DOI: 10.1002/chem.201404081
Communication
& Chirality Recognition
Artificial Neural Networks for Guest Chirality Classification through Supramolecular Interactions Jarosław M. Granda and Janusz Jurczak*[a]
Chem. Eur. J. 2014, 20, 12368 – 12372
12368
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication Abstract: A novel strategy for classification of guest chirality based on the combination of artificial neural networks and anion-receptor chemistry is reported. The receptor reported herein forms supramolecular complexes with a variety of biologically important carboxylates, in which the chemical shift changes during addition of anions result in complex guest-stereochemistry-dependent patterns as followed by 1H NMR spectroscopy. The neural network had learnt these patterns from a training set of 12 anions, and successfully identified the “unknown” chirality of 14 guests present in the test set. Additionally, principal component analysis could discriminate most of the guests studied (26) and allowed for identification of the receptor protons, which are responsible for information transfer of guest chirality.
The determination of absolute configurations is one of the most important challenges in chemistry. Several procedures have been developed to determine absolute configuration, including X-ray structure analysis,[1] chiroptical methods,[2] NMR with chiral additives,[3] chromatographic separation with chiral stationary phases,[4] and so forth. Apart from the X-ray and chiroptical approaches, these techniques are based on specific molecular interactions between a chiral host and one of the guest enantiomers.[5] Additionally, these methods provide an entry point for designing efficient sensors for chiral analytes.[6] Although those methods are well established and widely used, each of them has its own limitations, and thus there is still a need to develop new protocols providing improved chiral recognition. In some instances, cooperative chiral recognition/ sensing is of great importance. This can be achieved by incorporating a fluorescent or chromogenic signaling unit into the guest chiral structure.[6a, b] Recently, we described the chiral recognition of a-functionalized carboxylates by neutral anion receptor 1 containing d-glucuronic acid (Figure 1), which forms supramolecular complexes with a variety of guests through hydrogen bonds.[7] In this communication we will Figure 1. Structure of chiral seek to demonstrate that the difanion receptor 1 investigated ferent patterns of chemical-shift in this study changes of receptor 1 during NMR titrations with enantiomeric anions can be successfully used in predicting guest chirality. Sample patterns of chemical-shift
[a] J. M. Granda, Prof. J. Jurczak Institute of Organic Chemistry Polish Academy of Sciences Kasprzaka 44/52, 01-224, Warsaw (Poland) E-mail: [email protected] Supporting information for this article is available on the WWW under http://dx.doi.org/10.1002/chem.201404081. Chem. Eur. J. 2014, 20, 12368 – 12372
www.chemeurj.org
changes for receptor 1 during titrations with enantiomeric mandelates are shown in Figure 2. After analyzing the above phenomena, we anticipated that the anion-binding-induced chemical-shift changes for acetyl groups and for protons belonging to the pyranose ring are in some way related to the size and shape of the anionic guests. Therefore, we decided to analyze these patterns by using artificial neural networks (AANs), which are the most complex and sophisticated computational models used for pattern analysis in general.[8] Although AANs were introduced to chemistry a number of years ago[9] and are now widely used in predicting various physical[10] and chemical properties,[11] as yet there are only few examples of their application to chirality determination,[12] and to the best of our knowledge ANNs have not been used in chiral recognition utilizing anion-receptor chemistry. In this regard, we will here seek to demonstrate that ANNs can successfully be used in a classification task, namely, determining the configuration of biologically important a-carboxylate anions on the basis of receptor 1 chemical-shift changes (Dd)
Figure 2. Patterns of chemical shift changes for acetyl and methyl groups in anion receptor 1 during titration with S(+)-mandelate (A) and R()-mandelate (B).
