Nikolay P. Savchuk

Drug Discovery Using Support Vector Machines. The Case Studies of Drug-likeness, Agrochemical-likeness, and Enzyme Inhibition Predictions

Support Vector Machines (SVM) is a powerful classification and regression tool that is becoming increasingly popular in various machine learning applications. We tested the ability of SVM, in comparison with well-known neural network techniques, to predict drug-likeness and agrochemical-likeness for large compound collections. For both kinds of data, SVM outperforms various neural networks using the same set of descriptors. We also used SVM for estimating the activity of Carbonic Anhydrase II (CA II) enzyme inhibitors and found that the prediction quality of our SVM model is better than that reported earlier for conventional QSAR. Model characteristics and data set features were studied in detail.

Recent Progress in Discovery and Development of Antimitotic Agents

This review highlights structural diversity of antimitotic agents. In particular, we emphasized current antimitotic therapies based on modulation of microtubule dynamics. With several successful anticancer drugs on the market and numerous compounds in clinical developments, tubulin-binding agents remain among the most important categories of anticancer agents. Compounds targeting mitotic kinases and kinesins are also discussed.

KOHONEN MAPS FOR PREDICTION OF BINDING TO HUMAN CYTOCHROME P450 3A4

The drug development process utilizes the parallel assessment of activity at a therapeutic target as well as absorption, distribution, metabolism, excretion, and toxicity properties of molecules. The development of novel, reliable, and inexpensive computational methods for the early assessment of metabolism and toxicity is becoming increasingly an important part of this process. We have used a computational approach for the assessment of drugs and drug-like compounds which bind to the cytochromes P450 (P450s) with experimentally determined Km values. The physicochemical properties of these compounds were calculated using molecular descriptor software and then analyzed using Kohonen self-organizing maps. This approach was applied to generate a P450-specific classification of nearly 500 drug compounds. We observed statistically significant differences in the molecular properties of low Km molecules for various P450s and suggest a relationship between 33 of these compounds and their CYP3A4-inhibitory activity. A test set of additional CYP3A4 inhibitors was used, and 13 of 15 of these molecules were colocated in the regions of low Km values. This computational approach represents a novel method for use in the generation of metabolism models, enabling the scoring of libraries of compounds for their Km values to numerous P450s.

Comprehensive computational assessment of ADME properties using mapping techniques.

One strategy to potentially improve the success of drug discovery is to apply computational approaches early in the process to select molecules and scaffolds with ideal binding and physicochemical properties. Numerous algorithms and different molecular descriptors have been used for modeling ligand-protein interactions as well as absorption, distribution, metabolism and excretion (ADME) properties. In most cases a single data set has been evaluated with one approach or multiple algorithms that have been compared for a single dataset. These models have been primarily evaluated by leave-one out analysis or boot strapping with groups representing 25-50% of the training set left out of the final model. In a very few examples a test set of molecules not included in the model has been used for an external evaluation. In the present study we have applied Sammon non-linear maps, Support Vector Machines and Kohonen Self Organizing Maps to modeling numerous datasets for ADME properties including human intestinal absorption, blood brain barrier permeability, cytochrome P450 binding, plasma protein binding, P-gp inhibition, volume of distribution and plasma half life.

Computational mapping tools for drug discovery

During the past decade, computational technologies have become well integrated in the modern drug design process and have gained in influence. They have dramatically revolutionized the way in which we approach drug discovery, leading to the explosive growth in the amount of chemical and biological data that are typically multidimensional in structure. As a result, the irresistible rush towards using computational approaches has focused on dimensionality reduction and the convenient representation of high-dimensional data sets. This has, in turn, led to the development of advanced machine-learning algorithms. In this review we describe a variety of conceptually different mapping techniques that have attracted the attention of researchers because they allow analysis of complex multidimensional data in an intuitively comprehensible visual manner.

