Simultaneous prediction of antibody backbone and side-chain conformations with deep learning

Area of Science:

• Computational biology
• Structural biology
• Immunology

Background:

• Antibody engineering is vital for developing medical diagnostics and immunotherapies.
• Accurate modeling of antibody complementarity determining regions (CDRs) is essential for antibody function.
• Existing deep learning methods predict antibody backbone structures but often neglect side-chain conformations.

Purpose of the Study:

• To develop a deep learning method (DeepSCAb) for predicting antibody variable fragment structures, including side-chain conformations.
• To enable accurate antibody structure prediction using only sequence input.
• To improve antibody engineering and design applications.

Main Methods:

• DeepSCAb utilizes a deep learning network that predicts inter-residue geometries and side-chain dihedrals.
• The method employs an interpretable self-attention layer for rotamer predictions, identifying conserved anchor positions.
• Input is limited to the antibody sequence, requiring no prior backbone information.

Main Results:

• DeepSCAb accurately predicts side-chain conformations, outperforming methods that lack side-chain context.
• The model competitively predicts side-chain conformations compared to methods requiring input backbone structures.
• DeepSCAb demonstrates superior performance in discriminating near-native antibody structures from decoys.

Conclusions:

• DeepSCAb enhances antibody structure prediction through accurate side-chain modeling.
• The method is valuable for antibody engineering, particularly for antibodies with unknown backbone structures.
• DeepSCAb has potential applications in antibody-antigen docking and the design of novel therapeutic antibodies.