Benchmarking AlphaFold3's protein-protein complex accuracy and machine learning prediction reliability for binding...

Area of Science:

• Computational biology
• Structural biology
• Biochemistry

Background:

• AlphaFold 3 (AF3) is an advanced protein structure prediction tool.
• AF3 extends capabilities to predict protein-protein complexes, unlike its predecessors.
• Accurate prediction of protein complex structures is crucial for drug discovery and protein engineering.

Purpose of the Study:

• To perform an independent validation of AlphaFold 3's predictions for protein-protein complexes.
• To assess the accuracy of AF3 in predicting binding free energy changes upon mutation.
• To identify limitations and potential errors in AF3's complex structure predictions.

Main Methods:

• Utilized the SKEMPI 2.0 database, comprising 317 protein-protein complexes and 8338 mutations.
• Evaluated the Pearson correlation coefficient (PCC) between AF3 predictions and experimental data for binding free energy changes.
• Compared the Root Mean Square Error (RMSE) of AF3 complex structures against Protein Data Bank (PDB) structures.

Main Results:

• AF3 achieved a PCC of 0.86 for predicting binding free energy changes, closely matching PDB structure performance (0.88).
• AF3 complex structures resulted in an 8.6% increase in prediction RMSE compared to original PDB complex structures.
• Significant errors in some AF3 complex structures were not reflected in its ipTM performance metric, and predictions were unreliable for flexible regions.

Conclusions:

• While AF3 shows promise for protein complex prediction, its accuracy for binding free energy changes is comparable but not superior to existing PDB structures.
• The increased RMSE and unreliability for flexible regions indicate that AF3 complex structures require cautious interpretation and further refinement.
• Independent validation is essential to understand the practical utility and limitations of AF3 in biological and drug discovery applications.