). purposes. Concretely, 50 mM phosphate buffer saline (PBS) at pH For CV measurements, the potential was scanned in the range from 7.0 containing 100 mM potassium chloride was used for the −0.2 to +0.7 V at a scan rate of 50 mV·s−1, while for EIS, a frequency 3319 https://doi.org/10.1021/acssensors.2c01260 ACS Sens. 2022, 7, 3318−3325 ACS Sensors pubs.acs.org/acssensors Article Figure 1. TEM images of (a) fMWCNTs and (b) Au-fMWCNTs. SEM images of MIP(CFX)@Au-fMWCNTs on GEC (c) before and (d) after the extraction of the template. Electrochemical characterization of the different steps involved from the preparation of the sensors to the actual sensing by means of (e) CV and (f) EIS employing a 5 mM [Fe(CN)6]3−/4− solution in PBS: (from top to bottom) before and after extraction of the template; MIP rebinding; and NIP after the extraction process. range from 100 kHz to 100 mHz with a fixed AC amplitude of 10 mV where multiple antibiotics are used in combination, or for quality and an applied potential of +0.24 V was considered. control purpose during drug production, the above antibiotic drugs For the analysis of FQs, differential pulse voltammetry (DPV) were analyzed both individually or in mixtures. On the other hand, measurements were performed by scanning the potential from +0.6 to considering the urine excretion rate of FQ antibiotics,25 the analysis of +1.5 V with a step potential of 5 mV and a pulse amplitude of 50 mV FQs in human urine was also attempted. Urine samples were collected without stirring. Prior to each measurement, the electrodes were in sterile bottles from volunteers, diluted 40 times with acetate buffer incubated in the solution to be analyzed for 300 s under stirring to (pH 4.5), and then spiked with the FQ stocks. Again, the dilution of enhance the enrichment of the analytes into the MIP material. To urine samples was required so as to ensure that the concentration falls regenerate the electrodes, a fixed potential of +1.2 V was applied for within the experimental domain of the model. All real samples were 90 s in PBS (pH 10.0) after each measurement. measured under the same conditions as already described, without Sample Preparation. Given the different solubilities of the any further pretreatment. different FQs, the stock solutions of CFX, LFX, MFX, and NA were Data Processing. Multivariate data analysis was carried out in prepared in acetate buffer (pH 3.5), while the stock solution of FLQ Matlab 7.1 (MathWorks, Natick, MA, USA) by means of specific was prepared in 0.1 M NaOH solution. In order to characterize the routines developed by the authors, employing its Statistics and Neural analytical response of the developed sensors toward each of the Network toolboxes. The voltammetric responses of the sensor array analytes, individual calibration curves were constructed by measuring were combined and compressed by employing discrete cosine solutions of increasing concentrations prepared by proper dilution of transform (DCT), which allowed to reduce the dimensionality of the FQs’ stock solutions in buffer (pH 4.5). the input signals while preserving the relevant information.26 Next, For the simultaneous quantification of CFX, LFX, and MFX, two the obtained coefficients were either submitted to principal subsets of samples were prepared and measured under identical component analysis (PCA) or artificial neural networks (ANNs). conditions in a randomized order. On the one side, the training subset was composed by 27 samples based on a 33 tilted factorial design24 in which the concentrations of FQs were varied in the range from 2 to 300 μM for CFX, LFX, and MFX. On the other side, to evaluate the ■ RESULTS AND DISCUSSION Design of an MIP(FQs)@Au-fMWCNT-Based ET. The performance of the proposed model, the testing subset was composed main goal of this study is the development of a sensitive ET for by 10 samples that were randomly distributed in the concentration the simultaneous identification and quantification of FQs domain defined by the factorial design. To demonstrate its applicability, the MIP-based ET was applied to (Figure S2). In this regard, the fabrication of MIP-based the determination of the FQs under study both in commercial electrochemical sensors toward FQs (MIP(FQs)) is