Doi:10.1016/j.theochem.2005.06.032

Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 Prediction of antifungal activity by support vector machine approach Shi-Wei Chena, Ze-Rong Lib,*, Xiang-Yuan Lia,** aCollege of Chemical Engineering, Sichuan University, Chengdu 610064, People’s Republic of China bCollege of Chemistry, Sichuan University, Chengdu 610065, People’s Republic of China Received 13 December 2004; revised 30 June 2005; accepted 30 June 2005 A set of molecular descriptors, including electronic descriptors, topological descriptors, geometric descriptors and molecular shape indices, are calculated to characterize the structural and physicochemical properties for 94 chemical compounds: 42 antifungal active and 52inactive. Support Vector Machine (SVM) classification method is employed to model the discrimination between the antifungal activity andinactivity for these compounds. Leave-one-out (LOO) cross-validation method is used to optimize the SVM model and a genetic algorithm isused in variable selection, this reduces the number of molecular descriptors from 67 to 30. Five-fold cross-validation method and anindependent evaluation set are used to test SVM model. The training sets are effectively and evenly chosen in the descriptor space byclustering based on their chemical similarity, and both of the test methods give consistent results with the LOO method. Compared to theLOO method, 5-fold cross-validation method or the independent test method requires much less time for the SVM model optimization and itis impossible to do a LOO cross-validation for a very large data set. Our work suggests that a proper choice of training set for 5-fold cross-validation method or the independent test method can give consistent results with the LOO method. Comparison of the results by SVMmethod and those by other statistical classification methods, for example k-nearest neighbor (k-NN) and C4.5 decision tree that use the samepre-selected molecular descriptors, is also conducted. Our investigation indicates the potential of SVM in facilitating the prediction ofantifungal activity.
q 2005 Elsevier B.V. All rights reserved.
Keywords: Support vector machines; Antifungal activity; Variable selection; Training set design; Genetic algorithm compounds with a broad antifungal spectrum that can betaken as a starting point for the development and During the past two decades, the prevalence of systemic fungal infections increased significantly due to the Quantitative structure-activity relationship (QSAR) rep- increasing use of broad-spectrum antibiotics, immunosup- resents an attempt to correlate structure descriptors of pressive agents, hyperalimentation products and central compounds with their biological activity. Conventional venous catheters, intensive care of low birth weight infants, QSAR approaches have been challenged by combinatorial organ transplantation, and the acquired immunodeficiency chemistry and high throughput screening (HTS) which are syndrome (AIDS) epidemic Invasive fungal infections innovative techniques adopted by the pharmaceutical and have become the important causes of morbidity and agrochemical industries in an effort to reduce costs and mortality in immunocompromised patients None of shorten drug discovery timelines. HTS produces a large the existing systemic antifungals satisfies medical need amount of screening data, which in most cases identifies completely It is urgent to find new chemical compounds as either active or inactive Compounds inHTS assay have diverse structures which make it difficult to * Corresponding authors. Tel.: C86 28 8540 3231; fax: C86 28 8540 analyze HTS data using conventional QSAR methods and make reliable predictions . Classification and pattern ** Tel.: C86 28 8540 5233; fax: C86 28 8540 7797.
recognition have become a necessary step in QSAR to E-mail addresses: lizrscu@yahoo.com.cn (Z.-R. Li), xyli@scu.edu.cn analyze the large amount of data produced by combinatorial chemistry and HTS. Since Lipinski’s rule of 5 should not be 0166-1280/$ - see front matter q 2005 Elsevier B.V. All rights reserved.
applied to antibiotics, antifungals, vitamins and cardiac S.-W. Chen et al. / Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 glycosides, a correspondingly simple classification into inactive The antifungal compounds contain allylamines, antifungal active and inactive is not advocated.
imidazoles, thiocarbamtes, triazole and other derivatives, A lot of computational methods are available nowadays showing a great diversity in molecular structure.
to discover new drugs with antifungal activity, such asmultiple linear regression (MLR), linear discriminant analysis (LDA), comparative molecular similarityindices analysis (CoMSIA), genetic function approxi- In our statistical investigation, 67 descriptors are mation (GFA), artificial neural network (ANN), calculated to encode structural and physicochemical proper- and so forth. In particular, the LDA model showed ties of molecules, including topological descriptors, promising capability of antifungal activity prediction and electronic descriptors, geometrical descriptors, descriptors achieved a prediction accuracy of 60% for the actives and of based on charged-partial surface area and a series of 98% for inactive A new statistical learning method, molecular shape descriptors defined by us.
