Biosensors and Bioelectronics 198 (2022) 113807 Contents lists available at ScienceDirect Biosensors and Bioelectronics journal homepage: www.elsevier.com/locate/bios A novel electronic tongue using electropolymerized molecularly imprinted polymers for the simultaneous determination of active pharmaceutical ingredients Mingyue Wang, Xavier Cetó, Manel del Valle * Sensors and Biosensors Group, Department of Chemistry, Universitat Autònoma de Barcelona, Edifici Cn, 08193, Bellaterra, Barcelona, Spain A R T I C L E I N F O A B S T R A C T Keywords: The combination of chemometrics and electrochemical sensors modified with molecularly imprinted polymers Molecularly imprinted polymers (MIPs) towards the development of MIP-based electronic tongues (ETs) was explored herein. To demonstrate the Electronic tongues potential of such an approach, the simultaneous determination of paracetamol, ascorbic acid and uric acid Electrochemical sensors mixtures in pharmaceutical samples was evaluated. To this aim, MIP-based sensors for the different compounds Artificial neural networks Pharmaceutical analysis were prepared by in situ electropolymerization of pyrrole in the presence of p-toluenesulfonate anion (pTS− ), which acted as functional doping ion of the polypyrrole (PPy) MIP backbone. Morphological characterization of the MIPs was done by scanning electron microscopy (SEM), while functionalization of the electrodes was monitored electrochemically. Under the optimized measuring conditions, the developed sensors showed a good performance, with good linearity at the μM level (R2 > 0.992, limits of detection between 1 and 24 μM) as well as good repeatability (intra- and inter-day RSD values between 3 and 6% over 30 consecutive measurements). Finally, the quantification of the individual substances in different pharmaceutical samples was achieved by an artificial neural networks (ANNs) model, showing satisfactory agreement between expected and obtained values (R2 > 0.987). 1. Introduction overcome such scenarios, therapeutic drug monitoring (TDM) is required, so as to ensure individually-directed proper dosing. Therefore, The greatly improved life quality and the extended life expectancy of the accurate and rapid analysis of biomolecules and active pharma­ human being witnessed over the past decades can be partly attributed to ceutical ingredients (APIs) aroused keen interest of researchers, phar­ the ever-increasing use of pharmaceutical drugs against diseases and the maceutical industry and clinicians. more sophisticated methods for detecting key metabolite biomarkers Despite various analytical methods have already been applied for that allow the diagnosis of pathologies. TDM, including high performance liquid chromatography (HPLC), Traditionally, drugs are administered on the basis of factors such as spectrophotometry and titrimetry, drawbacks like being time- patients’ weight or age, which works for most of the cases. However, consuming, laborious, or expensive are inevitable. Compared with certain drugs have a narrower therapeutic window, which can result in a these conventional methods, the combination of electrochemical patient easily being under- or over-dosed. The same happens with methods with molecularly imprinted polymers (MIPs) can be a people who suffer altered metabolism or certain diseases. From this competitive alternative for the detection of biomolecules owing to their perspective, in comparison to mono-component drugs, multidrug for­ numerous advantages such as rapid response, low-cost and high sensi­ mulations are less harmful because of the smaller incorporated amounts tivity and selectivity (BelBruno, 2019; Gui et al., 2018). MIPs are of each active component. Nevertheless, certain combinations of drugs tailor-made biomimetic materials, sometimes referred to as plastic an­ are also likely to lead to unpredictable systemic levels (Hilberg et al., tibodies, not only possessing the “key and lock” specific recognition 2005; Meneghello et al., 2018). Thus, despite the great advantages that mechanism owing to the specific sites produced by template molecules, may offer, those still have some negative effects, especially in cases of but also with advantages over natural antibodies in terms of reusability, overdosing (Bergh et al., 2021; Root-Bernstein et al., 2020). To and higher mechanical and chemical stability in harsh conditions (Chen * Corresponding author. Sensors and Biosensors Group, Universitat Autònoma de Barcelona, Campus UAB, Edifici Cn, 08193, Bellaterra, Spain. E-mail addresses: manel.delvalle@uab.cat, manel.delvalle@gmail.com (M. del Valle). https://doi.org/10.1016/j.bios.2021.113807 Received 5 October 2021; Received in revised form 12 November 2021; Accepted 13 November 2021 Available online 17 November 2021 0956-5663/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). M. Wang et al. Biosensors and Bioelectronics 198 (2022) 113807 et al., 2011; Rebelo et al., 2021). Hence, MIP-based electrochemical the performance of the built sensors. Next, the registered voltammo­ sensors have been research hotspots for their broad-spectrum applica­ grams were analyzed by chemometric tools to achieve the simultaneous tions in environmental monitoring (Rebelo et al., 2021), disease diag­ quantification of the ternary mixtures. Benefiting from the comple­ nosis (Zaidi, 2018), food or medical safety evaluation (Cao et al., 2019) mentary advantages of MIPs structure and ETs approach, the MIP-based and so on (Herrera-Chacon et al., 2021; Wackerlig and Lieberzeit, 2015). sensor array exhibited superb sensitivity, selectivity and a wide detec­ Among the different approaches for the synthesis of MIPs, electro­ tion range in the simultaneous determination of PA, AA and UA. polymerization is simple, facile, highly reproducible and facilitates the Furthermore, the novel ET sensor was successfully applied in the anal­ direct integration of MIPs with voltammetric sensors (Arabi et al., 2021; ysis of real pharmaceutical samples. Herrera-Chacon et al., 2021). In this regard, pyrrole (Py) is a well-known functional monomer used to obtain imprinted structures without the 2. Experimental section necessity of cross-linkers or initiators (Kumar et