Model Quality Assessment for CASP16.

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• The CASP16 experiment builds upon previous CASP assessments, introducing new evaluation modes for protein model accuracy.
• Emphasis is placed on evaluating multimeric assemblies and the performance of large-scale model pools generated by tools like MassiveFold.

Purpose of the Study:

• To assess the ability of computational methods to estimate the accuracy of predicted protein models, with a focus on multimeric structures.
• To introduce and evaluate a new mode (QMODE3) for selecting high-quality models from large sets of predictions.
• To explore the utility of per-atom local confidence measures from AlphaFold3 for accuracy estimation.

Main Methods:

• Three evaluation tasks were conducted: QMODE1 (global accuracy), QMODE2 (interface residue accuracy), and QMODE3 (model selection).
• Predictors were evaluated using OpenStructure-based metrics and a novel penalty-based ranking scheme for QMODE3.
• The accuracy and utility of per-atom local confidence measures, particularly pLDDT from AlphaFold3, were investigated.

Main Results:

• Methods incorporating AlphaFold3-derived features, especially per-atom pLDDT, demonstrated superior performance in estimating local accuracy.
• These AlphaFold3-based features proved highly useful for experimental structure solution.
• Performance in QMODE3 varied across different target types (monomeric, homomeric, heteromeric), highlighting challenges in evaluating complex assemblies.

Conclusions:

• Per-atom confidence measures from AlphaFold3 are valuable for assessing local accuracy and guiding experimental structure determination.
• Evaluating the accuracy of predicted models, particularly complex multimeric assemblies, remains an ongoing challenge.
• The CASP16 framework provides a robust platform for assessing advancements in protein structure prediction and model evaluation.