support vector machine (SVM) is very useful for classification of systems with multiple mechanisms, such as the prediction of blood-brain barrier penetration, P-glycoprotein substrates, HIV protease cleavage sites in protein, and protein fold recognition In those cases, SVM is found to be superior to other with K Z n1 C n2 C n3, n1, n2, n3Z0,1,2,3., Edt represents an integral over the body which may be a volume or a Since thousands of molecular descriptors are available surface. As defined, these quantities depend on rotation and for QSAR analysis, and only a subset of them is statistically translation of the molecules. Therefore, Covell and co- significant in terms of correlation with biological activity workers defined K order invariant moment (KZ2n) as for a particular QSAR, deriving the optimal subset for a QSAR model through variable selection needs to be addressed. Several variable selection techniques, includingsimulated annealing recursion feature elimination G2n is invariant with respect to rotations and translation after , and genetic algorithm, have been used to translating the molecules to the center-of-mass of the body.
select variables to improve the classification performance of Based on the above illustrations, we define a new K order moment shape index as given in Eq. (3), in order to encode This work applies SVM as a pattern recognition method for the antifungal activity of compounds. A set of moleculardescriptors including a series of molecular shape descriptors defined by us are calculated to characterize the structural and physicochemical properties for known antifungals and K(real) is the moment of actual molecule and inactive compounds. The compounds are used for develop- K(sphere) the moment of a sphere with the same volume as the actual molecule. c(0) is the usually defined ovality.
ing the SVM system of prediction. A widely used feature These molecular shape descriptors encode the degree of the selection method, genetic algorithm (GA), is adopted for the deviation of the shape of one molecule from a sphere. The variable selection in order to find the informative features.
programs computing the new descriptors described here are The trained SVM system then is used to classify the written by the authors using Fortran, but they are still under chemical compounds as active or inactive. The classifi- cation accuracy of this system is evaluated using two In this work, atomic charges are calculated according to methods: an independent set and 5-fold cross-validation.
partial equalization of orbital electronegativity (PEOE) The training sets are effectively and evenly chosen in the method of Gasteiger et al. Electronical descriptors descriptor space by clustering based on the structure are all based on these PEOE charges. Descriptors used in similarity of the compounds, and the results by SVM this work are listed in All of these descriptors for method are compared with those by decision tree method each compound are computed using our designed molecular and k-nearest neighbor method, using the same sets of compounds and molecular descriptors.
SVM is a supervised machine learning technique for learning classification and regression rules from data.
A good introduction to SVM can be found in literature The compounds investigated in this work are listed in . Here we give only the main idea of SVM for , 42 compounds with antifungal activity and 52 S.-W. Chen et al. / Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 Table 1Prediction from SVM by LOO and 5-fold cross validation *Incorrectly classified compounds in the test set and independent evaluation set.
a The number of antifungal active compound in this paper.
c Tr, training set; Ts, test set; Ind, independent evaluation set.
d The number of inactive compound in this paper.
S.-W. Chen et al. / Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 Second order moment shapeindex (based on moleculesurface area) Given a training data set {xi, yi}, iZ1,2,.,N, where yi2{K1,1} represents the label of the classification of an arbitrary sample xi2Rd, d being the dimension of the input space. If the training data are linearly separable, the original which maximizes the margin between positive samples (yiZ1) and negative examples (yiZK1). This can be solved through the minimization of kwkunder the constraints of S.-W. Chen et al. / Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 work. Secondly, the fitness of each chromosome is evaluated by the cross-validated predictive accuracy of the SVM model. Thirdly, a new population is created by one- point crossover and mutation of the chromosomes that are selected from the population in an arbitrary proportion with highest fitness. Typical mutation rate of 1% and crossover where sgn is a sign function. A linear classifier may not be rate of 25% are chosen. Finally, we turn back to second step the most suitable hypothesis for the two classes. For non- until the number of generations reaches a given maximum.
linear classifier, the SVM maps the data to some higher After these steps, an optimal subset of molecular descriptors dimensional feature space by introducing a kernel function will be obtained for a given s. By the optimization of s, a k(x,y) and constructing a separating hyperplane in this SVM model, which gives the best prediction accuracy, is space. The resulting decision function becomes and the solution is obtained by maximizing There are various ways to measure the prediction performance, and some of them are more suitable than others, depending on the application considered .
