A renaissance of neural networks in drug discovery

ABSTRACT

Introduction: Neural networks are becoming a very popular method for solving machine learning and artificial intelligence problems. The variety of neural network types and their application to drug discovery requires expert knowledge to choose the most appropriate approach.

Areas covered: In this review, the authors discuss traditional and newly emerging neural network approaches to drug discovery. Their focus is on backpropagation neural networks and their variants, self-organizing maps and associated methods, and a relatively new technique, deep learning. The most important technical issues are discussed including overfitting and its prevention through regularization, ensemble and multitask modeling, model interpretation, and estimation of applicability domain. Different aspects of using neural networks in drug discovery are considered: building structure-activity models with respect to various targets; predicting drug selectivity, toxicity profiles, ADMET and physicochemical properties; characteristics of drug-delivery systems and virtual screening.

Expert opinion: Neural networks continue to grow in importance for drug discovery. Recent developments in deep learning suggests further improvements may be gained in the analysis of large chemical data sets. It’s anticipated that neural networks will be more widely used in drug discovery in the future, and applied in non-traditional areas such as drug delivery systems, biologically compatible materials, and regenerative medicine.     

Statistical evaluation of nonisothermal prediction of drug stability

Nonisothermal prediction of drug stability based on direct nonlinear estimation of the shelf-life was compared with the isothermal approach. The reliability of the statistics for the estimates of the shelf-life (the time period required for a drug to degrade to 90% remaining at 25 degrees C) and activation energy obtained by the two methods was evaluated by the Monte Carlo method of computer simulations. The accuracy and precision of the estimates obtained by the nonisothermal method depended largely on the experimental conditions, such as experimental periods, sampling time, and temperature rise programs. The uncertainty of the estimates was determined mainly by the extents of drug degradation and temperature change achieved during the experiment. The nonisothermal method needed suitable experimental designs and precise assay methods of drug contents to provide reliable parameter estimates.     

Statistical prediction of drug stability based on nonlinear parameter estimation

The classical approach in Arrhenius prediction of drug stability uses two sequential steps of linear regression involving (a) a function of drug content versus time to obtain the rate constants (k) at several elevated temperatures and (b) the relationship of logarithm of mean k versus reciprocal temperature to predict the room temperature rate constant and hence the shelf-life of the drug. Uncertainties in drug content determinations are often neglected in the second regression. The classical approach also provides a wide and unsymmetrical 95% confidence interval for the predicted shelf-life. We have developed equations which allow for direct statistical prediction of shelf-life using observed values of drug content, time, and temperature. Nonlinear regression analysis was employed to provide parameter estimates of drug shelf-life and the energy of activation. The developed approach was shown to provide good estimates of shelf-life with meaningful statistics of reactions over a wide range of stability and energetics, with various kinetic orders, with different levels of noise in the data, and with different types of data structure. Comparison between the nonlinear approach and the classical approach showed that the nonlinear approach provided better mean estimates of shelf-life with much smaller and more symmetrical 95% confidence intervals than the classical approach. The method appears sufficiently robust and wide-ranging as to be potentially applicable for the prediction of the drug stability of pharmaceutical products.     

Combining in vitro and in vivo pharmacokinetic data for prediction of hepatic drug clearance in humans by artificial neural networks and multivariate statistical techniques

Several statistical regression models and artificial neural networks were used to predict the hepatic drug clearance in humans from in vitro (hepatocyte) and in vivo pharmacokinetic data and to identify the most predictive models for this purpose. Cross-validation was performed to assess prediction accuracy. It turned out that human hepatocyte data was the best predictor, followed by rat hepatocyte data. Dog hepatocyte data and dog and rat in vivo data appear to be uncorrelated with human in vivo clearance and did not significantly contribute to the prediction models. Considering the present evaluation, the most cost-effective and most accurate approach to achieve satisfactory predictions in human is a combination of in vitro clearances on human and rat hepatocytes. Such information is of considerable value to speed up drug discovery. This study also showed some of the limitations of the approach related to the size of the database used in the present evaluation.     

