OPUS-TASS

OPUS-TASS predicts protein backbone torsion angles (ϕ and ψ) and secondary structure from amino acid sequences using deep learning to support accurate protein structural modeling.

Key Features:

• Modified Transformer architecture: Uses a modified Transformer to capture long-range interactions between residues regardless of sequence distance.
• Deep learning models: Employs advanced deep neural networks to improve prediction accuracy for torsion angles and secondary structure.
• Multitask learning: Trains on related tasks simultaneously to enhance generalization for torsion-angle and secondary-structure prediction.
• Ensemble models: Combines outputs from multiple models to reduce errors and increase robustness, analogous to approaches used by SPOT-1D.
• Post-processing refinement (OPUS-Refine): Applies OPUS-Refine to further reduce mean absolute errors for ϕ and ψ predictions.
• Benchmark performance on CAMEO93: Reports mean absolute errors of 16.56 and 22.56 for ϕ and ψ versus SPOT-1D's 16.89 and 23.02, and secondary-structure accuracies of 89.06% (3-state) and 78.87% (8-state) versus 87.72% and 77.15%; OPUS-Refine reduces MAEs to 16.28 and 21.98 for ϕ and ψ.

Scientific Applications:

• Protein structure prediction: Provides torsion-angle and secondary-structure inputs for modeling protein tertiary structure from sequence data.
• Drug discovery and design: Supplies structural predictions that facilitate identification of drug targets and design of interacting molecules.
• Functional annotation: Supports annotation of unknown proteins by enabling structural comparisons to infer function based on predicted secondary structure and torsion angles.

Methodology:

Trained and validated on comprehensive datasets and evaluated against benchmark datasets TEST2016, TEST2018, CASP12, CASP13, CASP-FM, and CAMEO93.