12369
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication in formed supramolecular complexes. Our approach is similar to indicator displacement arrays (IDA),[13] with the difference that various fragments of receptor 1 transfer information about structural features of the anionic guests, whereas IDA uses indicators to accomplish it. ANNs came into being as models of information processing of the nervous system.[9a, b] The typical tasks that ANNs can perform are: classification—assignment of the data to two or more categories; modeling—prediction of new properties by using existing ones; association—finding relations within data; mapping—finding simpler representation of the data. The basic building block of the ANNs is an artificial neuron, which mimics the behavior of the biological one. The input data (x), which is fed to the neuron is altered by proper weights (w), corresponding to synaptic strengths in biological systems, is summed [Eq. (1)] and then subjected to an acting function that gives an output out of the neuron [Eq. (2)]. inp ¼
X
xi wi
ð1Þ
1 1 þ eðinpÞ
ð2Þ
i
out ¼
The acting function most commonly used is the sigmoid one, known also as the logistic function [Eq. (2)]. Its use allows introducing nonlinearity to the network and its output value is always non-negative. The artificial neurons are stacked into the structure of a network. The most common architecture is a feed-forward network structure, in which a layer of the neurons are fully connected to the previous and next layer, but there is no connection within a given layer. The outputs from neurons in a given layer are the inputs for the next layer. The network also contains an input layer, which does not perform any mathematical operation on the data, it only propagates the data to the first layer of neurons. The output layer plays a role of the output of the network. To perform the above-mentioned tasks the weights of the network (w) must have the proper values. The process in which w of the network is set is known as a learning. The most common procedure used for learning is the algorithm of errors back-propagation[14] in which, at the same time, data is given to the input layer of the ANN while the output layer is fed with the desired output. The algorithm sets w by minimizing errors between subsequent layers of neurons starting from the output layer. Figure 3 shows the organization of the feed-forward neural network used in this study. The multilayer structure of the ANN allowed it to be tailored to the chirality-recognition problem. The number of input units in the network was equal to the number of protons whose chemical-shift changes were tracked. The network has two output units, corresponding to R or S chirality of the anionic guest. The training data needed for chirality identification was obtained from 1H NMR spectra of a mixture of receptor 1 with four equivalents of guest, as well as of a free receptor, both in [D6]DMSO. On that basis, the chemical-shift changes of acetyl and methyl groups of receptor 1 were calculated and normalized before being supplied to Chem. Eur. J. 2014, 20, 12368 – 12372
www.chemeurj.org
Figure 3. Schematic representation of the structure of the ANN used in this study (10 input units, two hidden layers, and two output units).
the ANNs.[15] This procedure produces a set of supramolecular fingerprints of anions. The ANN learning process was accomplished by using an errors back-propagation algorithm.[9a] The input layers were supplied with the chemical-shift changes for the appropriate protons, while at the same time the output layers were supplied with the chirality of the anionic guest: [1,0] for S- and [0,1] for R-carboxylate. After a number of trials, we set the number of hidden layers to two, each containing 10 components.[16] The training set of chemical-shift changes for guests, which was provided to the ANNs, consisted of various anions, chosen to allow the network to learn complex patterns for different combinations of a-carboxylate substituents (Table 1). For example, R/S-mandelate has a hydrogen-bond donor (the OH group) and an aromatic substituent, R/S-2-hydroxybutyrates have a hydrogen-bond donor and aliphatic substituents, R/S-2-phenyl butyrate have aliphatic and aromatic substituents, N-tert-butoxycarbonyl (Boc)-protected l/d-valine has
12370
Table 1. Anions used (as tetrabutylammonium salts) in the training set and the ANN outputs. Anion
ANNs output Anion chirality (S/R) S
[1.0, 0.0]
ANNs output chirality (S/R) R [0.97, 0.03]
R [0.0, 1.0]
S
[0.0, 1.0]
S
S
[0.98, 0.02]
[0.98, 0.02]
R [0.03, 0.97]
R [0.0, 1.0]
S
S
[0.96, 0.04]
R [0.05, 0.95]
[1.0, 0.0]
R [0.01, 0.99]
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication an aliphatic substituent and an NH hydrogen-bond donor, whereas l/d-phenylglycine has a free NH2 group and an aromatic phenyl substituent. A further 14 “unknown” anions, not included in the training set data, were then supplied to the trained network, to evaluate its ability to classify the chirality of the guest (Table 2). The ANN successfully classified the chirality of the test set of anions shown in Table 2, indicating that the ANN had learned the mechanisms responsible for chirality transfer from anion to receptor 1, and can therefore be successfully applied for new test cases.
