Antibody structure prediction using interpretable deep learning.

Area of Science:

• Biochemistry
• Computational Biology
• Immunology

Background:

• Therapeutic antibodies represent a significant and expanding market within biologics.
• Accurate antibody structure determination is crucial for rational antibody design but often relies on time-consuming experimental methods.

Purpose of the Study:

• To introduce DeepAb, a deep learning approach for predicting antibody Fv structures directly from amino acid sequences.
• To evaluate DeepAb's performance against existing methods and provide interpretability for its predictions.

Main Methods:

• Development of DeepAb, a deep learning model utilizing an interpretable attention mechanism.
• Validation of DeepAb on a diverse set of therapeutically relevant antibodies with known structures.
• Introduction of a novel mutant scoring metric based on network confidence.

Main Results:

• DeepAb consistently outperformed leading alternative methods in predicting antibody Fv structures.
• The attention mechanism revealed physically relevant residue interactions, offering insights into the model's predictions.
• The novel mutant scoring metric successfully identified mutations predicted to improve antibody binding affinity.

Conclusions:

• DeepAb provides an accurate and interpretable method for antibody structure prediction from sequence.
• The model's ability to identify beneficial mutations has significant implications for antibody engineering and drug discovery.
• DeepAb is a valuable tool for various antibody prediction and design applications.