# Fileset

[Advanced Sensor Research - 2024 - Ngo - Empirical Modification of Force Fields for the Development of Peptide‐Based Gas.pdf](https://mdr.nims.go.jp/filesets/7d5edffe-8001-41ef-8f4b-4b1a328ea4bb/download)

## Creator

[Thuc Anh Ngo](https://orcid.org/0000-0003-2458-7443), [Tanju Yildirim](https://orcid.org/0000-0002-0269-4718), Meng‐Qun Feng, [Kosuke Minami](https://orcid.org/0000-0003-4145-1118), [Kota Shiba](https://orcid.org/0000-0001-7775-0318), [Genki Yoshikawa](https://orcid.org/0000-0002-9136-8964)

## Rights

[Creative Commons BY Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)

## Other metadata

[Empirical Modification of Force Fields for the Development of Peptide‐Based Gas Sensors](https://mdr.nims.go.jp/datasets/76e0da48-e084-4675-afdc-949ddb1c49ac)

## Fulltext

Empirical Modification of Force Fields for the Development of Peptide‐Based Gas SensorsRESEARCH ARTICLEwww.advsensorres.comEmpirical Modification of Force Fields for the Developmentof Peptide-Based Gas SensorsThuc Anh Ngo,* Tanju Yildirim, Meng-Qun Feng, Kosuke Minami, Kota Shiba,and Genki Yoshikawa*Molecular dynamics models combined with computational approaches can beused as advanced screening techniques for finding highly efficientmaterial-molecule interactions based on binding affinity, including in thedevelopment of gas sensors. However, most models are originally designedfor liquid phase interactions, which do not align with gas sensing conditions,resulting in lower-than-expected performance. This study introduces anempirical modification method to adjust peptide interaction models for a gasphase, aiming to better accommodate the interaction between pentapeptidesand target gas molecules. By adapting the weights of terms in the Gibbs freeenergy equation given in an empirical force field model, we demonstrate asignificant increase in the absolute value of coefficient of determination (R02) ,from an average of 0.05 with conventional liquid phase models to 0.90 withproposed gas phase models. An empirical modification technique for gasphase interactions markedly enhances the prediction accuracy of models,facilitating the effective development of peptide-based gas sensors.1. IntroductionOne of the most important characteristics of a gas sensor isselectivity toward target molecules.[1,2] Among various materi-als, oligopeptides derived from the specific binding region of ol-factory receptors have gained significant interest as a receptorT. A. Ngo, M.-Q. Feng, K. Minami, K. Shiba, G. YoshikawaResearch Center for Macromolecules and BiomaterialsNational Institute for Materials Science (NIMS)1-1 Namiki, Tsukuba, Ibaraki 305-0044, JapanE-mail: ngo.thucanh@nims.go.jp; yoshikawa.genki@nims.go.jpT. A. Ngo, M.-Q. Feng, G. YoshikawaMaterials Science and EngineeringGraduate School of Pure and Applied ScienceUniversity of Tsukuba1-1-1 Tennodai, Tsukuba, Ibaraki 305-8577, JapanT. YildirimFaculty of Science and EngineeringSouthern Cross UniversityEast Lismore, NSW 2480, AustraliaThe ORCID identification number(s) for the author(s) of this articlecan be found under https://doi.org/10.1002/adsr.202400122© 2024 The Author(s). Advanced Sensor Research published byWiley-VCH GmbH. This is an open access article under the terms of theCreative Commons Attribution License, which permits use, distributionand reproduction in any medium, provided the original work is properlycited.DOI: 10.1002/adsr.202400122material for gas sensors, which is dueto their important role in attaininggas-specific recognition in the olfactorysystem.[3] To design and find effectiveoligopeptides with specific selectivity, theevaluation of binding affinity betweena peptide and a target gas molecule iscrucial.[4,5] Molecular docking is a com-putational technique aimed at predict-ing the affinity between a receptor and aligand. Molecular docking has been ap-plied to designing peptides as a mate-rial for gas sensors.[5–11] There are twomajor approaches to the engineering ofpentapeptides; i) based on the interactionsite of the olfactory proteins; and ii) com-prehensive screening of short oligopep-tides. For the former approach, oligopep-tides with lengths ranging from 5 to15 amino acids have been reported torecognize some specific gases such astrimethylamine, acetic acid, and butyric acid.[6,7,12–15] Thismethod, however, has the limitation of finding actual interac-tion sites of the proteins due to the lack of accurate 3D pro-tein structures. In contrast, the latter approach has been widelyused for designing oligopeptides toward a wide range of gasspecies based on simulated binding affinities.[5,16,17] The bind-ing affinity calculated based on Gibbs free energy (ΔG) be-tween oligopeptides and gases has been utilized as a guidelinefor designing peptide-based materials for olfactory sensors.[17]Three tripeptides (WWW, WAW, and WHW) containing aro-matic rings with the highest simulated binding affinities for aro-matic compounds showed good performance in detecting ex-plosive gases with aromatic structures. The limited accuracy to-ward some gases such as alcohols remains challenging using thisapproach.