DeepSacon

DeepSacon predicts solvent accessibility and contact number from one-dimensional amino acid sequences to provide constraints for protein folding models and three-dimensional (3D) protein structure construction.

Key Features:

• Prediction targets: Solvent accessibility (three-state and 15-state) and contact number predicted from one-dimensional amino acid sequences.
• Model architecture: Deep neural networks employing a stacked autoencoder architecture with dropout to enhance robustness and prevent overfitting.
• Performance on monomeric soluble globular proteins: On a dataset of 5,729 monomeric soluble globular proteins it achieves three-state accuracy of 0.70, 15-state accuracy of 0.33, and a Pearson Correlation Coefficient (PCC) of 0.74 for contact number predictions.
• CASP11 benchmark performance: On the CASP11 dataset it attains three-state accuracy of 0.68 and PCC of 0.69 for both solvent accessibility and contact number predictions.

Scientific Applications:

• Protein folding constraints: Provide quantitative constraints for protein folding models derived from predicted solvent accessibility and contact number.
• 3D structure modeling: Aid construction and refinement of three-dimensional protein structures using predicted residue-level structural properties.
• Stability and interaction analysis: Inform analyses of protein stability and intermolecular interactions through solvent accessibility and contact number information.

Methodology:

DeepSacon applies deep neural networks using a stacked autoencoder architecture combined with dropout to predict solvent accessibility and contact number from one-dimensional amino acid sequences.