based on pharmaceuticals and biological fluids. On the one hand, commercial the electropolymerization approach. To further improve the antibiotic drugs were bought from the local drug stores with the performance of the MIP-based sensors, fMWCNTs that have doctor’s prescription. Specifically, four different FQ antibiotic drugs high conductivity and Au NPs that have electrocatalytic were evaluated: Cetraxal plus (3 mg·mL−1 CFX and 0.25 mg·mL−1 properties were incorporated before the electropolymerization Fluocinolone acetonide, Laboratorios SALVAT, S.A., Barcelona, of MIP(FQs) films, as depicted in Scheme S1a. Upon synthesis Spain), Ciprofloxacin Normon (250 mg CFX, Normon Laboratories, of the different MIPs, the sensors were combined into a sensor S.A., Madrid, Spain), Levofloxacin Aurovitas (500 mg LFX, APL Swift Services Limited, Birzebbugia, Malta), and Moxifloxacin Qualigen array to achieve the simultaneous determination of the three (400 mg MFX, Pharmathen, S.A., Attiki, Greece). For the analysis, FQs in the mixtures (Scheme S1b). drugs were dissolved or diluted in acetate buffer (pH 4.5) directly to One of the key parameters for MIP-based sensors is the make the expected concentrations fit in the experimental domain of thickness of the MIP films, which is critical for mass and the built model. Moreover, as there might be cases in the clinical use electron transport. If electropolymerization is carried out under 3320 https://doi.org/10.1021/acssensors.2c01260 ACS Sens. 2022, 7, 3318−3325 ACS Sensors pubs.acs.org/acssensors Article Figure 2. (a−d) Baseline-corrected DPV curves of (a) MIP(CFX)@Au-fMWCNTs, (b) MIP(LFX)@Au-fMWCNTs, (c) MIP(MFX)@Au- fMWCNTs, and (d) NIP@Au-fMWCNTs toward 10 μM solutions of: (black) CFX, (red) LFX, and (blue) MFX. (e−g) Calibration plots of the three target FQs (e) CFX, (f) CFX, and (g) MFX with the MIP- and NIP-based sensors. potentiostatic conditions, the thickness of the film can be easily decorated with small particles on their surface, corresponding controlled by adjusting the deposition charge. However, in our to the Au NPs. case, polymerization was carried out by CV under nonstirred Since different MIPs were obtained by introducing different conditions, where the mass transfer is controlled by the FQ molecules during the electropolymerization process, the diffusion process and the creation of the molecular imprints is MIP(CFX) film grown onto Au-fMWCNTs will be presented favored by the diffusion of the electroactive template toward as an example. The morphology of the MIP(CFX)@Au- the electrode, generating a far higher number of recognition fMWCNTs before and after the extraction of the template sites.27,28 Thus, the film thickness was optimized by adjusting molecule was characterized by SEM, as shown in Figure 1c,d, the number of cycles of the electropolymerization in the range respectively. From Figure 1c, it can be seen how the diameter from 5 to 35 cycles. As shown in Figure S3a, the current peak of Au-fMWCNTs increased significantly due to the growth of of MIP(CFX)@Au-fMWCNTs is higher when 10 cycles were the MIP film on the surface as this is now larger than 100 nm. selected, as thicker films cannot ensure an efficient electron It is worth noting that a thin and uniform MIP film grew along and mass transport. Moreover, ensuring a sufficient number of the surface of the Au-fMWCNTs instead of integrally specific sites is critical for in the specific recognition of the aggregating as a top layer that covers the whole. The peculiar template by the MIP. While a low concentration of the morphology is attributed to the electropolymerization method template might lead to poor imprinting, a too large employed, which not only allows a better control over the distribution and thickness of the MIP film but also facilitates concentration is not good for the formation of the polymers. surface imprinting and formation of a core−shell structure.29 To this end, the molar ratio of Py to CFX in the Finally, from the comparison of Figure 1c,d, it can be seen how electropolymerization solution was also optimized (Figure there are no significant morphological changes in the S3b), from which a ratio of 5:1 was selected as optimal. MIP(CFX)@Au-fMWCNT structure after the