In-silico Prediction of the Sweetness of Aspartame Analogues from QSPR Analysis①

CHEN Ke-Xin SHEN Qin-Qin SHEN Shi-Yi ZHOU Xi-To LI Zu-Gung CHEN Zhong-Xiu

?

In-silico Prediction of the Sweetness of Aspartame Analogues from QSPR Analysis①

CHEN Ke-Xiana②SHEN Qian-QianaSHEN Shi-YiaZHOU Xia-TaoaLI Zu-GuangbCHEN Zhong-Xiua③

a(310018)b(310014)

The extensive utilization of the low-energy dipeptide sweetener aspartame in foods leads to various studies on searching for new sweeteners in series. However, the real mechanistic cause of their sweetness power is still not completely known owing to their complex interactions with human sweet receptor, which may be different from that of other sweeteners to some extent. In this contribution, predictive quantitative structure-property relationship (QSPR) models have been developed for diverse aspartame analogues using Materials Studio 5.0 software. The optimal QSPR model (2= 0.913,2=0.881 and2pred=0.730) constructed by the genetic function approximation method has been validated by the tests of cross validation, randomization, external prediction and other statistical criteria, which shows that their sweetness poweris mainly governed by their electrotopological-state indices (S_sssCH and S_ssNH), spatial descriptors (Shadow length: LX, ellipsoidal volume and Connolly surface occupied volume) and topological descriptors (Chi(3): cluster and Chi(0) (valence modified)), which partially supports both multipoint attachment theory proposed by Nofre and Tintiand B-X theory proposed by Kier. Present exploited results provide the key structural features for the sweetness power of aspartame analogues, supplement the mechanistic understanding of the sweet perception, and would be also helpful for the design of potent sweetener analogs prior to their synthesis.

structure-property relationship, genetic function approximation, sweetness potency, dipeptides, correlation;

1 INTRODUCTION

Recently, the taste perception of sweetness, bitter- ness, saltiness, sourness and umami as five accepted basic tastes hasreceived great attentiondue to their importance in many disciplines, specialfood che- mistry[1], and also the non-completely elucidated interaction mechanisms between the taste com- pounds and taste receptors[2]. Among these tastes, sweetness is the benchmark property of a sweet compound that produces pleasant and happy sense- tions[3], which remains one of the most desired tastes for human innately. However, the excessive intake of high calorific sweet foods containing sugars or saccharides has caused many adverse effects towards human health like metabolic disorders, cardiovascular diseases, obesity, type 2 diabetes and other pathologies[4]. The health requirements from consumers have led many academic institutions and food industries to discover alternative intense sweeteners with low or no caloriesnatural extraction[5, 6]or chemical synthesis[7, 8]to replace sugars that are always used at a much higher amount[9].

Though a large number of such sweeteners have been identi?ed, only eight highly potent sweeteners including aspartame, saccharin and neotame have been approved for use by US FDA currently[10,11], and none of them display the same pure sweet pro?le to that of sucrose[6,12]. Some most com- monly used sweeteners such as saccharin[13]and aspartame[8,14,15]have been the center of con- troversies regarding their toxic effects towards human health[16]. Therefore, the growing interest in this field remains the further discovery of new sweeteners with low or no calories, high-quality sweet perception and no toxicity based on the mechanistic understanding of sweet perception of available sweeteners.

Substantial efforts have been made worldwide to understand the mechanistic profile of sweet per- ception, which seems to vary from the chemical, psy-chological and pharmacological perspectives[12,17,18]due to its complexity in human, and is also found to be related to the functionality of receptor, sweetener concentration, multiple tastes of sweetenersor their interactions with coexistent substances[2,7,18]. The sweet chemoreceptor of G protein-coupled recep- tor[19]comprising T1R2 and T1R3 subunits[2,3,20]has been found to recognize sweet stimuli produced by sweeteners and undergoes conformational chan- ges to induce the sweet transduction processes[21]. However, their crystal structures have not been completely explored yet[12]. Consequently, exploringspecific active sites in sweet receptors responsible for the sweet perception remains a great challenge.

