CNNcon: improved protein contact maps prediction using cascaded neural networks

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Determining 3D protein structures lags behind sequence discovery, creating a significant gap.
• Protein modeling offers a computational approach to bridge this sequence-structure gap.
• Predicting residue contact maps is a key intermediate step for 3D structure prediction.

Purpose of the Study:

• To develop an improved computational method for predicting protein residue contact maps.
• To enhance the accuracy and applicability of protein structure prediction, particularly for longer sequences.

Main Methods:

• Developed CNNcon, a novel contact map predictor utilizing multiple neural networks.
• Employed a architecture with six sub-networks and a final cascade-network for prediction.
• Trained and tested sub-networks and the cascade-network with dedicated datasets.

Main Results:

• CNNcon achieved an average prediction accuracy of 58.86% for contacts within 8 Å for proteins up to 450 residues.
• Outperformed existing state-of-the-art contact map predictors in comparative analyses.
• Demonstrated consistent accuracy with increasing protein sequence length, overcoming the 'thin density' problem.

Conclusions:

• CNNcon effectively predicts residue contacts, facilitating more accurate 3D protein structure prediction.
• The method's robustness with longer sequences makes it valuable for predicting structures of large proteins.
• This advancement aids in bridging the sequence-structure gap and improving template-based structure prediction.