extraction of Morphological Characterization of MIP(FQs)@Au- the template and immersion in the EtOH−NaOH mixture, fMWCNT Sensors. Both SEM and TEM were used to confirming the stability of the developed nanocomposite characterize the morphology of the fabricated nanomaterials. material. Moreover, the SEM image of the NIP@Au- Figure S4a shows the SEM top view image of the bare GEC fMWCNTs (Figure S4c) also shows a similar morphology, electrode, showing a smooth and uniform surface, which is in confirming the expected similarities between the different contrast to the images after Au-fMWCNT casting (Figure S4b) synthesized MIP and NIP films. and electropolymerization (Figure S4c), undoubtedly confirm- Electrochemical Characterization of MIP(FQs)@Au- ing the modification of the electrode. As mentioned earlier, the fMWCNT Sensors. The electrochemical behavior of the combination of MWCNTs and Au NPs with MIP films allows sensors was investigated by CV and EIS in a 5 mM obtaining a synergistic effect that results in an improvement of [Fe(CN)6]3−/4− solution. First, the response of the bare the sensor’s performance. Figure 1a,b shows the TEM images GEC was evaluated (Figure S5a), from which the reversible of fMWCNTs before and after their decoration with Au NPs, redox peaks of [Fe(CN)6]3−/4− were observed. Next, this was from which it can be seen that the diameter of the CNTs in compared to the responses of fMWCNT- and Au-fMWCNT- both cases is certainly smaller than 100 nm. Furthermore, the modified GECs, demonstrating that loading of fMWCNTs surface of the CNTs seen in Figure 1a is smooth and without with Au NPs enhanced the electron transfer, which is reflected particles on them, while the CNTs in Figure 1b are clearly by the larger peaks in CV and a much lower charge-transfer 3321 https://doi.org/10.1021/acssensors.2c01260 ACS Sens. 2022, 7, 3318−3325 ACS Sensors pubs.acs.org/acssensors Article resistance in EIS. Subsequently, an analogous comparison was To this aim, solutions with increasing concentrations of conducted between the bare MIP sensor and that incorporat- different FQs were analyzed using the different sensors by DPV ing Au-fMWCNTs (Figure S5c,d). As it can be seen, under the optimized conditions. An extract of the voltammetric MIP(CFX)@Au-fMWCNTs provided a better electrochemical responses is provided in Figure 2a−d (extracted from Figure performance than MIP(CFX), but worse than Au-fMWCNTs S8a−d), where it can be seen how the three different MIP- (Figure S5c,d). This is due to the poor electrical conductivity based sensors show a voltammetric response to the three FQs and blocked electron transport originated in the insulating (CFX, LFX, and MFX, respectively) and so does the NIP- features of the polymeric matrix (Figures 2 and S5). After the based sensor. However, the differences in the peak position extraction of the template, we can observe a significant and height should be noted. As anticipated, the NIP@Au- improvement in the electrochemical behavior for the MIP fMWCNT-based sensor shows the lowest response to the sensor, both in the CV and EIS analyses, because the sites three FQs compared with the other three MIP-based sensors created during the extraction process facilitate electron and due to the lack of recognition sites toward the different FQs. mass transport, while the NIP sensor still showed quite a Furthermore, when comparing the responses of the specific resistive behavior. Lastly, rebinding of these sites in the MIP MIP(FQ)@Au-fMWCNT-based sensors to the different FQs, film with the template was evaluated, and the sluggish diffusion it can be found that each MIP shows the highest response to its of [Fe(CN)6]3−/4− due to the “gate-controlled effect” is clearly specific template (Figure 2). That is, MIP(CFX)@Au- reflected in the graph. Consequently, taking the above results fMWCNTs clearly shows the highest peak to CFX, while from CV and EIS, the suitability of the developed MIP-based MIP(LFX)@Au-fMWCNTs shows to LFX, confirming the electrochemical sensor was confirmed. specific recognition of the different MIP(FQ) films. The fact Optimization of the Measuring Conditions. In order to that the MIP(MFX)@Au-fMWCNT response to MFX is not enhance the voltammetric responses of the