Abstract
The success of AI-based pipelines in Drug Discovery (DD) heavily depends on three components: the dataset (size, content, imbalance ratio), the encoding system and the predictive model. This study focuses on optimizing these elements to develop robust predictive models for the biological activity of chemical molecules. We trained five machine learning and six deep learning algorithms to predict the anti-pathogen activity of chemical compounds, derived from PubChem Bioassays targeting four infectious diseases. The datasets exhibited a significant imbalance towards the inactive class. To address this, we implemented a K-ratio random undersampling approach (K-RUS) to determine optimal imbalance ratios (IRs), and compared them to conventional resampling approaches. Across all simulations, a moderate IR (1:10) significantly enhanced models' performance. Through external validation, we assessed the generalization power of the top-performing models and observed optimal balance between true positive and false positive rates with the 10-RUS configuration. Furthermore, we investigated the chemical similarity between active and inactive classes, which revealed the underlying mechanisms of misclassification. Our findings emphasized the need for optimizing imbalance ratios and leveraging chemical diversity to improve AI-driven DD against infectious diseases.