Discovery of Potential Tyrosinase Inhibitors via Machine Learning and Molecular Docking with Experimental Validation of Activity and Skin Permeation

Abstract

Tyrosinase, a copper-dependent oxidase, plays a critical role in melanin biosynthesis and is a target in skin-whitening cosmetics. Conventional inhibitors like arbutin and kojic acid are widely used but suffer from cytotoxicity, instability, and inconsistent efficacy, highlighting the need for safer, more effective alternatives. In this study, two ligand-based machine learning models were developed: one to predict the biological activity of compounds and the other to estimate specific pIC₅₀ values. These models were employed to screen potential tyrosinase inhibitors from natural product libraries and FDA-approved drug databases. Subsequently, the molecules identified through machine learning screening were subjected to more precise multi tyrosinase-like structures molecular docking to refine the selection. We identified three top-ranking inhibitors, rhodanine-3-propionic acid, lodoxamide, and cytidine 5'-(dihydrogen phosphate), with strong binding affinities mediated by metal ion coordination and π-π interactions at the enzyme's active site. In vitro assays revealed that all three compounds exhibited higher inhibitory activity against mushroom tyrosinase compared to arbutin (IC₅₀ = 38.37 mM), with rhodanine-3-propionic acid displaying the most potent inhibition (IC₅₀ = 0.7349 mM). Furthermore, transdermal permeation experimental results confirmed that these compounds achieved markedly better skin permeability than commercial arbutin-based formulations, highlighting their potential as next-generation agents for inhibiting melanin production in cosmetic applications.

1

Stepwise workflow for screening tyrosinase inhibitors. The process includes database compilation, ML-based activity prediction, molecular docking, and in vitro validation, identifying three promising tyrosinase inhibitors.

2

Evaluation of ML models for tyrosinase inhibitor prediction. (A) Accuracy of activity classification models for predicting tyrosinase inhibitory activity. (B) R ² performance of regression models for pIC₅₀ numeric prediction. All results are reported as mean ± standard deviation from 10-fold cross-validation.

3

Retrospective enrichment analysis of tyrosinase-related conformations for virtual screening. (A) ROC enrichment curves for two AbTYR structures. (B) ROC enrichment curves for nine HsTYRP1 structures.

4

Binding interactions and docking scores of three candidate tyrosinase inhibitors across multiple receptor conformations. (A) Cytidine 5′-(dihydrogen phosphate), (B) lodoxamide, and (C) rhodanine-3-propionic acid docked into 2Y9X, 5M8O, and 5M8R. Key molecular interactions are annotated, including hydrogen bonds, salt bridges, and π-related interactions.

5

IVPT profiles of cytidine 5′-(dihydrogen phosphate), rhodanine-3-propionic acid, and lodoxamide through excised pig skin over 24 h. (A) Cumulative permeation over time. (B) Permeation flux over time. Data are expressed as mean ± standard error (n = 3).