sensors, the comparatively larger is attributed to the intrinsically higher measuring conditions of FQs were optimized. As previously electrochemical response of the other FQs, as is seen for the stated, the response of the MIP(CFX)@Au-fMWCNT sensor NIP sensor. In this regard, it can be deduced that the similar is shown as an example, taking the peak height from the DPV structure and nature of FQs, all based on different substitutions measurement toward a stock solution of CFX as the parameter made to the quinolone ring, hinder the attainment of a highly to maximize. First, the influence of the pH of the buffer was specific MIP-based sensor, but lead to a class-selective material investigated in the range from 3.5 to 5.5 (Figure S6), obtaining rather than a highly specific one�a situation that is well suited pH 4.5 as the most suitable. for ET approaches, as chemometrics may improve its Second, the incubation time was also optimized as an specificity. Indeed, this cross-sensitivity observed is completely essential factor for the binding of the target analytes into the analogous to that of the “dummy” MIPs, where a relatively MIP film sites. Figure S7a shows how the response of the MIP- similar molecule to the target analyte is used for imprinting, based sensor increases with the incubation time, until reaching although later the MIP is used not to analyze the template a maximum at 300 s, indicating that the saturation had been molecule but the target analyte.30 reached. In addition, the time of the electrochemical treatment To further confirm the behavior of the MIP, the response of required for the sensor’s regeneration after the sensing was also MIP(CFX)@Au-fMWCNTs was compared to those of the studied, observing that for the measurement of a 10 μM CFX bare electrode and an electrode modified with Au-fMWCNTs, solution, 30 s were sufficient to recover the baseline (Figure but not MIP (Figure S9). Furthermore, the response toward an S7b). However, to ensure that proper regeneration was antibiotic from a different family (vancomycin) as well as other achieved even when higher target concentrations were tested, types of compounds that could also be found either in tablets 90 s were chosen as the regeneration time for further or urine such as glucose, ascorbic acid, and paracetamol were experiments. also evaluated. From those findings, it can be seen how the Lastly, the stability of the sensors under those conditions MIP contributes to increase the voltammetric responses was evaluated to ensure that they could withstand a significant obtained for FQs, while the response for other potentially number of measurements necessary for any analytical interfering compounds is decreased. Additionally, it has to application, but especially when dealing with ETs. To this been remarked, that apart, these current maxima may appear at aim, CFX was selected as the substance to evaluate the different potential, therefore facilitating its discrimination. variation on its voltammetric response upon successive After preliminary assessment of the responses of the measurements, assuming that a similar behavior will be different sensors toward the different FQs, appropriate obtained for the rest of the drugs. Employing three different calibration of each MIP(FQs)@Au-fMWCNT and NIP@Au- electrodes, a 5 μM stock solution was measured for 25 fMWCNT was carried out. From Figure S10, it can be seen consecutive times while measuring also a blank (acetate buffer how the peak currents increase steadily when FQs concen- solution) in between each measurement to evaluate the trations become greater. The observed peaks are attributed to repeatability of the sensors. Thus, each sensor was used for the oxidation of the piperazine ring common to the different 50 consecutive measurements. Under these conditions, the FQs (Figure S11).31,32 A hyperbolic response was observed as obtained relative standard deviation (% RSD) of the measured evidenced in Figure 2e−g, from which the dissociation peak height corresponding to CFX was 8.7% between the constant (kD) and the maximum binding response at saturation different sensors over all the measurements. (Bmax) were calculated by fitting those to a one site saturation Calibration of MIP(FQs)@Au-fMWCNTs to