DeepCNF-D: Predicting Protein Order/Disorder Regions by Weighted Deep Convolutional Neural Fields

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Intrinsically disordered proteins (IDPs) and regions (IDRs) are crucial for biological processes like transcription regulation and signal transduction.
• Accurate prediction of protein order/disorder regions from sequence is essential for understanding their functions.

Purpose of the Study:

• To develop an improved computational method for ab initio prediction of intrinsically disordered protein regions.
• To enhance prediction accuracy by leveraging sequential information and addressing data imbalance.

Main Methods:

• A learning method called weighted Deep Convolutional Neural Fields (DeepCNF) was developed.
• The method exploits long-range sequential information and interdependencies between adjacent order/disorder labels.
• Different weights were assigned during training and prediction to handle label imbalance.

Main Results:

• The weighted DeepCNF method achieved high performance on CASP9 and CASP10 targets.
• Achieved AUC values of 0.855 for CASP9 and 0.898 for CASP10.
• Outperformed existing state-of-the-art single ab initio predictors.

Conclusions:

• Weighted DeepCNF offers a significant improvement in predicting intrinsically disordered protein regions.
• The method's ability to handle long-range dependencies and label imbalance contributes to its enhanced accuracy.
• This advancement facilitates further functional and mechanistic studies of disordered regions.