[8] Thus, further implementation of molecular dock-ing to achieve better accuracy in predicting the interaction be-tween pentapeptides and gases requires highly accurate dockingcalculations.Several models exist for evaluating intermolecular energylandscapes in gas-peptide interactions, including the geomet-ric model, first-principles calculations, and empirical forcefields.[4,18] The geometric method using the GRAMM (GlobalRange Molecular Matching) docking program identified ligand-binding residues with the greatest exothermic interactionswith chosen gas ligands.[6,7,18–20] However, the method requiresmultiple steps and biophysics knowledge for modeling struc-tures and deriving peptide sequences based on interactionAdv. Sensor Res. 2025, 4, 2400122 2400122 (1 of 8) © 2024 The Author(s). Advanced Sensor Research published by Wiley-VCH GmbHhttp://www.advsensorres.commailto:ngo.thucanh@nims.go.jpmailto:yoshikawa.genki@nims.go.jphttps://doi.org/10.1002/adsr.202400122http://creativecommons.org/licenses/by/4.0/http://crossmark.crossref.org/dialog/?doi=10.1002%2Fadsr.202400122&domain=pdf&date_stamp=2024-11-29www.advancedsciencenews.com www.advsensorres.comFigure 1. Evaluation of binding affinity between pentapeptide library and 12 gas molecules with AutoDock Vina. a) Typical working space between apentapeptide and a gas molecule to compute binding energy (ΔG). b) Screening result of pentapeptide library (1.6 million candidates) toward differentgas molecules using AutoDock Vina developed with liquid phase data shown using ΔG. It is noteworthy that the lower ΔG is, the higher the frequencyshift (Δf) expected in QCM experiments. c) Predicted liquid phase binding energies of four selected pentapeptides (KVYYY, CDHWW, EHIPW, andEFFPW) toward four target gas molecules: toluene, acetic acid (AA), acetone, and n-hexane.sites. The first-principles method is another way to evaluate theintermolecular energies between oligopeptides and targetgases.[11,17] Even though the accuracy of the method is high,extensive computational work limits the length of a peptidein one simulation. Unlike physics-based models, empiricalforce fields rely on machine learning algorithms to calibrateexperimental data and import results into simulation models,improve gas-peptide prediction accuracy, and reduce computa-tional complexity. Empirical force fields such as chemgauss havebeen applied to design oligopeptides for specific biomarkers.[5,8]Peptides selected from the screening with chemgauss4 showa good agreement between binding affinities and experimentvalues for various gases except for alcohols. AutoDock modelsare widely known empirical models in designing peptide-based materials for gas sensors.[9,10,17,21,22] Whilst the empiricalmodel showed high performance in some gases that havearomatic structures, the model lacks accuracy for other gasspecies such as aldehydes.[9] Limitations of empirical modelsmay be attributed to their original calibration being basedpurely on liquid phase data. Herein, we addressed this issuewith gas phase data together with optimizing AutoDock Vinamodel.The aim of this work is to adapt the empirical force fields bymodifying the weights in the model of each gas based on gas-solidphase data for better correlations with sensing signals. Four pen-tapeptides, namely KVYYY, CDHWW, EIHPW, and EFFPW, wereinitially selected via the liquid phase model and used as referencedata for evaluating the gas-peptide binding affinity via measur-ing the mass of adsorbed gas molecules on pentapeptides usinga Quartz Crystal Microbalance (QCM). First, we adapted the em-pirical force fields considered in AutoDock Vina model for gasphase binding affinity by modifying the weights of the empiri-cal force fields with the genetic algorithm (GA), which increasesthe predictability of the model from no correlation to strong cor-relation with the experimental results. Further improvement ofthe weights using the stepwise algorithm (SA) shows high lin-earity between the predicted normalized binding affinity andthe sensor output signal. This adaptive approach demonstratessignificant potential for enhancing the precision of empiricalforce fields for gas phase measurements, paving the way formore accurate and reliable applications of gas sensors in variousfields.2. Results and Discussion2.1. Virtual Screening ResultsThe virtual screening was conducted to evaluate the bindingproperties of all possible pentapeptides bound to 12 differentgases using the liquid phase AutoDock Vina model (Figure 1a).These gases include aromatic hydrocarbons, amines, acids, alka-nes, ketones, aldehydes, esters, and alcohols. These chosen gasesexhibit diverse physicochemical properties including varyingmolecular weights, solubility, and polarity. The first step is find-ing the most stable state of interaction in which a peptide anda gas molecule are placed in close contact inside a unit cell tocompute the lowest binding energy. It should be noted that thelower binding affinity of the complex corresponds to the strongerinteraction of the peptide and gas. Thus, the values of the bind-ing affinity and the sensor frequency shift with the increasedgas molecule absorption on peptide material are negatively cor-related. The overall binding affinities of the pentapeptide librarytoward the 12 gases are diverse depending on the physicochem-ical characteristics of both peptides and gas molecules. For ex-ample, the gases containing phenyl groups, such as toluene andaniline, exhibit strong affinities toward pentapeptides because of𝜋–𝜋 stacking (Figure 1b).