So far, several sweet perception theories have been proposed to explain the sweetness caused by sweeteners, such as AB-H theory by Shallenberger and Acree in 1969[22], B-X theory by Kier in 1972[23], and multipoint attachment theory by Nofre and Tintiin 1982[24]and 1996[25]. Many studies have showed that mostsweeteners have more than one hydrogen-bonding AH-B sites and hydrophobic X sites, and not all sweeteners interact with sweet receptoreight interaction sites[1, 18]. Therefore, the complete description of interactions between sweeteners and sweet receptors by these theories to elicit the sweetness is quite difficult. Based on these theories, several arti?cial model sweet receptors that possess both hydrogen-bonding and hydrophobic characteristics[26-28],fullerenol, have been employed to explore the physicochemical basis between the sweeteners and model receptors with or without coexistent substances using isothermal titration calorimetry technique, and some interesting correlations between the thermodynamic parameters and sweetness were established. Nevertheless, many sweeteners have more than one tastes[29], and subtile difference in their chemical structures may change their taste perceptions from sweetness to either tastelessness or other tasteness[18, 30]. Some com- pounds are not sweet but can bind to sweet receptor (false positives), and some are sweet but do not bind to sweet receptor (false negatives)[31]. Therefore, the urgent demand for novel sweeteners provides an impetus for their rational design based on the structural requirements explored by highly efficient methods, not only this mechanistic understanding at molecular level.

The great progress in computational chemistry achieved for the drug design and discoveryin pharmaceutical field offers the mechanistic unders- tanding of sweetness and the rational design of sweeteners efficient and convenient alternative methods, as the sweetness is structure-dependent[32]. The quantitative structure-activity/properties rela- tionship (QSAR/QSPR) studies pioneered by Hansch[33, 34]are one of the useful approaches that statistically correlate the biological activities or properties with the structures of a series of mole- cules in terms of descriptors, which can be quickly done within a very short period of time to screen a large number of compounds irregardless of their complex interactions with their receptors. Many robust and validated QSPR models have been built for sweeteners by statistical methods like genetic function approximation (GFA), support vector machine and artificial neural network to rationalize the structure-sweetness relationships[2, 6, 12, 17, 18, 32, 36-38], which are summarized in some references[2, 18]. Most of these studies focus on the classi?cation-based strategy for sweet, non-sweet and bitter taste using the statistical methods like classi?cation and regres- sion tree, discriminant analysis, k-nearest neighbors and linear learning machine[2, 18], or expert system[2]. Though there are many predictive multiclass QSPR modeling studies on a diverse set of sweete-ners[2, 6, 12, 17, 18, 38], the predicted sweetness residuals of some sweeteners are still large, the chance for over?tting or noise correlations always exist for a large number of sweeteners[17], and the quantitative predictions of sweetness of unknown compounds remainquite rare[17]. Besides, the indispensable molecular descriptors appearing in the established optimal QSPR models vary from the utilizations of different combinations of molecular descriptors generated from different softwares[2, 6, 12, 17, 18, 32, 36-38], which has not been compared and explained detai- ledly for the observed sweetness. Moreover, the cal- culated digital values of some molecular descriptors like spatial descriptorsare largely dependent on the conformations of concerned molecules. However, the sapophoric conformations of sweeteners interac- ting with the sweet receptor remain unresolved. In addition, the statistically significant QSPR models for the specific types of sweeteners are relatively limited[32, 36, 37], and there is still a long way towards the experimental validations from the theoretical predictions.

The object of this paper is to develop statistically significant linear QSPR models for aspartame analogues by GFA method to elucidate the indis- pensable structural requirements for the sweet receptor binding of these sweetenersunder the principles of Organization for Economic Co- operation and Development (OECD)[39]. The high sweet power and extensive utilization of low-energy nutritive aspartame in foods[40], though under some debates[8, 14, 15, 41], has led to the identification of many structurally diversified analogues with sweetness[36,37,42-51]. The signi?cance of QSPR models was evaluated by cross-validation, randomi- zation and external prediction test, which hightlights the structural requirements for the specific aspartame-type sweeteners and their mechanistic interpretations. The results derived can assist food chemists to design new sweeteners in series prior to synthesis.

2 MATERIALS AND METHODS

Unless stated otherwise, all QSPR modeling experiments were carried out by using the QSAR module of Materials Studio 5.0 software (Accelrys Inc)[52].

Table 1. Structures of Aspartame Analogues and Their Experimental and Predicted Sweetness Power (SP) Based on the Best QSPR Modela

aSQ-(= 61～70) are sweeteners in test set. Residual =SP – SP..

