Validation of machine learning supervised algorithms applied for drug-target identification in cannabinoid receptors

Abstract

The exploration of natural products (NPs) forms a fundamental framework in the pharmaceutical industry. In recent decades, the efficiency of discovering new drugs from NPs has been dramatically enhanced through the use of computational approaches. In this context, Machine Learning (ML) methods stand out as valuable tools for classifying, categorizing, and predicting the properties of NPs. ML methods play a crucial role in expediting drug discovery processes, offering enhanced insights and efficiency in harnessing the potential of natural compounds. In this thesis, we adopted a target-based approach paradigm guided by a hypothesis that delineates the molecular mechanisms of interaction between drugs and their respective targets. The molecular targets of interest in this study are the cannabinoid receptors type 1 (CB1) and type 2 (CB2), which are of interest in the context of energy metabolism and food intake. We utilized the COCONUT database, containing approximately 400,000 molecules, to explore the chemical space of natural products (NPs). In particular, we optimized ML algorithms such as Support Vector Machine, Random Forest, and Deep Neural Networks by employing various training and testing splits, along with the respective parameters for each method. The algorithms underwent validation based on scoring metrics such as the F1_Score, accuracy, and Receiver Operating Characteristics curves to evaluate the performance of binary classification models. The Random Forest (RF) algorithm was chosen as the top-performing model based on scoring metrics. It was then utilized to predict active molecules associated with the cannabinoid CB1 and CB2 receptors. The predictions were guided by selecting NPs based on the criteria of functional similarity.