[17,23] For the experimental verification,we chose four pentapeptides consisting of KVYYY, CDHWW,EHIPW, and EFFPW because of their lowestΔG to toluene, aceticacid, acetone, and n-hexane, respectively (Figure 1c). To assess theoverfitting and ensure the accuracy of the prediction, two pen-tapeptides with similarities in structural properties (DFIPW andRTYYY) were selected as test samples.2.2. Liquid Phase Model Performance Assessment withEvaluation MetricsTo evaluate the performance of the phase model, a gas sens-ing array comprising of the six pentapeptides was tested with allAdv. Sensor Res. 2025, 4, 2400122 2400122 (2 of 8) © 2024 The Author(s). Advanced Sensor Research published by Wiley-VCH GmbH 27511219, 2025, 4, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/adsr.202400122 by National Institute For, Wiley Online Library on [11/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://www.advancedsciencenews.comhttp://www.advsensorres.comwww.advancedsciencenews.com www.advsensorres.comFigure 2. QCM responses of four selected pentapeptides based on comprehensive screening. a) Experiment setup b) QCM response of four pentapep-tides toward toluene, acetic acid (AA), acetone, and n-hexane.12 gases, including toluene, aniline, methylcyclohexane (MCH),acetic acid (AA), propionic acid (PA), acetone, n-hexane, n-heptane, hexanal, ethyl acetate (EtOAc), ethanol, and 1-hexanol(Figure 2a). The responses of each peptide-coated QCM sen-sor to four representative gases are shown in Figure 2b.The average absolute frequency shift (|Δf|) of replicate experi-ments is assumed as an experimental indicator for the affin-ity between the peptides and the gases (Figure S1, SupportingInformation).To assess the accuracy of the liquid phase model in predict-ing the gas response of the peptide-coated sensors with bindingaffinities, we compare the normalized Gibbs free energy (ΔGnorm)with the absolute change in frequency (|Δf|) in terms of Pear-son’s correlation coefficient (R), Spearman’s rank correlation co-efficient (𝜌), and the coefficient of determination through the ori-gin (R02) as described in the Experimental Section. The predic-tion results based on the liquid phase AutoDock Vina agree withthe experimental results in the case of toluene with an R value of−0.9. This finding aligns with the previous screening results oftriple pentapeptides, where aromatic amino acids were preferredfor enhancing binding affinities with aromatic ring-containingmolecules.[17] It is, however, less effective for other gases, partic-ularly acids, alcohols, and ketones (Figure 3; Figures S1 and S2,Supporting Information). The inverse relationship between thebinding affinities and the response of peptide-coated sensorswith the positive values of R clearly illustrates the limitation ofthe liquid phase model in the cases of acetic acid, propionic acid,and ethanol. The earlier attempt to use the empirical force fieldsfor choosing peptides as the sensor receptor also resulted in apoor performance for short alkyl chain alcohols.[21] The incon-sistency between the predicted binding affinities obtained fromthe liquid phase model and the experimental results stems fromthe weights that were empirically determined solely based on theliquid phase data. To improve the efficiency of this liquid phaseAutoDock Vina model for the gas-solid phase, it is necessary toscale the physical terms with appropriate weights using gas phasedata.2.3. Concept Demonstration with GAWe demonstrate the approach to the development of the gasphase model by scaling each physical term with experimentaldata and multiplying it with optimal weights. The gas phasemodel was developed for each gas, in which all pentapeptideswere exposed to the same vapors with the same conditions, re-sulting in 12 corresponding models. GA is employed to adjustweights to achieve optimal agreement in the trend between thebinding affinities and experimental results. In this modified ap-proach, a group of a certain number of different solutions withsix weight values each was initialized. In this study, a group oftwenty solutions was employed to balance the sufficient diversityof each group and the feasible computation. Each weight valuewas randomly set between 0 and 1. After the molecular dynamicsimulation, each solution was evaluated to have a score ( ) con-sidering Ri, 𝜌i, and R0i2 for determining the performance of themodel with the updated weights. New solutions were created bycrossing over the weight values from their parent sets. To keepsolutions diverse and avoid the local minimum, some random-ness was introduced by changing some values in these new solu-tions. New sets were ranked based on how well they performedin the Si for generating the next groups. This cycle of selection,crossover, introducing randomness, and evaluation was repeatedfor 1000 cycles. The optimal weights entail iteratively decreasingthe score (Si) in Equation (4), with the optimal weights emergingfrom the most successful weights of the preceding generation.The final weights for all 12 gases after 1000 generations with theminimum Si are shown in Table 1.By applying GA to the weight optimization process, good R and𝜌 