XP-GCN: Extreme learning machines and parallel graph convolutional networks for high-throughput prediction of blood-brain barrier penetration based on feature fusion.

Accurate prediction of molecular properties plays a crucial role in drug analysis and discovery. Especially for neurotherapeutic drugs developed for the treatment of neurological disorders, one of today's most critical challenges, this process becomes even more crucial. The prediction of the blood-brain barrier penetration (BBBP) ability of the molecules contained in the development process of a drug is the primary condition for knowing the effect of drugs on the central nervous system. In this study, we developed an innovative deep-learning architecture for predicting BBBP using well-characterised molecular properties. The new method, called XP-GCN (Extreme Parallel Graph Convolutional Network), combines Graph Convolutional Networks (GCN) and Extreme Learning Machines (ELM) to predict molecular properties. It achieves multidimensional feature fusion for BBBP prediction without relying on back-propagation algorithms. XP-GCN offers a comprehensive approach by combining graphical representations and molecular fingerprints while evaluating molecules through multidimensional analysis of SMILES strings, enhanced with data augmentation techniques. The model analyses various attributes of molecules through parallel GCN layers. Trained on a notably extensive molecular dataset compared to many literature studies, it has achieved high performance and computational efficiency by fusing information obtained from molecular fingerprints. Due to ELM's forward random weight and bias assignment features, the classification task without backpropagation-based training showed a classification performance of over 95 % in less than one millisecond, even in models with a low number of hidden neurons. XP-GCN, which shows a peak ROC-AUC performance of 0.9846 across different hyperparameters, offers a novel approach to molecular feature prediction with its computational efficiency, high accuracy, and classification capability. Furthermore, the robustness of the XP-GCN model has been demonstrated by Friedman and Nemenyi's post-hoc statistical analyses.