<p>Aberrant activation of Receptor Tyrosine Kinases (RTKs) is a well-established trigger of tumorigenesis, and the overuse of RTK inhibitors often leads to drug resistance and tumor recurrence. While current Drug-Target Interaction (DTI) prediction methods (including those based on heterogeneous information networks) have shown promise, they remain limited in their ability to fully capture the nature of DTIs and often lack interpretability. To overcome these limitations, this study introduces a novel hybrid optimization model termed MDBO-RF, which integrates a Modified Dung Beetle Optimizer (MDBO) with Random Forest (RF). The key innovation lies in the enhancement of the DBO algorithm through a quaternion-based learning mechanism and the Cauchy mutation strategy, specifically designed to overcome the slow convergence and susceptibility to local optima that plague traditional metaheuristic algorithms used for hyperparameter tuning. The model leverages commonly used molecular descriptors to enhance the prediction of Tyrosine Kinase (TK) inhibitory activity and enable efficient compound screening. Our results demonstrate that MDBO-RF achieves a 3.41% increase in prediction accuracy compared to the standard RF model and outperforms several other contemporary machine learning approaches. The model effectively streamlines the RTK inhibitor screening process by improving prediction accuracy in multi-target competitive binding scenarios and reducing false-positive screening due to off-target effects. This work underscores the value of hybrid optimization strategies in bioinformatics and provides a robust, interpretable tool for accelerating drug discovery.</p>