Avaliação de modelos de IA para prever a solubilidade de medicamentos em sistemas de co-solventes

The low solubility of drugs presents a significant challenge in pharmaceutical development. This study evaluates the ability of five artificial intelligence models — Artificial Neural Network (ANN), Random Forest, K-Nearest Neighbors (KNN), Linear Regression, and XGBoost — to predict drug solubility in cosolvent systems. The study uses a dataset that integrates the physicochemical properties of the solute and solvent. The objective was to identify the most accurate model and analyze the relative importance of the predictor variables. The results demonstrate a clear hierarchy in model performance. Tree-based algorithms XGBoost and Random Forest showed exceptional predictive power with coefficients of determination (R²) of 0.9737 and 0.9752, respectively, and the lowest mean squared errors. The ANN model also demonstrated robust performance (R² = 0.9585), whereas the linear regression model was unable to capture the nonlinear relationships inherent in the system (R² = 0.4463).  Feature importance analysis revealed significant differences: tree models prioritized the drug's molar mass, while ANN attributed considerable importance to the cosolvent system's properties, such as the solvents' molar mass and volume. These differences highlight the complexity of solubility as a multifactorial phenomenon. We conclude that ensemble models, particularly XGBoost, are highly effective and accurate tools for in silico solubility prediction. Additionally, we find that a comparative modeling approach is crucial for gaining a holistic understanding of the physicochemical interactions that govern the process.