Traditional machine learning (ML) and deep learning (DL) play a key role in the prediction of the selectivity of small molecule inhibitors. Many models based on existing datasets are available for predicting the bioactivity of compounds. However, there remains controversy regarding the relative performance of ML and DL for such predictions. In this study, ten metaheuristic algorithms are applied to optimize the hyperparameters of eleven ML and DL models based on different molecular characterizations, aiming at systematically comparing the model prediction performance and identifying the optimal models. The results demonstrate that the ML and DL models optimized using metaheuristic hyperparameter optimization algorithms significantly outperform traditional grid search models in terms of predictive performance. Additionally, in low-dimensional feature spaces, molecular graph-based DL models, such as SSA-GAT and SSA-Attentive FP, are capable of autonomously extracting relevant features from data through an end-to-end learning mechanism, which outperforms the ML models. Conversely, in high-dimensional feature spaces (e.g., RDKit+ECFP+AtomPairs+Maccs-XGBoost), ML methods leverage the complementarity of molecular features along with the high-order optimization capabilities of metaheuristic algorithms to effectively capture intricate feature interactions, often leading to higher accuracy and robustness in high-dimensional modeling. These findings provide valuable information to guide the selection of ML and DL methods for activity prediction of target inhibitors.