al., 2021; Zhang et al., 2018). However, anions must be introduced into the polypyrrole (PPy) 2.1. Reagents and instruments backbones to ensure its structural stability. Although perchlorate (ClO4− ) is the most commonly used, different anions can greatly influ­ All the reagents were analytical reagent grade and used in the ex­ ence the properties of the resulting polymer and its morphology (Bolat periments without any further purification process. Solutions were et al., 2019; Carquigny et al., 2008; Raudsepp et al., 2008). It has been prepared using deionized water from a Milli-Q System (Millipore, Bill­ reported that p-toluenesulfonate (pTS− ) doped PPy film holds high erica, MA, USA). Pyrrole (Py), p-toluenesulfonic acid (pTS− ) sodium salt, electrical conductivity, good mechanical properties, higher flexibility potassium chloride, potassium hydrogenphosphate, potassium dihy­ and durability, and a micro-porous structure (Rajesh et al., 2004; drogenphosphate, paracetamol (PA), ascorbic acid (AA) and uric acid Raudsepp et al., 2008; Solanki et al., 2007). (UA) were purchased from Sigma-Aldrich (St. Louis, MO, USA). Phos­ Despite the improved sensitivity and selectivity of MIP-based elec­ phate buffered saline (PBS 50 mM, pH 7.0) used in this work contained trochemical sensors, those can still show some cross-response towards 100 mM potassium chloride. Graphite powder (particle size <50 μm) other molecules, especially with chemically analogous structures. was bought from BDH Laboratory Supplies (Poole, UK) and epoxy resin Moreover, when attempting the multi-determination of several com­ kit was supplied by Resineco green composites (Barcelona, Spain). pounds, the use of multivariate regression methods might demonstrate Scanning electron microscopy (SEM) was observed by EVO MA10 advantageous. In this direction, electronic tongues (ETs) approach may (Zeiss, Oberkochen, Germany) operated at 30 kV. help to overcome such issues (Cetó et al., 2016; del Valle, 2010; Her­ rera-Chacon et al., 2021). ETs are based on the combination of an array of (bio)sensors able to provide a wide and complete response of the 2.2. Preparation of the pTS− /PPy MIPs and NIP sensors analyzed species, plus a modelling tool able to interpret and extract meaningful data from the complex readings. However, even with the Hand-made graphite-epoxy composite electrodes (GEC, geometric large number of publications reporting the development of MIP-based area of 28.3 mm2) acted as platforms for the electropolymerization of electrochemical sensors and the great benefits that could be derived pTS− /PPy MIPs and non-imprinted polymer (NIP). The construction from the combination of MIPs with ET approach, few publications procedures for GEC are based on previous reports from the group exploiting this approach can be found (Herrera-Chacon et al., 2018; (Alegret et al., 1996). Firstly, the epoxy resin was mixed with its cor­ Huynh and Kutner, 2015). responding weight of hardener and graphite powder until completely In order to demonstrate the potential of MIP-based ETs for TDM, the homogenized. Then, the mixture was loaded into the cavity of a PVC quantitative analysis of paracetamol (PA), ascorbic acid (AA) and uric tube, in which a copper disk soldered to a connector had been previously acid (AA) mixtures is investigated herein as a proof-of-concept of what introduced, and dried in the oven at 40 ◦ C for 48h. Finally, the dried can be achieved. Moreover, having as ultimate goal, to provide a generic GECs are polished with sandpapers of decreasing grain size until a protocol for the development of MIP-based ETs towards almost any smooth surface was obtained. other analyte. For the electropolymerization of pTS− /PPy MIPs onto the GEC In this context, overdose of PA, which is widely used as an analgesic electrodes, 20 mL of an aqueous solution containing 0.05 M pyrrole and antipyretic drug, could lead to fatal hepato- and nephro-toxic effects monomer, 0.1 M pTS− and 0.02M of the template molecule (either PA, (Vale and Proudfoot, 1995). Precisely, its wide availability in conjunc­ AA or UA) were purged with nitrogen for 10 min before cycling the tion with its relatively high toxicity lead to PA being one of the most potential from − 0.6 V to +0.8 V vs. Ag/AgCl at a scan rate of 100 mV common causes of poisoning worldwide (Prince et al., 2000). Similarly, s− 1. In this manner, replacing the template solution, the different pTS− / AA is a valuable vitamin that is extensively employed in combination PPy MIPs could be obtained. For PA and AA, those were dissolved in with other medicines (such as PA) in the treatment of the common cold Milli-Q water directly. But in the case of UA, due to its lower solubility, (Naidu, 2003). It has been proven that the deficiency of AA is closely UA was firstly dissolved in a NaOH solution (20 mL, pH 10.0) by stirring related to terrible scurvy, the disturbance in the collagen metabolism and sonicating, and then pyrrole and pTS− were added into the filtered and the reduction of women’s fertility (Brambilla et al., 2018; Dosedel UA solution. Next, in order to extract the template molecules from the et al., 2021; Grewal et al., 2019). On the contrary, AA abuse can cause PPy film, the electrodes were dipped into PBS and cycled 35 times at a deep vein thrombosis (Bergh et al., 2021). Moreover, some studies have scan rate of 50 mV s− 1 in the range of − 0.6 V to +1.0 V vs. Ag/AgCl. suggested that AA intake is related to UA concentration in body fluids Finally, the electrodes were rinsed and kept dry at 4 ◦ C in the fridge (Huang et al., 2005; Juraschek et al., 2011); the latter being associated when not in use. with various diseases, such as gout, hyperuricemia or Lesch–Nyhan For control purposes, the pTS− /PPy NIP modified GEC was prepared syndrome, arthritis, diabetes, high cholesterol, renal, neurological, following the same procedure as above-mentioned, but without the cardiovascular and kidney diseases (Sieminska et al., 2020). addition of the template molecules into the polymerization solution. The Based on the aforementioned, herein we report on the preparation of lack of the template molecule during the polymerization leads to the three PPy MIP-based sensors for PA, AA and UA in the presence of pTS− formation of a polymer with non-specific and non-homogeneous sites, anion through a generic, facile and simple electropolymerization what should clearly demote its selectivity towards the target analyte. method, and their subsequent integration into a sensor array to Thus, by benchmarking the performance of the MIP against the NIP, it is construct a voltammetric ET for the simultaneous determination of those possible to demonstrate and evaluate the imprinting effect, as it provides three analytes in pharmaceutical samples. The voltammetric responses a measurement of the specific interactions in the generated MIP sites to of the developed MIP-based sensors were firstly characterized to assess the unspecific interactions between the analyte and polymeric material. 