AlphaFold3 at CASP16.

Area of Science:

• Structural bioinformatics
• Computational biology
• Protein structure prediction

Background:

• AlphaFold2 revolutionized protein structure prediction.
• The new AlphaFold3 model promises enhanced capabilities, including non-protein molecule prediction.
• Benchmarking is crucial to validate performance claims.

Purpose of the Study:

• To evaluate the performance of AlphaFold3 against AlphaFold2 using CASP16 data.
• To assess AlphaFold3's accuracy for various targets, including protein complexes and nucleic acids.
• To identify areas for improvement in AlphaFold3's prediction and model selection strategies.

Main Methods:

• Utilized the AlphaFold3 web server for predictions (AF3-server).
• Performed manual predictions for large targets and ligands (Elofsson group).
• Compared AlphaFold3 results against AlphaFold2-based methods and CASP rankings.

Main Results:

• AlphaFold3 shows marginal improvement over AlphaFold2 for protein complexes, especially when AlphaFold2 uses massive sampling.
• AF3-server outperforms AlphaFold2 on easier CASP targets but not harder ones.
• Accuracy for predicting stoichiometry of large complexes and nucleic acid interactions is limited, though competitive with top methods.

Conclusions:

• AlphaFold3 offers a user-friendly tool with predictions comparable to state-of-the-art methods.
• Further development is needed for better model selection from generated predictions.
• AlphaFold3 represents a significant advancement in structure prediction, particularly for non-protein components.