are quickly achieved with average values of −0.83 and −0.87,Adv. Sensor Res. 2025, 4, 2400122 2400122 (3 of 8) © 2024 The Author(s). Advanced Sensor Research published by Wiley-VCH GmbH 27511219, 2025, 4, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/adsr.202400122 by National Institute For, Wiley Online Library on [11/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://www.advancedsciencenews.comhttp://www.advsensorres.comwww.advancedsciencenews.com www.advsensorres.comFigure 3. Performance comparisons of the AutoDock Vina, developed us-ing liquid phase data and gas phase data, with optimal weights obtainedwith GA and SA. a) Pearson’s correlation coefficient (Ri) implies the lin-earity between ΔG and |Δf|. A value closer to −1 indicates high linearity.b) Spearman’s rank correlation coefficient (𝜌i) implies the correct rankingpredicted binding affinity of one gas molecule to four different peptides.A value closer to −1 indicates high selectivity. c) Coefficient of determina-tion through the origin (R0 i2) implies the linearity between ΔG and |Δf|considering gas-solid phase interaction. An R02 value closer to 1 indicatesgreater linearity with zero-intercept models.respectively (Figures 3a,b and 4). In the cases of acetic acid, pro-pionic acid, and ethanol, all of which previously show the inverserelationships between the binding affinities and the experimentalresults, Pearson’s correlation coefficients increase considerably,reaching almost −1 for acetic acid (Figure 3a,b). Considering theTable 1. GA optimized weights for 12 gas molecules.Gas molecules wg1 wg2 wrep. wHP wHB wrot .Toluene −0.4 −0.2 1.8 −0.8 −0.8 −2.7Aniline 0.0 −0.4 −0.2 −1.0 −0.4 −0.8MCH −0.4 0.0 −1.9 −0.7 −0.4 7.3AA −1.0 −0.2 0.7 −0.1 −0.2 −1.6PA −0.6 −0.2 −0.8 −0.4 −0.1 0.1Acetone −0.4 −0.3 5.0 −0.4 0.0 0.9n-Hexane 0.0 0.0 −0.2 −1.0 −0.4 −1.0n-Heptane −0.9 −0.6 −0.2 −0.6 −0.4 −0.9Hexanal −0.1 −0.2 −0.7 −0.8 −0.2 −0.4EtOAc −1.0 −0.5 −1.3 −0.8 0.0 −0.2Ethanol −0.9 −0.6 −0.7 −0.7 −0.8 −0.51-Hexanol −0.7 −0.1 −0.3 −0.3 −0.2 4.9R02, a notable increase with the GA optimized gas phase modelsis observed, particularly for EtOAc and 1-hexanol. The R02 values,however, remain low for other gases, such as toluene, aniline,and ethanol. GA optimization method only provides R02 valueswith less than 0.5 (Figure 3c; Figure S3, Supporting Information).This limitation of GA-optimized models is possibly due to the al-terations in specific weight values within the GA functions, influ-encing the overall performance (Figure S4, Supporting Informa-tion). The GA optimization method mostly modifies the rotationand repulsion weights as shown in Figures S5–S8, SupportingInformation. A correlation test among weight values across 1000generations and Si, however, suggests that variations in weightvalues of rotation and repulsion terms do not significantly in-fluence the overall performance of the model in some instances(Figure S4, Supporting Information). For example, in the caseof MCH, the weight value of the gauss1 term largely affects theimprovement of Si but is not modified intensively further in theoptimization process. Since gauss1 represents the short-range re-pulsive interactions, the results imply that two Gaussian termsmay be critical components of steric interactions between gasmolecules and peptides in the gas phase, affecting sensor perfor-mance. The strength of intermolecular forces—including bothsteric and electrostatic interactions—is largely influenced by theelectron distribution within the molecules. Charge density plays acrucial role in determining this electron distribution, which leadsto variations in the strength of the steric interactions. A previousstudy involving first-principles molecular modeling of sensingmaterial selection demonstrated that changes in charge densityare correlated with the sensor’s response to different gases.[21]Therefore, we decided to comprehensively analyse all the chang-ing directions of the weights including the Gaussian terms usingthe SA method.2.4. Further Investigation with SA Optimization Method for allWeight Changing DirectionsSA optimization method systematically explores a wide range ofweight values from −1.0 to 0.0 with a step increment of 0.2. Eachsolution was evaluated to have the Si as explained in the previ-ous section. All combinations were ranked according to Si, andAdv. Sensor Res. 2025, 4, 2400122 2400122 (4 of 8) © 2024 The Author(s). Advanced Sensor Research published by Wiley-VCH GmbH 27511219, 2025, 4, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/adsr.202400122 by National Institute For, Wiley Online Library on [11/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://www.advancedsciencenews.comhttp://www.advsensorres.comwww.advancedsciencenews.com www.advsensorres.comFigure 4. Parity plots of absolute frequency shift versus the normalized optimized binding energy using GA. The binding energy is normalized to itsmaximum value. The training dataset, which includes four pentapeptides (KVYYY, cyan; CDHWW, blue; EHIPW, purple; and EFFPW, magenta), is shownusing open circles. The validating dataset, which includes pentapeptides DFIPW (red) and RTYYY (yellow), is shown with filled circles. Circles representthe average values calculated from three measurements, with error bars