Evaluation of AlphaFold-Multimer prediction on multi-chain protein complexes

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• Accurate prediction of single protein structures is achievable.
• Current methods like AlphaFold-Multimer and FoldDock excel at modeling protein dimers.
• The performance of these tools on larger, multi-chain protein complexes and their evaluation remain challenging.

Purpose of the Study:

• To evaluate the performance of AlphaFold-Multimer on diverse protein complexes.
• To investigate discrepancies in existing evaluation metrics for multimeric structures.
• To introduce a novel scoring metric for assessing protein complex quality.

Main Methods:

• Analysis of AlphaFold-Multimer performance on a homology-reduced dataset of homo- and heteromeric protein complexes.
• Comparison of pairwise and multi-interface evaluation metrics (e.g., TM-score, DockQ).
• Development and proposal of a new scoring metric, Predicted DockQ version 2 (pDockQ2).
• Modeling of protein complexes from the CORUM database.

Main Results:

• AlphaFold-Multimer performance varies across different evaluation metrics, with some complexes scoring well on one metric but poorly on another.
• The study highlights the differences between pairwise and multi-interface evaluation of protein chains within a complex.
• Two highly confident protein complex structures with no sequence homology to existing structures were identified.
• The new pDockQ2 score aims to provide a more reliable estimation of interface quality in multimers.

Conclusions:

• Existing evaluation methods for protein complexes may not fully capture the quality of predictions.
• The proposed pDockQ2 metric offers a potentially more accurate assessment of multimeric complex quality.
• Further development and validation of computational tools and evaluation metrics are crucial for advancing protein structure prediction of large complexes.