2.1 Experimental data selection

Data sets of a series of aspartame analogues (Table 1) in this study were selected from the available publications[36, 37, 42-51], where their relative sweetness to sucrose were converted to the corres- ponding sweetness power () in terms of logarithm of sweetness to reduce theskewness of the data set[53]. The highestvalue for the same sweetener collected from different bibliographic sources was used. The sweeteners were then divided into the training set (60) and test set (10) mainly based on the suggestions by Oprea[54], where (a) thevalues of test set must span some time but should not be larger than that of the training set by more than 10%; (b)the test set must get a balanced number of sweeteners with both strong and weak sweetness power for the uniform sampling of the data. The division was also assisted from the Hierarchical cluster analysis using their generated molecular descriptors and consideration of the structural diversity between the test and training sets, respectively. The resultant meanvalues of training and test sets were 2.756 and 2.251, respec- tively, and the distribution ofvalues of the test set (1.041～3.732) was among that of training set (0.301～4.669). Therefore, the test set captured the common features of training set, and theirvalues could be relatively well predicted.

2.2 Molecular modeling of sweeteners

The initial structures of sweeteners were con- structed using the ChemDraw module of Chem- Office 2004 software, which were transformed into the molfiles with Chem3D module. The/configurations of structures were checked and adjusted using the materials visualizer of Materials Studio 5.0 software, followed by geometrical minimizations in Forcite module to get the lower energy conformations by using smart algorithm, Universal force field, Gasteiger charge method and other default settings. The molecular structures of most aspartame analogues were constructed based on the low-energy conformation of aspartame before their geometrical minimizations. The resultant xsd files of sweeteners were then used for the calcula- tions of molecular descriptors.

2.3 Calculation of descriptors in different categories

Different categories of descriptors, namely E-state keys, spatial, thermodynamic, structural, topological and information were generated in the study table for the above optimized structures using the default settings within QSAR module of Materials Studio 5.0 software, and the descriptors with zero values for the most sweeteners were discarded. Prior to the QSPR model generation, the resultant 149 descriptors were considered for their inter-correlation, and 73 descriptors (Table 2) were kept for the following study after removing the highly pairwise-correlated descriptorstheir correlation matrix analysis, and also to make sure that the descriptors in the final optimal QSPR model are orthogonal.

Table 2. Descriptors and Their Categories Used to Build the QSPR Models

2.4 Development of QSPR models and their validation methods

All QSPR models were generated by the GFA methoddeveloped by Rogers and Hopfinger[55]. GFA method is an excellent statistical methodology to build significant correlation between the physicochemical property or bioactivity of a series of compounds and their molecular parameters and also improve theconvenient interpretation of the resultant correlations. In this work, the length of QSPR model was initially fixed to 5, which was changed from 1 to 8 to evaluate the statistical evolu- tion of QSPR models. After several preliminary calculations, the maximum generations of 5000, population size of 50, mutation probability of 0.1, scoring function of, the scaledsmoothness parameter of 0.5and other default settings were set to give reasonable convergence. The cross-validated test (2, Eq. 1)[56, 57], randomi- zation test[58, 59], and external prediction test (2, Eq. 2)[60]were performed to determine the reliability, significance and predictive power of the generated models.

3 RESULTS AND DISCUSSION

3.1 Generation of the QSPR models

Various QSPR models were generated to correlate the SP values of sweeteners and their remaining different descriptors by GFA method within Materials Studio 5.0 software. The optimal QSPR model was mainly selected on the basis of statistical parameters like, square of the correlation coefficient (2), adjusted square of the correlation coefficient (2),cross-validated2(2), significance-of-regression F-value (), and criticalat 95% confi- dence level.denotes the slightly modified Friedman’s lack of fit score, which is used to evaluate the excellence of each progeny QSPR model[61]. The smaller theis, the more excellent the QSPR model will be.2is an internal validation indicator of the QSPR model data fit,2is an adjusted2estimation, and2is a critical measurement of the predictive power of a QSPR model[56]. The closer the values of2,2and2are to 1.0, the better the regression QSPR model is generated, where2should be fairly close to2to eliminate the probable data overfitting in the QSPR model[62]. The largerover the corresponding criticalat 95% con- fidence level indicates the statistically significant regression of the QSPR model.

Table 3. Statistical Evolution of QSPR Models for the Sweetness Power with Different Numbers of Descriptorsa

avalue denotes the significance-of-regression F-values. The values in the parenthesis refer to the criticalvalues at 95% confidence level.