FQs. One of ligand binding model (Langmuir model, Table S1). Fur- the fundamental steps when developing sensors is the thermore, a logarithmic relationship could be employed characterization of their response in terms of linearity, between the concentrations and the corresponding peak sensitivity, limit of detection (LOD), reproducibility, etc. heights for low concentrations (from which the LOD could Furthermore, when developing an ET, it is also important to be calculated). In all the cases, a good linear relationship for assess the cross-responses toward the different target analytes. the three FQs is obtained with R2 values close to 1. 3322 https://doi.org/10.1021/acssensors.2c01260 ACS Sens. 2022, 7, 3318−3325 ACS Sensors pubs.acs.org/acssensors Article Qualitative Analysis of FQs. Despite MIP-based sensors manner, we ensured a more realistic assessment as different are meant to be highly specific, those might still show certain data subsets are used for each step. cross-response when analogous molecules are analyzed, for After an iterative trial-and-error process in which different instance, when analyzing different FQs. One of the common neural network parameters were tuned, the final architecture solutions to the limited selectivity may be to synthesize more had 48 input neurons (12 coeffs. × 4 sensors) in the input advanced MIP-based materials. For example, in this case, layer, 7 neurons and logsig transfer function in the hidden layer, limited selectivity for FQs may be partially attributed to the use and 3 neurons and tansig transfer function in the output layer of electropolymerized Py instead of a more sophisticated (one for each of the FQs under study). Subsequently, polymerization method employing carefully screened mono- comparison graphs of predicted versus expected concentrations mers and cross-linkers. However, the truth is that a similar were built both for the training and testing subsets to easily behavior has also been reported with other polymerization assess their performance (Figure 4). Moreover, the regression strategies.33,34 In this direction, herein we propose the parameters were also calculated to numerically assess the combination of chemometrics with MIPs as a straightforward goodness of the modeling. As can be observed from Figure 4, solution to further improve its selectivity. Such an approach for each FQ, a good trend was obtained with regression lines allows to shift the complexity from the chemical to the data close to the ideal ones (y = x, slope, intercept, and correlation processing side; while in turn further advantages might also be coefficient values of 1, 0, and 1, respectively). Consequently, attained, that is, to cope with nonlinear systems or to we confirmed the potential of the approach not only to counterbalance possible matrix effects.9,35 discriminate between different FQs (as shown previously with In order to demonstrate such a statement, the analysis of the PCA), but also to individually quantify them even in stock solutions of five different FQs was carried out under the mixtures. optimized conditions, employing the three developed MIP- To somehow assess the improvement derived from the and NIP-based sensors. Next, the voltammetric responses were usage of the MIP-based sensor array, the same set of samples compressed by means of DCT from 114 points down to 12 was also measured by employing a bare GEC electrode. The coefficients with high fidelity as evidenced from Figure S12 and recorded data were then processed in the same way, and the submitted to PCA (Figure 3). First, it should be noted that ca. normalized root mean square error (NRMSE) for both approaches was compared. After optimization of the DCT− ANN topology, a total NRMSE of 0.134 was obtained for the bare electrode; a value that is significantly larger than that of the MIP-based ET, which was 0.033. Finally, to further demonstrate the applicability of the proposed MIP-based ET, some pharmaceutical and human urine samples were analyzed. For both of them, no other sample pretreatment than its proper dilution to fit within the evaluated range of concentrations was carried out. As before, the corresponding readings of the sensors were compressed with DCT and interpolated into the built ANN