From interaction networks to interfaces, scanning intrinsically disordered regions using AlphaFold2.

Area of Science:

• Structural Biology
• Computational Biology
• Bioinformatics

Background:

• Protein-protein interaction networks are crucial but challenging to analyze, particularly for intrinsically disordered regions.
• Current methods often fail to precisely delimit interaction sites, especially for small motifs in disordered regions.

Purpose of the Study:

• To improve the accuracy of AlphaFold2-Multimer in predicting protein-protein interaction sites, particularly for intrinsically disordered regions.
• To investigate strategies for enhancing the prediction of binding interfaces and assess the role of confidence scores in partner discrimination.

Main Methods:

• Utilized a non-redundant dataset of protein-peptide complexes with intrinsically disordered regions.
• Implemented input sequence fragmentation and integrated evolutionary information strategies.
• Tested performance on a larger dataset from the ELM database and analyzed AlphaFold2 confidence scores.

Main Results:

• AlphaFold2-Multimer achieved only a 40% success rate with full protein sequences for disordered region interactions.
• Input fragmentation and evolutionary data integration improved prediction accuracy to 90%.
• Similar high success rates were observed on the larger ELM database dataset.

Conclusions:

• Delineating interaction regions into fragments and incorporating evolutionary data significantly enhances AlphaFold2-Multimer's accuracy for protein-protein interaction prediction.
• The study highlights the limitations of AlphaFold2's confidence scores for discriminating binding partners, especially with small interaction motifs.