S-Pred: protein structural property prediction using MSA transformer.

Area of Science:

• Computational biology
• Protein structure prediction
• Bioinformatics

Background:

• Predicting protein local structural features from amino acid sequences is crucial for function prediction and 3D structural modeling.
• The increasing sequence-structure gap necessitates advanced prediction methods.
• Advances in structural databases and computing power have improved prediction method performance.

Purpose of the Study:

• To introduce S-Pred, a novel tool for predicting eight-state secondary structures (SS8), accessible surface areas (ASAs), and intrinsically disordered regions (IDRs) from protein sequences.
• To leverage advanced deep learning techniques for enhanced protein structure prediction.

Main Methods:

• S-Pred utilizes multiple sequence alignment (MSA) as input.
• MSA data is processed by the MSA Transformer, a protein language model employing an attention mechanism.
• A long short-term memory (LSTM) network generates the final predictions.

Main Results:

• S-Pred achieved approximately 76% accuracy for SS8 prediction.
• A Pearson's correlation of 0.84 was obtained between experimental and predicted ASAs.
• Intrinsically disordered regions (IDRs) were predicted with an F1-score of 0.514.

Conclusions:

• S-Pred demonstrates consistent and accurate performance in predicting key protein structural features.
• The tool effectively bridges the sequence-structure gap, aiding functional and structural analysis.
• S-Pred is available as open-source code and a web server for broader accessibility.