model. Those are also plotted in Figure 4, where a good agreement between the expected and predicted values is also observed despite the relatively complex matrixes of the pharmaceutical tablets (excipients, fillers, disintegrants, binders, etc.) and urine (urea, inorganic salts, creatinine, uric acid, proteins, vitamins, etc.). Again, confirming the suitability and advantages derived Figure 3. Score plot obtained from DCT−PCA of the voltammetric from the combination of MIPs and ETs, which allowed to responses of the MIP-based ET toward five different FQs. Ellipses address not only the lack of selectivity of MIPs, but the plotted correspond to 95% confidence limits for each of the clusters. successful quantification of FQs mixtures in real samples. 72.4% of the original data variance is now represented by only those two first principal components (PCs), providing a better ■ CONCLUSIONS Herein, a sensitive ET based on different MIPs incorporating representation of sample measurements that allows an easier Au NP-decorated MWCNTs has been developed for the assessment of its (dis)similarities. Second, we can see how analysis of FQs. In contrast to the conventional methods for different clusters corresponding to each of the FQs appear and MIP synthesis, in this study, MIP films were electro- how these clusters separate from each other. Consequently, the polymerized directly on the surface of Au-fMWCNTs. The superb ability of the developed ET system to discriminate FQs use of electropolymerization facilitates the integration of MIPs is well denoted. with voltammetric sensors, while the combination of CNT’s Quantification of FQs. In view of the above results, the large surface/mass ratio and high conductivity and the catalytic quantification of mixtures of different FQs was attempted, effect of Au NPs with the specific recognition of MIPs leads to namely, mixtures of CFX, LFX, and MFX. To this aim, the set a synergistic effect, which leads to an improved performance of of samples described in the Experimental Section was analyzed, the developed sensors. and, as before, the obtained responses were compressed by The developed sensors were morphologically and electro- means of DCT. In this case, ANNs were chosen for data chemically characterized to confirm the modification of the modeling, making use of the samples of the training subset to electrodes and assess their analytical responses. Despite the optimize the topology of the neural network, while those of the good performance obtained, the similar structure of the test subset were used to assess the model performance. In this different FQs hinders the attainment of highly specific MIPs, 3323 https://doi.org/10.1021/acssensors.2c01260 ACS Sens. 2022, 7, 3318−3325 ACS Sensors pubs.acs.org/acssensors Article Figure 4. Modeling ability of DCT−ANN. Comparison graphs of obtained versus expected concentrations for (a) CFX, (b) LFX, and (c) MFX, for both the training (solid circle, solid line) and testing (open circle, dotted line) subsets. The dashed line corresponds to the ideal comparison line (y = x). The results of the analysis of the pharmaceutical samples (solid diamond, blue) and spiked human urine samples (solid square, orange) are also plotted. leading to a class-selective material rather than a highly specific Complete contact information is available at: one. In this direction, to improve their specificity, the https://pubs.acs.org/10.1021/acssensors.2c01260 developed MIP-based sensors were combined into a sensor array to develop an MIP-based ET. Thanks to the use of Notes chemometrics, it was possible not only to discriminate between The authors declare no competing financial interest. different FQs, but also to achieve the simultaneous quantification of CFX, LFX, and MFX. Finally, the developed ET was applied to the analysis of pharmaceutical and biological samples, with satisfactory recovery values obtained, confirming ■ ACKNOWLEDGMENTS The authors acknowledge the support of the Spanish Ministry the lack of matrix effects. of Science and Innovation (project PID2019-107102RB-C21/ Overall, the proposed approach is an appealing and AEI/10.13039/501100011033). promising tool for the determination of FQs, but more importantly, it paves the way for the development of similar applications. On the one side, electropolymerization is a facile ■ REFERENCES (1) Majdinasab, M.; Mitsubayashi, K.; Marty, J. L. Optical and approach for the fabrication of MIP-based sensors, while the Electrochemical Sensors and Biosensors for the Detection of integration of nanomaterials leads to an improved electro- Quinolones. Trends Biotechnol. 