Application of Artificial Neural Network in Drug Discovery

                   

                                         

                                        Bansilal Ramnath Agarwal Charitable Trust's

                                Vishwakarma Institute of Technology

                                     (An Autonomous Institute affiliated to Savitribai                                       Phule Pune University)

Vishwakarma Institute of Technology, 666, Upper Indiranagar, Bibwewadi , Pune, Maharashtra, INDIA - 411 037.

   

ARTIFICIAL NEURAL NETWORKS MODELING

ANN is biologically inspired computational model, capable of simulating neurological processing ability of the human brain. Average human brain contains about 100 billions of neurons with each neuron being connected with 1000-10,000 connections to others. A single neuron consists of three major parts—dendrites (fine branched out threads) carrying signals into the cell, the cell body receiving and processing the information, and the axon (a single longer extension). The axon carries the signal away and relays it to the dendrites of the next neuron or receptor of a target cell. The signals are conducted in all-or-none fashion through the cells. All the connections in the brain enable it to learn, recognize patterns, and predict outcomes. Similarly to the brain, ANN is composed of numerous processing units, artificial neurons. The connections among all the units vary in strength, which is defined by coefficients or weights. The ANN mimics working of human brain and potentially fulfills the cherished dream of scientists to develop machines that can think like human beings. ANNs simulate learning and generalization behavior of the human brain through data modeling and pattern recognition for complex multidimensional problems. A significant difference between an ANN model and a statistical model is that the ANN can generalize the relationship between independent and dependent variables without a specific mathematical function. Thus, an ANN works well for solving nonlinear problems of multivariate and multiresponse systems such as space analysis in quantitative structure-activity relationships in drugs discovery and drugs development.

 

APPLICATION OF ANNS MODELING IN DRUG DELIVERY AND PHARMACEUTICAL RESEARCH

 

ANN has the ability to investigate complex, nonlinear relationships. Neural networks find their application in many diverse fields such as engineering, pharmaceutical sciences, and medicine. ANNs are frequently being used for regression and discriminant data analysis. ANNs are being increasingly applied for the screening of large inhibitor libraries, also referred to as virtual screening, assessment of the properties of the ligand in terms of their pharmacophoric features, docking, quantitative structure activity relationship (QSAR) studies, and prediction of the ADME-Tox properties. The potential applications of ANN methodology in the pharmaceutical sciences are broad as ANNs capabilities can be summarized by modeling, pattern recognition and prediction. Thus, applications of ANNs include drug modeling, dosage design, protein structure and function prediction, pharmacokinetics and pharmacodynamics modeling as well as interpretation of analytical data.

Virtual Screening

VS represents a computational way to predict biological and pharmacological properties of compounds. Computational filtering of the billions of compounds that exist in the chemical libraries is an essential cost- and time-cutting measure that many companies now look forward to. The focus would be to first identify a virtual space in the vast chemical library and then screen the selected compounds using high-throughput technologies. An important prerequisite for VS is the identification of target proteins. This can be done by employing systems biology approaches to mine drug targets in various metabolic pathways. VS can then be used to identify lead compounds that bind selectively to the target protein. Screening of ligands can be either ligand based or target protein based. Target-based drug design usually requires information about the structure of the target protein. This information is usually obtained using X-ray crystallography or nuclear magnetic resonance. Ligand-based VS is based on a similarity search between the preexisting active compound and chemical library. Neural networks are increasingly being used for VS as they are able to work around the interactions of the ligand and the target protein. ANNs are good at deciphering nonlinear relationships and have fewer prerequisites. Neural networks are advantageous as compared to statistical models because they can recognize patterns. Moreover, they do not require complete datasets nor is any experimental information required. ANNs work on patterns and are thus capable of deciphering nonlinear dependencies of the output variables (properties, biological activities) based on the input variables (descriptors). The complexity of the ANN increases as the number of descriptors increases. Hence, ANN tends to work with the descriptors that are most relevant in the study; the objective of VS would be to map and focus on a specific part of the library, screening the ligands based on the proposed descriptors. Pharmacophores describe the features needed by the ligand to bind actively to the target protein. These features include information about the number of aromatic groups, the hydrogen bond donors and acceptors, and the hydrophobic groups. By analyzing different descriptors, ANN can predict the complex relationships between the structure of the molecules and their physicochemical properties.

Quantitative Structure Activity Relationship (QSAR)

QSAR models summarize the relationship between the chemical structural features of a compound and its physicochemical properties or biological active properties such as molecular weight, molar volume, electronegativity, partition coefficient, number of hydrogen bond donors and acceptors, etc. They essentially relate the topology of the molecule with the physicochemical descriptors with the biological activity of the molecule. QSAR models define the mathematical relationship between the descriptors and biological activities of unknown ligands with that of known ligands. These were based on the presence or absence of certain physicochemical properties. Traditional QSAR methods are based on multiple linear regression and partial least square regression. These are capable of deciphering linear relationships. ANNs have the added advantage of resolving nonlinear relationships. This flexibility of ANN enables it to discover more complex relationships. To overcome these limitations, ANNs are being used to predict the biological activities of unknown compounds. Multilayered feedforward networks can be used to estimate the solute solvent interactions by using several connectivity indices such as constitutional electrostatic, quantum, and topological descriptors. The application of ANN in QSAR studies has been increasing due to the speed and accuracy with which the processing of information takes place; as a result, nearly 35 different types of ANN are being used in drug design.

