Integrating graph attention networks and deep learning with multi-stage computational screening: Bisdemethoxycurcumin...

Parkinson's disease (PD) remains a progressive neurodegenerative disorder with limited disease modifying therapies, as current treatments such as L-DOPA offer only symptomatic relief and are associated with severe adverse effects. This study presents an integrated multi-stage computational pipeline that combines quantitative structure activity relationship (QSAR) modelling enhanced by graph attention networks (GATs), molecular docking, density functional theory (DFT), molecular dynamics (MD) simulations, and ADMET profiling to identify natural inhibitors targeting four key PD related proteins: adenosine A₂A receptor (A₂AR), gamma aminobutyric acid type A receptor (GABA-A), nuclear factor erythroid 2- related factor 2 (Nrf2), and (α-synuclein). A comprehensive dataset of 1393 structurally diverse compounds derived from 6- hydroxybenzothiazole 2-carboxamide and related scaffolds was assembled and used to train and validate seven regression algorithms. GATs achieved superior predictive performance (R² = 0.991, Q² = 0.987, RMSE = 0.487), outperforming Extra Trees (R² = 0.742), Random Forest (R² = 0.761), and SVR-RBF (R² = 0.65). The large dataset size enhanced model generalizability, reduced overfitting risk, and enabled robust feature selection from an expanded descriptor pool of 312 molecular descriptors. Y-randomization with 100 permutations confirmed that all models with meaningful predictive performance were free from chance correlation (ΔQ² > 0.50), validating the robustness of the QSAR workflow. Molecular docking of six phytochemicals derived from Curcuma longa (ar-turmerone, α -turmerone, bisdemethoxycurcumin) and Salvia officinalis (camphor, 1,8-cineole, α, β-thujone) revealed that bisdemethoxycurcumin BDMC exhibited the strongest binding affinity to the A₂A receptor (-9.1 kcal/mol) among the evaluated phytochemicals. L-DOPA, the current clinical standard of care, displayed weak affinity to the A₂A receptor (-7.2 kcal/mol), consistent with its known mechanism as a dopamine precursor rather than an A₂A ligand. DFT calculations at the B3LYP/6311G(d,p) level showed that BDMC possessed the lowest HOMO- LUMO energy gap (0.1428 eV) and the highest electrophilicity index (ω = 0.1477), indicating superior chemical reactivity and electron accepting capacity. A 200 ns MD simulation confirmed stable BDMC A₂A receptor binding, with 3-8 (mean ≈ 5) persistent hydrogen bonds and no dissociation events throughout the trajectory, as evidenced by centre of mass distance analysis and principal component analysis. ADMET profiling using SwissADME and ProTox II classified BDMC as blood brain barrier (BBB) permeable, compliant with Lipinski's Rule of Five, and safe (LD₅₀ = 2560 mg/kg, toxicity Class 5), compared to L-DOPA (Class 4, LD₅₀ = 1460 mg/kg). Comparative analysis demonstrated that BDMC offers superior predicted binding affinity to the A₂A receptor and a more favourable predicted toxicity profile compared to L-DOPA, positioning it as a compelling A₂A receptor-targeting lead candidate for PD therapy. Collectively, these findings establish a robust computational framework for accelerating drug discovery in neurodegenerative diseases and highlight BDMC as a promising A₂A Receptor inhibitor warranting further in-vitro and in- vivo experimental validation.