AUCpreD: proteome-level protein disorder prediction by AUC-maximized deep convolutional neural fields

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Intrinsically disordered regions (IDRs) are crucial for biological processes but challenging to predict due to their prevalence and low ratio of disordered residues.
• Existing prediction methods often rely on sequence profiles, limiting their proteome-wide applicability due to computational cost.

Purpose of the Study:

• To develop a highly accurate and computationally efficient method for predicting intrinsically disordered regions (IDRs) across entire proteomes.
• To address the challenge of class imbalance in predicting disordered residues.

Main Methods:

• Formulated IDR prediction as a sequence labeling problem using Deep Convolutional Neural Fields (DeepCNF), integrating deep convolutional neural networks (DCNN) and conditional random fields (CRF).
• Developed a novel training approach by maximizing the area under the ROC curve (AUC) to handle the imbalanced ratio of ordered to disordered residues, instead of using maximum-likelihood estimation.

Main Results:

• The developed method, AUCpreD, demonstrated superior performance compared to existing popular disorder predictors.
• AUCpreD achieved high accuracy comparable to or exceeding methods that utilize sequence profiles, even when sequence profiles were not used.
• The method is suitable for efficient proteome-wide disorder prediction.

Conclusions:

• AUCpreD offers a significant advancement in predicting intrinsically disordered regions, providing a balance of accuracy and computational efficiency.
• The approach is effective for large-scale proteome analysis, overcoming limitations of previous methods.
• The AUC maximization training strategy is robust for imbalanced biological sequence data.