B Shuriya¹ , TN Prabhu², S Kowsalya¹, A Suganya³

¹Department of CSE, United Institute of Technology,; ²Department of Computer Science and Business Systems, Dr. N.G.P. Institute of Technology; ³Department of CSE (AI ML), Sri Eshwar College of Engineering, Coimbatore, India.

For correspondence:-  B Shuriya   Email: shuriyasmile@gmail.com

Received: 3 September 2025        Accepted: 19 March 2026        Published: 30 March 2026

Citation: Shuriya B, Prabhu T, Kowsalya S, Suganya A. Integrating graph neural networks, molecular descriptors, and Pharmacophore features for robust LogK prediction. Trop J Pharm Res 2026; 25(3):329-338 doi: https://dx.doi.org/10.4314/tjpr.v25i3.5

© 2026 The authors.
This is an Open Access article that uses a funding model which does not charge readers or their institutions for access and distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0) and the Budapest Open Access Initiative (http://www.budapestopenaccessinitiative.org/read), which permit unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited..

Abstract

Purpose: An interpretation molecular activity-prediction framework was studied based on an ensemble of 3D graph neural networks with hybrid molecular features. Method: Molecular graphs and 3D conformers were generated from a dataset of 15,000 molecules with LogK values, along with 2D descriptors and Pharmacophore features. Random Forest, Gradient Boosting, Support Vector Machine Learning, Deep Neural Networks, and 3D-Graphical Neural Network ensemble models were trained using Bayesian optimized hyperparameters. Results: The ensemble beat all baselines with an ROC- AUC of 0.94 and F1-score of 0.85 and scaffold-split validation demonstrated its generalization to new chemical scaffolds (ROC- AUC = 0.91, Balanced Accuracy = 0.88, F1 = 0.84). On the positive side, hydrogen bond acceptor, aromatic rings, and moderate hydrophilic were predictive features, whereas hysteric bulk and high logP were negative predictors of activity according to SHapley Additive exPlanations and attention analyses. Conclusion: This framework provides a promising computational approach for virtual screening and rational drug design.

Keywords: Support Vector, Machine Learning, Deep Neural Network, 3D- Graphical, Ensemble model, Drug discovery

Impact Factor
Thompson Reuters (ISI): 0.6 (2023)
H-5 index (Google Scholar): 49 (2023)