Fig. 1. Plot of correlation coefficients2and2as a function of number of descriptors

The number of descriptors necessary and ade- quate to appear in the optimal QSPR model was firstly evaluated by a brute force approach[57, 58, 63]. The relevant results are shown in Table 3, and the relationship between both2and2and the number of descriptors was plotted in Fig. 1. Obviously, the2,2and2increased anddecreased with increasing the number of descriptors appearing in QSPR models, which become almost constant after the number of descriptors over 7. As the QSPR models with more descriptors would be more difficult to interpret and also cause over-fitting regression, and the final optimal QSPR model generated in this paper consisted of 7 descriptors. The statistically significant QSPR model with 7 descriptors is shown below, and these descriptors (given in Table S1) were found to be almost orthogonaltheir inter-correlation analysis. The summary of the best QSPR models with 7 descriptors is given in Table S2.

The optimal QSPR model:

= –1.160881180 × Chi(3): cluster + 1.125360407 × Chi(0) (valence modified) + 1.285393123 ×

S_sssCH + 0.118681048 × S_ssNH – 0.002021353 × ellipsoidal volume + 0.167221851 ×

Shadow length: LX – 0.030549168 × Connolly surface occupied volume + 1.157529317

N= 60,= 0.622,2= 0.913,2= 0.901,= 78.10,2= 0.881

n= 10,2= 0.730,2= 0.800

whereNandnare the number of sweeteners in the training and test sets, respectively;2is the correlation coefficient between the experimental and predictedvalues of test set. The meanings of the other statistical parameters are given in the above text.

3.2 Validation of the QSPR model

To determine the reliability, predictivity and significance of the above obtained optimal QSPR model, the LOO cross-validated test[56, 57], randomi- zation test[58, 59]and external prediction test[60]were employed. The2of 0.881 resulting from the cross-validation test (Eq. 1) indicated that the yielded QSPR model was not obtained by chance correlation. The randomization tests at 90%, 95%, 98% and 99% confidence levels were performed by using the randomization of response variables (9, 19, 49 and 99 trials) with multiple linear regression (MLR) method, where the same SP values of training set and the descriptors of the yielded QSPR model were used for this calculation. The randomi- zation test results (Table 4 and Fig. 2) showed that the random2produced from all permuted data sets (maximum2of 0.250) was significantly lower than that of the nonrandom2of the primary QSPR model (2= 0.913), which suggested that the yielded QSPR model was statistically signi?cant and pre- dominant[64].

Fig. 2. Histogram of the results of randomization tests for QSPR models at 99% confidence level

Table 4. Results of Randomization Tests for QSPR Models Generated by MLR Method

Besides, the predictive power of the yielded QSPR model was calculated by2(Eq. 2) based on the external prediction of thevalues of ten sweeteners in the test set. The predictedvalues of test set were given in Table 1, and the correlation (2= 0.896) between the experimental and predictedvalues of all sweeteners studied in this paper were depicted in Fig. 3. The modest2of 0.730 accounted for the relatively good predictive ability of theyielded QSPR model. The absolute residuals of training and test sets were ranging from 0.002 to 0.802 (0.281 ± 0.212) and from 0.099 to 0.939 (0.436 ± 0.272), respectively, which mean that the predicted SP values of sweeteners in series from yielded QSPR model should consider this residual range.

Moreover, the yielded QSPR model was also validated by the statistical criteria of a predictive QSPR model proposed by Golbraikh and Tropsha[62]. A QSPR model could be considered acceptable if it satis?ed the following criteria:2CV> 0.5,2> 0.6, [(2–R2)/2] < 0.1, 0.85≤≤1.15, where2is the correlation coefficient between the experimental and predictedvalues of test set, andandR2were computed by using the following Eqs. 3 and 4, respectively.

Fig. 3. Correlation between the experimental and predicted sweetness power

In addition, the novelr2metrics test proposed by Roy group[65]was employed to validate the true predictivity of yielded QSPR model, which was calculated based on the correlation coefficient between the experimentalvalues in the-axis and the predictedvalues in-axis for the least- squares regression lines using the following Eq. 5.

Table 5. Occurrences of Selected Significant Descriptors in Population during the Generation of QSPR Models by GFA Method with the Number of Descriptors Ranging from 5 to 8a

aPopulation size was set at 50 in this work.

3.3 Interpretation of the optimal QSPR model

The optimal QSPR model with 7 descriptors could explain and predict 91.3% and 88.1% of the variance, respectively. The presence of three descriptors with negative coef?cients in the QSPR model indicates that the large value of Chi(3): cluster, ellipsoidal volume or Connolly surface occupied volume decreases the sweetness power, while Chi(0) (valence modified), Shadow length: LX, S_sssCH and S_ssNH could increase the sweetness power due to their positive correlations with SP values. The signi?cance of these descriptors can be seen in both Tables 3 and 5 in terms of their occurrence during the QSPR model generations with different numbers of descriptors.