2019, 37, 898−915. chemical behavior. On the other side, applied machine learning (2) Kergaravat, S. V.; Gagneten, A. M.; Hernandez, S. R. provides a straightforward approach to tackle MIP’s cross- Development of an Electrochemical Method for the Detection of response. Quinolones: Application to Cladoceran Ecotoxicity Studies. Micro- ■ ASSOCIATED CONTENT * Supporting Information sı chem. J. 2018, 141, 279−286. (3) Rusu, A.; Hancu, G.; Uivaroşi, V. Fluoroquinolone Pollution of Food, Water and Soil, and Bacterial Resistance. Environ. Chem. Lett. 2014, 13, 21−36. The Supporting Information is available free of charge at (4) Marchant, J. When Antibiotics Turn Toxic. Nature 2018, 555, https://pubs.acs.org/doi/10.1021/acssensors.2c01260. 431−433. Molecular structure of the analytes; SEM images of the (5) Leder, C.; Suk, M.; Lorenz, S.; Rastogi, T.; Peifer, C.; sensors; voltammograms recorded during electropoly- Kietzmann, M.; Jonas, D.; Buck, M.; Pahl, A.; Kümmerer, K. merization; optimization of the experimental conditions; Reducing Environmental Pollution by Antibiotics through Design for characterization of the voltammetric responses; and Environmental Degradation. ACS Sustainable Chem. Eng. 2021, 9, detailed results of the data treatment (PDF) 9358−9368. (6) Rudnicki, K.; Sipa, K.; Brycht, M.; Borgul, P.; Skrzypek, S.; ■ AUTHOR INFORMATION Corresponding Author Poltorak, L. Electrochemical Sensing of Fluoroquinolone Antibiotics. TrAC, Trends Anal. Chem. 2020, 128, No. 115907. (7) Gayen, P.; Chaplin, B. P. Selective Electrochemical Detection of Manel del Valle − Sensors and Biosensors Group, Department Ciprofloxacin with a Porous Nafion/Multiwalled Carbon Nanotube of Chemistry, Universitat Autònoma de Barcelona, Faculty of Composite Film Electrode. ACS Appl. Mater. Interfaces 2016, 8, 1615−1626. Sciences, 08193 Bellaterra, Barcelona, Spain; orcid.org/ (8) Ye, C.; Chen, X.; Zhang, D.; Xu, J.; Xi, H.; Wu, T.; Deng, D.; 0000-0002-1032-8611; Email: manel.delvalle@uab.cat Xiong, C.; Zhang, J.; Huang, G. Study on the Properties and Reaction Authors Mechanism of Polypyrrole@Norfloxacin Molecularly Imprinted Electrochemical Sensor Based on Three-Dimensional CoFe-MOFs/ Mingyue Wang − Sensors and Biosensors Group, Department AuNPs. Electrochim. Acta 2021, 379, No. 138174. of Chemistry, Universitat Autònoma de Barcelona, Faculty of (9) Cetó, X.; Voelcker, N. H.; Prieto-Simón, B. Bioelectronic Sciences, 08193 Bellaterra, Barcelona, Spain Tongues: New Trends and Applications in Water and Food Analysis. Xavier Cetó − Sensors and Biosensors Group, Department of Biosens. Bioelectron. 2016, 79, 608−626. Chemistry, Universitat Autònoma de Barcelona, Faculty of (10) Cui, F.; Yue, Y.; Zhang, Y.; Zhang, Z.; Zhou, H. S. Advancing Sciences, 08193 Bellaterra, Barcelona, Spain Biosensors with Machine Learning. ACS Sens. 2020, 5, 3346−3364. 3324 https://doi.org/10.1021/acssensors.2c01260 ACS Sens. 2022, 7, 3318−3325 ACS Sensors pubs.acs.org/acssensors Article (11) Wasilewski, T.; Kamysz, W.; Gebicki, J. Bioelectronic Tongue: (31) de Faria, L. V.; Lisboa, T. P.; Campos, N.; Alves, G. F.; Matos, Current Status and Perspectives. Biosens. Bioelectron. 2020, 150, M. A. C.; Matos, R. C.; Munoz, R. A. A. Electrochemical methods for No. 111923. the determination of antibiotic residues in milk: A critical review. (12) Herrera-Chacón, A.; Cetó, X.; del Valle, M. Molecularly Anal. Chim. Acta 2021, 1173, No. 338569. imprinted polymers - towards electrochemical sensors and electronic (32) Akilarasan, M.; Tamilalagan, E.; Chen, S.-M.; Maheshwaran, S.; tongues. Anal. Bioanal. Chem. 2021, 413, 6117−6140. Fan, C.-H.; Habila, M. A.; Sillanpä, M. Rational synthesis of rare-earth (13) Huynh, T. P.; Kutner, W. Molecularly imprinted polymers as lanthanum molybdate covered reduced graphene oxide nano- recognition materials for electronic tongues. Biosens. Bioelectron. 