Discriminating physiological from non-physiological interfaces in structures of protein complexes: A community-wide...

Area of Science:

• Structural Biology
• Computational Biology
• Protein Science

Background:

• Accurate scoring and ranking of protein complex models, including determination of oligomeric state from crystal lattice structures, remain significant challenges in structural biology.
• Distinguishing between physiological (biological) and non-physiological (crystal packing) protein-protein interactions is crucial for understanding protein function and assembly.

Purpose of the Study:

• To develop and evaluate methods for reliably scoring and ranking protein complex models, specifically focusing on differentiating homodimers.
• To create a robust benchmark dataset for assessing the performance of protein-protein interface scoring functions.

Main Methods:

• Compiled a benchmark dataset of 1677 homodimer protein crystal structures, including both physiological and challenging non-physiological complexes with large interface areas.
• Collected and evaluated 252 existing protein-protein interface scoring functions.
• Developed a consensus score using top-performing individual scores and a cross-validated Random Forest (RF) classifier.

Main Results:

• The consensus score and RF classifier achieved high performance, with areas under the Receiver Operating Characteristic (ROC) curve of 0.93 and 0.94, respectively.
• Both developed approaches outperformed individual scoring functions in discriminating between physiological and non-physiological complexes.
• AlphaFold2 models demonstrated higher accuracy in recalling physiological dimers, validating the benchmark dataset's annotations.

Conclusions:

• Optimizing combined interface scoring functions and evaluating them on challenging benchmark datasets is a promising strategy for improving protein complex analysis.
• The developed benchmark dataset and scoring approaches provide valuable tools for assessing and advancing computational methods in protein complex modeling.