In the optimal QSPR model, both Chi (3): cluster and Chi (0) (valence modified) belong to the topological descriptors of Kier & Hall chi con- nectivity indices[66], where the former is the third- order subgraph molecular connectivity index that involves four skeletal atoms in a trigonal rela- tionship, and the latter is the zero-order valence- modified connectivity index that takes into account the electron configuration of the atom represented by the vertex. Their larger correlation coef?cients indicate that such atom connectivity within a sweetener distinguished by order and subgraph type largely affects the sweetness power. The S_sssCH and S_ssNH are the descriptors of E-state keys, which encodes the sum information of both atomic topological environment and electronic interaction in given molecule[67]. The former index represents the atomic type of -CH< in alkanes, and the latter index stands for the counter descriptor for a N bonded to one hydrogen and a single bond (atomic type of -NH-). The shadow indices[68]represent a set of spatial descriptors that characterize the molecular shape on the basis of molecular conformation and orientations. Shadow length: LX is the length of molecule in the X dimension, and its positive correlation withvalue means that sweetness power can increase when such molecular length can be designed longer. Ellipsoidal volume[52]describes the volume of the ellipsoid of inertia derived from the inertia tensor of the molecule, which has axes proportional to the inverse of the square root of the principal moments of inertia that is aligned along the principal axes. Connolly surface occupied volume[52]could reflect the accessible region that a sweetener interacts with the sweet receptor. Their negative correlation withvalues indicates that relatively small interacting volume between sweeteners and the sweet receptor could improve the sweetness power.

The appearance of these significant descriptors in optimal QSPR model partially supports both multi- point attachment theory proposed by Nofre and Tinti[24, 25]and also the B-X theory proposed by Kier[23]for sweetness. MPA theory assumes that a sweetener interacts with human sweetness receptor at recognition site at least eight binding sites namely B, AH, XH, G(= 1～24) and D, where Gn represents the punctual intermo- lecular steric interaction of nonpolar or weakly polar groups like -CH2-, >CH- and -CH3, B denotes an anionic group or H-bond acceptor atoms; AH and XH refer to the H-bond donor groups; and D is often the H-bond acceptor 4-cyanophenyl group, while B-X theory assumes that a hydrophobic binding site (X) is also needed for the interaction between a sweetener and sweetness receptor except for the essential H-bond acceptor-donor interactions (AH···B). In the optimal QSPR model, we find that Chi(0) (valence modified) and S_ssNH contribute to the H-bond acceptor-donor interaction; S_sssCH contributes to the hydrophobic interaction; Chi(3): cluster steric, Shadow length:LX, ellipsoidal volume and Connolly surface occupied volume favor the steric interaction. Therefore, three types of elementary interactions have been extracted, which are useful for better understanding of the sweetness of aspartame analogues that they may have more than one AH-B site with sweet receptors.

4 CONCLUSION

In this study, statistically significant linear QSPR models were built based on GFA statistical method using various descriptors comprising E-state keys, spatial, thermodynamic, structural, topological and information with the purpose to derive the indis- pensable structural requirements for the sweetness power of aspartame analogues in series. The optimal QSPR model was validated by LOO cross-validation test[56, 57], randomization test at four confidence levels[58, 59, 69], external prediction test[60]and other statistical criteria[62, 65], which shows that the sweetness power of aspartame analogues can be enhanced based on the following strategy: relatively smaller ellipsoidal volume and their accessible volume with the sweet receptor; relatively longer molecular length in X length; larger sum of the E-state values of atomic types of both -CH< and –NH-; larger molecular Chi (0) (valence modified) indice and smaller molecular Chi (3): cluster indice. The outcome in this study can supplement the availa- ble design strategies of sweeteners[2, 6, 12, 17, 18, 32, 36-38], partially fit in with the multipoint attachment theory and B-X theory for sweetness, and also provide a helpful guidance for the design or screen of more sweet aspartame analogues prior to their synthesis. As the interactions between sweeteners and sweet receptor can cause their dynamic mutual confor- mational changes, and revealing the conformational flexibility of aspartame is interesting to obtain the shape of active site of the sweet receptor[40], in our future work, we will address the conformational effect of aspartame analogues on the sweet mecha- nism by using theoretical methods like molecular dockingand molecular dynamics simulation to get a mature understanding of the causes and differences of their sweetness.

(1) Damodaran, S.; Parkin, K. L.; Fennema, O. R.. CRC Press, Boca Raton 2008, p1-35.

