Solvation Meta Predictor

Abstract

Predicting the solubility of aqueous mixtures is a critical task in cheminformatics, impacting fields such as drug discovery, chemical engineering, and environmental science. This study aims to enhance the predictive accuracy of machine learning models for solubility by employing advanced ensemble techniques. We evaluated the performance of three individual models: SMI, MDM, and GNN, and compared them to ensemble methods including simple averaging and an Optuna-optimized ensemble. Our results indicate that the Optuna-optimized ensemble model achieved the highest predictive accuracy, with an R2 value of 0.8117, outperforming individual models and simple ensemble techniques. To further improve model performance, we propose the implementation of the Mixture of Experts (MoE) approach. This advanced ensemble technique leverages specialized experts and a gating network to optimize model predictions based on input features. MoE promises to enhance model flexibility, scalability, and specialization, making it a robust tool for handling complex and heterogeneous datasets. Future work will involve integrating additional models and exploring other ensemble strategies to further improve predictive accuracy. The findings of this study highlight the potential of ensemble methods to significantly improve the prediction of solubility in aqueous mixtures, offering valuable insights for various scientific and industrial applications.