Leveraging Machine Learning and ChemBERTa for Efficient Identification of CB1 Receptor-Active Compounds

The CB1 receptor, a key component of the endocannabinoid system, influences mood, pain perception, and appetite regulation, making it a valuable yet complex target for therapeutic development. Traditional drug discovery methods are time-intensive and resource-heavy, often with limited success. This study demonstrates how machine learning, particularly transformer-based models like ChemBERTa, can analyze molecular datasets, uncover structural patterns, and predict properties with high precision. We explore shallow machine learning models, deep learning models and hybrid approaches to accelerate the discovery of CB1 receptor-active compounds. Using a curated ChEMBL dataset, we trained both traditional models such as Random Forest and deep learning architectures like ChemBERTa, achieving accuracies of 83.29% and 85.45%, respectively. By prioritizing highpotential compounds, our approach can reduce the traditional bottlenecks of time and cost in drug discovery. Our findings highlight the synergy between traditional cheminformatics and advanced machine learning. Random Forest models provided interpretability and feature-driven insights, while ChemBERTa offered superior generalizability and scalability, showcasing the advantages of transfer learning. Despite challenges like data sparsity and the absence of experimental validation, this work establishes the foundation for a scalable, AI-driven pipeline to identify selective $\mathbf{C B 1}$ receptor-targeted therapies. By integrating computational advancements with domain expertise, this study highlights a practical framework for advancing targeted therapies. Combining traditional cheminformatics with machine learning, particularly transformer-based models like ChemBERTa, streamlines the discovery of CB1 receptor-active compounds, emphasizing efficiency, scalability, and costeffectiveness.