My Scientific Work

My main scientific activity is devoted to the development and application of informatics and machine learning methods (neural networks, SVM, etc) to solve different problems in theoretical chemistry, as well as to molecular modeling. The covered scientific areas include chemoinformatics, SAR/QSAR/QSPR studies, neuroinformatics, mathematical chemistry, medicinal chemistry, quantum chemistry and force-field molecular modeling in organic chemistry, biological and supramolecular chemistry. The most important of my scientific achievements are: (a) formulation of the “inverse” QSAR/QSPR problem; (b) neural network for “direct” structure-property correlations; (c) mathematical theory of fragmental descriptors, (d) universal approach for predicting properties of organic compounds by using neural networks in combination with fragmental descriptors; (e) molecular models of glutamate receptors and their interactions with drug molecules, (f) one-class classification as a universal approach to conducting virtual screening, (g) the method of continuous molecular fields; (h) defining chemoinformatics as theoretical chemistry discipline.

Developments in theoretical conformational analysis

As a part of my Diploma work (Master Thesis), I developed a method for describing conformational spaces of 8-, 9- and 10-membered rings in the framework of the Cremer-Pople approach and constructed the corresponding maps.¹ Using original approach, I also derived the complete set of all canonical types of conformations for cyclic molecules.²

Development of methods for mathematical description, systematic enumeration and computer generation of organic reactions and their mechanisms

As a part of my PhD studies, I developed methodology, algorithms and computer program SYMBEQ for systematic generation and search for novel types of organic reaction on the basis of Zefirov-Tratch formal-logical approach.³ Using this program, I systematically explored all organic reactions with open⁴ and linear-cyclic⁵ topologies of bond redistribution. Novel reactions designed using the SYMBEQ program were verified experimentally, and a new reaction (reaction of dimethyl selenoxide with the ester of acetylenedicarboxylic acid leading to formation of furan tetracarboxylic acid) was found.³ I also extended the formal-logical approach to description and computer generation of reaction mechanisms, especially catalytic mechanisms.⁶ In the course of this work, I also developed an algorithm and computer software for structural-fragment searches,⁷ along with the universal computer graphics program for organic chemistry MOLED,⁸ which were used by me further for developing the STAR software for building correlations structure-property.⁹

Molecular modeling of self-assembling photo-switchable supramolecular devices

The first part of my postdoc work in the Photochemistry Center of Russian Academy of Sciences dealt with molecular modeling and quantum chemistry studies of crown-containing styryl dyes. I simulated conformations and electronic absorption spectra of crown-containing styryl dyes and their complexes with metal cations.^10-12 Then I conducted a molecular mechanics study of regio- and stereo-selectivity of cation-dependent photochemical [2+2]-autocycloaddition of crown-containing styryl dyes.¹³^,¹⁴ As a further development of these studies, I studies, by combining force-field molecular modeling with quantum chemistry calculations, regio- and stereo-selectivity of the cation-dependent photochemical [2+2]-autocycloaddition leading to formation of several types of self-assembling photo-switchable molecular devices as well as their electronic absorption spectra: multiphotochromic 15-crown-5 ethers with rigid spacers and their anion-capped complexes,¹⁵^,¹⁶ photoswitchable molecular pincers,¹⁷ photoswitchable receptors based on dimeric complexes of styryl dyes containing 15-crown-5 ether unit.¹⁸ Another of my works related to supramolecular chemistry dealt with molecular dynamics study of cyclodextrine complexation.¹⁹

