A sparse autoencoder-based deep neural network for protein solvent accessibility and contact number prediction.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Predicting three-dimensional (3D) protein structures from one-dimensional (1D) sequences is a fundamental challenge in structural biology.
• Accurate prediction of protein structural characteristics, such as solvent accessibility and contact number, is crucial for modeling protein folding and determining 3D structures.
• These predicted features serve as essential restraints for computational protein structure building.

Purpose of the Study:

• To introduce DeepSacon, a novel computational method for the accurate prediction of protein solvent accessibility and contact number.
• To leverage deep neural networks, specifically stacked autoencoders with dropout, for enhanced prediction of these critical structural features.
• To demonstrate the superior performance of DeepSacon compared to existing state-of-the-art methods.

Main Methods:

• Development of DeepSacon, a deep neural network architecture based on stacked autoencoders.
• Integration of a dropout technique to improve the generalization and robustness of the model.
• Training and evaluation of the model on a large dataset of 5729 monomeric soluble globular proteins and the CASP11 benchmark dataset.

Main Results:

• DeepSacon achieved a three-state accuracy of 0.70 for solvent accessibility and a Pearson Correlation Coefficient (PCC) of 0.74 for contact number on the monomeric soluble globular protein dataset.
• On the CASP11 benchmark dataset, DeepSacon demonstrated a three-state accuracy of 0.68 for solvent accessibility and a PCC of 0.69 for contact number.
• The method showed significant improvements in prediction quality over current state-of-the-art techniques.

Conclusions:

• DeepSacon reliably predicts protein solvent accessibility and contact number using a stacked sparse autoencoder and dropout approach.
• The computational method offers a significant advancement in predicting key structural features essential for 3D protein structure determination.
• The findings highlight the potential of deep learning for addressing complex problems in structural bioinformatics.