Protein intrinsically disordered region prediction by combining neural architecture search and multi-objective...

Area of Science:

• Protein bioinformatics
• Computational biology
• Structural biology

Background:

• Intrinsically disordered regions (IDRs) are crucial for protein function but challenging to identify accurately.
• Existing predictors struggle with varying ratios of long (LDRs) and short (SDRs) disordered regions due to limitations in capturing length-dependent features.
• Current deep learning models often overlook the complementarity of different prediction approaches.

Purpose of the Study:

• To develop a novel and stable predictor for identifying intrinsically disordered regions (IDRs) in protein sequences.
• To overcome the performance limitations of existing methods on datasets with diverse LDR/SDR ratios.
• To leverage neural architecture search (NAS) for automated network construction and feature extraction.

Main Methods:

• Employed neural architecture search (NAS) to automatically design protein sequence analysis networks.
• Integrated NAS-generated models with length-dependent and general models to capture unique and common features of LDRs and SDRs.
• Developed a new predictor named IDP-Fusion for enhanced IDR identification.

Main Results:

• The proposed IDP-Fusion predictor demonstrated superior and more stable performance compared to existing methods.
• Achieved consistent accuracy across independent test sets with varying proportions of SDRs and LDRs.
• Successfully captured hidden length-dependent features in protein sequences through automated network architecture.

Conclusions:

• IDP-Fusion provides a more robust solution for identifying intrinsically disordered regions (IDRs).
• The combined approach effectively addresses the challenges posed by different lengths of disordered regions.
• This advancement aids in more accurate protein structure and function analysis.