Protein–ligand interactions (PLIs) play a crucial role in drug discovery, and accurately predicting protein–ligand binding affinity (PLA) remains a central challenge in computer-aided drug design. Although graph neural networks (GNNs) have demonstrated considerable potential in molecular modeling, existing methods still face several limitations, including excessive reliance on hand-crafted chemical features, loss of spatial information, and difficulties in integrating heterogeneous affinity labels, which restrict their generalization capability in PLA prediction. To address these challenges, we propose GAMT-GINE, a graph isomorphism network that integrates continuous spatial awareness with multi-task learning. The model employs minimalist atomic features and a batch-normalization-free mechanism, together with a multi-task branch that uses a large amount of half-maximal inhibitory concentration (IC₅₀) data as an auxiliary prediction target. Experimental results show that GAMT-GINE achieves a Pearson’s correlation coefficient (R_p) of 0.791 and a root mean square error (RMSE) of 1.403 on the CASF-2013 benchmark dataset. In the generalization evaluation on CASF-2016, R_p further increases to 0.831, while RMSE decreases to 1.227, demonstrating performance comparable to that of current State-of-the-Art models. Furthermore, comprehensive evaluations, including ablation studies, feature importance analysis, analysis of the effects of data filtering on model performance and data composition, and analysis of the influence of training–test data similarity on prediction results, indicate that GAMT-GINE can effectively utilize continuous spatial information and heterogeneous affinity labels, achieving good predictive accuracy and cross-dataset generalization capability.

In practical beef cattle monitoring scenarios, automated body condition scoring (BCS) based on single-frame visual information can be affected by posture variation, occlusion, and image quality changes. This study proposed an episode-level BCS assessment framework using synchronized RGB-D videos, integrating multi-frame visual information, attention-based multiple instance learning (Attention-MIL), and a dorsal geometry-based quality control mechanism to improve reliability under complex acquisition conditions. The publicly available Raramuri Criollo dataset was reorganized into a dataset containing 3815 RGB-D frame pairs, 52 animal identity groups, and 151 episodes, and evaluated using animal-wise three-fold cross-validation. Predictions were generated only when an episode contained at least three quality-filtered frames and reliable angle estimation. With an average coverage of 0.7345 after quality gating, the final model achieved 61.66% accuracy, a Macro-F1 score of 0.5680, and a quadratic weighted kappa (QWK) of 0.6339 for five-class BCS classification. Additionally, 89.71% of predictions were within one BCS category of the reference scores. Three-class management classification achieved a Macro-F1 score of 0.8059. Ablation experiments showed complementary contributions from RGB and depth information, and the combined angle and quality-statistic features improved several metrics in the original ablation. Additional matched experiments did not establish an independent angle benefit but supported multi-frame aggregation and within-Episode prediction consistency. The proposed RGB-D multi-frame fusion approach demonstrates potential for automated beef cattle body condition monitoring under controlled acquisition conditions.