Prediksi Aktivitas Antioksidan Senyawa Bioaktif Menggunakan Pendekatan Machine Learning Berbasis Deskriptor Molekul
Purpose: This study builds machine learning models that predict the DPPH radical-scavenging activity of phenolic acids from simple, mechanistically designed molecular descriptors, and examines which structural features drive the activity. Design/Methodology: Activity data (pTEAC) for 21 phenolic acids measured under a single protocol were taken from Jing et al. (2012), keeping the original split of 15 training and 6 test compounds. Fifteen two-dimensional descriptors were computed, including the number and position of hydroxyl and methoxy groups, electron-donating groups ortho/para to the phenolic OH, conjugation, TPSA, and Hammett constants. Multiple linear regression, partial least squares, k-nearest neighbours, support vector regression, and random forest were compared using leave-one-out cross-validation, an external test set, 100-fold Y-randomization, permutation importance, and a leverage-based applicability domain. Findings: Random forest performed best (Q²LOO = 0.892; external R²pred = 0.767) and passed Y-randomization, while k-NN and SVR overfitted. The number of electron-donating groups ortho/para to the phenolic OH and TPSA dominated the model. A two-descriptor linear equation was interpretable but statistically weak (Q²LOO = 0.409). The 2D models remained below published 3D CoMFA/CoMSIA models (R²pred ≈ 0.98); most of the test error came from one compound flagged as outside the applicability domain. Practical Implications: The workflow can be used to rank phenolic compounds from natural extracts before laboratory testing, provided that predictions are restricted to the applicability domain. Originality/Value: The study combines transparent mechanism-based descriptors with a complete OECD-oriented validation on real experimental data and reports honestly where the 2D approach falls short of 3D models.
Authors
- Qomariah Hasanah (ORCID: https://orcid.org/0000-0003-3176-7143)
Institutions
- Universitas Islam Negeri Fatmawati Sukarno Bengkulu (ID)
Publication Details
- Published
- 2026-11-15
- DOI
- https://doi.org/10.67319/kuadran.v1i1.4
- Primary Topic
- Computational Drug Discovery Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00