QUANTITATIVE STRUCTURE-METABOLISM RELATIONSHIP MODELING OF METABOLIC N-DEALKYLATION REACTION RATES

It is widely recognized that preclinical drug discovery can be improved via the parallel assessment of bioactivity, absorption, distribution, metabolism, excretion, and toxicity properties of molecules. High-throughput computational methods may enable such assessment at the earliest, least expensive discovery stages, such as during screening compound libraries and the hit-to-lead process. As an attempt to predict drug metabolism and toxicity, we have developed an approach for evaluation of the rate of N-dealkylation mediated by two of the most important human cytochrome P450s (P450), namely CYP3A4 and CYP2D6. We have taken a novel approach by using descriptors generated for the whole molecule, the reaction centroid, and the leaving group, and then applying neural network computations and sensitivity analysis to generate quantitative structure-metabolism relationship models. The quality of these models was assessed by using the cross-validated correlation coefficients of 0.82 for CYP3A4 and 0.79 for CYP2D6 as well as external test molecules for each enzyme. The relative performance of different neural networks was also compared, and modular neural networks with two hidden layers provided the best predictive ability. Functional dependencies between the neural network input and output variables, generalization ability, and limitations of the described approach are also discussed. These models represent an initial approach to predicting the rate of P450-mediated metabolism and may be applied and integrated with other models for P450 binding to produce a systems-based approach for predicting drug metabolism.

In Silico Estimation of DMSO Solubility of Organic Compounds for Bioscreening

Solubility of organic compounds in DMSO is an important issue for commercial and academic organizations handling large compound collections or performing biological screening. In particular, solubility data are critical for the optimization of storage conditions and for the selection of compounds for bioscreening compatible with the assay protocol. Solubility is largely determined by the solvation energy and the crystal disruption energy, and these molecular phenomena should be assessed in structure-solubility correlation studies. The authors summarize our long-term experimental observations and theoretical studies of physicochemical determinants of DMSO solubility of organic substances. They compiled a comprehensive reference database of proprietary data on compound solubility (55,277 compounds with good DMSO solubility and 10,223 compounds with poor DMSO solubility), calculated specific molecular descriptors (topological, electromagnetic, charge, and lipophilicity parameters), and applied an advanced machine-learning approach for training neural networks to address the solubility. Both supervised (feed-forward, back-propagated neural networks) and unsupervised (Kohonen neural networks) learning methods were used. The resulting neural network models were validated by successfully predicting DMSO solubility of compounds in independent test selections. (Journal of Biomolecular Screening 2004:22-31)

New diversity calculations algorithms used for compound selection.

Some modifications were introduced into the previously described Centroid diversity sorting algorithm, which uses cosine similarity metric. The modified algorithm is suitable for the work with large databases on personal computers. For example, for diversity sorting of the database with the size greater than a million of records, less than 9 h are required (Pentium III, 800 MHz). The problem of selecting new compounds into the existing collection is examined to reach the maximum diversity of the collection. The article describes the new algorithm for the selection of heterocyclic compounds.

Recent Progress in Development of Non-ATP Competitive Small-Molecule Inhibitors of Protein Kinases

The majority of marketed and late stage development kinase inhibitors are reported to be ATP-competitive. As a result, many promising drug candidates display non-specific activity that results in undesired physiological effects. There is growing interest towards non-ATP competitive kinase inhibitors, as they are expected to yield highly specific and efficacious molecules devoid of non-mechanistic toxicity. Recent developments in this area are summarized in our review.

Advanced exact structure searching in large databases of chemical compounds.

Efficient recognition of tautomeric compound forms in large corporate or commercially available compound databases is a difficult and labor intensive task. Our data indicate that up to 0.5% of commercially available compound collections for bioscreening contain tautomers. Though in the large registry databases, such as Beilstein and CAS, the tautomers are found in an automated fashion using high-performance computational technologies, their real-time recognition in the nonregistry corporate databases, as a rule, remains problematic. We have developed an effective algorithm for tautomer searching based on the proprietary chemoinformatics platform. This algorithm reduces the compound to a canonical structure. This feature enables rapid, automated computer searching of most of the known tautomeric transformations that occur in databases of organic compounds. Another useful extension of this methodology is related to the ability to effectively search for different forms of compounds that contain ionic and semipolar bonds. The computations are performed in the Windows environment on a standard personal computer, a very useful feature. The practical application of the proposed methodology is illustrated by several examples of successful recovery of tautomers and different forms of ionic compounds from real commercially available nonregistry databases.