Evaluating deep learning based structure prediction methods on antibody-antigen complexes.

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• AlphaFold2 advanced protein complex structure prediction but struggled with interactions lacking coevolutionary signals, such as host-pathogen and antibody-antigen complexes.
• Strategies to improve prediction accuracy include massive sampling and architectural changes, like AlphaFold3's pairformer, which enhances structural reasoning without relying on coevolution.

Purpose of the Study:

• To benchmark protein structure prediction methods on novel antibody-antigen complexes.
• To evaluate the impact of sampling and model quality estimation on prediction accuracy.
• To compare the performance of AlphaFold3 against other state-of-the-art methods.

Main Methods:

• Benchmarking structure prediction tools on a dataset of unseen antibody-antigen complexes.
• Analyzing the effect of increased sampling on the generation of accurate protein models.
• Assessing the reliability of internal quality estimates in identifying the best predicted structures.
• Comparing AlphaFold3 with AlphaFold2, Boltz-1, and Chai-1.

Main Results:

• Increased sampling improves protein model generation in a log-linear fashion.
• AlphaFold's internal quality estimates often fail to identify the best predicted models, impacting performance.
• AlphaFold3 significantly outperforms AlphaFold2, Boltz-1, and Chai-1.
• AlphaFold3's performance decreases for complexes dissimilar to the training set, suggesting an ability to recognize remote structural similarities.

Conclusions:

• AlphaFold3 represents a significant advancement in protein complex structure prediction, particularly for challenging interactions.
• Accurate identification of the best predicted model remains a critical challenge across all evaluated methods.
• The ability of AlphaFold3 to generalize to complexes with low structural similarity to training data warrants further investigation.