ChemSol AI: A Deep Learning Approach to Predict Molecular Solubility

The ability to predict molecular solubility is an important factor in drug development. Solubility dictates how readily a drug compound will be absorbed into the body and ultimately, how effective that compound will be as a therapeutic agent. Conventional feature-based approaches to solubility prediction have been limited by their inability to adequately model the complexities of a molecule’s structure or capture the nonlinear interactions occurring within those structures. A new framework for solubility prediction utilizing deep learning techniques (specifically transformers and graph neural networks) has been designed to generate more complex and nuanced molecule representations. A new framework was trained on carefully curated dataset consisting of verified solubility values combined with extensive structural information about each molecule. Using common regression and classification criteria, six predictive models Random Forest, XGBoost, Gradient Boosting, ChemBERTa, Graph Transformer and GAT were compared. With an MAE of 1.40 and R2 of 0.69, Graph Transformer demonstrated the best regression performance, whereas ChemBERTa demonstrated a balanced performance with 80.6% accuracy in solubility categorization. This method facilitates quicker and more effective early-stage screening of possible drug candidates by providing more precise and trustworthy solubility estimations. The results demonstrate the increasing influence of sophisticated deep learning methods in computational chemistry and their capacity to greatly speed up data-driven pharmaceutical development.