The accurate prediction of small molecule-target binding affinity and the design of scoring functions are pivotal in structure-based drug discovery. Classical methods, including free energy perturbation and thermodynamic integration, offer high accuracy but suffer from low computational efficiency, limiting their practical utility. Alternative approaches like molecular mechanics/Poisson-Boltzmann surface area and molecular mechanics/generalized born surface area achieve a balance between efficiency and accuracy but rely on time-intensive molecular dynamics simulations. Traditional scoring functions—categorized as physics-based, empirical, or knowledge-based—are widely used in molecular docking but often fail to reliably distinguish active compounds from inactive ones. The emergence of artificial intelligence (AI) technologies has transformed this field, introducing AI-based scoring functions that surpass traditional linear regression methods by directly learning interaction patterns from data. Furthermore, AI-driven models such as RF-Score, Pafnucy, and InteractionGraphNet have demonstrated improved prediction accuracy by leveraging advanced feature extraction techniques, including graph representations and deep learning frameworks. Despite these advancements, challenges such as limited generalization ability, dataset variability, and model interpretability persist. This chapter explores the evolution of scoring function design, highlighting the transformative role of AI in predicting binding affinities and addressing the methodological and practical challenges faced in integrating AI into drug design workflows.