indicating the corresponding standard deviations. Black lines show the resultsof a linear regression.the best set of the six weight values was determined. The finalweights for all 12 gases after 46656 (= 66) iterations are shown inTable 2. As a result, the multidirectional changes in the weightsderived from the SA optimization method exhibit substantial en-hancements in the overall model performances.Table 2. SA optimized weights for 12 gas molecules.Gas molecules wg1 wg2 wrep. wHP wHB wrot.Toluene −0.2 −1.0 −0.4 −0.6 −0.6 −0.2Aniline −1.0 −0.8 −1.0 −0.6 −1.0 −1.0MCH −0.4 −1.0 −0.8 0.0 −0.2 −0.4AA −0.2 −0.4 0.0 0.0 −1.0 −0.2PA 0.0 −0.8 −1.0 −1.0 −1.0 0.0Acetone −0.6 −0.8 −0.4 −0.4 −1.0 −0.6n-Hexane −0.8 −0.2 −1.0 −1.0 −1.0 −0.8n-Heptane −1.0 −0.4 −0.4 −1.0 −1.0 −1.0Hexanal −0.2 −0.6 −0.2 0.0 −1.0 −0.2EtOAc 0.0 −0.2 −0.8 −0.6 −1.0 0.0Ethanol −0.2 −0.8 −0.2 −0.8 −1.0 −0.21-Hexanol −1.0 −0.4 −0.4 −1.0 −1.0 −1.0In comparison to the liquid phase model and the GA-optimized gas model, the SA-optimized gas models greatly im-prove R and 𝜌 (Figures 3a,b and 5). An average R value of −0.97demonstrates the adequate ability of the models to predict thepeptide behavior with gases. Furthermore, the optimized weightsfrom the SA method resolve the issue of assessing polar gasesusing the AutoDock Vina model, especially in the case of aniline.For instance, the dominant effect of 𝜋–𝜋 stacking in the liquidphase and GA optimized gas phase models results in the samebest affinities toward KVYYY for both toluene and aniline despitethe difference in the polarity of the gas molecules. The SA opti-mized gas phase model accurately predicts EFFPW as the mostaffinitive peptide toward aniline as observed in the experiments.The SA optimized gas phase models further achieve the R02 val-ues of above 0.9 in many gases (Figure 3c). The improvement ofthe SA optimized gas phase models in discriminating peptidesthat gave similar gas responses can be attributed to the consid-eration of R02 in the weight optimization process. Furthermore,the SA optimized models can accurately predict which pentapep-tide has a higher response than the other one for all 12 gases inthe validating peptide set (DFIPW and RTYYY), whilst the liquidphase model can predict only for toluene, acetic acid, propionicacid, and n-heptane. The success of the SA optimized model canAdv. Sensor Res. 2025, 4, 2400122 2400122 (5 of 8) © 2024 The Author(s). Advanced Sensor Research published by Wiley-VCH GmbH 27511219, 2025, 4, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/adsr.202400122 by National Institute For, Wiley Online Library on [11/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://www.advancedsciencenews.comhttp://www.advsensorres.comwww.advancedsciencenews.com www.advsensorres.comFigure 5. Parity plots of absolute frequency shift versus the normalization values of optimized binding energy SA. The binding energy is normalized toits maximum value. The training dataset, which includes four pentapeptides (KVYYY, cyan; CDHWW, blue; EHIPW, purple; and EFFPW, magenta), isshown using open circles. The validating dataset, which includes pentapeptides DFIPW (red) and RTYYY (yellow), is shown with filled circles. Circlesrepresent the average values calculated from three measurements, with error bars indicating the corresponding standard deviations. Black lines showthe results of a linear regression.be attributed to the improvement of two Gaussian terms (wg1 andwg2) describing the steric interaction and the hydrogen bondingterms. All models increase the effects of hydrogen bonding bydecreasing the value of the hydrogen bonding weight, in mostcases to the minimum value of −1. Overall, the implementationof the SA optimization method improves all three metrics, lead-ing to the best agreement in trends between binding affinitiesand experimental results.3. ConclusionThis study introduces empirical modification techniques for theadaptation of the AutoDock Vina model based on gas phase data,significantly enhancing the accuracy of predicting the gas re-sponses of pentapeptides coated on QCM sensors. We also pro-pose the coefficient of determination through the origin (R02) asan evaluation metric to support the weight optimization processwith respect to the nature of the gas sensing phenomenon. De-spite the significant improvements made by tuning weights, thegas phase model still falls short in predicting absolute values ofΔG and the behavior of aniline and ethanol for peptide-coatedsensors that have similarities in the gas responses. These lim-itations of the current gas phase AutoDock Vina models couldbe attributed to the constraints of their ability to capture the in-teractions with small molecules, such as gas molecules. Theseproblems could be addressed with further comprehensive opti-mization of the model by modifying the energetic terms based onthe parameters involved in addition to the tuning of their weightvalues. By addressing these factors, future models may be ableto predict molecular interactions more accurately in gas phaseenvironments.Our work has improved the ability of the AutoDock Vina modelto predict gas response of peptide-coated sensors by address-ing gas phase limitations and optimizing weight values. Thisstudy paves the way for the efficient application