Table 2. Anions used ( as tetrabutylammonium salts) in the test set. Anion
ANNs predicted chirality (S/R)
Anion
ANNs predicted chirality (S/R)
S [1.0, 0.0]
S [1.0, 0.0]
R [0.06, 0.94]
R [0.0, 1.0]
S [1.0, 0.0]
S [0.99, 0.01]
R [0.06, 0.95]
S [0.98, 0.02]
S [1.0, 0.0]
R [0.01, 0.99]
S [1.0, 0.0]
S [1.0, 0.0]
R [0.0, 1.0]
R [0.0, 1.0]
Figure 4. PCA for receptor 1 chemical-shift changes upon addition of various anions. PC1 variance = 59.8 %; PC2 variance = 24.5 %. Chem. Eur. J. 2014, 20, 12368 – 12372
www.chemeurj.org
To gain further insight into the mechanism of chiral recognition, we decided to apply principal component analysis (PCA),[17] which is used to decompose multidimensional data into a set of successive orthogonal “components” with maximum variance. On that basis, it is possible to reduce such multidimensional data into a lower number of dimensions while retaining as much information as possible. A scikit-learn package[18] was used to generate a PCA plot in which two principal components contained 84.3 % of the discriminatory information. Figure 4 shows visual discrimination for most of the enantiomeric anions on a two-dimensional plot. PCA poorly discriminated the enantiomers of Mosher acid, 2-phenylbutyric acid, and Boc-N-valine. Although they were not correctly distinguished by PCA, the ANN managed to predict their chirality correctly. On the other hand, analysis of the contribution of chemicalshift changes of receptor 1 protons to the first two components (Figure 5) showed that chemical-shift changes for acetyl
Figure 5. Receptor 1 proton contributions to the first and the second component (PC1 and PC2).
12371
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication groups at the C4 position in the glucopyranose ring and that the methyl groups of diindolylmethane play a crucial role in the chirality-transfer process. The contributions of these two protons can be easily explained, namely, the acetyl groups at C4 are in the proximity of the receptor 1 anion-binding pocket and can directly sense the chirality of the anionic guest, whereas the chemical-shift changes of the methyl groups can be correlated to the chirality of the diindolylmethane moiety, which can presumably adopt a helical-like structure that is folded/unfolded after binding of a chiral guest. In conclusion, the combination of anion-receptor chemistry and pattern-recognition protocols (ANNs and PCA) yields a useful strategy for chirality recognition. The neural network is able to learn complex patterns of chemical-shift changes and then classify the chirality of carboxylates on that basis. Additionally, PCA analysis allows for identification of the protons that transfer information about chirality from the anionic guest to receptor 1. On the other hand, the results obtained show that sugar moieties play a crucial role in the recognition processes. Their proper preorganization and spatial complexity allowed for recognition of such a subtle difference in guest structure as chirality.
Acknowledgements
[4] [5] [6]
[7] [8] [9]
[10]
[11]
[12]
[13]
This research was financed by the European Union within the European Regional Development Fund, Project POIG.01.01.02.14-102/09.
