Practical Outcomes From CASP16 for Users in Need of Biomolecular Structure Prediction

Area of Science:

• Computational biology and structural bioinformatics.
• Artificial intelligence in molecular modeling.

Background:

• The Critical Assessment of Structure Prediction (CASP) benchmarks progress in computational protein structure prediction.
• Advancements in deep learning, exemplified by AlphaFold 2 (AF2) and AlphaFold 3 (AF3), have revolutionized biomolecular modeling.

Purpose of the Study:

• To evaluate the performance of state-of-the-art biomolecular modeling tools, including AlphaFold 3, in the 16th CASP.
• To assess the capabilities and limitations of AI-driven structure prediction for proteins, protein complexes, and nucleic acids.
• To provide guidance on tool selection and interpretation of AI-generated structural models.

Main Methods:

• Benchmarking of various computational methods against experimentally determined structures in the CASP16 dataset.
• Focus on AlphaFold 2 and AlphaFold 3 performance across different prediction tasks: protein monomers, assemblies, and protein-ligand complexes.
• Analysis of prediction accuracy, confidence metrics, and limitations for various biomolecular systems.

Main Results:

• Protein monomer and domain prediction accuracy is near-maximal, with remaining challenges in specific secondary structures and mutational effects.
• AlphaFold-based methods show progress in protein assembly prediction, though complex topologies and antibody-antigen interactions remain difficult.
• AlphaFold 3 demonstrates strong potential in protein-ligand co-folding pose prediction but shows unreliable ligand affinity prediction; nucleic acid prediction capabilities are limited.

Conclusions:

• AlphaFold 3 represents the state of the art in biomolecular modeling, offering improvements over AlphaFold 2 and enhanced confidence metrics.
• Significant challenges persist in predicting complex protein assemblies, antibody-antigen interactions, and nucleic acid structures.
• Accurate interpretation of AI-generated models necessitates understanding their limitations and leveraging confidence metrics for reliable structural insights.