A Fusion-based Sequence-Structure Protein Encoder of α-synuclein under Structural Uncertainty

Parkinson’s disease (PD) is the second most common neurodegenerative disorder, and it is classified as a movement disorder. When neurons in the brain that produce dopamine slowly die off, irregular brain activity and motor symptoms are formed. The primary contributor to this is the α-synuclein protein, which plays a crucial role in neuroscience through its misfolding and clumping, leading to the development of PD. Due to intrinsically disordered and structural heterogeneity, it presents a unique challenge for computational modelling and protein representation learning. This research work presents an AI-driven, disorder-aware protein feature extraction framework that combines sequence-based contextual embeddings with confidence-filtered 3D structural representations to generate a unified, biologically informed encoding of α-synuclein. Sequence-only representations lack explicit spatial and conformational constraints, while structure-only representations are sensitive and perform poorly in intrinsically disordered regions when used independently. To address these limitations, a fusion strategy that combines residue-level sequence embeddings with a structure-aware graph representation derived from AlphaFold-predicted coordinates, which merges per-residue confidence (pLDTT) to selectively down-weight unreliable structural regions, is introduced. The experimental analysis demonstrates that fused representation is more stable, compact, and informative than single-modality embeddings. This study establishes a principled feature extraction paradigm for α-synuclein and other intrinsically disordered proteins, providing a foundation for AI-driven target modelling and other predictive tasks in PD.