The most common measure of overall performance is Qtotal which is the fraction of correctly predicted antifungal active compound and inactive compound among all predictions.
As for the kernel function, the most widely used forms are where TP, TN, FP and FN are the numbers of the truepositive, the true negative, the false positive and the false negative, respectively. To get a measure on the sensitivity of For the model selection in SVM, the SVM minimizes a prediction performance, SEZ TP=ðTPC FNÞ, the fraction of bound on the expected generalization error for a test set, correctly predicted antifungal compounds among observed instead of minimizing error on the training set. This is antifungal compounds, is used. Similarly, to get a measure accomplished by minimizing a composite error, comprised of the training error plus a regularization term relating to the SPZ TN=ðTNC FPÞ, the fraction of correctly predicted inactive compounds among observed inactive compounds,is used.
ability in the SVM model optimization through the LOOcross-validation.
Genetic algorithm (GA) models the process of natural evolution in which species with a high fitness can prevailand survive to the next generation, and the best species can be adapted by crossover and/or mutation to search for thebetter individuals In this work, a chromosome and Because of the limitation of LOO method for the SVM its fitness in the species represent the encoding of a set of model optimization for a large data set, two widely used molecular descriptors and the predictive accuracy of additional methods are employed in the present work to the SVM model, respectively. We use the leave-one-out evaluate SVM model. In the first method, the compounds (LOO) cross-validation method to evaluate the average are divided into three sets: a training set, a test set and an generalization ability of the SVM model. The optimization independent set. The training set are used to train the SVM, of the model includes the optimization of the exponent s in and the test set are used to optimize the exponent in the Gaussian kernel and the optimal choice of molecular Gaussian kernel function by maximizing the generalization descriptors. Our algorithm for the model optimization ability, and the independent set are used to test the consists of four steps for a given s parameter in SVM.
prediction ability of the final model.
The first step is the creation of an initial population. The In the selection of compounds in a training set, one initial population of chromosomes is created by setting all molecular descriptor corresponds to one dimension of the bits in each chromosome to a random value (1 or 0). Bit ‘1’ multidimensional space. In order to make the space sampled denotes a selection of a variable, and bit ‘0’ denotes a non- effectively and evenly, we use the k-means clustering selection. The size of the initial population is 30 in this algorithm for the training set design. At first the compounds S.-W. Chen et al. / Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 are divided into a given number of clusters according to the size of the training set, using the k-means clustering algorithm. The distances between two compounds in chemical space is the Euclidean distance. In this way, compounds that are close in the space are in the same cluster. Then the compound nearest to the center of each cluster is chosen as a member of the training set. The compounds left after the selection of the training set are finally divided into test set and independent set at random.
The compounds of training set, the test set and the independent set are listed in and labeled as Tr, Ts The second method is the 5-fold cross-validation method. The data set of 94 compounds are divided into five subsets of nearly equal size. One subset is used as thetest set and all samples in the other four subsets are used as See for the definition of the compound.
the training set. This procedure is repeated until every subsetis once used as the test set.
(k-NN) and C4.5 decision tree to model the antifungal Clustering technique is also used for the subset design in activity of these compounds for comparison.
the 5-fold cross-validation. However, here the clustering is In k-NN, the Euclidean distance between the done manually rather than by computer, based on the unclassified vector x and each individual vector xi in the structure similarity. The antifungal compounds are divided training set is measured, and k nearest vectors to the into five clusters (The first cluster contains unclassified vector x are used to determine the class of imidazoles and triazoles, the second cluster is composed of unclassified vector x. The class of the majority of the k- derivatives of benzothiazoles and benzimidazoles, allyla- nearest neighbors is chosen as the predicted class of the mines and thiocarbamates are classified into the third unclassified vector x. The important parameter, the number cluster, the fourth cluster contains the derivatives of of nearest neighbors, can affect the outcome. We found that naphthyl and quinoline, and finally the compounds left the nearest neighbor (kZ1) can give the best prediction in enter the fifth cluster. In the same way, the inactive compounds are divided into eight clusters based on their C4.5 decision tree is formalism expressing mappings structural similarity, as listed in . The compounds in from feature values to classes (predictions), and consists of every cluster are averagely divided into five subsets at attribute nodes that link to subtrees and leaves labeled with a random, so the compounds in every subset can span the class C4.5 decision tree uses recursive partitioning chemical space well. In this way, each training set in the to examine every attribute of the data and rank them 5-fold cross-validation can sample the space evenly.
