IDP-Seq2Seq

IDP-Seq2Seq predicts intrinsically disordered regions (IDRs) in protein sequences by applying sequence-to-sequence learning to map sequences into a semantic space that reflects structural patterns.

Key Features:

• Sequence-to-sequence learning (Seq2Seq): Maps protein sequences into a semantic space to capture structural patterns beyond linear sequence information.
• Predicted contacts (CCMs) and sequence-based features: Integrates predicted residue-residue contacts (CCMs) and other sequence-derived features to enhance prediction accuracy.
• Attention mechanism: Employs an Attention mechanism to capture global associations between all residue pairs within proteins.
• Length-specific predictors: Provides three specialized models—IDP-Seq2Seq-L for long disordered regions, IDP-Seq2Seq-S for short disordered regions, and IDP-Seq2Seq-G for both long and short regions.
• Fusion strategy: Fuses the specialized predictors into a single comprehensive predictor to improve discriminative power and generalization.
• Benchmark evaluation: Validated on four independent test datasets and the CASP test dataset, showing robustness across varying ratios of long and short disordered regions.

Scientific Applications:

• IDR prediction: Prediction of intrinsically disordered regions (IDRs) in protein sequences across varied length scales.
• Protein structure–function analysis: Characterization of protein structural patterns beyond linear sequence to inform studies of protein structure and function.
• Benchmarking and comparative evaluation: Comparative assessment of IDR prediction performance on independent datasets and the CASP test dataset.

Methodology:

Sequence-to-sequence learning maps protein sequences into a semantic space using predicted residue-residue contacts (CCMs) and sequence-based features; an Attention mechanism captures global residue associations; three predictors (IDP-Seq2Seq-L, IDP-Seq2Seq-S, IDP-Seq2Seq-G) trained for different disorder lengths are fused into a single predictor.