Traditional machine learning (ML) and deep learning (DL) play a key role in predicting the activity of target inhibitors. Many models based on existing datasets can predict compound bioactivity. However, debate persists regarding whether ML or DL performs better for such prediction tasks. In this study, datasets were constructed based on different molecular representations. Ten metaheuristic algorithms were applied to optimize the hyperparameters of eleven ML and DL models, aiming to systematically compare their predictive performance and identify the optimal ones. The results show that ML and DL models whose hyperparameters were optimized by metaheuristic algorithms significantly outperformed those optimized using the traditional grid search method. Furthermore, in low-dimensional feature spaces, graph-based DL models, such as SSA-GAT and SSA-Attentive FP, can automatically extract informative features from data via an end-to-end learning mechanism, yielding better performance than ML models. In contrast, in high-dimensional feature spaces (e.g., the feature space formed by combining RDKit descriptors with ECFP, AtomPairs, and MACCS fingerprints), ML methods, leveraging the complementary information in molecular features and the powerful optimization capability of metaheuristic algorithms, can effectively capture complex feature interactions. Consequently, ML methods often demonstrate higher accuracy and robustness in high-dimensional modeling. These findings provide valuable guidance for selecting between ML and DL approaches for target inhibitor activity prediction.