ADOPT

ADOPT predicts intrinsic protein disorder using a self-supervised deep bidirectional transformer combined with a supervised classifier to identify ordered and disordered residues from amino acid sequences for structural and functional analysis.

Key Features:

• Deep Bidirectional Transformer (ESM-derived): Employs a self-supervised encoder based on a deep bidirectional transformer derived from Facebook's Evolutionary Scale Modeling library to extract rich sequence representations.
• Self-Supervised and Supervised Learning: Integrates self-supervised representation learning with a supervised disorder prediction model to improve generalization and predictive accuracy.
• Residue-level Dense Representations: Produces dense, residue-level embeddings that capture complex sequence patterns and evolutionary/structural signals.
• Balanced NMR-derived Training Dataset: Trains the supervised predictor on a carefully constructed dataset derived from nuclear magnetic resonance (NMR) chemical shifts with balanced representation of disordered and ordered residues.
• Performance and Feature Efficiency: Reports superior accuracy and fast processing compared to existing predictors and achieves high accuracy using fewer than 100 relevant features.

Scientific Applications:

• Drug Discovery and Development: Predicting intrinsic disorder assists investigation of proteins and intrinsically disordered proteins (IDPs) implicated in disease mechanisms relevant to therapeutic targeting.
• Protein Function Annotation: Identification of disordered regions supports annotation of functional roles for sequence regions lacking fixed tertiary structure.
• Structural Biology Research: Disorder predictions inform studies of protein dynamics, interactions, and regions unsuitable for stable tertiary structure determination.

Methodology:

Uses a self-supervised encoder employing a deep bidirectional transformer to learn dense representations of amino acid sequences without explicit labels; a supervised disorder predictor is trained on a balanced dataset derived from NMR chemical shifts to classify residues as disordered or ordered based on the learned features.