Molecular modeling of different proteins and protein-ligand complexes

The experience in modeling supramolecular complexes allowed proceeding to the modeling of proteins and their complexes with organic ligands. Using combination of comparative protein modeling with molecular mechanics and dynamics, I started with modeling of a neurotrophic factor fragment from pigment epithelium.²⁰^,²¹ After that I conducted a big cycle of studies aimed at building spatial models of human glutamate receptors, understanding structure-activity relationships from the structural point of view, performing virtual screening and drug discovery.²² The following macromolecules were explored: aminoterminal domain of glutamate metabotropic receptor mGluR1,^23-25 ion channel inside the NMDA receptor,²⁶^,²⁷ glutamate-binding sites of the metabotropic glutamate receptors mGluR1-mGluR8,²⁸ ligand-binding domain of the kainite receptor.²⁹ glutamate-binding,³⁰ glycine-binding³¹^,³² and N-terminal³³ domains of the NMDA receptor, ligand-binding site of the GluR2 subunit of the AMPA-receptor.³⁴ Results of these studies were used in virtual screening³⁵ and design of neuroprotective and cognition-enhancing drugs.²⁶ Predictions concerning the structure of the aminoterminal domain of the mGluR1 receptor and the binding modes of its ligands were further validated in experimental studies. Based on these studies, I have also formulated the selectivity fields concept,³² which is currently rather popular in medicinal chemistry. I also took part in molecular modeling of adenosine,^36-39 melatonin,^40-42 and WNT-protein binding FZD^43-45 receptors and their interactions with ligands. Another important object of my molecular modeling studies was angiotensin-converting enzyme.^46-49 My recent studies dealt with modeling interactions between taxol and colchicine analogs with the goal of discovering anti-cancer drugs.⁵⁰^,⁵¹

Development of algorithms and programs for handling chemical information in databases

In parallel with modeling photochromic dyes, I also developed software for handling spectral databases of dyes based on my original approaches and algorithms for handling chemical structures, properties and their light absorption spectra.⁵²^,⁵³ Besides, I created Java-based three-tier client-server system for handling databases of chemical structures, along with Java library and Tcl/Tk platform for handling chemical information based on original approaches.

Inverse problem in QSPR studies

In the joint work with Dr. M. Skvortsova, we formulated for the first time the “inverse problem” – the task of generation of chemical structures for given values of their properties, and solved it for several cases.^54-58 All solutions were based on correlations of properties with topological indexes, and combination of graph generation with application of graph theory to minimize the space of solutions was used to generate chemical structures. Nowadays this is very hot area in chemoinformatics, and the article⁵⁶ is my most cited paper.

Neural device for finding direct structure-property correlations

In 1993 I developed a special neural device for building direct correlations between structures and properties of chemical compounds without the need to pre-calculate molecular descriptors.⁵⁹^,⁶⁰ This neural network of special architecture works directly with molecular graphs and extracts from them latent non-linear features which are the most useful for predicting the property under study. Very good performance of this device was demonstrated for predicting 7 different physicochemical properties. This work can be considered innovate not only in chemoinformatics, but also in machine learning, because first publications on graph-based data mining and machine learning with structured data appeared later.

Theorems on the basis of invariants of molecular graphs and fragmental approach

To address the problem of designing or choosing the optimal sets of molecular descriptors for predicting properties of chemical compounds, I formulated and proved (in cooperation with Dr. M. Skvortsova) several theorems on the basis of invariants of labeled graphs.^61-63 According to these theorems, any molecular graph invariant (that is any molecular descriptor or scalar property) can be uniquely represented as (1) a linear combination of the occurrence numbers of some substructures (fragments), both connected and disconnected, or (2) a polynomial on occurrence numbers of connected substructures of corresponding molecular graph (which are the values of fragmental descriptors). These results were used for formulating a methodology of constructing general models for structure-property relationships at the topological level⁶⁴ and a unified method to construct linear equations for structure-property relations.⁶⁵ In a later publication, it was also shown that any metric in chemical space (that is any similarity measure between chemical structures) can be expressed through fragmental descriptors.⁶⁶^,⁶⁷ Besides, a subset of fragmental descriptors providing unique coding of chemical structures was found.⁶⁸