(2) Rojas, C.; Todeschini, R.; Ballabio, D.; Mauri, A.; Consonni, V.; Tripaldi, P.; Grisoni, F. A QSTR-based expert system to predict sweetness of molecules.2017, 5, 53-12.

(3) Carocho, M.; Morales, P.; Ferreira, I. C. F. R. Sweeteners as food additives in the XXI century: a review of what is known, and what is to come.2017, 107, 302-317.

(4) Lustig, R.; Schmidt, L.; Brindis, C. Public health: the toxic truth about sugar.2012, 482, 27-29.

(5) Behrens, M.; Meyerhof, W.; Hellfritsch, C.; Hofmann, T. Sweet and umami taste: natural products, their chemosensory targets, and beyond.2011, 50, 2220-2242.

(6) Chéron, J. B.; Casciuc, I.; Golebiowski, J.; Antonczak, S.; Fiorucci, S. Sweetness prediction of natural compounds.2017, 221, 1421-1425.

(7) DuBois, G. E.; Prakash, I. Non-caloric sweeteners, sweetness modulators, and sweetener enhancers.2012, 3, 353-380.

(8) Chattopadhyay, S.; Raychaudhuri, U.; Chakraborty, R. Artificial sweeteners-a review.. 2014, 51, 611-621.

(9) Asadi, M.Wiley Press, New York 2006, p15-30.

(10) Godshall, M. A. The expanding world of nutritive and non-nutritive sweeteners.2007, 69, 12-20.

(11) http://www.fda.gov/Food/IngredientsPackagingLabeling/FoodAdditivesIngredients/ucm397725.htm.

(12) Goel, A.; Gajula, K.; Gupta, R.; Rai, B. In-silico prediction of sweetness using structure-activity relationship models.2018, 253, 127-131.

(13) Ellwein, L.; Cohen, S. The health risks of saccharin revisited.1990, 20, 311-326.

(14) Magnuson, B. A.; Burdock, G. A.; Doull, J. Aspartame: a safety evaluation based on current use levels, regulations, and toxicological and epidemiological studies.2007, 37, 629-727.

(15) Mahnam, K.; Raisi, F. A theoretical and experimental study of calcium, iron, zinc, cadmium, and sodium ions absorption by aspartame2017, 43, 87-103.

(16) Suez, J.; Korem, T.; Zeevi, D.; Zilberman-Schapira, G.; Thaiss, C. A.; Maza, O.; Elinav, E. Arti?cial sweeteners induce glucose intolerance by altering the gut microbiota.2014, 514, 181-186.

(17) Ojha, P. K.; Roy, K. Development of a robust and validated 2-QSPR model for sweetness potency of diverse functional organic molecules.2018, 112, 551-562.

(18) Rojas, C.; Ballabio, D.; Consonni, V.; Tripaldi, P.; Mauri, A.; Todeschini, R. Quantitative structure-activity relationships to predict sweet and non-sweet tastes.2017, 135, 66-78.

(19) Trzaskowski, B.; Latek, D.; Yuan, S.; Ghoshdastider, U.; Debinski, A.; Filipek, S. Action of molecular switches in GPCRs-theoretical and experimental studies.2012, 19, 1090-1109.

(20) Sainz, E.; Korley, J. N.; Battey, J. F. Identi?cation of a novel member of the T1R family of putative taste receptors.2001, 77, 896-903.

(21) Digby, G.; Lober, R.; Sethi, P.; Lambert, N. Some G protein heterotrimers physically dissociate in living cells.2006, 103, 17789-17794.

(22) Shallenberger, R.; Acree, T. Molecular structure and sweet taste.1969, 17, 701-703.

(23) Kier, L. A molecular theory of sweet taste.1972, 61, 1394-1397.

(24) Tinti, J.; Nofre, C.; Peytavi, A. Interaction of suosan with the sweet taste receptor.1982, 175, 266-268.

(25) Nofre, C.; Tinti, J.; Glaser, D. Evolution of the sweetness receptor in primates. II. Gustatory responses of non-human primates to nine compounds known to be sweet in man.1996, 21, 747-762.

(26) Chen, Z.; Guo, G.; Deng, S. Isothermal titration calorimetry study of the interaction of sweeteners with fullerenols as an arti?cial sweet taste receptor model.2009, 57, 2945-2954.

(27) Chen, Z.; Wu, W.; Zhang, W.; Deng, S. Thermodynamics of the interaction of sweeteners and lactisole with fullerenols as an arti?cial sweet taste receptor model.2011, 128, 134-144.