2 M. Wang et al. Biosensors and Bioelectronics 198 (2022) 113807 2.3. Electrochemical measurements the chosen technique. The potential was scanned from − 0.4 V to +0.8 V with a step potential of 5 mV and a pulse amplitude of 50 mV without Electrochemical measurements were all carried out at room tem­ stirring. Before each measurement, the electrodes were immersed for perature, in a multichannel configuration using an Autolab PGSTAT30 180 s in the solution to be analyzed under stirring to allow for the (Ecochemie, Netherlands) controlled by GPES Multichannel and FRA enrichment of the analytes in the MIP sites. After each measurement, the software packages. A standard three-electrode cell composed of an Ag/ electrodes were electrochemically cleaned in PBS by applying a fix po­ AgCl (3 M KCl) reference electrode, a Pt wire as counter electrode and tential of +0.8 V for 180 s. the modified GECs with pTS− /PPy MIPs and NIP as working electrodes. Electrochemical characterizations of bare and modified GECs were 2.4. Samples under study performed by cyclic voltammetry (CV) and electrochemical impedance spectroscopy (EIS) in a 5 mM K3[Fe(CN)6]/K4[Fe(CN6)] solution in PBS. On the one hand, stock solutions of each of the different APIs were CV measurements were conducted in the potential range from − 0.2 V to prepared in PBS. From those, intermediate working solutions at different +0.7 V at a scan rate of 50 mV s− 1. Impedance spectra were recorded in concentrations were prepared by appropriate dilution in PBS, and used the frequency range between 50 kHz and 50 mHz with a fixed AC to characterize the analytical response of each of the developed MIP- amplitude of 10 mV and an applied potential of +0.23 V. For the analysis based sensors in terms of linearity, sensitivity, limit of detection of the impedance spectra, data was fitted to the Randles equivalent (LOD), etc. circuit (inset of Fig. 1f), composed of the solution resistance (Rs), the On the other hand, for the simultaneous determination of APIs, a set charge transfer resistance between the solution and the electrode surface of 38 samples divided into two subsets were prepared by appropriate (Rct), the Warburg impedance (Zw, related to the diffusion of the elec­ dilution and mixing of the stock solutions. The first subset used to troactive species from the solution towards the electrode surface), and establish the response model (training subset), was based on a 33 tilted the double-layer capacitance (Cdl). factorial design, as shown in Fig. S1. The concentrations for each of the For the analysis of APIs, differential pulse voltammetry (DPV) was compounds were in the range 0–50 μM for PA, 0–1000 μM for AA and Fig. 1. Cyclic voltammograms taken during the electropolymerization of pTS− /PPy onto a GEC (a) with paracetamol (MIP(PA)) and (b) without template molecule (NIP). Cyclic voltammograms taken during the template extraction of (c) pTS− /PPy MIP(PA) and (d) pTS− /PPy NIP. (e) Voltammetric and (f) impedimetric responses towards a 5 mM [Fe (CN)6]3-/4- solution in PBS after each modi­ fication step and rebinding experiments: (black) bare electrode, (red and green) after Py electropolymerization, (blue and purple) after extraction of the template, (green and brown) after rebinding of the template. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.) 3 M. Wang et al. Biosensors and Bioelectronics 198 (2022) 113807 50–500 μM for UA. Next, in order to assess the performance of the built simultaneous multi-determination of the different compounds in model, a set of 11 samples (testing subset) randomly distributed along mixtures. the experimental domain defined by the factorial design was also pre­ pared and measured under the same conditions. 3. Results and discussion Lastly, in order to demonstrate the applicability of the developed ET in real-world scenario, the analysis of commercial tablets bought from As already introduced, the aim of this work is to develop a voltam­ local drugstores was attempted. To this aim, tablets were dissolved in metric ET consisting of an array of MIP-based sensors targeting a PBS so that the expected concentrations fall within the experimental mixture of oxidizable APIs. In this direction, herein we propose a facile domain of the built model, without requiring further pre-treatment. and generic strategy for the synthesis of MIPs based on the electro­ Furthermore, some of them were also spiked with UA to evaluate its polymerization of Py, an approach that can be easily adapted to different recovery values as well. Concretely, the three following pharmaceutical target analytes (the templates). Moreover, another main asset of this preparations were evaluated: Efferaldol (330 mg paracetamol and 200 approach is that it’s a one-step electrochemical-synthesis, which makes mg ascorbic acid (vitamin C); Bristol Myers Squibb, S.A., Madrid), the integration of the MIP onto the electrode straightforward, while Multicentrum (800 μg vitamin A, 15 mg vitamin E, 100 mg vitamin C, 30 allowing to control the thickness of the MIP by changing the total charge μg vitamin K, 1.4 mg thiamine (vitamin B1), 1.75 mg riboflavin (vitamin flown at the electrode, the polymerization media or the electrolyte B2), 2 mg vitamin B6, 2.5 μg vitamin B12, 5 μg vitamin D, 62.5 μg biotin nature. (vitamin B7), folic acid (vitamin B9), 20 mg niacin (vitamin B3), 7.5 mg Next, the developed