人工神经网络在药物控释系统研究中的应用

目的介绍人工神经网络在药物控释系统研究中的应用.方法查阅相关文献,总结、归纳国内外人工神经网络在药物控释系统中的应用.结果人工神经网络能优化处方组成和工艺过程,使其在控释片剂、控释微粒以及透皮吸收中得到应用.结论人工神经网络在设计和开发药物控释系统中具有广阔的前景.     

Quantitative Structure-Pharmacokinetic Relationships for drug distribution properties by using general regression neural network

Quantitative Structure-Pharmacokinetic Relationships (QSPkR) have increasingly been used for developing models for the prediction of the pharmacokinetic properties of drug leads. QSPkR models are primarily developed by means of statistical methods such as multiple linear regression (MLR). These methods often explore a linear relationship between the pharmacokinetic property of interest and the structural and physicochemical properties of the studied compounds, which are not applicable to those agents with nonlinear relationships. Hence, statistical methods capable of modeling nonlinear relationships need to be developed. In this work, a relatively new kind of nonlinear method, general regression neural network (GRNN), was explored for modeling three drug distribution properties based on diverse sets of drugs. The three properties are blood-brain barrier penetration, binding to human serum albumin, and milk-plasma distribution. The prediction capability of GRNN-developed models was compared to those developed using MLR and a nonlinear multilayer feedforward neural network (MLFN) method. For blood-brain barrier penetration, the computed r(2) and MSE values of the GRNN-, MLR-, and MLFN-developed models are 0.701 and 0.130, 0.649 and 0.154, and 0.662 and 0.147, respectively, by using an independent validation set. The corresponding values for human serum albumin binding are 0.851 and 0.041, 0.770 and 0.079, and 0.749 and 0.089, respectively, and that for milk-plasma distribution are 0.677 and 0.206, 0.224 and 0.647, and 0.201 and 0.587, respectively. These suggest that GRNN is potentially useful for predicting QSPkR properties of chemical agents.     

Application of artificial neural networks in the design of controlled release drug delivery systems

Controlled release drug delivery systems offer great advantages over the conventional dosage forms. However, there are great challenges to efficiently develop controlled release drug delivery systems due to the complexity of these delivery systems. Traditional statistic response surface methodology (RSM) is one of the techniques that has been employed to develop and formulate controlled release dosage forms. However, there are some limitations to the RSM technique. Hence, another technique called artificial neural networks (ANN) has recently gained wide popularity in the development of controlled release dosage forms. In this review, the basic ANN structure, the development of the ANN model and an explanation of how to use ANN to design and develop controlled release drug delivery systems are discussed. In addition, the applications of ANN in the design and development of controlled release dosage forms are also summarized in this review.     

药剂学药物有效期预测及稳定性实验改革及体会

针对药剂学药物有效期预测及稳定性实验进行改革,建立了新的教学模式,改进原有实验方案,取得了良好效果,从而更好地培养了药学本科生的信息利用能力、独立思考能力、综合分析和解决问题的能力、交流合作能力,为日后的科研工作奠定了基础,对药学生的科研素质提高和能力培养有着积极的意义。     

Rank regression in stability analysis

Stability data are often collected to determine the shelf life of certain characteristics of a pharmaceutical product, for example, a drug's potency over time. Statistical approaches such as the linear regression models are considered as appropriate to analyze the stability data. However, most of these regression models in both theory and practice rely heavily on their underlying parametric assumptions, such as normality of the continuous characteristics or their transformations. In this article, we propose and study some rank-based regression procedures for the stability data when the linear regression models are semiparametric with unspecified error structure. Numerical studies including Monte Carlo simulations and practical example are demonstrated with the proposed procedures as well.     

Application of artificial neural networks (ANNs) and genetic programming (GP) for prediction of drug release from solid lipid matrices

The aim of the present study was to develop a semi-empirical mathematical model, which is able to predict the release profiles of solid lipid extrudates of different dimensions. The development of the model was based on the application of ANNs and GP. ANNs' abilities to deal with multidimensional data were exploited. GP programming was used to determine the constants of the model function, a modified Weibull equation. Differently dimensioned extrudates consisting of diprophylline, tristearin and polyethylene glycol were produced by the use of a twin-screw extruder and their dissolution behaviour was studied. Experimentally obtained dissolution curves were compared to the calculated release profiles, derived from the semi-empirical mathematical model.     

