Statistical learning approach for predicting specific pharmacodynamic, pharmacokinetic, or toxicological properties of pharmaceutical agents

Abstract

Pharmaceutical agents have been developed and tested for possessing desirable pharmacodynamic, pharmacokinetic, and minimal level of toxicological properties. Computational methods have been explored for predicting these properties aimed at the discovery of promising leads and the elimination of unsuitable ones in early stages of drug development. Statistical learning methods have shown their potential for predicting these properties for structurally diverse sets of agents by using both conventional (quantitative structure-activity and structure-property relationships) and more recently explored (such as neural networks and support vector machines) statistical models. These methods have been used for predicting agents of a variety of pharmacodynamic (such as inhibitors or agonists of a therapeutic target), pharmacokinetic (such as P-glycoprotein substrates, human intestine absorption, and blood-brain barrier penetrating capabilities), and toxicological (such as genotoxicity) properties. The strategies, current progresses, and the underlying difficulties and future prospects of the application of the recently explored statistical learning methods are discussed. Drug Dev. Res. 66:245-259, 2006.

Prediction models of human plasma protein binding rate and oral bioavailability derived by using GA–CG–SVM method

“…In fact, the effect of the feature selection on the SVM model has been widely discussed in literature [18,34,35]. But few of studies have analyzed the possible impact of the SVM parameter optimization [20].…”

“…On the other hand, most of the prediction models developed so far are based on the quantitative structure-activity relationship (QSAR) and quantitative structure-property relationship (QSPR) [7,[13][14][15][16]. However, there is an inherent deficiency for the QSAR and QSPR, namely, the limitation of the diverse range of chemical structures [17,18], which may be another reason why the BIO and PPBR are difficult to be predicted. An alternative way to overcome this problem is the use of nonlinear supervised learning method, such as support vector machine (SVM) [8,12,19], which can cover more diverse range of structures than those described by the QSAR and QSPR models.…”

A support vector machines approach for virtual screening of active compounds of single and multiple mechanisms from large libraries at an improved hit-rate and enrichment factor

“…Therefore, alternative approaches that enhance screening speed and compound diversity without relying on target structural information are highly desired. ML methods have been explored for developing www.elsevier.com/locate/JMGM Machine learning-BKD [12,9,11,13,14] 101 such alternative VS tools [7][8][9] because of their high-CPU speed (100 K data points per hour on 3 GHz PC) [11] and capability for covering highly diverse spectrum of compounds [32]. The reported performance of various LBVS and SBVS tools in screening compound libraries of >90,000 compounds is summarised in Table 1.…”

Machine Learning Techniques and Drug Design

“…ANNs techniques have been also used in the fields of robotics, pattern identification, psychology, physics, computer science, biology and many others [37][38][39][40]. In addition, ANNs have been applied to the modeling of several systems in a wide range of applications such as animal science [41], cancer imaging extraction and classification [42,43], pharmacodynamic and pharmacokinetic modeling [44,45] and mapping of dose-effect relationships on pharmacological response [46], to predict secondary structures [47] and transmembrane segments [48], simulation of C13 nuclear magnetic resonance spectra [49], prediction of drug resistance of HIV-1 protease ligands [50], prediction of toxicity of chemicals to aquatic species [51], and as well as predicting physicochemical properties from the perspective of pharmaceutical research [52].…”