A Hybrid ChemBERTa–Morgan XGBoost Ensemble for Computational Multi-Endpoint Toxicity Prioritization of Blumea balsamifera Compounds
Authors:
- Kusnaeni KusnaeniKKKusnaeni Kusnaeni
- Agus Mohamad SolehASAgus Mohamad Soleh
- Bagus SartonoBSBagus Sartono
- Taufik Adhi Prasetya WardanaTWTaufik Adhi Prasetya Wardana
- Fitriagustiani FitriagustianiFFFitriagustiani Fitriagustiani
- Muhammad Fadhil NurahmadMNMuhammad Fadhil Nurahmad
- Muhammad RahmatullahMRMuhammad Rahmatullah
- Muhammad Rizqy FirmansyahMFMuhammad Rizqy Firmansyah
- Afif Budi Andy BABAfif Budi Andy B
Preprint
- 17 August 2026
- preprint
- Published by Elsevier in SSRN Electronic Journal
Abstract
Biological investigations of Blumea balsamifera remain dominated by extract or fraction-based testing; consequently, the activities of individual compounds across various biological mechanisms have not been consistently characterized. This study developed a computational framework to prioritize 31 Blumea balsamifera compounds based on their predicted activities across 12 Tox21 endpoints. The framework integrates Morgan fingerprints, RDKit descriptors, and ChemBERTa embeddings through two endpoint-specific XGBoost branches combined using an equal-weight soft voting ensemble of calibrated probabilities. The models were developed using 7,820 Tox21 structures after three structures identical to the target compounds were excluded. Evaluation was performed using repeated five-fold scaffold cross-validation, Platt scaling, endpoint-specific thresholds, and applicability domain analysis. The ensemble achieved mean ROC-AUC, average precision, balanced accuracy, and MCC values of 0.784±0.063, 0.372±0.130, 0.712±0.052, and 0.273±0.103, respectively. Luteolin-7-methyl-ether ranked highest, with a score of 37.10 and predicted activity at 11 endpoints. This framework supports the selection of compounds for more targeted biological validation.• Develops a hybrid framework for computational prioritization of toxicity-related activity across 12 Tox21 endpoints.• Integrates Morgan fingerprints, RDKit descriptors, and ChemBERTa embeddings into a calibrated XGBoost ensemble.• Identifies luteolin-7-methyl ether as the highest priority candidate for multi-endpoint biological validation.Keywords
This publication has 13 references indexed in Scilit:
- MolPROP: Molecular Property prediction with multimodal language and graph fusionJournal of Cheminformatics, 2024
- Exploring the effectiveness of flavone derivatives for treating liver diseases: Utilizing DFT, molecular docking, and molecular dynamics techniquesMethodsX, 2023
- A systematic study of key elements underlying molecular property predictionNature Communications, 2023
- Chemical representation learning for toxicity predictionDigital Discovery, 2023
- Chemical constituents and bioactivities of Blumea balsamifera (Sembung): a systematic reviewFood Science and Technology, 2023
- Chemical constituents of Blumea balsamiferaPhytochemistry Letters, 2021
- Antibacterial effect of Blumea balsamifera (L.) DC. essential oil against Staphylococcus aureusArchives of Microbiology, 2021
- Phytochemical composition and health properties of Sembung plant (Blumea balsamifera): A reviewVeterinary World, 2021
- Why is Tanimoto index an appropriate choice for fingerprint-based similarity calculations?Journal of Cheminformatics, 2015
- Extended-Connectivity FingerprintsJournal of Chemical Information and Modeling, 2010