sensors are combined into an array, and the pantothenic acid (vitamin B5), 162 mg calcium, 125 mg phosphorous, registered responses are analyzed by means of different chemometric 100 mg magnesium, 5 mg iron, 100 μg iodine, 500 μg copper, 2 mg methods that allow the discrimination or quantification of the sub­ manganese, 40 μg chromium, 50 μg molybdenum, 30 μg selenium and 5 stances in the samples. The general procedure for the synthesis and mg zinc; GlaxoSmithKline España, Madrid) and Redoxon (1000 mg usage of pTS− /PPy MIPs with different template molecules is illustrated vitamin C, 10 mg zinc, 10 μg vitamin D; Bayer Hispania, S.L., Sant Joan in Scheme 1. Despí). Such an approach seeks to combine the advantages of MIPs to address the lack of selectivity of sensors used in ETs, with the modelling capabilities of ETs to address possible interferences or matrix effects 2.5. Data analysis from MIPs (Herrera-Chacón et al., 2021). The idea is similar to bioETs, but with the advantage of MIPs being cheaper, more stable, more Processing of the data was done in Matlab 7.1 (MathWorks, Natick, resistant to harsh conditions and tailor-made, offering a suitable re­ MA, USA) with the aid of its Neural Network toolbox, while final rep­ ceptor to almost any analyte of interest. resentation and analysis were done in Origin 2019 (OriginLab Corpo­ ration, Northampton, MA, USA). The potential of the ET to discriminate between the different APIs was assessed by means of principal compo­ 3.1. Synthesis and morphological characterization nent analysis (PCA), whereas artificial neural networks (ANNs) were chosen as the modelling tool for the quantification of their mixtures. To CV was used for the electropolymerization of Py monomer in the this end, recorded voltammograms were first compressed by means of presence of the template molecules (PA, AA or UA) and pTS− as sup­ discrete cosine transform (DCT) (Cetó and Pérez, 2020), and the ob­ porting electrolyte and doping counter ion to maintain the electro- tained coefficients were then submitted to PCA or ANNs. neutrality of the synthetized polymer. Such procedure includes both Briefly, the use of DCT was motivated to reduce the dimensionality the synthesis and immobilization of the MIP onto the electrode in a and complexity of the voltammetric signals, which allows to gain ad­ single step (Özcan and Şahin, 2007). Production of specific MIP sites is vantages in the modelling stage such as a reduction in training time, to favored by hydrogen bonding interactions between the N–H groups of avoid redundancy in input data and to obtain a model with better PPy and electron lone pairs of the template molecules. Moreover, generalization ability (Cetó et al., 2013a,b). PCA allows to evaluate compared with small-sized inorganic anions such as ClO4− , samples (dis)similarities and assess initial patterns in the data, whilst medium-sized organic ones such as pTS− , play an important role in ANNs represent a high-performance modelling tool able to achieve the improving the performance of PPy MIPs films. Firstly, with the benzene Scheme 1. (a) Preparation of the different pTS− /PPy MIPs by electropolymerization of Py in the presence of the template molecules and (b) incorporation of the fabricated MIP-based sensors into the ET sensor array and its application to the analysis of APIs in pharmaceutical samples. 4 M. Wang et al. Biosensors and Bioelectronics 198 (2022) 113807 ring, it helps enhancing its binding ability because of the π-π interactions low conductivity of the film. Thus, for the selection of the optimal between pTS− and the analytes. Secondly, PPy films deposited in the number of cycles, the ratio between the MIP and NIP responses was presence of such aromatic anions exhibit an anisotropic molecular or­ taken as the selection criteria, selecting 24 cycles as the optimum. ganization, resulting in higher conductivity and crystallinity (Raudsepp Next, the morphology of the modified electrodes was observed by et al., 2008). What’s more, as a sulfonate surfactant, pTS− provides SEM (Fig. 2). Low magnification images show a homogenous and uni­ additional benefits by generating higher degree of electrostatic form distribution of pTS− /PPy MIPs and NIP films over the surface of the cross-linking and PPy films with a much more microporous structure, GEC. From their comparison, it is conspicuous that pTS− /PPy MIPs are which favors a larger sensing surface and higher mass transfer coeffi­ rougher than pTS− /PPy NIP, which provides a larger active area for cient (Solanki et al., 2007; Suematsu et al., 2000). charge and mass transportation. Moreover, to further characterize the The use of amperometric methods allows the control of the poly­ synthesized material and confirm the adhesion of the film, elemental meric films’ thickness by tuning the charge (Coulombs), but even more, composition analysis of the cross-section area for pTS− /PPy MIP(PA) CV allows to monitor the entrapment of the template molecules during and NIP were measured by energy-dispersive X-ray spectrometry (EDX). its growth. As shown in Fig. 1a and b, the capacitive currents increase Concretely, the analysis of nitrogen, oxygen and sulfur were targeted to during the process of electropolymerization with the increase of the confirm the expected composition of the material. A clear change can be number of cycles as a result of the growing thickness of pTS− /PPy film. observed for nitrogen and oxygen, which confirms the extent of the There are remarkable differences between Fig. 1a and b, where the surface modification of both MIP and NIP as N can be attributed to PPy oxidation peak potential shifts to higher potentials in the presence of PA, while sulfur to the doping pTS− anions. suggesting that it is being entrapped into the polymeric chains (Özcan and Şahin, 2007). 