DeepInterAware: Deep Interaction Interface-Aware Network for Improving Antigen-Antibody Interaction Prediction from...

Area of Science:

• Biochemistry
• Immunology
• Computational Biology

Background:

• Accurate prediction of antigen-antibody interactions (AAI) is crucial for developing effective human therapeutics.
• Limited structural data for AAIs presents a significant challenge in prediction.
• Recent advances show structural information can be inferred from sequence data, enabling sequence-based prediction.

Purpose of the Study:

• To propose DeepInterAware, a framework that incorporates interaction interface information learned from sequence data for AAI prediction.
• To evaluate DeepInterAware's performance against existing methods in interaction prediction.
• To explore DeepInterAware's capabilities in identifying binding sites, detecting mutations, and predicting binding free energy changes.

Main Methods:

• Developed DeepInterAware, a deep learning framework that integrates sequence-derived interaction interface information and inherent sequence specificity.
• Applied DeepInterAware to interaction prediction tasks.
• Utilized a HER2-targeting antibody screening experiment to demonstrate practical application.

Main Results:

• DeepInterAware outperforms existing methods in interaction prediction.
• Demonstrated promising inductive and transfer capabilities for predicting interactions with unseen antigens/antibodies and for similar tasks.
• Showcased ability to identify potential binding sites and detect mutations within antigens/antibodies.
• Successfully applied in a HER2-targeting antibody screening experiment, identifying binding antibodies.

Conclusions:

• DeepInterAware is an effective tool for predicting antigen-antibody interactions, leveraging sequence data.
• The framework offers unique advantages, including mechanistic insights into AAIs and mutation impact assessment.
• DeepInterAware shows significant potential for advancing antibody screening and therapeutic development.