Differential performance of RoseTTAFold in antibody modeling

Area of Science:

• Structural Biology
• Immunology
• Computational Biology

Background:

• Accurate antibody structure prediction is crucial for understanding antibody-antigen interactions.
• The H3 loop, a key region for antigen binding, presents a significant challenge in antibody modeling.
• Existing prediction methods struggle with atomic accuracy, especially without homologous structures.

Purpose of the Study:

• To evaluate the performance of RoseTTAFold, a deep learning algorithm, in predicting the 3D structures of antibodies.
• To compare RoseTTAFold's antibody modeling capabilities against established tools like SWISS-MODEL and ABodyBuilder.

Main Methods:

• Collected sequences of 30 antibodies for structure prediction using RoseTTAFold.
• Compared RoseTTAFold models with SWISS-MODEL and ABodyBuilder.
• Stratified model quality assessment using Global Model Quality Estimate (GMQE) scores.

Main Results:

• RoseTTAFold demonstrated comparable performance to SWISS-MODEL in modeling most CDR loops, particularly with GMQE scores below 0.8.
• RoseTTAFold showed improved H3 loop prediction accuracy compared to ABodyBuilder.
• Overall, RoseTTAFold's accuracy was not superior to SWISS-MODEL or ABodyBuilder but showed specific strengths.

Conclusions:

• RoseTTAFold is a capable tool for antibody 3D structure prediction, with notable accuracy for the H3 loop.
• Further development of RoseTTAFold could enhance its antibody modeling capabilities.
• The study provides insights into the strengths and limitations of deep learning approaches for antibody structure prediction.