3.2. Electrochemical characterization Upon synthesis of the MIP, the next step was the removal of the template molecules, which was also done by CV, triggering the recog­ Apart from the morphological characterization, each of the different nition at the MIP specific sites. Correspondingly, it can be seen from modification stages was confirmed electrochemically by CV and EIS Fig. 1c that as the number of cycle increases, the oxidation peak of PA measurements using [Fe(CN)6]3-/4- as the redox probe (Fig. 1e and f). disappears gradually, demonstrating that PA molecules are being Initially, we can see how the bare GEC displays the typical reversible released. Similarly, electropolymerization and extraction of pTS− /PPy peaks of [Fe(CN)6]3-/4-, while the overall current increases slightly due MIP(AA) and pTS− /PPy MIP(UA) are shown in Fig. S2. Due to the to the higher capacitive component after the electropolymerization, different nature and electrochemical behavior of the different template confirming the attachment of PPy onto the electrode surface. The worse molecules, different voltammetric profiles are obtained during the electrochemical behavior of the pTS− /PPy MIP was attributed to the electropolymerization, but following the same trend overall. presence of the template affecting the conductivity of PPy. Similarly, Undoubtedly, the thickness of the modified film plays a crucial role analyzing the EIS spectra, it can be seen how initially the charge transfer in the performance of the MIP-based sensor, which can be controlled by resistance (Rct) is rather small, but still observing the typical semicircle changing the number of cycles during the electropolymerization. shape, while after the electropolymerization, the Rct decreased signifi­ Therefore, the thickness of the film was optimized by changing the cantly as could be expected from the voltammetric measurement. The number of scans from 8 to 40 cycles (Fig. S3). As can be seen, a low next step was the extraction of the template, which should release the number of cycles leads to higher responses both for the MIP and the NIP sites so that the redox probe obtains easier access to the electrode sur­ as the films are too thin. Oppositely, when this is too high, the MIP face. However, the PPy film was over-oxidized during this step, which behaves similarly as the NIP a fact that can be attributed to the difficulty resulted in a decrease of its conductivity. This is why there is such a large of the molecules to diffuse through the thicker film, and the difficulty of increase in the Rct and, accordingly, a drastic decrease in the voltam­ being oxidized when being too far from the electrode surface due to the mogram current if comparing to the one before the removal of the Fig. 2. (a) Top and (b) cross-section SEM images of pTS− /PPy MIP(PA). (d) Top and (e) cross-section SEM images of pTS− /PPy NIP. EDX pattern of (c) pTS− /PPy MIP (PA) and (f) pTS− /PPy NIP; (blue) oxygen, (brown) nitrogen and (purple) sulfur. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.) 5 M. Wang et al. Biosensors and Bioelectronics 198 (2022) 113807 template. Finally, after the incubation with the template, which is ex­ Table 1 pected to block the diffusion of [Fe(CN)6]3-/4-, a decrease of the vol­ Analytical parameters of the different pTS− /PPy MIPs and NIP towards the three tammetric current and an increase of the Rct upon rebinding of the target analytes in PBS extracted from the DPV measurements (results averaged template molecules with the specific MIP sites was observed. From the from three different electrodes; the limit of detection (LOD) was calculated on voltammetric data, it may hardly be seen how the diffusion of the redox the basis of typical error of the regression line (S/N = 3)). probe was blocked to some extent, but it is much clearer from the Analyte LOD (μM) Sensitivity (nA⋅μM− 1) Linear range (μM) R2 Nyquist plot due to the higher sensitivity of the impedimetric mea­ pTS /PPy MIP(PA)/GEC ¡ surements. Overall, CV and EIS measurements are in agreement with Paracetamol 1.8 139 0.5–50 0.992 each other, both confirming the functionalization of the electrode. Ascorbic acid 177 0.8 500–2000 0.988 Uric acid 1.9 91 2–50 0.994 pTS-/PPy MIP(AA)/GEC 3.3. Voltammetric characterization Paracetamol 177 57 1–1000 0.999 Ascorbic acid 19 18 5–1000 0.997 Uric acid 17 53 0.5–1000 0.997 3.3.1. Calibration curves - pTS /PPy MIP(UA)/GEC The next step before proceeding with the ET approach was the Paracetamol 1.3 82 0–50 0.997 characterization of the electrochemical responses of each sensor towards Ascorbic acid 320 0.2 100–2500 0.983 each considered compound to assess their linear range and sensitivity, as Uric acid 24 26 50–500 0.996 pTS-/PPy NIP/GEC well as confirming that the sensors show certain cross-response and that Paracetamol 1.1 78 0.5–50 0.997 their responses are stable over time, as otherwise those are not suitable Ascorbic acid 157 0.1 500–2500 0.999 to be used in ETs. Uric acid 3.1 39 5–100 0.996 To this aim, calibration curves were built for each of the APIs by DPV due to its higher sensitivity in comparison to CV. However, before that there is still some degree of cross-response. However, the fact that proceeding with the electrochemical characterization, the incubation pTS− /PPy MIP(UA) doesn’t show the highest sensitivity to UA may be time had to be optimized, taking a compromise between the maximum attributed to the lower solubility of UA in the electrolyte solution, which signal and shorter analysis time (Fig. S4). As could be expected, there is a ultimately affects the number of specific sites generated in pTS− /PPy logarithmic increase of the signal with the incubation time, reaching MIP(UA). almost saturation after incubating in the analytes’ solution for 180s. Consequently, 180s was taken as the suitable incubation time for further 3.3.2. Repeatability study experiments. Another important feature when developing a sensor, and especially The voltammetric responses of the three pTS− /PPy MIPs towards significant from the perspective of ETs, is the repeatability of the increasing concentrations of their respective template molecules are developed devices. In order to assess the repeatability, 30 consecutive shown in Fig. 3a–c, with current peaks increasing as the concentration measurements for each pTS− /PPy MIP-based sensor were conducted in increases. From those, the peak heights were calculated and the cali­ their respective template solutions (PA solution was used for NIP-based bration curves were built (Fig. 3d–f), evidencing a linear trend with high sensors). After