Predicting Protein-Peptide Interactions: Benchmarking Deep Learning Techniques and a Comparison with Focused Docking.

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Deep learning (DL) has revolutionized protein structure prediction, impacting structural biology.
• AlphaFold2 demonstrated early success in predicting protein-peptide interactions, outperforming other methods.
• Recent advancements include new AlphaFold2 models for multimeric assemblies and the OmegaFold ab initio folding model.

Purpose of the Study:

• To evaluate the docking success rates of new deep learning folding models, including AlphaFold2 and OmegaFold.
• To compare their performance against the specialized peptide-docking software AutoDock CrankPep (ADCP).
• To assess the efficacy of a consensus docking approach combining DL models and ADCP.

Main Methods:

• Utilized a dataset of 99 nonredundant protein-peptide complexes for evaluation.
• Assessed docking success rates using a consistent performance metric across all tested methods.
• Compared AlphaFold2, OmegaFold, and ADCP, including a consensus strategy.

Main Results:

• The latest AlphaFold2 model significantly outperformed other deep learning approaches for peptide docking.
• ADCP achieved a 62% success rate when considering sampled solutions, though modest for top-ranking results alone.
• A consensus approach using ADCP and AlphaFold2 yielded 60% top-ranking and 66% top-5 success rates.

Conclusions:

• New deep learning models, particularly AlphaFold2, show significant promise for predicting protein-peptide interactions.
• Combining complementary methods like ADCP and AlphaFold2 in a consensus approach enhances overall docking accuracy.
• These findings advance the capabilities for understanding and predicting protein-peptide complex structures.