Effect of source variation on drug release from HPMC tablets: linear regression modeling for prediction of drug release

The aim of this study was to investigate the effect of source variation of hydroxypropyl methylcellulose (HPMC) raw material on prediction of drug release from HPMC matrix tablets. To achieve this objective, the flow ability (i.e., angle of repose and Carr's compressibility index) and apparent viscosity of HPMC from 3 sources was investigated to differentiate HPMC source variation. The physicochemical properties of drug and manufacturing process were also incorporated to develop the linear regression model for prediction of drug release. Specifically, the in vitro release of 18 formulations was determined according to a 2 × 3 × 3 full factorial design. Further regression analysis provided a quantitative relationship between the response and the studied independent variables. It was found that either apparent viscosity or Carr's compressibility index of HPMC powders combining with solubility and molecular weight of drug had significant impact on the release behavior of drug. The increased drug release was observed when a greater in drug solubility and a decrease in the molecular weight of drug were applied. Most importantly, this study has shown that the HPMC having low viscosity or high compressibility index resulted in an increase of drug release, especially in the case of poorly soluble drugs.     

药物有效期预测及稳定性影响因素实验设计

目的探索替莫唑胺作为药物稳定性学生实验的可行性,建立预测其水溶液室温有效期的实验方法,并考察pH对稳定性的影响。方法用HPLC法测定替莫唑胺水溶液有效成分含量;用一级反应速率方程及阿雷尼乌斯公式进行计算预测。结果本品水溶液在高温条件下不稳定;同时受pH影响较大,在酸性条件下稳定,在碱性条件下不稳定;室温有效期约为21d。结论本次实验药物选择合适,实验操作简便,快速准确,适合学生操作,可用于药剂学药物稳定性的实验教学。     

An alternative approach based on artificial neural networks to study controlled drug release

An alternative methodology based on artificial neural networks is proposed to be a complementary tool to other conventional methods to study controlled drug release. Two systems are used to test the approach; namely, hydrocortisone in a biodegradable matrix and rhodium (II) butyrate complexes in a bioceramic matrix. Two well-established mathematical models are used to simulate different release profiles as a function of fundamental properties; namely, diffusion coefficient (D), saturation solubility (C(s)), drug loading (A), and the height of the device (h). The models were tested, and the results show that these fundamental properties can be predicted after learning the experimental or model data for controlled drug release systems. The neural network results obtained after the learning stage can be considered to quantitatively predict ideal experimental conditions. Overall, the proposed methodology was shown to be efficient for ideal experiments, with a relative average error of <1% in both tests. This approach can be useful for the experimental analysis to simulate and design efficient controlled drug-release systems.     

Application of dynamic neural networks in the modeling of drug release from polyethylene oxide matrix tablets

The main objective of this study was to demonstrate the possible use of dynamic neural networks to model diclofenac sodium release from polyethylene oxide hydrophilic matrix tablets. High and low molecular weight polymers in the range of 0.9–5 × 10⁶ have been used as matrix forming materials and 12 different formulations were prepared for each polymer. Matrix tablets were made by direct compression method. Fractions of polymer and compression force have been selected as most influential factors on diclofenac sodium release profile. In vitro dissolution profile has been treated as time series using dynamic neural networks. Dynamic networks are expected to be advantageous in the modeling of drug release. Networks of different topologies have been constructed in order to obtain precise prediction of release profiles for test formulations. Short-term and long-term memory structures have been included in the design of network making it possible to treat dissolution profiles as time series. The ability of network to model drug release has been assessed by the determination of correlation between predicted and experimentally obtained data. Calculated difference (f₁) and similarity (f₂) factors indicate that dynamic networks are capable of accurate predictions. Dynamic neural networks were compared to most frequently used static network, multi-layered perceptron, and superiority of dynamic networks has been demonstrated. The study also demonstrated differences between the used polyethylene oxide polymers in respect to drug release and suggests explanations for the obtained results.