Protein–ligand interactions are fundamental to many biological processes and central to drug discovery. The binding affinity (BA) between a small molecular ligand and its target protein largely determines therapeutic efficacy. Therefore, accurately predicting BA is vital for rational drug design [1]. Traditionally, experimental methods such as isothermal titration calorimetry (ITC) [2] and surface plasmon resonance (SPR) [3], aided by X-ray crystallography, have been employed to measure BAs [4]. Although these techniques provide detailed thermodynamic and structural insights, they are often expensive, time-consuming, and low-throughput, limiting their scalability in modern drug discovery [5].

In recent years, machine learning (ML) has emerged as a promising solution to this challenge and has shown promise by learning complex patterns from large datasets [6,7]. One of the first ML models was developed by Ballester et al. [8]. These ML models range from traditional regressors to advanced deep learning frameworks, including graph neural networks and transformers. They often incorporate multimodal inputs such as protein sequences, three-dimensional (3D) structures, and molecular properties—to enhance predictive accuracy [9]. Despite significant progress, several challenges of ML in the prediction of protein-ligand interactions persist. Key obstacles include the limited quality and diversity of training data, the lack of interpretability in model predictions, and the poor generalizability to novel proteins or ligands—all of which continue to impede the full integration of ML–based approaches into real-world drug development pipelines [10].

This review summarizes recent advances in ML-based BA prediction. Sections Machine learning in structure-based binding affinity prediction and Machine learning in ligand-based binding affinity prediction discuss ML approaches based on protein and ligand features, using case studies to highlight model effectiveness. Section Mathematical AI approaches for protein-ligand binding affinity prediction reviews mathematical artificial intelligence (AI) applied to protein-ligand binding affinities prediction. Section Machine learning in other directions of binding affinity prediction extends the discussion to broader applications like lead optimization, ADMET (absorption, distribution, metabolism, excretion, and toxicity) prediction, mutation analysis, and multi-target modeling. Section Free energy perturbation method for binding affinity prediction addresses current limitations in data quality, generalizability, interpretability, and computational cost. Finally, Section Challenges and limitations outlines future directions, including explainable AI, transfer learning, foundation models, and quantum computing, etc. By reviewing these developments, we aim to provide a forward-looking perspective that supports continued innovation in BA prediction and its integration into drug discovery workflows.