Progress and Bottlenecks for Deep Learning in Computational Structure Biology: CASP Round XVI

Area of Science:

• Computational structural biology
• Bioinformatics
• Structural modeling

Background:

• Community-wide experiments like CASP rigorously assess computational structural biology methods.
• Recent years have seen significant advancements, with deep learning methods, particularly AlphaFold variants, dominating many areas.
• The focus has shifted from method utility to achieving experimental accuracy.

Purpose of the Study:

• To evaluate the state-of-the-art in computational structure prediction at CASP16.
• To assess the performance of deep learning and traditional methods across various structural biology targets.
• To identify current limitations and future trends in structure prediction accuracy.

Main Methods:

• Analysis of results from the Critical Assessment of protein Structure Prediction (CASP16) experiment.
• Comparison of deep learning methods (e.g., AlphaFold variants) against traditional approaches.
• Evaluation of structure prediction accuracy for monomer proteins, protein complexes, RNA, macromolecular ensembles, and ligand-protein interactions.

Main Results:

• Deep learning methods show high accuracy for monomer proteins, approaching experimental uncertainty limits.
• Significant accuracy gains were observed for protein complexes, with room for further improvement.
• Deep learning methods were unsuccessful for RNA structure prediction, performing similarly to traditional methods.
• Ligand-protein structure and affinity predictions improved substantially with deep learning, though not reaching experimental accuracy.
• Methods for accuracy estimation proved effective in selecting high-quality models for protein complexes.

Conclusions:

• Deep learning dominates protein structure prediction, but RNA remains a challenge.
• Combining physics-inspired methods with deep learning and increasing training data are promising future directions.
• Continued advancements are needed to bridge the gap between predicted and experimental accuracy in all categories.