[14] [15] [16]
Keywords: anions · chirality · receptors · structure–activity relationships · supramolecular chemistry [1] H. D. Flack, G. Bernardinelli, Chirality 2008, 20, 681 – 690. [2] P. L. P. Nina Berova, Koji Nakanishi, Robert W. Woody, Comprehensive Chiroptical Spectroscopy, Applications in Stereochemical Analysis of Synthetic Compounds, Natural Products, and Biomolecules, Vol. 2, Wiley, Hoboken, 2012. [3] a) T. J. Wenzel, J. D. Wilcox, Chirality 2003, 15, 256 – 270; b) A. Claesson, L. I. Olsson, G. R. Sullivan, H. S. Mosher, J. Am. Chem. Soc. 1975, 97, 2919 – 2921; c) M. Kainosho, S. D. Beare, K. Ajisaka, W. H. Pirkle, J. Am. Chem. Soc. 1972, 94, 5924 – 5926; d) M. D. McCreary, D. W. Lewis, D. L. Wernick, G. M. Whitesides, J. Am. Chem. Soc. 1974, 96, 1038 – 1054;
Chem. Eur. J. 2014, 20, 12368 – 12372
www.chemeurj.org
[17] [18]
e) N. A. Shaath, T. O. Soine, J. Org. Chem. 1975, 40, 1987 – 1988; f) G. R. Sullivan, D. Ciavarel, H. S. Mosher, J. Org. Chem. 1974, 39, 2411 – 2412; g) G. M. Whitesides, D. W. Lewis, J. Am. Chem. Soc. 1971, 93, 5914 – 5916. V. Meyer, Practical High-Performance Liquid Chromatography, 5edth edWiley, Chichester, 2010. W. H. Pirkle, T. C. Pochapsky, Chem. Rev. 1989, 89, 347 – 362. a) G. A. Hembury, V. V. Borovkov, Y. Inoue, Chem. Rev. 2008, 108, 1 – 73; b) L. Pu, Chem. Rev. 2004, 104, 1687 – 1716; c) D. Leung, S. O. Kang, E. V. Anslyn, Chem. Soc. Rev. 2012, 41, 448 – 479. J. M. Granda, J. Jurczak, Org. Lett. 2013, 15, 4730 – 4733. a) H. M. Cartwright, Methods Mol. Biol. 2008, 458,1 – 13; b) S. Curteanu, H. Cartwright, J. Chemometrics 2011, 25, 527 – 549. a) J. A. Burns, G. M. Whitesides, Chem. Rev. 1993, 93, 2583 – 2601; b) J. Gasteiger, J. Zupan, Angew. Chem. 1993, 105, 510 – 536; Angew. Chem. Int. Ed. Engl. 1993, 32, 503 – 527; c) P. A. Jansson, Anal. Chem. 1991, 63, A357 – A362. a) G. Astray, J. F. Galvez, J. C. Mejuto, O. A. Moldes, I. Montoya, J. Comput. Chem. 2013, 34, 355 – 359; b) R. M. Balabin, E. I. Lomakina, J. Chem. Phys. 2009, 131, 074104. a) M. Alonso, B. Herradon, Chem. Eur. J. 2007, 13, 3913 – 3923; b) S. C. McCleskey, P. N. Floriano, S. L. Wiskur, E. V. Anslyn, J. T. McDevitt, Tetrahedron 2003, 59, 10089 – 10092; c) D. W. Elrod, G. M. Maggiora, R. G. Trenary, J. Chem. Inf. Comput. Sci. 1990, 30, 477 – 484; d) M. Alonso, C. Miranda, N. Martin, B. Herradon, Phys. Chem. Chem. Phys. 2011, 13, 20564 – 20574; e) V. Simon, J. Gasteiger, J. Zupan, J. Am. Chem. Soc. 1993, 115, 9148 – 9159; f) V. Hall, A. Nash, E. Hines, A. Rodger, J. Comput. Chem. 2013, 34, 2774 – 2786. a) J. Aires-de-Sousa, J. Gasteiger, J. Comb. Chem. 2005, 7, 298 – 301; b) S. H. Shabbir, L. A. Joyce, G. M. da Cruz, V. M. Lynch, S. Sorey, E. V. Anslyn, J. Am. Chem. Soc. 2009, 131, 13125 – 13131. J. W. Lee, J.-S. Lee, Y.-T. Chang, Angew. Chem. 2006, 118, 6635 – 6637; Angew. Chem. Int. Ed. 2006, 45, 6485 – 6487. D. E. Rumelhart, G. E. Hinton, R. J. Williams, Nature 1986, 323, 533 – 536. The patterns of chemical-shift changes for anions used in this study can be found in the Supporting Information. There is no algorithm to determine the optimal ANNs structure. The most common approach is based on trial and error procedures. However, it is known that by adding a hidden layer of neurons the network is able to learn more advanced representations of training data, see ref. [14]. On the other hand, too complicated network structures may lead to poor generalization in the test set (so-called “overfitting”). S. Stewart, M. A. Ivy, E. V. Anslyn, Chem. Soc. Rev. 2014, 43, 70 – 84. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, JMLR 2011, 12, 2825 – 2830.