The main impact of the theorems on the basis of invariants of molecular graphs is that they allowed substantiating the fragmental approach in chemoinformatics.⁶⁹^,⁷⁰ Based on these results, I developed a set of fragmental descriptors based on hierarchical scheme of atom classification,⁷¹^,⁷² as well as algorithms and software (program FRAGMENT) for its efficient computation. Although the primary goal of these descriptors was to predict physico-chemical properties of organic compounds, they can successfully be applied to predicting biological activity, including the assignment of organic compounds to pharmacological groups⁷³ and mutagenicity.⁷⁴ This set of descriptors was successfully used by us for building models to predict numerous properties of organic compounds, including the enthalpy of sublimation,⁷⁵ flash point,⁷⁶ molecular polarizability,⁷⁷ magnetic susceptibility,⁷⁸ affinity of dyes for cellulose fiber,⁷⁹ the stability of complexes with α-cyclodextrin,⁸⁰ the enthalpy of vaporization,⁸¹ lipophilicity,⁸² Abraham constants,⁸² melting points of ionic liquids.⁸³ As a further development of the fragmental approach, I developed fragmental descriptors with labeled atoms.⁸⁴ Using this type of fragmental descriptors, I succeeded in developing models to make quantitative predictions of several types of biological activity,^84-86 reaction rate constant for ester hydrolysis⁸⁴ and several types of local properties, including ³¹P NMR shift,⁸⁴ ionization constants,⁸⁷ different substituent constants.⁸⁸ As another direction of the development of the fragmental approach, I put forward the concept of pseudofragmental descriptors (FragProp), the values of which are combination of certain properties of atoms forming the fragments.⁸⁹ The advantages of using pseudofragmental descriptors were demonstrated by the example of predicting physical properties of polymers.⁸⁹

Methodological developments in the use of neural networks in chemoinformatics

I started to work with artificial neural networks in 1993 with an article,⁹⁰ in which for the first time neural networks was used to predict physical properties of organic compounds. At that time neural network was a new tool, and it was not clear how to: (1) prevent overtraining; (2) handle stochastic properties of neural networks; (3) work with a big number of descriptors; (4) interpret neural network models; (4) build models with required symmetry properties. To solve the first problem, in 1995 I suggested splitting data into three sets: (i) a training set used for learning; (ii) a validation set used to define a point for early stopping of learning; (iii) a test set used for assessing predictive performance of the neural network model.⁹¹ The stochastic properties of neural networks were addressed by using ensembles of neural networks, advantages of which over separate neural network models were demonstrated by us for the case of predicting physical properties of organic compounds.⁷² Further development of the ensemble modeling idea in the framework of the three-set approach resulted in the creation of the double cross-validation procedure,⁸⁴ which became a standard routine for all out neural network studies. To address the problem of a big quantity of descriptors I have developed the Fast Stagewise Multiple Linear Regression (FSMLR) procedure,⁸⁴ which can efficiently select descriptors even from a huge number of them (millions). Besides, FSMLR can be used as independent machine learning method for the rapid construction of linear regressions. Currently this method is available as a part of the OCHEM project.⁹² The problem of the interpretability of neural network regression models was solved by means of putting forward an approach based on rapid calculation of first and second derivatives of outputs with regard to inputs.⁹³ The problem of proper symmetry of neural network models was solved by suggesting the concept of the learned symmetry,⁹⁴ in accordance with which neural networks learn the required symmetry properties during training. I wrote several reviews concerning the use of neural networks in chemoinformatics^95-97 and developed software NASAWIN,⁹⁸ which was used in most of my studies in this field.