Docking:

Docking is a method of finding the preferred orientation of one molecule with another such that they form a stable complex. The strength of the binding of the two molecules is based on the affinity they have toward each other. Docking methods can be based on a complementary shape or the interactions between the ligand and the target. The interactions between the ligand and the target protein can be Vander Waal’s interactions, electrostatic interactions, pi bonding, interaction with metal ions, etc. Based on the ligand protein pairwise interaction, the binding energies are calculated. The docking algorithm scores the binding poses and ranks them is expected to have the strongest affinity toward the target protein. Docking methods rely on information about the binding site in terms of its solvent accessibility. The surface description of the binding site of the target protein as well as the ligand’s molecular surface are prerequisites for the match-based methods. ANNs can be used to predict the binding energy of the final docked complex by using the match surface descriptors, which would fasten the process of ligand screening. In addition to predicting the binding affinity of the docked complexes, neural networks can be trained to identify the protein conformations that are relevant for inhibitor binding. This is usually done by using the protein conformations that have high binding affinities as a training set. Several features such as the docking score, ligand efficiency, similarity searching, structural information of the ligand’s pose, and the position of the ligand’s pose have been fed as an input to the ANN. The relationship between the features and the ligand poses is then compared by ANN, and the existing trade-offs between the features are estimated.

 

ADME-Tox Prediction:

 

ADME-Tox properties relate to the absorption, distribution, metabolism, excretion, and toxicity of a drug molecule. Traditionally, these properties were predicted at the end of the drug discovery pipeline; however, with the advancement of in-silico tools such properties can be predicted in the early phase ADMET properties play an important role as they account for the failure of 60% of drug molecules during the drug development process. Early prediction of these properties would lead to a significant cost reduction in the field of drug research. Chemical libraries must be filtered and only the compounds that have acceptable physical and chemical properties must be selected during further drug design. Much of the bioavailability of a drug is dependent on its solubility and ability to cross the intestinal membranes, and this in turn relates to the physicochemical properties of a compound such as water solubility, acidity, number of rotatable bonds, nonpolar surface area, etc. The compounds that fail to comply with the famous Lipinski’s rule of five and the Verber’s rules generally have poor pharmacokinetic properties. Such drugs may show poor absorption, faster rate of metabolism and excretion, unfavorable distribution, and might be toxic in nature. The drug likeness of a compound can easily be predicted by filtering the compounds based on the Lipinski’s rule of five. The rule of five (RO5) was compiled after analyzing nearly 2500 compounds that were under Phase II clinical trials. ANNs are being used to filter off the candidates that show Mwt >500 and log P >5, hydrogen bond donors >5, and hydrogen bond acceptors >10. These properties correspond to poor absorption of the drug in the membrane and its ability to cross intestinal barriers. It is now possible to develop predictive models that are sophisticated enough to predict the ADMET properties. Such models can thereby replace traditional in vitro assays and in vivo experiments. ANNs are good at making predictions and hence they are being trained with experimental data to mine compounds that have feasible ADMET properties.

 

Formulation Development:

 In the field of formulation development, ANNs are being used to deal with the influx of datasets from different types of variables (binomial, discrete, and continuous) and nominal factors. Mathematical models (mechanistic, empirical, semiempirical, continuum, and discrete) have been employed to understand the behavior of formulations or processes. They are used for estimating the release time of a drug from drug formulations. Along with mathematical models, statistical experiments are now being recognized as useful techniques for designing tablets, microspheres, microparticles, nanoparticles, emulsions, hard capsules, and gels. The release rate and the dissolution rate of any formulation can be estimated in order to obtain a sustained drug delivery within a shorter time period. ANN’s can be used for predicting the ranking the formulations and processing variables that influenced the drug release. The main variables were the coating parameters like the blend size, blend time with the lubricant, the amounts of matrix in forming the polymer, and the direct compression filter. ANNs are trained based on these parameters to reduce the number of input parameters in order to optimize the drug release from the mini tablets. ANNs function as decision support systems in prospective ketoprofen SD formulation. Pharmaceutical optimization is thus generally done by first determining and quantifying the relationship between the formulation’s response and the variables and then finding the exact composition of these variables. The optimization procedure involves a series of experiments that will measure the response variables. The data obtained are then fit into a mathematical model and statistical tests are performed. ANNs are quicker at optimizing the amount of each variable in the formulation and thus are increasingly being used for the development of pharmacological formulations and even for the synthesis of liposomes, which act as drug delivery systems

 

CONCLUSION

ANNs have the capability to resolve nonlinear relationships and thus are advantageous over other statistical techniques during pattern analysis. They can be applied to model complex relationships between different physicochemical properties, mapping the correlation between several variables with the chemical structures. ANNs are therefore increasingly being used in the drug discovery process. The typical drug discovery process proves to be lengthy and expensive. It would be useful to apply different neural networks for a faster, efficient, cost-effective, and safer drug delivery.