Received: June 23, 2014 Published online on September 1, 2014
12372
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication
& Chirality Recognition
Artificial Neural Networks for Guest Chirality Classification through Supramolecular Interactions Jarosław M. Granda and Janusz Jurczak*[a]
Chem. Eur. J. 2014, 20, 12368 – 12372
12368
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication Abstract: A novel strategy for classification of guest chirality based on the combination of artificial neural networks and anion-receptor chemistry is reported. The receptor reported herein forms supramolecular complexes with a variety of biologically important carboxylates, in which the chemical shift changes during addition of anions result in complex guest-stereochemistry-dependent patterns as followed by 1H NMR spectroscopy. The neural network had learnt these patterns from a training set of 12 anions, and successfully identified the “unknown” chirality of 14 guests present in the test set. Additionally, principal component analysis could discriminate most of the guests studied (26) and allowed for identification of the receptor protons, which are responsible for information transfer of guest chirality.
The determination of absolute configurations is one of the most important challenges in chemistry. Several procedures have been developed to determine absolute configuration, including X-ray structure analysis,[1] chiroptical methods,[2] NMR with chiral additives,[3] chromatographic separation with chiral stationary phases,[4] and so forth. Apart from the X-ray and chiroptical approaches, these techniques are based on specific molecular interactions between a chiral host and one of the guest enantiomers.[5] Additionally, these methods provide an entry point for designing efficient sensors for chiral analytes.[6] Although those methods are well established and widely used, each of them has its own limitations, and thus there is still a need to develop new protocols providing improved chiral recognition. In some instances, cooperative chiral recognition/ sensing is of great importance. This can be achieved by incorporating a fluorescent or chromogenic signaling unit into the guest chiral structure.[6a, b] Recently, we described the chiral recognition of a-functionalized carboxylates by neutral anion receptor 1 containing d-glucuronic acid (Figure 1), which forms supramolecular complexes with a variety of guests through hydrogen bonds.[7] In this communication we will Figure 1. Structure of chiral seek to demonstrate that the difanion receptor 1 investigated ferent patterns of chemical-shift in this study changes of receptor 1 during NMR titrations with enantiomeric anions can be successfully used in predicting guest chirality. Sample patterns of chemical-shift
[a] J. M. Granda, Prof. J. Jurczak Institute of Organic Chemistry Polish Academy of Sciences Kasprzaka 44/52, 01-224, Warsaw (Poland) E-mail: [email protected] Supporting information for this article is available on the WWW under http://dx.doi.org/10.1002/chem.201404081. Chem. Eur. J. 2014, 20, 12368 – 12372
www.chemeurj.org
changes for receptor 1 during titrations with enantiomeric mandelates are shown in Figure 2. After analyzing the above phenomena, we anticipated that the anion-binding-induced chemical-shift changes for acetyl groups and for protons belonging to the pyranose ring are in some way related to the size and shape of the anionic guests. Therefore, we decided to analyze these patterns by using artificial neural networks (AANs), which are the most complex and sophisticated computational models used for pattern analysis in general.[8] Although AANs were introduced to chemistry a number of years ago[9] and are now widely used in predicting various physical[10] and chemical properties,[11] as yet there are only few examples of their application to chirality determination,[12] and to the best of our knowledge ANNs have not been used in chiral recognition utilizing anion-receptor chemistry. In this regard, we will here seek to demonstrate that ANNs can successfully be used in a classification task, namely, determining the configuration of biologically important a-carboxylate anions on the basis of receptor 1 chemical-shift changes (Dd)
Figure 2. Patterns of chemical shift changes for acetyl and methyl groups in anion receptor 1 during titration with S(+)-mandelate (A) and R()-mandelate (B).