according to their ability to partition the remaining data, andthus a decision tree is constructed. Instances are sorted downthe tree from the root node to some leaves. Each node in the 2.7. Comparison with other statistical classification tree specifies a test of some attribute of the instance, and each branch descending from that node corresponds to oneof the possible values from this attribute.
In order to evaluate the SVM method for the antifungal activity prediction, we also used the k-nearest neighbor The overall prediction accuracy (Qtotal) is 89.4% after 242 GA generations using LOO cross-validation method for an optimal Gaussian exponent and 30 molecular descriptors are selected and marked with asterisks in . These descriptors mainly consist of topological and shape descriptors. To evaluate the effect of variable selection on the classification accuracy of the SVM model, a 5-fold cross-validation is conducted using the selected 30 descriptors (termed as SVMCGA model) and then all the See for the definition of the compound.
67 descriptors (termed as SVM model). The results are S.-W. Chen et al. / Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 Table 5SVM and SVMCGA prediction accuracy by using 5-fold cross-validation given in The average accuracies for antifungals and selected using GA gives an overall prediction accuracy of inactive compounds, and the overall accuracy are, respect- 89.4% and the prediction accuracy of 78.6% for the active ively, 91.0, 81.8 and 84.0% for SVM model and 97.1, 85.2 and 98% for the inactive by LOO. Results show that for both and 89.4% for SVMCGA model, showing an obvious of the validation methods, nine antifungal compounds, improvement with GA variable selection. Our investi- Chlordantoin, Ciclopirox, Cloxyquine, Fluconazole, Flucy- gations suggest that GA is useful for removing redundant tosine, Hexetidine, Itraconazole, Nifuratel, Saperconazole, descriptors, and helpful for the computational efficiency of are misclassified as inactive compounds, and one inactive compound, Tamoxifen, is misclassified as antifungal(The misclassified compounds by LOO cross- These ten compounds are always misclassified by the In this work, the test with SVM is performed using the SVM system using 30 descriptors selected. Therefore, our methods described in Section 2.6, i.e. the independent set analysis suggests that the incorrect classification of them and 5-fold cross-validation methods. As shown in arises from an inadequate description of the detailed , for independent set and 5-fold cross-validation configuration. In the same way, the six compounds of methods, the prediction accuracies of the antifungal activity incorrect classification using the independent validation are are found 100 and 97.1% and those of inactive 77.8 and included in the ten compounds except Clonidine.
85.0%, respectively. Both methods give consistent resultsand demonstrate the stability of the SVM system.
3.3. Effect of cluster analysis on classification accuracy Only a tentative comparison can be made to provide some crude estimate regarding the approximate level of In 5-fold cross-validation, each subset built through accuracy of our method with respect to those obtained by clustering and reflects its distribution approximately in the other studies because of differences in the use of descriptors same manner. With SVMCGA, the prediction accuracies of and classification methods. Garcı´a–Domenech et al.
antifungal activity are 100.0, 100.0, 100.0, 85.7 and 100.0%, modeled the antifungal activity for these compounds using whereas the accuracies of inactive groups are 91.7, 83.3, LDA with topological descriptors by LOO and got the 83.3, 83.3 and 83.3% for the five subsets. It is clear that prediction accuracy of over 60% for the active and 98% for prediction for antifungal activity is better than for inactive the inactive . Our SVM modeling with the descriptors groups. Also, it can be seen from that SVM without Table 6Performance comparison of different classification methods by using independent evaluation set See the text and for the definitions of symbols.
