QSPR modeling to predict the Partition Coefficient (logP) of psychoanaleptic drugs using ARKA descriptors

Abstract

A Quantitative Structure Property Relationship (QSPR) model was developed for predicting the partition coefficient (logP) values of 121 psychoanaleptic drugs using four machine learning algorithms: Random Forest (RF), XGBoost Regressor (XGBR), Support Vector Regression (SVR), and a Dragonfly Algorithm combined with the Support Vector Regressor (DA-SVR). Ten pertinent molecular descriptors were selected using the genetic algorithm (GA) within the AlvaModel software and used as input features to build the model. Subsequently, these descriptors were transformed into ARKA descriptors to achieve dimensionality reduction, particularly beneficial for small datasets, and to test the data's modelability. Both AlvaDesc descriptors and ARKA descriptors were used as input features. The combination of ARKA descriptors with the DA SVR algorithm produced the best-performing model, achieving R² = 0.971 and RMSE = 0.311, thereby demonstrating robust predictive capability. Benchmarking against the RDKit Crippen logP predictor further confirmed the superiority of the proposed approach, with test set results of R² = 0.82 and RMSE = 0.58 compared to R² = 0.72 and RMSE = 0.72 for RDKit. This result highlights the effectiveness of ARKA descriptors in improving model performance and interpretability for predicting logP values.