12369
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication in formed supramolecular complexes. Our approach is similar to indicator displacement arrays (IDA),[13] with the difference that various fragments of receptor 1 transfer information about structural features of the anionic guests, whereas IDA uses indicators to accomplish it. ANNs came into being as models of information processing of the nervous system.[9a, b] The typical tasks that ANNs can perform are: classification—assignment of the data to two or more categories; modeling—prediction of new properties by using existing ones; association—finding relations within data; mapping—finding simpler representation of the data. The basic building block of the ANNs is an artificial neuron, which mimics the behavior of the biological one. The input data (x), which is fed to the neuron is altered by proper weights (w), corresponding to synaptic strengths in biological systems, is summed [Eq. (1)] and then subjected to an acting function that gives an output out of the neuron [Eq. (2)]. inp ¼
X
xi wi
ð1Þ
1 1 þ eðinpÞ
ð2Þ
i
out ¼
The acting function most commonly used is the sigmoid one, known also as the logistic function [Eq. (2)]. Its use allows introducing nonlinearity to the network and its output value is always non-negative. The artificial neurons are stacked into the structure of a network. The most common architecture is a feed-forward network structure, in which a layer of the neurons are fully connected to the previous and next layer, but there is no connection within a given layer. The outputs from neurons in a given layer are the inputs for the next layer. The network also contains an input layer, which does not perform any mathematical operation on the data, it only propagates the data to the first layer of neurons. The output layer plays a role of the output of the network. To perform the above-mentioned tasks the weights of the network (w) must have the proper values. The process in which w of the network is set is known as a learning. The most common procedure used for learning is the algorithm of errors back-propagation[14] in which, at the same time, data is given to the input layer of the ANN while the output layer is fed with the desired output. The algorithm sets w by minimizing errors between subsequent layers of neurons starting from the output layer. Figure 3 shows the organization of the feed-forward neural network used in this study. The multilayer structure of the ANN allowed it to be tailored to the chirality-recognition problem. The number of input units in the network was equal to the number of protons whose chemical-shift changes were tracked. The network has two output units, corresponding to R or S chirality of the anionic guest. The training data needed for chirality identification was obtained from 1H NMR spectra of a mixture of receptor 1 with four equivalents of guest, as well as of a free receptor, both in [D6]DMSO. On that basis, the chemical-shift changes of acetyl and methyl groups of receptor 1 were calculated and normalized before being supplied to Chem. Eur. J. 2014, 20, 12368 – 12372
www.chemeurj.org
Figure 3. Schematic representation of the structure of the ANN used in this study (10 input units, two hidden layers, and two output units).
the ANNs.[15] This procedure produces a set of supramolecular fingerprints of anions. The ANN learning process was accomplished by using an errors back-propagation algorithm.[9a] The input layers were supplied with the chemical-shift changes for the appropriate protons, while at the same time the output layers were supplied with the chirality of the anionic guest: [1,0] for S- and [0,1] for R-carboxylate. After a number of trials, we set the number of hidden layers to two, each containing 10 components.[16] The training set of chemical-shift changes for guests, which was provided to the ANNs, consisted of various anions, chosen to allow the network to learn complex patterns for different combinations of a-carboxylate substituents (Table 1). For example, R/S-mandelate has a hydrogen-bond donor (the OH group) and an aromatic substituent, R/S-2-hydroxybutyrates have a hydrogen-bond donor and aliphatic substituents, R/S-2-phenyl butyrate have aliphatic and aromatic substituents, N-tert-butoxycarbonyl (Boc)-protected l/d-valine has
12370
Table 1. Anions used (as tetrabutylammonium salts) in the training set and the ANN outputs. Anion
ANNs output Anion chirality (S/R) S
[1.0, 0.0]
ANNs output chirality (S/R) R [0.97, 0.03]
R [0.0, 1.0]
S
[0.0, 1.0]
S
S
[0.98, 0.02]
[0.98, 0.02]
R [0.03, 0.97]
R [0.0, 1.0]
S
S
[0.96, 0.04]
R [0.05, 0.95]
[1.0, 0.0]
R [0.01, 0.99]
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication an aliphatic substituent and an NH hydrogen-bond donor, whereas l/d-phenylglycine has a free NH2 group and an aromatic phenyl substituent. A further 14 “unknown” anions, not included in the training set data, were then supplied to the trained network, to evaluate its ability to classify the chirality of the guest (Table 2). The ANN successfully classified the chirality of the test set of anions shown in Table 2, indicating that the ANN had learned the mechanisms responsible for chirality transfer from anion to receptor 1, and can therefore be successfully applied for new test cases.