of empirical forcefields, as in this case, AutoDock Vina, in engineering peptide ma-terials with the desired selectivity toward gas molecules for thefurther development of gas sensors.4. Experimental SectionMaterial Preparations: Synthetic pentapeptides (KVYYY, CDHWW,EHIPW, EFFPW, RTYYY, and DFIPW) were purchased from Cosmo BioCo., Ltd. with 80% purity. Toluene, aniline, methylcyclohexane (MCH),acetic acid (AA), propionic acid (PA), acetone, n-hexane, n-heptane, hex-Adv. Sensor Res. 2025, 4, 2400122 2400122 (6 of 8) © 2024 The Author(s). Advanced Sensor Research published by Wiley-VCH GmbH 27511219, 2025, 4, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/adsr.202400122 by National Institute For, Wiley Online Library on [11/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://www.advancedsciencenews.comhttp://www.advsensorres.comwww.advancedsciencenews.com www.advsensorres.comanal, ethyl acetate (EtOAc), ethanol, and 1-hexanol were purchased fromSigma–Aldrich, Fujifilm Wako Pure Chemical Corporation, and KantoChemical Co., Ltd. MilliQ water (Millipore) was used for preparing thepeptide stock solution.QCM measurements: An AT-cut quartz crystal resonator (QA-A9M-AU(M), SEIKO EG&G Co., Ltd.) was used as the sensing element in whichpentapeptides can be coated onto the top surface. Each peptide (1 mg)was dissolved in 1 mL of distilled water to make a stock solution. To de-posit each peptide on the QCM, the QCM was cleaned with O2 plasmaand immersed in the peptide solution for 48 h at room temperature. Thepeptide-coated QCMs were stored at room temperature in the dark untiluse.The QCMs were placed in a Teflon chamber, which was connected totwo gas lines: an inlet and an outlet. The inlet was connected to a gassystem, which consisted of two mass flow controllers, a mixing chamber,a purging line, and a sampling line with a vial (40 mL) containing a sol-vent liquid (5 mL). The sensing setup was placed in an incubator and thetemperature was maintained at 25 °C. Before measuring the QCM out-put, pure nitrogen gas was introduced into the QCM chamber at the flowrate of 50 mL min−1 for 5 min to remove any contaminants. Nitrogen gas,which was one of the inert gases, was bubbled into a pure solvent, and theresulting headspace vapor was directed into the QCM chamber for 1 minfollowed by 1 min of purging with nitrogen and repeated five times. Theconcentration of vapors was precisely controlled at 20% (i.e., 20 mL min−1of sample vapors mixed with 80 mL min−1 of nitrogen) with two mass flowcontrollers. The total flow rate was 100 mL min−1 for both sampling andpurging cycles. Data was recorded using a Crystal oscillator measurementsystem (QCM922A, SEIKO EG&G Co., Ltd.) in the form of frequency shift(Δf).Molecular Docking Analysis: The peptide libraries and ligands weresubjected to molecular docking using AutoDock Vina.[24] Structures ofpentapeptides were constructed by PyMol (The PyMOL Molecular Graph-ics System, Version 1.2r3pre, Schrödinger, LLC.). Each peptide and gasmolecule conformation were visualized and checked to guarantee theiraccuracy in terms of hydrogen atoms and bond orders. The energies ofall molecules were minimized based on a chemistry forcefield functionnamed MMFF94s integrated with AutoGrid4. Possible stable conformerswere precalculated to consider the molecule’s flexibility. In docking, a rect-angular search space of 20 Å × 20 Å × 20 Å was adapted, enclosing botha pentapeptide and a corresponding gas molecule. The flexibility of bothpartners was considered indirectly by employing nine different conforma-tions in ensemble docking results. A docking solution with the lowest rootmean square deviation (RMSD) was selected as the best conformation be-tween a pentapeptide and a gas molecule to calculate the intermolecularfree energy.[25] Based on the scoring function of Vina, ΔG was predictedas the sum of distance-dependent atom pair interactions, f(d):ΔG =∑atom pairfinter (dinter)1 + wrotNrot+ finter (dintra) , (1)where dinter and dintra are the intermolecular and intramolecular surfacedistance of each atom pair, Nrot is the number of active rotatable bonds,and wrot is the associated weight. Each atom pair interacts through stericinteractions (i.e., gauss1, gauss2, and repulsion) and could be hydropho-bic interaction and non-directional hydrogen bonding (Hbond). The atompair interactions, f(d) is given byf (d) = wg1gauss1(d) + wg2gauss2(d) + wreprepulsion(d)+ wHPhydrophobic(d) + wHBHbond(d), (2)where wg1, wg2, wrep., wHP, and wHB are associated weights for each term.Detailed scoring function can be found in the original paper by Trott andOlson, 2010.[21] For virtual screening, all possible pentapeptides with 12gas molecules including toluene, aniline, MCH, AA, PA, acetone, n-hexane,n-heptane, hexanal, EtOAc, ethanol, and 1-hexanol were analyzed and thecorresponding binding affinity was estimated using the reported weightsfor liquid phase as shown in Table S1, Supporting Information.