[Paper Review] End-to-End Full-Atom Antibody Design

Xiangzhe Kong, Wenbing Huang+1 more|arXiv (Cornell University)|Feb 1, 2023

Monoclonal and Polyclonal Antibodies Research11 citations

TL;DR

The paper introduces dyMEAN, an end-to-end E(3)-equivariant multi-channel graph network that designs full-atom antibodies from an epitope and an incomplete antibody sequence, enabling simultaneous sequence and structure generation and docking.

ABSTRACT

Antibody design is an essential yet challenging task in various domains like therapeutics and biology. There are two major defects in current learning-based methods: 1) tackling only a certain subtask of the whole antibody design pipeline, making them suboptimal or resource-intensive. 2) omitting either the framework regions or side chains, thus incapable of capturing the full-atom geometry. To address these pitfalls, we propose dynamic Multi-channel Equivariant grAph Network (dyMEAN), an end-to-end full-atom model for E(3)-equivariant antibody design given the epitope and the incomplete sequence of the antibody. Specifically, we first explore structural initialization as a knowledgeable guess of the antibody structure and then propose shadow paratope to bridge the epitope-antibody connections. Both 1D sequences and 3D structures are updated via an adaptive multi-channel equivariant encoder that is able to process protein residues of variable sizes when considering full atoms. Finally, the updated antibody is docked to the epitope via the alignment of the shadow paratope. Experiments on epitope-binding CDR-H3 design, complex structure prediction, and affinity optimization demonstrate the superiority of our end-to-end framework and full-atom modeling.

Motivation & Objective

• Address the limitations of stage-wise antibody design methods by providing an end-to-end framework that handles full-atom geometry (including side chains) and complete antibody context.
• Develop a model that can generate the complete antigen–antibody complex from the epitope and an incomplete antibody sequence.
• Incorporate structural initialization, a shadow paratope mechanism, and adaptive multi-channel equivariant encoding to manage variable atom counts per residue.
• Ensure E(3)-equivariance to generalize across arbitrary epitope poses and orientations.

Proposed method

• Introduce structural initialization (SI) based on conserved residues to provide a knowledgeable antibody structure guess.
• Attach a shadow paratope around the epitope to bridge epitope–paratope interactions in an E(3)-invariant fashion.
• Employ an Adaptive Multi-channel Equivariant Encoder (AME) that updates both hidden states and full-atom coordinates across residues with variable channel sizes.
• Dock the native paratope to the shadow paratope using a Kabsch-based alignment to produce the final docked complex.
• Train with a joint objective comprising sequence loss, structure loss (coordinate and bond length), and docking loss (shadow paratope coordinates and cross-interface distances).
• Prove E(3)-equivariance of the design process and docking outcomes (Theorem 4.1).

Experimental results

Research questions

• RQ1 Can an end-to-end model generate complete CDR-H3 sequences and full-atom antibody structures given only an epitope and incomplete antibody sequence?
• RQ2 Does incorporating full-atom geometry and shadow paratope-based docking improve design quality over staged pipelines?
• RQ3 Is the proposed dyMEAN framework robust to arbitrary epitope poses due to E(3)-equivariance?
• RQ4 How does dyMEAN perform on tasks including epitope-binding CDR-H3 generation, complex structure prediction, and affinity optimization?

Key findings

• dyMEAN outperforms pipeline-based baselines across multiple metrics for epitope-binding CDR-H3 design, including amino acid recovery and docking quality.
• In complex structure prediction tasks, dyMEAN yields superior TMscore, lDDT, and DockQ scores compared to staged approaches.
• The end-to-end full-atom design enables simultaneous generation of the paratope sequence and the complete antibody structure, with folding and docking outcomes that are invariant/equivariant under E(3) transformations.
• Shadow paratope attachments enable robust information transfer between epitope and antibody, facilitating docking that aligns the native paratope with the shadow paratope.
• The approach demonstrates effectiveness in affinity optimization tasks, illustrating practical utility beyond static structure prediction.