each measurement, electrodes were cleaned in PBS as correlation coefficients. Furthermore, not only the response of each MIP described in Section 2.3. Furthermore, a blank was also measured in was evaluated towards its template, but also the cross-responses towards fresh PBS after cleaning as control to evaluate any possible fouling ef­ the other analytes were evaluated (Figs. S5, S6 and S7 as well as fect. The day-to-day repeatability was evaluated in the same way, using Table 1). From the comparison of the calibration curves of the three the same MIPs and NIP electrodes during three consecutive days. From different pTS− /PPy MIPs and the NIP against PA and AA, it can be those, the relative standard deviation (expressed as percentage, RSD%) clearly seen that both MIP-based sensors show higher sensitivity towards were calculated, with values below 6.5% for all the sensors under the their own template molecule than the other sensors (other MIPs and two different scenarios (Table S1), asserting that all MIPs and NIP NIP), which demonstrates the specificity of the developed MIPs, but also Fig. 3. (Top) DPV responses of the MIP-based sensors towards their respective templates: (a) pTS− /PPy MIP(PA) to paracetamol, (b) pTS− /PPy MIP(AA) to ascorbic acid and (c) pTS− /PPy MIP(PA) to uric acid. (Bottom) Calibration graphs derived from the measured peak heights of pTS− /PPy MIP(PA), pTS− /PPy MIP(AA), pTS− / PPy MIP(UA) and pTS− /PPy NIP based sensors for the three analytes: (d) paracetamol, (e) ascorbic acid and (f) uric acid. 6 M. Wang et al. Biosensors and Bioelectronics 198 (2022) 113807 sensors render great repeatability. models (Cetó et al., 2013a,b). 3.4. MIP-based ET 3.4.1. Qualitative analysis Under the described conditions in Section 2.3, five replicate samples As already mentioned, the use of chemometric methods could assist of 50 μM solutions of each of the APIs were measured employing the in improving the selectivity of MIP-based sensors by shifting the developed sensor array (3 MIPs plus the NIP) and then submitted to complexity from the chemical to the modelling side (Cetó et al., 2016; DCT-PCA. del Valle, 2010). To demonstrate such a statement, discrimination of the The main advantage of PCA is allowing to summarize the informa­ three considered APIs was first attempted by means of PCA from tion contained in large data sets (e.g., the four voltammograms regis­ different stock solutions. This represents a preliminary step to confirm tered with the different sensors for each sample) down to a two- or three- that the MIP-based ET is able to discriminate the different compounds, coordinates point in a scores plot, where all the considered samples are and therefore has the potential to achieve its simultaneous quantifica­ also represented. Thus, it provides a better representation of samples tion by developing a proper quantitative model (in this case, by means of (dis)similarities and allows to rapidly detect and assess initial patterns in ANNs) using an appropriate set of samples. the data. In both cases, the set of samples were measured by the MIP-based The obtained 2D score plot is shown in Fig. 4a, with a cumulative sensor array and the recorded voltammograms were then compressed variance of ca. 91.2%, a large value denoting that almost all the variance by means of DCT down to 16 coefficients prior to the qualitative analysis present in the original data is now summarized with only the first two by PCA or quantitative modelling by ANNs. This step, apart from de- principal components (PCs). More significantly, we can see how clear noising the voltammetric data, assists in reducing the complexity of clusters are obtained for each of the APIs as well as for the buffer solu­ the input signal, which in turn allows to obtain simpler and more robust tion, all of them clearly separated with very small confidence intervals Fig. 4. (a) Score plot of the first two principal components obtained from the DCT-PCA of the voltametric responses of the pTS− /PPy MIPs and NIP modified sensors towards: ( ) PBS, ( ) paracetamol, ( ) ascorbic acid and ( ) uric acid. Ellipses plotted correspond to 95% confidence limits for each of the clusters. (b,c,d) Modeling ability of the optimized DCT-ANNs. Comparison graphs of obtained vs. expected concentrations for (b) paracetamol, (d) ascorbic acid and (d) uric acid, for both the training (●, solid line) and testing subsets ( , dotted line). Dashed line corresponds to the ideal comparison line (y = x). Additionally, the results of the analysis of the pharmaceutical samples are also plotted ( ). 7 M. Wang et al. Biosensors and Bioelectronics 198 (2022) 113807 for each of the groups. The clustering is so clear that either PC1 or PC2 To evaluate the improvement derived from the usage of the proposed would be enough to distinguish the different compounds, with the buffer MIP-based sensor array, the same set of samples were also measured being in one of the extremes of the plot, further from any other cluster. employing a bare GEC electrode, and the data was processed in the same Moreover, from the comparison of the current PCA plot with the ones way. After optimization of the ANN topology, a total NRMSE of 0.084 reported in previous publications based on arrays of graphite epoxy was obtained; a value that is significantly larger than the one of the MIP- composites (GECs)(Sarma et al., 2020) or screen-printed electrodes based ET, which was 0.020. (SPEs)(Ortiz-Aguayo et al., 2019) modified with different electro­ Lastly, to benchmark the performance of the MIP-based developed catalysts such as Prussian blue or metallic nanoparticles (representing a ET, the obtained results were compared to the ones reported in previous “conventional” voltammetric ET), we can see how the combination of works in which the same mixtures were analyzed (Table S2) (Cetó et al., MIPs and ETs significantly improves the former. 