NetSurfP-2.0: Improved prediction of protein structural features by integrated deep learning.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Predicting protein local structure from primary sequences is crucial for understanding protein function without experimental data.
• Existing prediction tools face challenges with accuracy and runtime, limiting their utility with rapidly increasing sequencing data.
• Accurate and efficient prediction of local structural features is essential for bioinformatics and structural biology.

Purpose of the Study:

• To introduce NetSurfP-2.0, a novel computational tool for predicting local protein structural features.
• To achieve unprecedented accuracy and runtime in predicting solvent accessibility, secondary structure, disorder, and dihedral angles.
• To provide a reliable and fast method for structural and functional annotation of proteins.

Main Methods:

• Developed NetSurfP-2.0, a sequence-based tool utilizing convolutional and long short-term memory neural networks.
• Trained the model on experimentally solved protein structures.
• Integrated a single model to predict multiple local structural features simultaneously.

Main Results:

• NetSurfP-2.0 demonstrates state-of-the-art prediction accuracy across all output features on independent test datasets.
• Achieved 80% correlation for solvent accessibility and 85% precision for 3-class secondary structure predictions.
• Optimized processing time allows prediction of over 1000 proteins in under 2 hours and complete proteomes in under 1 day.

Conclusions:

• NetSurfP-2.0 offers a significant advancement in predicting local protein structural features with high accuracy and efficiency.
• The tool's speed and precision facilitate rapid analysis of large-scale sequencing data.
• NetSurfP-2.0 is a valuable resource for unraveling protein function and structure in computational biology.