This paper reviews mainstream structure-based deep-learning approaches for predicting protein-ligand binding affinity (PLBAP), focusing on molecular representations, learning architectures, and model interpretability. The authors generate a model taxonomy and evaluate representatives from a uniform basis to address the lack of valid comparisons among these models. The review highlights the strengths and weaknesses of each approach, aiming to benefit structure-based drug discovery. Key models discussed include Atomic Convolutional Neural Networks (ACNNs), Intermolecular Contact Profile-CNNs (IMC-CNNs), Molecular Grid-CNNs (Grid-CNNs), and Graph Convolutional Networks (Graph-GCNs). The evaluation of scoring performances and screening powers of these models is conducted using the PDBbind Refined Set and CSAR-HiQ datasets, as well as the DUD-E database. The results show that Graph-GCN models are the most promising for PLBAP tasks due to their balance between prediction accuracy and computational efficiency. The paper also discusses the interpretability of these models, providing insights into their structural and physicochemical mechanisms. Overall, the review underscores the potential of deep-learning models in PLBAP and suggests directions for future research.This paper reviews mainstream structure-based deep-learning approaches for predicting protein-ligand binding affinity (PLBAP), focusing on molecular representations, learning architectures, and model interpretability. The authors generate a model taxonomy and evaluate representatives from a uniform basis to address the lack of valid comparisons among these models. The review highlights the strengths and weaknesses of each approach, aiming to benefit structure-based drug discovery. Key models discussed include Atomic Convolutional Neural Networks (ACNNs), Intermolecular Contact Profile-CNNs (IMC-CNNs), Molecular Grid-CNNs (Grid-CNNs), and Graph Convolutional Networks (Graph-GCNs). The evaluation of scoring performances and screening powers of these models is conducted using the PDBbind Refined Set and CSAR-HiQ datasets, as well as the DUD-E database. The results show that Graph-GCN models are the most promising for PLBAP tasks due to their balance between prediction accuracy and computational efficiency. The paper also discusses the interpretability of these models, providing insights into their structural and physicochemical mechanisms. Overall, the review underscores the potential of deep-learning models in PLBAP and suggests directions for future research.