2013a,b; Gutés et al., 2007; Ortiz-Aguayo et al., 2019; Sarma et al., To numerically quantify this improvement, Silhouette clustering 2020). Despite it has to be reckoned that some of the differences attained metric was considered (Kaufman and Rousseeuw, 1990). Silhouette in the final performance originate from the modelling stage itself, that is, index provides a measure of how good the clustering is by comparing the depending on the choice of the machine learning algorithms, it can be intra-cluster distances (cohesion) with respect to the intra-cluster dis­ neglected that the richer and more relevant the raw data is, the better tances (separation). Its value is normalized between − 1 and +1, where performance will be attained. From the comparison, we can see how the high values are indicative of good clustering (highly cohesioned groups current approach clearly outperforms all the previous reported ETs, with good separation between them), and oppositely, low values indi­ therefore, further confirming the benefits derived from the combination cate very poor separation between clusters. The cohesion can be inter­ of MIPs and ETs. preted as a measure of sensors’ reproducibility, while separation as a measure of the sensor capability to discriminate the considered 3.4.3. Real sample analysis compounds. Finally, after generating and validating the performance with the In our case, the Silhouette index calculated from the PCA score plot samples of the training and testing subsets, respectively, the developed was 0.998 (a high value very close to the index maximum), while this ET was applied to the analysis of some pharmaceutical preparations. The was 0.947 and 0.838 for the above-mentioned GEC- and SPE-based ETs. pharmaceutical samples were prepared as described in Section 2.4 and The highest value of the current approach is related to the highest measured as previous, without performing any sample pre-treatment selectivity and specificity of MIPs, while in the case of the GECs, the step. Obtained voltammograms were compressed with DCT and fed to improved performance arises from a preliminary optimization on the the ANN model which provided the actual amount of API present in the sensors that constituted the ET. Thus, these values confirm the higher pharmaceutical formulation after multiplying by the proper dilution performance of the MIP-based ET approach and its potential to succeed factor. in the simultaneous quantification of APIs mixtures. In this manner, the analysis of seven different samples prepared from raw and spiked dilutions of the three different pharmaceutical formu­ 3.4.2. Quantitative analysis lations was carried out, and the results are summarized in Fig. 4 and The set of samples described in section 2.4 were prepared and Table S3. As can be observed, a good agreement between obtained and analyzed under the same conditions as earlier, recording a voltammo­ expected values is obtained, again being the values for the slope and gram for each of the sensors. As can be seen from Fig. S8, despite each intercept close to the ideal ones and within the confidence interval. This MIP-based sensor responds mainly to its target analyte, there is still is highly relevant as pharmaceutical formulations have a relatively some cross-response and peak overlapping for certain samples, complex matrix with many compounds (excipients, fillers, disintegrants, hampering the quantification of each API from the univariate regression binders, etc.) or other APIs not considered when building the chemo­ curves. Nevertheless, as demonstrated with PCA, the use of chemometric metric model. However, it is evident that those do not affect the per­ methods can greatly assist in improving MIPs’ selectivity. Thus, an ANN formance of the model, as no significant differences are found between model was used to achieve the simultaneous determination of the three the results derived from the stock solutions and the pharmaceutical APIs. The whole set of voltammograms were compressed with DCT and tablets. This can be attributed not only to the use of ETs, which are the ANN model was built employing the data of the train subset, using known to allow minimizing possible matrix effects, but especially the data of the test subset to assess its performance. This data splitting favored by the higher selectivity of the used MIP-based sensors. was done to ensure that the system was not over-fitted, providing a more realistic evaluation of the model performance as test data was not used 4. Conclusions at all during the modelling stage. The topology of the neural network was optimized by systematically The potential and advantages derived from the incorporation of MIPs varying the number of neurons in the hidden layer as well as the transfer into the ET sensor array have been demonstrated herein through the functions between the different layers. After the optimization, the analysis of APIs mixtures. For the development of the MIP-based ET, a selected ANN model had 64 neurons in the input layer (corresponding to generic facile electropolymerization method of Py in the presence of the DCT coefficients for each of the sensors), 4 neurons and purelin pTS− and the target analytes (the templates) has been proposed, which transfer function in the hidden layer, and 3 neurons (one for each of the lead to the synthesis of a new kind of pTS− /PPy MIPs. Doping of pTS− APIs) and purelin transfer function in the output layer. anions into the MIPs structures was confirmed from the SEM-EDX To visualize the performance of the model, the comparison graphs of analysis. Apart from the SEM morphological characterization, the obtained vs. expected concentrations were built for each of the analytes modification of the electrode was also confirmed electrochemically after (Fig. 4b–d), and the equations of the linear regressions as well as their the different steps involved in the synthesis of the MIP as well as its confidence intervals calculated (Table S2). As shown, an excellent application for sensing. agreement is attained with regression lines almost indistinguishable Upon completing the analytical characterization of the pTS− /PPy from the theoretical ones, demonstrating the superb specificity derived MIPs, those were combined into a sensor array, and the developed ET from the combination of MIPs and chemometrics. This good behavior is was applied towards the simultaneous accurate determination of PA, AA also confirmed numerically from the slope, intercept and correlation and UA mixtures. Recorded voltammograms were compressed by means coefficient values, which are all close to the ideal values (1, 0 and 1, of DCT and an ANN model was built, which showed excellent perfor­ respectively). Additionally, the root mean square error (RMSE) and its mance in comparison to previously reported ETs attempting the quan­ normalization (NRMSE) were also calculated to obtain a unique metric tification of the same mixtures. of the overall model performance. On the one hand, the obtained results prove the high specificity, 8 M. Wang et al. Biosensors and Bioelectronics 198 (2022) 113807 great stability and superb resolution of the developed MIP-based sen­ Bolat, G., Yaman, Y.T., Abaci, S., 2019. Sens. Actuators B-Chem. 299, 127000–1270008. Brambilla, A., Pizza, C., Lasagni, D., Lachina, L., Resti, M., Trapani, S., 2018. Front. sors. On the other hand, it also demonstrates that the combination of Pediatr. 