A Novel Methodology for Human Plasma Protein Binding:Prediction, Validation, and Applicability Domain

Abstract

Background: Plasma protein binding is a key component in drug therapy as it affects the pharmacokinetics and pharmacodynamics of drugs. Objectives: This study aimed to predict the fraction of plasma protein binding. Methods: A quantitative structure-activity relationship, convolutional neural network, and feed-forward neural network (QSAR-CNN-FFNN) methodology was used. CNN was used for feature selection, which is known as a difficult task in QSAR studies. The values of the descriptors acquired without the preprocessing procedures were rearranged into matrices, and features from a deep fully connected layer of a pre-trained CNN (ALEXNET) were extracted. Then, the latest features learned from the CNN layers were flattened out and passed through an FFNN to make predictions. Results: The external accuracy of the validation set (Q2