Optimizing Drug Screening with Machine Learning

Drug screening is the process by which potential drugs are identified and optimized before the selection of a candidate drug to progress to clinical trials. To find drug candidates with good pharmacokinetic properties and adequate safety in the human body, pharmaceutical researchers need to comprehensively consider the biological activity of compounds and their influence on the human body. More specifically, only when the compound has good biological activity and ADMET (i.e., absorption, distribution, metabolism, excretion, and toxicity) properties can it qualify as a drug candidate.To improve the efficiency of drug screening, we propose a drug candidate screening approach based on machine learning methods, which not only discovers appropriate compounds but also reveals the potential effects of molecular descriptor (i.e., features) values on the properties of compounds. First, an accurate prediction model is trained based on independent variables (i.e., feature values) and dependent variables (i.e., bioactivity values or ADMET properties). Second, we use a feature interpretation algorithm to pick out features with a significant impact on the dependent variables. Third, we search for the approximate optimal values of these important features and analyze their numerical ranges that are beneficial to obtaining better bioactivity and ADMET properties. Experimental results demonstrate that our scheme is accurate, efficient, and reliable.