6, 126–129. chemometric methods with MIPs represents a very promising approach Cao, Y., Feng, T., Xu, J., Xue, C., 2019. Biosens. Bioelectron. 141, 111447–111464. as evidenced from both the PCA and ANN models. More importantly, the Carquigny, S., Segut, O., Lakard, B., Lallemand, F., Fievet, P., 2008. Synthetic Met. 158 main advantage of the proposed approach is that MIPs towards other (11), 453–461. Cetó, X., Céspedes, F., del Valle, M., 2013a. Microchim. Acta 180, 319–330. molecules might be generated following the same procedure providing a Cetó, X., Gutés, A., del Valle, M., 2013b. Acta Manilana 61, 39–49. generic protocol for the development of MIP-based ETs towards other Cetó, X., Pérez, S., 2020. Talanta 219, 121253–121260. analytes. Overall, the development of MIP-based ETs stands for an Cetó, X., Voelcker, N.H., Prieto-Simón, B., 2016. Biosens. Bioelectron. 79, 608–626. Chen, L., Xu, S., Li, J., 2011. Chem. Soc. Rev. 40 (5), 2922–2942. extremely promising route for a broad-spectrum of sensing applications del Valle, M., 2010. Electroanalysis 22, 1539–1555. in TDM, but also in many other fields. Dosedel, M., Jirkovsky, E., Macakova, K., Krcmova, L.K., Javorska, L., Pourova, J., Mercolini, L., Remiao, F., Novakova, L., Mladenka, P., 2021. Nutrients 13 (2), 615–648. CRediT authorship contribution statement Grewal, A.S., Singh, S., Sharma, N., Rani, L., 2019. Int. J. Pharm. Qual. Assur. 10, 605–612, 04. Mingyue Wang: Investigation, Methodology, Writing – original Gui, R., Jin, H., Guo, H., Wang, Z., 2018. Biosens. Bioelectron. 100, 56–70. Gutés, A., Calvo, D., Céspedes, F., del Valle, M., 2007. Microchim. Acta 157, 1–6. draft. Xavier Cetó: Conceptualization, Methodology, Software, Writing Herrera-Chacon, A., Cetó, X., del Valle, M., 2021. Anal. Bioanal. Chem. 413, 6117–6140. – original draft. Manel del Valle: Conceptualization, Methodology, Herrera-Chacon, A., González-Calabuig, A., Campos, I., del Valle, M., 2018. Sens. Writing – review & editing, Funding acquisition. All authors have read Actuators B-Chem. 258, 665–671. Hillberg, A.L., Brain, K.R., Allender, C.J., 2005. Adv. Drug Deliver. Rev. 57 (12), and agreed to the published version of the manuscript. 1875–1889. Huang, H.Y., Appel, L.J., Choi, M.J., Gelber, A.C., Charleston, J., Norkus, E.P., Declaration of competing interest Miller 3rd, E.R., 2005. Arthritis Rheum. 52 (6), 1843–18477. Huynh, T.P., Kutner, W., 2015. Biosens. Bioelectron. 74, 856–864. Juraschek, S.P., Miller 3rd, E.R., Gelber, A.C., 2011. Arthritis Care Res. 63 (9), The authors declare that they have no known competing financial 1295–1306. interests or personal relationships that could have appeared to influence Kaufman, L., Rousseeuw, P.J., 1990. Finding Groups in Data: an Introduction to Cluster the work reported in this paper. Analysis. John Wiley & Sons, Inc, NJ. Kumar, D.R., Dhakal, G., Nguyen, V.Q., Shim, J.J., 2021. Anal. Chim. Acta 1141, 71–82. Meneghello, A., Tartaggia, S., Alvau, D.M., Polo, F., Toffoli, G., 2018. Curr. Med. Chem. Acknowledgments 25 (34), 4354–4377. Naidu, A.K., 2003. Nutr. J. 2, 1–10. Ortiz-Aguayo, D., Bonet-San-Emeterio, M., del Valle, M., 2019. Sensors 19 (15), The authors acknowledge the support of the Spanish Ministry of 3286–3298. Science and Innovation (MCINN) through project PID2019-107102RB- Özcan, L., Şahin, Y., 2007. Sens. Actuators B-Chem. 127 (2), 362–369. C21. M.d.V. thanks the program ICREA Academia. Prince, M.I., Thomas, S.H.L., James, O.F.W., Hudson, M., 2000. The Lancet 355 (9220), 2047–2048. Rajesh, Takashima, W., Kaneto, K., 2004. React. Funct. Polym. 59 (2), 163–169. Appendix A. Supplementary data Raudsepp, T., Marandi, M., Tamm, T., Sammelselg, V., Tamm, J., 2008. Electrochim. Acta 53 (11), 3828–3835. Rebelo, P., Costa-Rama, E., Seguro, I., Pacheco, J.G., Nouws, H.P.A., Cordeiro, M., Supplementary data to this article can be found online at https://doi. Delerue-Matos, C., 2021. Biosens. Bioelectron. 172, 112719–112736. org/10.1016/j.bios.2021.113807. Root-Bernstein, R., Churchill, B., Turke, M., 2020. Int. J. Mol. Sci. 21 (17), 6230–6250. Sarma, M., Romero, N., Cetó, X., del Valle, M., 2020. Sensors 20 (17), 4798-4712. References Sieminska, E., Sobczak, P., Skibinska, N., Sikora, J., 2020. Med. Hypotheses 142, 109791–109801. Solanki, P.R., Arya, S.K., Singh, S.P., Pandey, M.K., Malhotra, B.D., 2007. Sens. Actuators Alegret, S., Alonso, J., Bartrolí, J., Céspedes, F., Martínez-Fàbregas, E., del Valle, M., B-Chem. 123 (2), 829–839. 1996. Amperometric biosensors based on bulk-modified epoxy graphite Suematsu, S., Oura, Y., Tsujimoto, H., Kanno, H., Naoi, K., 2000. Electrochim. Acta 45, biocomposites. Sens. Mater. 8 (3), 147–154. https://myukk.org/SM2017/article.ph 3813–3821. p?ss=10231. Vale, J.A., Proudfoot, A.T., 1995. The Lancet 346, 547–552. Arabi, M., Ostovan, A., Li, J., Wang, X., Zhang, Z., Choo, J., Chen, L., 2021. Adv. Mater. Wackerlig, J., Lieberzeit, P.A., 2015. Sens. Actuators B-Chem. 207, 144–157. 33, 2100543–2100575. Zaidi, S.A., 2018. Sens. Actuators B-Chem. 265, 488–497. BelBruno, J.J., 2019. Chem. Rev. 119 (1), 94–119. Zhang, W., Duan, D., Liu, S., Zhang, Y., Leng, L., Li, X., Chen, N., Zhang, Y., 2018. Bergh, M.S., Oiestad, A.M.L., Baumann, M.H., Bogen, I.L., 2021. Int. J. Drug Policy 90, Biosens. Bioelectron. 118, 129–136. 103065–103069. 9