[25] For thegas phase model, the binding affinity was estimated using the optimizedweights determined using optimization algorithms described in the fol-lowing section.Evaluation Methods: To assess the performance of the obtained set ofweights as a scoring function for the solid-gas phase, a score of a scor-ing function was evaluated for the ith gas molecule (Si) based on threemetrics: namely, Pearson’s coefficient correlation, Spearman’s rank cor-relation coefficient, and coefficient of determination through the origin.These three metrics were the direct measurements of linearity betweennormalized binding affinity ∆Gnorm calculated by diving ∆G to the maxi-mum absolute value of the system, |∆G|max, and the absolute experimen-tal QCM frequency shift, |∆f|, for each gas molecule. Pearson’s coefficientcorrelation (R) assesses the linear correlation between normalized valuesof the predicted binding affinities and the experimental results.[26] Pear-son’s correlation coefficient of the ith gas molecule (Ri) was calculatedbased on ∆Gnorm and, |∆f| for each scoring function. Spearman’s rankcorrelation coefficient assesses the ability of a scoring function to cor-rectly rank the predicted binding affinity of one gas molecule to four dif-ferent pentapeptides.[26] Spearman’s rank correlation coefficient of the ithgas molecule (𝜌i) was calculated based on ∆Gnorm and |∆f| for each scor-ing function. To consider the nature of the solid-gas phase interaction, anadditional parameter known as the coefficient of determination throughthe origin was introduced for evaluating the model’s accuracy, indicatingthat no binding state between a pentapeptide and a gas molecule leadsto no signal response toward the gas molecule. The coefficient of deter-mination through the origin evaluates the linear correlation to the linearregression given by ŷij = ai xij + bi, where xij is the absolute experimentalfrequency shift, |∆f|, ŷij is the normalization of estimated binding affinity,and ai and bi are the slope and the intercept of a linear regression betweenpredicted binding affinity and experimental frequency shift. If ∆Gnorm waszero, meaning no binding affinity, the gas molecules do not adsorb on thepeptide, resulting in a zero value of |∆f|. Thus, the intercept bi in this studywas set to zero. The coefficient of determination through the origin wasdetermined between predicted binding affinity and linear regression:R20i = 1 −∑nj=1(yij − ŷij)2∑nj=1(yij − ȳij)2, (3)where yij and ȳij are the predicted binding affinity and the mean of thepredicted binding affinity. It should be noted that the range of Ri is from–1 to 1, while R20i in this study is always in the range from 0 to 1. Based onthe three metrics, the score, Si for a scoring function is defined asSi = (1 + Ri) + (1 + 𝜌i) + (1 − R20i). (4)The optimization algorithms, described in the following section, mini-mize the Si evaluated for each scoring function based on the set of weights.Optimization Algorithms: In the AutoDock Vina models adapted to thesolid-gas phase, the associated weights of each term in Equations (1) and(2) were optimized for each gas species. As an initial proof of concept,it was started with the same magnitude for all weights in the preliminaryevaluation of the ability of AutoDock Vina for obtaining good agreementbetween predicted binding properties and the response of peptide-coatedsensors. Two optimization algorithms: GA[27] and comprehensive screen-ing via SA were applied.GA was used as the first approach to obtaining optimal linear weightsfor a quick demonstration. In this modified approach, a group of twentydifferent solutions with six weight values each was initialized. Each weightvalue was randomly set between 0 and 1. After molecular dynamics sim-ulation and evaluation with the above metrics, the top ten solutions wereselected as parent solutions for the next generation based on their Si as ex-plained in the previous section. New solutions were generated by perform-ing crossovers on the weight values from the parent solutions. This pro-cess involved selecting two parent solutions and combining their weightvalues to create offspring solutions for the next generation. To maintainAdv. Sensor Res. 2025, 4, 2400122 2400122 (7 of 8) © 2024 The Author(s). Advanced Sensor Research published by Wiley-VCH GmbH 27511219, 2025, 4, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/adsr.202400122 by National Institute For, Wiley Online Library on [11/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://www.advancedsciencenews.comhttp://www.advsensorres.comwww.advancedsciencenews.com www.advsensorres.comdiversity among the solutions and to avoid convergence to a local mini-mum, randomness was introduced by modifying one weight value in thisnew solution. Specifically, a random value between −1 and 1 was added toa weight value, effectively implementing mutation. The new solution setswere then evaluated based on their performance using Si. These new so-lutions were ranked according to their score, and the best performing so-lutions were selected to form new groups for subsequent iterations. Thisiterative cycle of selection, crossover, introduction of randomness (mu-tation), and evaluation was repeated for 1000 cycles. Extensive iterationsallowed the algorithm to thoroughly explore the solution space and in-creased the likelihood of finding the global optimal solution rather thansetting on a local optimum.As the second approach to determining the optimal weight values, acomprehensive screening of all possible combinations of six weight valuesranging from 0 to −1 was conducted