(M0930-09-48) Data-Driven Prediction of Translational Formulation Strategies for Poorly Water-Soluble Drug Molecules: Machine Learning to Minimize Trial and Error in Early Formulation Development

Purpose: The development of an orally bioavailable formulation of Biopharmaceutics Classification System (BCS) class II drugs represents significant challenges. Poor solubility often results in poor absorption and therapeutic failure. There are various formulation approaches for enhancing the dissolution, but it is difficult to predict which formulation strategy will be successful for a given molecule. These complex considerations decide the process of determining optimal formulation strategies for novel drugs both time-consuming and resource-intensive, often necessitating extensive experimental studies. This study aims to address these challenges by developing and validating an innovative machine-learning approach for predicting optimal drug formulation technology. The research focuses on creating a model capable of predicting the most feasible formulation strategies for drug compounds based on their molecular descriptors and graph-based signatures. Molecular descriptors provide a numerical representation of chemical structures and properties, encompassing constitutional, topological, geometrical, and electronic characteristics. These descriptors offer a comprehensive profile of each drug compound, capturing the physicochemical properties that influence formulation behavior. Complementing these, graph-based signatures represent molecular structures as graphs, with atoms as nodes and bonds as edges. This approach encodes complex structural information, including specific substructures and overall molecular topology, crucial for understanding spatial arrangements and connectivity patterns that influence pharmaceutical properties. The combination of molecular descriptors and graph-based signatures provides a comprehensive computational representation of drug compounds. This rich feature set enables the machine learning model to discern subtle patterns and relationships between molecular properties and optimal formulation strategies. The overall goal is to minimize the trial and error in developing formulation at preclinical stages.

Methods: As an input for the machine learning model, the study evaluated six formulation strategies (Amorphous Solid Dispersion, Nanoparticles, Liposomes, Self-Emulsifying Drug Delivery Systems, Cyclodextrin complexes, and Solid Lipid Nanoparticles) for their feasibility with 50 drug molecules with diverse physicochemical characteristics. Each strategy was ranked from 1 to 6 for each drug, with 1 indicating the highest feasibility. Rankings were based on multiple factors, including solubility enhancement, permeability improvement, commercial viability, stability, drug loading capacity, and release profile characteristics. Molecular descriptors and graph-based signatures were calculated using various nodes on the KNIME analytical platform, including RDkit descriptor calculation. Data preprocessing involved removing low variance, constant, and linearly correlated variables (Pearson correlation). The dataset was then separated into test and training subsets. Various models, including random forest, gradient booster, decision tree, and SOTA, were tested and compared for accuracy. Further optimization of selected models employed feature selection loops such as random, backward selection, forward selection, and genetic algorithm. Model performance was evaluated using accuracy, F1-score, and confusion matrices.

Results: The initial analysis revealed ASD, nanosuspension, and SEDDS are the most frequently top-ranked strategies, suggesting their high feasibility for most drugs. The preprocessing phase reduced the number of variables from 119 to 74. Comparative analysis of various models identified random forest as the best-performing model, achieving the highest overall accuracy of 0.733, followed by the decision tree model with an accuracy of 0.6. Further optimization of the random forest model through feature selection loops yielded significant improvements. The genetic algorithm-based feature selection demonstrated the best performance, achieving an accuracy of 0.9 (Fig. 1). The ROC curve analysis (Fig. 2) provided insights into the model's performance across different formulation strategies. The model showed good discrimination ability for ASD and LIP formulations, with AUC values of 0.623 and 0.581, respectively. Performance for SEDDS was marginally better than random (AUC = 0.522). However, the model struggled with NP and CD classifications (AUC = 0.343 and 0.335, respectively), indicating potential misclassification or confusion with other classes. The majority of the drugs suitable for NP and CD are brick-dust molecules.

Conclusion: This study demonstrates the potential of machine learning approaches in predicting optimal formulation strategies for poorly water-soluble drugs. The developed model, particularly after optimization with genetic algorithm-based feature selection, shows promising accuracy in predicting the feasibility of various formulation strategies. These findings suggest that the integration of molecular descriptors and graph-based signatures can significantly enhance the efficiency of formulation strategy selection in early-stage drug development, potentially reducing time and resource requirements in pharmaceutical research and development processes.

Accuracy of various random forest models using feature selection loops. The genetic algorithm showed the best accuracy of 0.9.

ROC curve of the best model. The model's performance varies significantly across different formulation strategies. There's room for improvement, especially for NP and CD predictions. The model is the most reliable for predicting ASD and LIP formulations.