Integrative Protein Assembly With LZerD and Deep Learning in CAPRI 47-55

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Protein complex prediction is crucial for understanding biological functions.
• The CAPRI (Critical Assessment of PRedicted Interactions) experiment benchmarks computational methods for protein structure prediction.
• Evaluating and improving protein complex prediction models is an ongoing challenge.

Purpose of the Study:

• To report the performance of our group's protein complex prediction methods in recent CAPRI rounds (47-55).
• To assess the integration of classical and deep learning approaches in modeling protein complexes.
• To analyze the success and limitations of our modeling strategies through case studies.

Main Methods:

• Integration of established computational pipelines with novel deep learning models.
• Inclusion of literature-derived data, such as assayed interface residues, for human group predictions.
• Model selection via rank aggregation of scoring functions, generative model confidence, and expert evaluation.

Main Results:

• Successful modeling of eight protein interfaces in the evaluated CAPRI rounds.
• Achieved top quality level for all modeled interfaces, outperforming other groups.
• Demonstrated superior performance in two cases where no other group succeeded.

Conclusions:

• Our combined classical and deep learning approaches are highly effective for protein complex prediction.
• The integration of diverse data sources and advanced selection strategies enhances prediction accuracy.
• Continuous refinement of modeling pipelines, particularly towards deep learning unification, is key for future advancements.