Deep learning methods for protein torsion angle prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Deep learning (DL) offers state-of-the-art performance in various scientific domains.
• DL has been successfully applied to bioinformatics tasks like protein contact prediction and fold recognition since 2012.
• This study focuses on enhancing protein torsion angle prediction using DL.

Purpose of the Study:

• To develop and evaluate deep learning methods for improved prediction of protein torsion (dihedral) angles.
• To investigate the efficacy of different DL architectures for this task.
• To identify novel input features that enhance prediction accuracy.

Main Methods:

• Four DL architectures were designed: deep neural network (DNN), deep restricted Boltzmann machine (DRBN), deep recurrent neural network (DRNN), and deep recurrent restricted Boltzmann machine (DReRBM).
• Input features included existing protein data plus two novel features: predicted residue contact number and error distribution from sequence fragments.
• The models predicted backbone phi and psi angles.

Main Results:

• The developed DL models achieved a mean absolute error (MAE) of approximately 20-21° for phi angles and 29-30° for psi angles on an independent dataset.
• The MAE for psi angle prediction was 2° lower than existing methods.
• Performance on CASP12 targets was comparable or superior to state-of-the-art methods.

Conclusions:

• Deep learning is a powerful and valuable approach for predicting protein torsion angles.
• Deep recurrent network architectures showed slightly better performance than deep feed-forward architectures.
• The novel features, predicted residue contact number and error distribution, proved beneficial for improving prediction accuracy.