Application of neural networks in conjunction with fragmental descriptors

It follows from the aforementioned theorems on the basis of invariants of molecular graphs and the Kolmogorov theorem in neural network interpretation that any scalar property of chemical compound can be approximated by a multi-layered neural network with inputs fed by fragmental descriptors. This suggests theoretically the best combination of machine learning method and molecular descriptor type. The first evidence that such combination could be optimal for predicting physical properties of hydrocarbons was obtained by me in the paper.⁹⁰ The same conclusion was drawn for big datasets for the case of predicting magnetic susceptibility,⁷⁸ the enthalpy of vaporization,⁸¹ enthalpy of sublimation,⁷⁵ flash point.⁷⁶ The same conclusion was also made for very diverse datasets with physical properties of organic compounds.⁷² Finally, this was supported in the benchmark study of predicting the melting point of ionic liquids.⁸³

New integrated concepts in the use of neural networks in chemoinformatics

In order to more fully disclose the potential applications of neural networks in chemoinformatics, I proposed three integrated approaches. The first one, called QSCPR (Quantitative Structure-Conditions-Property Relationships),⁹⁹^,¹⁰⁰ is designed to predict properties of chemical compounds under at different conditions (e.g. the temperature, the pressure, the properties of solvents, etc). I have shown that by mixing descriptions of chemical structures and conditions with the help of backpropagation neural networks, it is possible to model the “structure-pressure-boiling point”, the “structure-temperature-density” and the “structure-temperature-viscosity” relationships for hydrocarbons,⁹⁹ as well as the “structure – reaction conditions – rate constants” relationships for the acid hydrolysis of esters.¹⁰⁰ As a further development, a concept of the “bimolecular” QSPR,¹⁰¹ in which descriptions of a pair of structures are combined with the help of neural networks, was suggested. The efficiency of this approach was demonstrated by predicting the solvation free energy of different compounds in different solvents with a single model.¹⁰¹

The second integrated approach deals with integration of the QSPR approach based on the use of neural networks with the results of molecular modeling taken in the form of quantum chemical descriptors. I have shown that in certain cases, when the amount of experimental data is not too small and not too big, such combination can lead to very efficient solutions. The advantages of this integrated approach were demonstrated for predicting: (i) the position of the long-wave absorption band of symmetrical cyanine dyes in alcohol solution,¹⁰² ionization constants for different classes of organic compounds,⁸⁷ and mutagenicity.^103-105

The third integrated approach deals with integration of different neural network QSPR models built for different but mutually interrelated endpoints in the framework of the “inductive knowledge transfer” concept. Efficiency of the “horizontal” integration of models (implemented through the use of neural networks and multi-task learning) was demonstrated (in cooperation with Prof. A. Varnek) for predicting a set of ADME properties.¹⁰⁶ Efficiency of the “vertical” integration of neural network QSPR models in the frame of the “multilevel approach to the prediction of properties of organic compounds”⁸² was shown for predicting soil sorption coefficients of organic compounds and solubility of fullerene C₆₀ in different organic solvents.⁸²

Main directions of my current work in chemoinformatics

The first direction of my current work deals with the use of one-class classification machine learning technique in chemoinformatics. The possibility to use the one-class SVM method for defining applicability domains for QSPR models was shown (in cooperation with Prof. A. Varnek) for predicting stability constants of organic ligands with alkaline-earth metals in water.¹⁰⁷ I have also demonstrated high efficiency of this approach for conducting ligand-based virtual screening.^108-110 The next direction of my work deals with development of the “continuous molecular fields” approach,^110-112 which is based on representing chemical structures in SAR/QSAR studies by means of continuous molecular field functions instead of traditional discrete sets of descriptors. The third direction of my work deals with the use of the Generative Topographic Mapping (GTM) technique as a universal tool for chemical data visualization, structure-property modeling and dataset comparison (in cooperation with Prof. A. Varnek).¹¹³ The fourth direction of my work deals with defining the chemical space, development of methods to work with it, and defining chemoinformatics as theoretical chemistry discipline (in cooperation with Prof. A. Varnek).⁶⁹^,¹¹⁴ We have formulated the main mathematical problems in chemoinformatics and the ways to solve them in the perspective review article.¹¹⁵ In 2011-2012, the papers¹¹⁴^,¹¹⁵ were the most accessed articles in two top journals in the field of chemoinformatics.