with a step size of 0.2. As each of thesix weights could have six possible values, this resulted in over 40,000possible combinations in total. Each combination represented a potentialsolution and was evaluated using Si. After evaluating all combinations, allsolutions were ranked according to their performance scores. Rankingswere then used to determine the best solution for the six weights. Thecomprehensive search method ensured a wide range of weight values toenhance the probability of identifying the global optimum.Statistical Analysis: All statistical analyses were performed usingPython 3.7.11 with the following libraries: pandas (v1.3.5) for data manip-ulation, NumPy (v1.21.6) for numerical computations, SciPy for statisticaltests (v1.7.3), and matplotlib (v3.1.1) for data visualization. Experimentaldata sets were collected from at least three independent replicates andwere expressed as means ± standard deviations.Supporting InformationSupporting Information is available from the Wiley Online Library or fromthe author.AcknowledgementsT.A.N. and M.-Q.F thank NIMS Joint Graduate School Program, NIMS.The calculations in this study were partially performed on the NumericalMaterials Simulator at the NIMS. This work was financially supported par-tially by the Public/Private R&D Investment Strategic Expansion Program(PRISM), Cabinet Office, Japan.Conflict of InterestThe authors declare no conflict of interest.Data Availability StatementThe data that support the findings of this study are available from the cor-responding author upon reasonable request.Keywordsempirical force fields, gas phase model, gas sensor, parameter optimiza-tion, volatile organic compoundsReceived: August 22, 2024Revised: October 18, 2024Published online: November 29, 2024[1] J. Vessman, R. I. Stefan, J. F. van Staden, K. Danzer, W. Lindner, D. T.Burns, A. Fajgelj, H. Müller, Pure Appl. Chem. 2001, 73, 1381.[2] P. Barik, M. Pradhan, Analyst 2022, 147, 1024.[3] J. Wang, E. J. Murphy, J. C. Nix, D. N. M. Jones, Sci. Rep. 2020, 10,3300.[4] A. J. M. Barbosa, A. R. Oliveira, A. C. A. Roque, Trends Biotechnol.2018, 36, 1244.[5] M. Mascini, D. Pizzoni, G. Perez, E. Chiarappa, C. Di Natale, P. Pittia,D. Compagnone, Biosens. Bioelectron. 2017, 93, 161.[6] T. Z. Wu, Y. R. Lo, J. Biotechnol. 2000, 80, 63.[7] T. Z. Wu, Y. R. Lo, E. C. Chan, Biosens. Bioelectron. 2001, 16, 945.[8] D. Pizzoni, M. Mascini, V. Lanzone, M. Del Carlo, C. Di Natale, D.Compagnone, Biosens. Bioelectron. 2014, 52, 247.[9] T. Wasilewski, B. Szulczyński, M. Wojciechowski, W. Kamysz, J.Gębicki, Sensors 2019, 19, 4284.[10] T. Wasilewski, B. Szulczyński, M. Wojciechowski, W. Kamysz, J.Gębicki, Microchem. J. 2020, 154, 104509.[11] M. Blanco, M. C. McAlpine, J. R. Heath, Computational Methods forSensor Material Selection (Eds: M. A. Ryan, A. V. Shevade, C. J. Taylor,M. L. Homer, M. Blanco, J. R. Stetter), Springer US, New York, NY,USA 2010, pp. 135–148.[12] Z. Kuang, S. N. Kim, W. J. Crookes-Goodson, B. L. Farmer, R. R. Naik,ACS Nano 2010, 4, 452.[13] H.-H. Lu, Y. K. Rao, T.-Z. Wu, Y.-M. Tzeng, Sens. Actuators, B 2009,137, 741.[14] S. Panigrahi, S. Sankaran, S. Mallik, B. Gaddam, A. A. Hanson, Mater.Sci. Eng. C Mater. Biol. Appl. 2012, 32, 1307.[15] M. Son, D. Kim, J. Kang, J. H. Lim, S. H. Lee, H. J. Ko, S. Hong, T. H.Park, Anal. Chem. 2016, 88, 11283.[16] S. Ju, K.-Y. Lee, S.-J. Min, Y. K. Yoo, K. S. Hwang, S. K. Kim, H. Yi, Sci.Rep. 2015, 5, 9196.[17] J. Park, J.-M. Lee, H. Chun, Y. Lee, S. J. Hong, H. Jung, Y.-J. Kim, W.-G. Kim, V. Devaraj, E. J. Choi, J.-W. Oh, B. Han, Biosens. Bioelectron.2021, 177, 112979.[18] T. Wasilewski, J. Gębicki, W. Kamysz, Sens. Actuators, B 2018, 257, 511.[19] E. Katchalski-Katzir, I. Shariv, M. Eisenstein, A. A. Friesem, C. Aflalo,I. A. Vakser, Proc. Natl. Acad. Sci. USA 1992, 89, 2195.[20] A. Singh, M. M. Copeland, P. J. Kundrotas, I. A. Vakser, Methods Mol.Biol. 2024, 2714, 101.[21] O. Trott, A. J. Olson, J. Comput. Chem. 2010, 31, 455.[22] J. Eberhardt, D. Santos-Martins, A. F. Tillack, S. Forli, J. Chem. Inf.Model. 2021, 61, 3891.[23] M. Brylinski, Chem. Biol. Drug Des. 2018, 91, 380.[24] R. Quiroga, M. A. Villarreal, PLoS One 2016, 11, e0155183.[25] E. W. Bell, Y. Zhang, J. Cheminform. 2019, 11, 40.[26] C. Wang, Y. Zhang, J. Comput. Chem. 2017, 38, 169.[27] C. S. de Magalhães, D. M. Almeida, H. J. C. Barbosa, L. E. Dardenne,Inf. Sci. 2014, 289, 206.Adv. Sensor Res. 2025, 4, 2400122 2400122 (8 of 8) © 2024 The Author(s). Advanced Sensor Research published by Wiley-VCH GmbH 27511219, 2025, 4, Downloaded from https://advanced.onlinelibrary.wiley.com/doi/10.1002/adsr.202400122 by National Institute For, Wiley Online Library on [11/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons Licensehttp://www.advancedsciencenews.comhttp://www.advsensorres.com Empirical Modification of Force Fields for the Development of Peptide-Based Gas Sensors 1. Introduction 2. Results and Discussion 2.1. Virtual Screening Results 2.2. Liquid Phase Model Performance Assessment with Evaluation Metrics 2.3. Concept Demonstration with GA 2.4. Further Investigation with SA Optimization Method for all Weight Changing Directions 3. Conclusion 4. Experimental Section Supporting Information Acknowledgements Conflict of Interest Data Availability Statement Keywords