2015, composites for the voltammetric detection of Moxifloxacin hydro- 74, 856−864. chloride. Bioelectrochemistry 2022, 146, No. 108145. (14) Ahmad, O. S.; Bedwell, T. S.; Esen, C.; Garcia-Cruz, A.; (33) Benito-Pena, E.; Martins, S.; Orellana, G.; Moreno-Bondi, M. Piletsky, S. A. Molecularly Imprinted Polymers in Electrochemical and C. Water-Compatible Molecularly Imprinted Polymer for the Optical Sensors. Trends Biotechnol. 2019, 37, 294−309. Selective Recognition of Fluoroquinolone Antibiotics in Biological (15) BelBruno, J. J. Molecularly Imprinted Polymers. Chem. Rev. Samples. Anal. Bioanal. Chem. 2009, 393, 235−245. 2019, 119, 94−119. (34) Urraca, J. L.; Castellari, M.; Barrios, C. A.; Moreno-Bondi, M. (16) Gui, R.; Jin, H.; Guo, H.; Wang, Z. Recent Advances and C. Multiresidue Analysis of Fluoroquinolone Antimicrobials in Future Prospects in Molecularly Imprinted Polymers-Based Electro- Chicken Meat by Molecularly Imprinted Solid-Phase Extraction and chemical Biosensors. Biosens. Bioelectron. 2018, 100, 56−70. High Performance Liquid Chromatography. J. Chromatogr. A 2014, (17) Rebelo, P.; Costa-Rama, E.; Seguro, I.; Pacheco, J. G.; Nouws, 1343, 1−9. H. P. A.; Cordeiro, M.; Delerue-Matos, C. Molecularly Imprinted (35) del Valle, M. Electronic Tongues Employing Electrochemical Polymer-Based Electrochemical Sensors for Environmental Analysis. Sensors. Electroanalysis 2010, 14, 1539−1555. Biosens. Bioelectron. 2021, 172, No. 112719. (18) Karimian, N.; Stortini, A. M.; Moretto, L. M.; Costantino, C.; Bogialli, S.; Ugo, P. Electrochemosensor for Trace Analysis of Perfluorooctanesulfonate in Water Based on a Molecularly Imprinted Poly(O-Phenylenediamine) Polymer. ACS Sens. 2018, 3, 1291−1298. (19) Ozcelikay, G.; Kaya, S. I.; Ozkan, E.; Cetinkaya, A.; Nemutlu, E.; Kır, S.; Ozkan, S. A. Sensor-Based MIP Technologies for Targeted Metabolomics Analysis. TrAC, Trends Anal. Chem. 2022, 146, No. 116487. (20) Anirudhan, T. S.; Athira, V. S.; Nair, S. S. Detection of Chlorpyrifos Based on Molecular Imprinting with a Conducting Polythiophene Copolymer Loaded on Multi-Walled Carbon Nano- tubes. Food Chem. 2022, 381, No. 132010. (21) Surya, S. G.; Khatoon, S.; Ait Lahcen, A.; Nguyen, A. T. H.; Dzantiev, B. B.; Tarannum, N.; Salama, K. N. A Chitosan Gold Nanoparticles Molecularly Imprinted Polymer Based Ciprofloxacin Sensor. RSC Adv. 2020, 10, 12823−12832. (22) Alegret, S.; Alonso, J.; Bartroli, J.; Céspedes, F.; Martinez Fabregas, E.; del Valle, M. Amperometric Biosensors Based on Bulk- Modified Epoxy Graphite Biocomposites. Sensor. Mater. 1996, 8, 147−153. (23) Sá, A.; Cipri, A.; González-Calabuig, A.; Ramos Stradiotto, N.; del Valle, M. Resolution of Galactose, Glucose, Xylose and Mannose in Sugarcane Bagasse Employing a Voltammetric Electronic Tongue Formed by Metals Oxy-Hydroxide/MWCNT Modified Electrodes. Sens. Actuators B Chem. 2016, 222, 645−653. (24) Cetó, X.; Céspedes, F.; Pividori, M. I.; Gutiérrez, J. M.; del Valle, M. Resolution of Phenolic Antioxidant Mixtures Employing a Voltammetric Bio-Electronic Tongue. Analyst 2012, 137, 349−356. (25) Rodvold, K. A.; Piscitelli, S. C. New Oral Macrolide and Fluoroquinolone Antibiotics: An Overview of Pharmacokinetics, Interactions, and Safety. Clin. Infect. Dis. 1993, 17, S192−S199. (26) Cetó, X.; Pérez, S. Voltammetric Electronic Tongue for Vinegar Fingerprinting. Talanta 2020, 219, No. 121253. (27) Ö zcan, L.; Ş ahin, Y. Determination of paracetamol based on electropolymerized-molecularly imprinted polypyrrole modified pen- cil graphite electrode. Sens. Actuators B Chem. 2007, 127, 362−369. (28) Sharma, P.; Pietrzyk-Le, A.; D’Souza, F.; Kutner, W. Electrochemically synthesized polymers in molecular imprinting for chemical sensing. Anal. Bioanal. Chem. 2012, 402, 3177−3204. (29) Ansari, S. Combination of Molecularly Imprinted Polymers and Carbon Nanomaterials as a Versatile Biosensing Tool in Sample Analysis: Recent Applications and Challenges. TrAC, Trends Anal. Chem. 2017, 93, 134−151. (30) Chen, R.-N.; Kang, S.-H.; Li, J.; Lu, L.-N.; Luo, X.-P.; Wu, L. Comparison and recent progress of molecular imprinting technology and dummy template molecular imprinting technology. Anal. Methods 2021, 13, 4538−4556. 3325 https://doi.org/10.1021/acssensors.2c01260 ACS Sens. 2022, 7, 3318−3325