S.-W. Chen et al. / Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 Clustering technique is employed to design the subsets for Performance comparison of different classification methods by 5-fold 5-fold cross-validation and for independent set validation, and our calculations show that the clustering technique is very helpful for improving the efficiency of SVM model building. Comparison of SVM method with k-NN and C4.5 decision tree shows that SVM method is a potential computation method for screening antifungal drugcandidates.
feature selection, i.e. the SVM model, shows the same trend.
Therefore, it is suggested that the subsets designed throughclustering can span effectively and evenly the sample space.
The misclassified compounds through LOO cross- validation and 5-fold cross-validation by SVMCGAs This work is supported by the National Natural Science model are the same, although LOO takes more time than 5-fold cross-validation. This implies that cluster analysiscan reduce the time for SVM model building withoutreducing the accuracy.
3.4. Comparison to k-NN and C4.5 decision tree methods [1] I. Al-Mohsen, W.T. Hughes, Ann. Saudi. Med. 18 (1998) 28.
[2] H.G. Nafsika, Curr. Opin. Microbiol. 1 (1998) 547.
In we compare the results of SVM with those of [3] A.H. Groll, T.f. Walsb, Swiss Med. Wkly. 132 (2002) 303.
k-NN method and C4.5 decision tree methods by 5-fold [4] R. Garcı´a-Domenech, I. Rios-Santamarina, A. Catala´, C. Calabuig, cross-validation. The same compound sets and descriptors L. del Castillo, J. Ga´lvez, J. Mol. Struct. (Theochem) 624 (2003) 97.
[5] M.C. Mar, G.R. Fe´lix, Lancet Infect. Dis. 2 (2002) 550.
are used. The overall prediction accuracy is 89.4, 76.5, [6] R. Garcı´a-Domenech, A. Catala´-Gregori, C. Calabuig, G. Anto´n-Fos, 75.6% for SVM, k-NN and decision tree, respectively, L. del Castillo, J. Ga´lvez, Internet Electron J. Mol. Des. 1 (2002) 339– obviously showing the prior classification ability of SVM.
For the present data set of 42 antifungals and 52 inactive [7] M.D. De-Bacher, P.T. Magee, J. Pla, Annu. Rev. Microbiol. 54 compounds, SVM, k-NN and C4.5 decision tree lead to false [8] J. Wo¨lcke, D. Ullmann, Drug Discov. Today 6 (2001) 637.
negative of 2, 27, and 21%, and false positive of 21, 19, and [9] H. Gao, M.S. Lajiness, J.V. Drie, J. Mol. Graph. Model. 20 (2002) 29%, respectively. This shows the considerably lower probability of false negative than false positive with the [10] C.A. Lipinski, F. Lombardo, B.W. Dominy, P.J. Feeney, Adv. Drug.
SVM model. Furthermore, from the perspective of a computerized screening of molecular structures for sub- [11] A.A.C. Pinheiro, R.S. Borges, L.S. Santos, C.N. Alves, J. Mol. Struct.
sequent synthesis and experimental testing, a greater risk to [12] V.M. Gokhale, V.M. Kulkarni, J. Med. Chem. 42 (1999) 5348.
overlook an active structure is probably preferred to a [13] V.M. Gokhale, V.M. Kulkarni, Bioorgan. Med. Chem. 8 (2000) 2487.
greater risk for false prediction of biological activities. In [14] K. Hasegawa, T. Deushi, O. Yaegashi, Y. Miyashita, S. Sasaki, Eur.
the prediction results of SVM compared with those of k-NN and C4.5 decision tree methods by using [15] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer, independent evaluation set. The prediction accuracy for [16] C.J.C. Burges, Data Min. Knowl. Disc. 2 (1998) 127.
antifungal activity are 100.0, 75, 80.0%, respectively for [17] M.W.B. Trotter, B.F. Buxton, S.B Holden, Meas. Control 34 SVM, k-NN and decision tree. Therefore, we can conclude that SVM method is superior to the other methods when [18] Y. Xue, C.W. Yap, L.Z. Sun, Z.W. Cao, J.F. Wang, Y.Z. Chen, considering the risk for prediction.
J. Chem. Inf. Comput. Sci. 44 (2004) 1497.
[19] R. Czermin´ski, A. Yasri, D. Hartsough, Quant. Struct.-Act. Relat. 20 [20] Y.D. Cai, X.J. Liu, X.B. Xu, K.C. Chou, J. Comput. Chem. 23 [21] C.H.Q. Ding, I. Dubchak, Bioinformatics 17 (2001) 349.