(28) Dong, W.; Chen, G.; Chen, Z.; Deng, S. Thermodynamics of the enantiomers of amino acid and monosaccharide binding to fullerenol used as an arti?cial sweet taste receptor model.2013, 141, 3110-3117.

(29) Shamil, S.; Birch, G.; Mathlouthi, M.; Clifford, M. Apparent molar volumes and tastes of molecules with more than one sapophore.1987, 12, 397-409.

(30) deMan, J. M.. Springer Press, Berlin 1999, p1-10.

(31) Assoli, A.; Laureati, M.; Borgonovo, G.; Morini, G.; Servant, G.; Pagliarini, E. Isovanillic sweeteners: sensory evaluation and in vitro assays with human sweet taste receptor.2008, 1, 174-183.

(32) Nunes, C. A.; Freitas, M. P. Aug-MIA-QSPR study of guanidine derivative sweeteners2013, 237, 565-570.

(33) Hansch, C.; Fujita, T. Correlation of biological activity of phenoxyacetic acids with Hammett substituent constants and partition coefficient.1962, 194, 178-180.

(34) Hansch, C.; Fujita, T.--analysis. A method for the correlation of biological activity and chemical structure.1964, 86, 1616-1626.

(35) Douguct, D.; Thoreau, E.; Grassy, G. A genetic algorithm for the automated generation of small organic molecules: drug design using an evolutionary algorithm.2000,14, 449-466.

(36) Iwamura, H. Structure-sweetness relationship of L-aspartyl dipeptide analogs. A receptor site topology.1981, 24, 572-583.

(37) Tarko, L.; Lupescu, I. QSAR Studies on amino-succinamic acid derivatives sweeteners.2006, 13, 22-40.

(38) Zhong, M.; Chong, Y.; Nie, X.; Yan, A.; Yuan, Q. Prediction of sweetness by multilinear regression analysis and support vector machine.2013, 78, S1445-S1450.

(39). http://www.oecd.org/env/ehs/risk-assessment/validationofqsarmodels.htm.

(40) Toniolo, C.; Temussi, P. Conformational flexibility of aspartame.2016, 106, 376-384.

(41) Kirkland, D.; Gatehouse, D. Aspartame: a review of genotoxicity data.2015, 84,161-168.

(42) Mazur, R. H.; Goldkamp, A. H.; James, P. A.; Schlatter, J. M. Structure-taste relation of aspartic acid amides.1970, 13, 1217-1221.

(43) Mazur, R. H.; Reuter, J. A.; Swiatek, K. A.; Schlatter, J. M. Synthetic sweeteners. 3. Aspartyl dipeptide esters from L- and D-alkylglycines.1973, 16, 1284-1287.

(44) Kawai, M.; Nyfeler, R.; Berman, J. M.; Goodman, M. Peptide sweeteners. 5. Side chain homologs relating zwitterionic and trifluoroacetylated amino acid anilide and dipeptide sweeteners.1982, 25, 397-402.

(45) Yamashita, Y.; Takahashi, Y.; Takayama, C.; Sumi, K.; Nakatsuka, K.; Ohkubo, T.; Abe, H.; Sasaki, S. Structure-taste correlation of L-aspartyl dipeptides using the SIMCA method.1986, 29, 906-912.

(46) Zeng, G.; Chen, J.; He, H.; Wang, Z.; Yan, J. In the pursuit of a better sweetener.1991, 39, 782-785.

(47) Polinelli, S.; Broxterman, Q. B.; Schoemaker, H. E.; Boesten, W. H. J.; Crisma, M.; Valle, G.; Toniolo, C.; Kamphuis, J. New aspartame-like sweeteners containing L-(Me)phen.1992, 2, 453-456.

(48) Hooper, N. M.; Hesp, R. J.; Tieku, S.; Boesten, W. H. J.; Toniolo, C.; Kamphuis, J. Stability of n-derivatized and .alpha.-methyl analogs of aspartame to hydrolysis by mammalian cell-surface peptidases.1994, 42, 1397-1401.

(49) Mossel, E.; Formaggio, F.; Crisma, M.; Toniolo, C.; Broxterman, Q. B.; Boesten, W. H. J.; Kamphuis, J.; Quaedflieg, P. J.; Temussi, P. Aspartame dipeptide analogues: effect of number of side-chain methylene group spacers and-methylation in the second position.1997, 8, 1305-1314.

(50) Frank, M.; Aitken, D. J. On the sweetness of-(trifluoroacetyl)aspartame.2000, 64, 1982-1984.