Table 2. Anions used ( as tetrabutylammonium salts) in the test set. Anion
ANNs predicted chirality (S/R)
Anion
ANNs predicted chirality (S/R)
S [1.0, 0.0]
S [1.0, 0.0]
R [0.06, 0.94]
R [0.0, 1.0]
S [1.0, 0.0]
S [0.99, 0.01]
R [0.06, 0.95]
S [0.98, 0.02]
S [1.0, 0.0]
R [0.01, 0.99]
S [1.0, 0.0]
S [1.0, 0.0]
R [0.0, 1.0]
R [0.0, 1.0]
Figure 4. PCA for receptor 1 chemical-shift changes upon addition of various anions. PC1 variance = 59.8 %; PC2 variance = 24.5 %. Chem. Eur. J. 2014, 20, 12368 – 12372
www.chemeurj.org
To gain further insight into the mechanism of chiral recognition, we decided to apply principal component analysis (PCA),[17] which is used to decompose multidimensional data into a set of successive orthogonal “components” with maximum variance. On that basis, it is possible to reduce such multidimensional data into a lower number of dimensions while retaining as much information as possible. A scikit-learn package[18] was used to generate a PCA plot in which two principal components contained 84.3 % of the discriminatory information. Figure 4 shows visual discrimination for most of the enantiomeric anions on a two-dimensional plot. PCA poorly discriminated the enantiomers of Mosher acid, 2-phenylbutyric acid, and Boc-N-valine. Although they were not correctly distinguished by PCA, the ANN managed to predict their chirality correctly. On the other hand, analysis of the contribution of chemicalshift changes of receptor 1 protons to the first two components (Figure 5) showed that chemical-shift changes for acetyl
Figure 5. Receptor 1 proton contributions to the first and the second component (PC1 and PC2).
12371
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Communication groups at the C4 position in the glucopyranose ring and that the methyl groups of diindolylmethane play a crucial role in the chirality-transfer process. The contributions of these two protons can be easily explained, namely, the acetyl groups at C4 are in the proximity of the receptor 1 anion-binding pocket and can directly sense the chirality of the anionic guest, whereas the chemical-shift changes of the methyl groups can be correlated to the chirality of the diindolylmethane moiety, which can presumably adopt a helical-like structure that is folded/unfolded after binding of a chiral guest. In conclusion, the combination of anion-receptor chemistry and pattern-recognition protocols (ANNs and PCA) yields a useful strategy for chirality recognition. The neural network is able to learn complex patterns of chemical-shift changes and then classify the chirality of carboxylates on that basis. Additionally, PCA analysis allows for identification of the protons that transfer information about chirality from the anionic guest to receptor 1. On the other hand, the results obtained show that sugar moieties play a crucial role in the recognition processes. Their proper preorganization and spatial complexity allowed for recognition of such a subtle difference in guest structure as chirality.
Acknowledgements
[4] [5] [6]
[7] [8] [9]
[10]
[11]
[12]
[13]
This research was financed by the European Union within the European Regional Development Fund, Project POIG.01.01.02.14-102/09.
