Interaction profile-based protein classification of death domain

Area of Science:

• Proteomics and structural biology
• Computational biology and bioinformatics

Background:

• Genomic initiatives generate vast protein sequence and structure data, driving focus on proteomics.
• High-throughput protein-protein interaction data necessitates in silico methods for predicting complex structures.
• Predicting protein complexes is crucial for understanding interaction networks, especially when experimental structure determination lags.

Purpose of the Study:

• To develop and evaluate an in silico system for classifying protein family members based on predicted protein-protein interactions.
• To assess the accuracy of homology modeling and docking simulations for protein classification.

Main Methods:

• Developed HODOCO (Homology modeling, Docking and Classification Oracle) system.
• Utilized protein Residue Potential Interaction Profiles (RPIPS) to characterize protein-protein interactions.
• Applied heuristic and support vector machine learning methods with 5-fold cross-validation on a dataset of 64 death domain superfamily proteins.

Main Results:

• The HODOCO system successfully classified death domain superfamily proteins into their subfamilies.
• The heuristic classification approach achieved 61% average accuracy.
• The support vector machine learning approach achieved 89% average accuracy.

Conclusions:

• Classifying proteins based on predicted interactions is reliable and valuable.
• The accuracy of this in silico approach is comparable to methods using experimentally determined structures.
• Results align with functional and sequence-based classifications, despite not directly using sequence information.