In this work, SVM method is employed for modeling the [22] J.M. Sutter, S.L. Dixon, P.C. Jurs, J. Chem. Inf. Comput. Sci. 35 discrimination between the antifungal activity and inactiv- [23] J.M. Sutter, J.H. Kalivas, Microchem. J. 47 (1993) 60.
ity, and GA is applied for variable selection in order to [24] Y. Xue, Z.R. Li, C.W. Yap, L.Z. Sun, X. Chen, Y.Z. Chen, J. Chem.
reduce the noise generated by the use of overlapping and redundant molecular descriptors. Our investigations show [25] I. Guyon, J. Weston, S. Barnhill, V. Vapnik, Mach. Learn. 46 that variable selection is helpful in enhancement of the [26] H. Yu, J. Yang, W. Wang, J. Han, Proc. IEEE Comput. Soc.
prediction ability for the prediction of antifungal activity of chemical agents. The prediction ability of SVM is tested by [27] T.J. Hou, J.M. Wang, N. Liao, X.J. Xu, J. Chem. Inf. Comput. Sci. 39 the independent validation and 5-fold cross-validation.
S.-W. Chen et al. / Journal of Molecular Structure: THEOCHEM 731 (2005) 73–81 [28] K. Hasegawa, T. Kimura, K. Funatsu, Quant. Struct. Act. Relat. 18 [41] J.H. Holland, Adaptation in Natural and Artificial Systems, University [29] J.K. Wegner, H. Fro¨hlich, A. Zell,, J. Chem. Inf. Comput. Sci. 44 [42] D.E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley, Reading, MA, 1989.
[30] J.K. Wegner, H. Fro¨hlich, A. Zell,, J. Chem. Inf. Comput. Sci. 44 [43] I.P. Androulakis, V.A. Venkatasubramanian, Comput. Chem. Eng. 15 [31] H. Fro¨hlich, J.K. Wegner, A. Zell, QSAR Comb. Sci. 23 (2004) 311.
[44] J.E. Roulston, Mol. Pharmacol. 20 (2002) 153.
[32] M.L. Mansfield, D.G. Covell, J. Chem. Inf. Comput. Sci. 42 [45] P. Baldi, S. Brunak, Y. Chauvin, C.A. Andersen, H. Nielsen, [33] J. Gasteiger, M. Marsili, Tetrahedron 36 (1980) 3219.
[46] C.J. Huberty, Applied Discriminant Analysis, Wiley, New York, [34] M. Mortier, K. Van Genechten, J. Gasteiger, J. Am. Chem. Soc. 107 [47] R.A. Johnson, D.W. Wichern, Applied Multivariate Statistical [35] C.J.C. Burges, Tutorial on Support Vector Machines for Pattern Recogni- Analysis, Prentice Hall, Englewood Cliffs, NJ, 1982.
[48] L. Breiman, J. Friedman, R. Olshen, P. Stone, Classification and [36] B. Scho¨lkop, A. short, A Short Tutorial on Kernels 2000 Regression Trees, Wadsworth, Belmont, CA, 1984.
[49] J.R. Quinlan, C4.5: Programs for Machine Learning, Morgan [37] B. Scho¨lkopf, Support Vector Learning 1999 ( [50] R. Todeschini, V. Consonni, Handbook of Molecular Descriptors, [38] M.W.B. Trotter, B.F. Buxton, S.B. Holden, Meas. Control 34 (2001) 235.
[39] R. Burbidge, M. Trotter, B. Buxton, S. Holden, Comput. Chem. 26 [51] H. Wiener, J. Am. Chem. Soc. 69 (1947) 17.
[52] L.H. Hall, L.B. Kier, Issues in representation of molecular structure, [40] R. Czerminski, A. Yasri, D. Hartsough, Quant. Struct.-Act. Relat. 20 [53] K.B. Lipkowitz, D.B. Boyd, Rev. Comput. Chem. 2 (1991) 401.

Source: http://ce.scu.edu.cn/xueyuanjiaoshizhuye/LiXiangyuan/script/PDF/Theochem-Chen.pdf