[14] [15] [16]
Keywords: anions · chirality · receptors · structure–activity relationships · supramolecular chemistry [1] H. D. Flack, G. Bernardinelli, Chirality 2008, 20, 681 – 690. [2] P. L. P. Nina Berova, Koji Nakanishi, Robert W. Woody, Comprehensive Chiroptical Spectroscopy, Applications in Stereochemical Analysis of Synthetic Compounds, Natural Products, and Biomolecules, Vol. 2, Wiley, Hoboken, 2012. [3] a) T. J. Wenzel, J. D. Wilcox, Chirality 2003, 15, 256 – 270; b) A. Claesson, L. I. Olsson, G. R. Sullivan, H. S. Mosher, J. Am. Chem. Soc. 1975, 97, 2919 – 2921; c) M. Kainosho, S. D. Beare, K. Ajisaka, W. H. Pirkle, J. Am. Chem. Soc. 1972, 94, 5924 – 5926; d) M. D. McCreary, D. W. Lewis, D. L. Wernick, G. M. Whitesides, J. Am. Chem. Soc. 1974, 96, 1038 – 1054;
Chem. Eur. J. 2014, 20, 12368 – 12372
www.chemeurj.org
[17] [18]
e) N. A. Shaath, T. O. Soine, J. Org. Chem. 1975, 40, 1987 – 1988; f) G. R. Sullivan, D. Ciavarel, H. S. Mosher, J. Org. Chem. 1974, 39, 2411 – 2412; g) G. M. Whitesides, D. W. Lewis, J. Am. Chem. Soc. 1971, 93, 5914 – 5916. V. Meyer, Practical High-Performance Liquid Chromatography, 5edth edWiley, Chichester, 2010. W. H. Pirkle, T. C. Pochapsky, Chem. Rev. 1989, 89, 347 – 362. a) G. A. Hembury, V. V. Borovkov, Y. Inoue, Chem. Rev. 2008, 108, 1 – 73; b) L. Pu, Chem. Rev. 2004, 104, 1687 – 1716; c) D. Leung, S. O. Kang, E. V. Anslyn, Chem. Soc. Rev. 2012, 41, 448 – 479. J. M. Granda, J. Jurczak, Org. Lett. 2013, 15, 4730 – 4733. a) H. M. Cartwright, Methods Mol. Biol. 2008, 458,1 – 13; b) S. Curteanu, H. Cartwright, J. Chemometrics 2011, 25, 527 – 549. a) J. A. Burns, G. M. Whitesides, Chem. Rev. 1993, 93, 2583 – 2601; b) J. Gasteiger, J. Zupan, Angew. Chem. 1993, 105, 510 – 536; Angew. Chem. Int. Ed. Engl. 1993, 32, 503 – 527; c) P. A. Jansson, Anal. Chem. 1991, 63, A357 – A362. a) G. Astray, J. F. Galvez, J. C. Mejuto, O. A. Moldes, I. Montoya, J. Comput. Chem. 2013, 34, 355 – 359; b) R. M. Balabin, E. I. Lomakina, J. Chem. Phys. 2009, 131, 074104. a) M. Alonso, B. Herradon, Chem. Eur. J. 2007, 13, 3913 – 3923; b) S. C. McCleskey, P. N. Floriano, S. L. Wiskur, E. V. Anslyn, J. T. McDevitt, Tetrahedron 2003, 59, 10089 – 10092; c) D. W. Elrod, G. M. Maggiora, R. G. Trenary, J. Chem. Inf. Comput. Sci. 1990, 30, 477 – 484; d) M. Alonso, C. Miranda, N. Martin, B. Herradon, Phys. Chem. Chem. Phys. 2011, 13, 20564 – 20574; e) V. Simon, J. Gasteiger, J. Zupan, J. Am. Chem. Soc. 1993, 115, 9148 – 9159; f) V. Hall, A. Nash, E. Hines, A. Rodger, J. Comput. Chem. 2013, 34, 2774 – 2786. a) J. Aires-de-Sousa, J. Gasteiger, J. Comb. Chem. 2005, 7, 298 – 301; b) S. H. Shabbir, L. A. Joyce, G. M. da Cruz, V. M. Lynch, S. Sorey, E. V. Anslyn, J. Am. Chem. Soc. 2009, 131, 13125 – 13131. J. W. Lee, J.-S. Lee, Y.-T. Chang, Angew. Chem. 2006, 118, 6635 – 6637; Angew. Chem. Int. Ed. 2006, 45, 6485 – 6487. D. E. Rumelhart, G. E. Hinton, R. J. Williams, Nature 1986, 323, 533 – 536. The patterns of chemical-shift changes for anions used in this study can be found in the Supporting Information. There is no algorithm to determine the optimal ANNs structure. The most common approach is based on trial and error procedures. However, it is known that by adding a hidden layer of neurons the network is able to learn more advanced representations of training data, see ref. [14]. On the other hand, too complicated network structures may lead to poor generalization in the test set (so-called “overfitting”). S. Stewart, M. A. Ivy, E. V. Anslyn, Chem. Soc. Rev. 2014, 43, 70 – 84. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, JMLR 2011, 12, 2825 – 2830.
Received: June 23, 2014 Published online on September 1, 2014
12372
2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
