Systematic evaluation of data preprocessing and model selection strategies for reliable pIC 50 prediction of acetylcholinesterase inhibitors

Predicting acetylcholinesterase (AChE) inhibitory activity is important in drug discovery. This study evaluates molecular descriptor - based machine learning models to predict AChE activity as pIC50 values. The primary objective was to comparatively investigate the impact of different data preprocessing strategies on prediction performance and model selection under challenging chemical datasets exhibiting low correlation structures. Tree based gradient boosting algorithms, namely CatBoost and XGBoost, together with sensitive regression models including Support Vector Regression and Multilayer Perceptron, were examined, and model specific data preparation pipelines were applied according to their structural assumptions. The target variable was stabilized through logarithmic transformation and winsorization of IC50 values. Model performance was assessed using both a 70-15-15 train-validation-test split and a 10-fold cross validation protocol. Furthermore, stacking based ensemble learning strategies were explored to enhance generalization capability. The results demonstrate that predictive performance is predominantly constrained by intrinsic dataset characteristics rather than algorithmic selection. Optimized tree-based models achieved the highest accuracy, while stacking provided only marginal improvements over the best individual learners. To improve interpretability, SHAP based explainable artificial intelligence analysis was conducted, highlighting the contributions of biologically meaningful molecular descriptors, and offers guidance for future studies addressing comparable biochemical modelling challenges.