(51) Goodman, M.; Del Vale, J. R.; Amino, Y.; Benedetti, E. Molecular basis of sweet taste in dipeptide taste ligands... 2002, 74, 1109-1116.

(52) Materials Studio Tutorials,, Accelrys Inc., San Diego, CA, USA.

(53) Chen, K. X.; Xie, H. Y.; Li, Z. G. QSAR analysis of 1,1-dioxoisothiazole and benzo[b]thiophene-1,1-dioxide derivatives as novel inhibitors of Hepatitis C Virus NS5B polymerase... 2009, 56, 684-693.

(54) Oprea,T. J.; Waller, G. L.; Marshall, G. R. Three-dimensional quantitative structure-activity relationship of human immunodeficiency virus(I) protease inhibitors. 2. Predictive power using limited exploration of alternate binding modes.... 1994, 37, 2206-2215.

(55) Rogers, D.; Hopfinger, A. J. Application of genetic function approximation to quantitative structure-activity relationships and quantitative structure-property relationships....1994, 34, 854-866.

(56) Schüürmann, G.; Ebert, R. U.; Chen, J.; Wang, B.; Kühne, R. External validation and prediction employing the predictive squared correlation coefficient-test set activity meantraining set activity mean.2008, 48, 2140-2145.

(57) Wu, W.; Zhang, C.; Lin, W.; Chen, Q.; Guo, X.; Qian, Y.; Zhang, L. Quantitative structure-property relationship (QSPR) modeling of drug-loaded polymeric micelles via genetic function approximation.2015, 10, e0119575-16.

(58) Chen, K. X.; Xie, H. Y.; Li, Z. G.; Gao, J. R. Quantitative structure-activity relationship studies on 1-aryl-tetrahydroisoquinoline analogs as active anti-HIV agents2008, 18, 5381-5386.

(59) Zhang, S.; Wei, L.; Bastow, K.; Zheng, W.; Brossi, A.; Lee, K. H.; Tropsha, A. Antitumor agents 252. Application of validated QSAR models to database mining: discovery of novel tylophorine derivatives as potential anticancer agents.. 2007, 21, 97-112.

(60) Roy, P. P.; Roy, K. On some aspects of variable selection for partial least squares regression models.2008, 27, 302-313.

(61) Friedman, J. H. Multivariate adaptive regression splines.1991, 19, 1-67.

(62) Alexander, G.; Alexander, T. Beware of2.2002, 20, 269-276.

(63) Chen, K. X.; Li, Z. G.; Xie, H. Y.; Gao, J. R.; Zou, J. W. Quantitative structure-activity relationship analysis of aryl alkanol piperazine derivatives with antidepressant activities.2009, 44, 4367-4375.

(64) Nair, P. C.; Sobhia, M. E. Quantitative structure-activity relationship studies on thiourea analogues as influenza virus neuraminidase inhibitors.2008, 43, 293-299.

(65) Ojha, P. R.; Mitra, I.; Das, I.; Narayan, R.; Roy, K. Further exploring rm2metrics for validation of QSPR models.2011, 107, 194-205.

(66) Kier, L. B.; Hall, L. H.; Murray, W. J.; Randic, M. Molecular connectivity 1: relationship to nonspecific local anesthesia.1975, 64, 1971-1974.

(67) Hall, L. H.; Kier, L. B. Electrotopological state indices for atom types: a novel combination of electronic, topological, and valence state information.1995, 35, 1039-1045.

(68) Rohrbaugh, R.; Jurs, P. C. Descriptions of molecular shape applied in studies of structure-activity and structure-property relationships.1987, 199, 99-109.

(69) Chen, K. X.; Xie, H. Y.; Li, Z. G. 2D-QSAR studies on anthranilic acid derivatives: a novel class of allosteric inhibitors of hepatitis C NS5B polymerase.2009, 28, 1217-1225.

7 May 2018;

16 July 2018

the National Natural Science Foundation of China (No. 21673207), Special Fundamental Research Fund for the Central Public Scientific Research Institutes (No. 562018Y-5983) and Zhejiang Provincial Collaborative Innovation Center of Food Safety and Nutrition (No. 2017SICR115, 2017SICR101). Dr. Kexian Chen gratefully acknowledges Prof. Zhang Li-Juan of South China University of Technology for kindly providing the QSAR module of Materials Studio 5.0 software

. Chen Ke-Xian. E-mail: ckx_chem@zju.edu.cn

. Chen Zhong-Xiu. E-mail: zhxchen@zjsu.edu.cn

10.14102/j.cnki.0254-5861.2011-2062