Towards More Accurate Full-Atom Antibody Co-Design

Jiayang Wu, Xingyi Zhang, Xiangyu Dong, Kun Xie, Ziqi Liu, Wensheng Gan, Sibo Wang, Le Song · 2025

Read Towards More Accurate Full-Atom Antibody Co-Design by Jiayang Wu, Xingyi Zhang, Xiangyu Dong, Kun Xie, Ziqi Liu, Wensheng Gan, Sibo Wang, Le Song (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-19391.txt.

Introduction

Wu and co-authors present Igformer, a system that designs antibodies by predicting both their letter sequence and their three-dimensional shape. The gripping loops of an antibody, called CDRs, decide what it binds, so getting their structure right is the heart of computer-aided antibody design for new medicines.

Igformer models the interface between antibody and target with a mix of local message passing and global attention, capturing both close chemical contacts and long-range shape effects. On standard tests of loop design and structure prediction it beat earlier methods by a clear margin. The authors argue that modeling interactions at several scales at once is the key advance for therapeutic antibody design.

Catalog record

CREATORJiayang Wu, Xingyi Zhang, Xiangyu Dong, Kun Xie, Ziqi Liu, Wensheng Gan, Sibo Wang, Le Song
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE90 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-19391

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Towards More Accurate Full-Atom Antibody Co-Design Jiayang Wu, Xingyi Zhang, Xiangyu Dong, Kun Xie, Ziqi Liu, Wensheng Gan, Sibo Wang, Le Song Abstract Antibody co-design represents a critical frontier in drug development, where accurate prediction of both 1D sequence and 3D structure of complementarity-determining regions (CDRs) is essential for targeting specific epitopes. Despite recent advances in equivariant graph neural networks for antibody design, current approaches often fall short in capturing the intricate interactions that govern antibody-antigen recognition and binding specificity. In this work, we present Igformer, a novel end-to-end framework that addresses these limitations through innovative modeling of antibody-antigen binding interfaces. Our approach refines the inter-graph representation by integrating personalized propagation with global attention mechanisms, enabling comprehensive capture of the intricate interplay between local chemical interactions and global conformational dependencies that characterize effective antibody-antigen binding. Through extensive validation on epitope-binding CDR design and structure prediction tasks, Igformer demonstrates significant improvements over existing methods, suggesting that explicit modeling of multi-scale residue interactions can substantially advance computational antibody design for therapeutic applications. 1 Main Antibodies are Y-shaped proteins that play a pivotal role in the immune system, capable of binding with high specificity to antigens, including pathogens and foreign molecules Murphy & Weaver (2016) . The design of antibodies for specific epitopes in antigens is an essential yet challenging task with broad applications in therapeutic development and drug discovery Feldmann & Maini (2003) ; Adams & Weiner (2005) ; Mullard (2022) ; Wang et al. (2024b) . The challenges arise primarily from the variability and structural complexity of complementarity-determining regions (CDRs), where binding primarily occurs He et al. (2024) . Understanding and modeling the intricate interactions between antibodies and antigens is crucial for developing an effective computational antibody design approach Baran et al. (2017) ; Tennenhouse et al. (2024) ; Yan et al. (2024) . Figure 1: End-to-end antibody co-design task. The past decade has witnessed a significant evolution in antibody design approaches. Traditional methods, including sequence-based models Alley et al. (2019) ; Shin et al. (2021) and energy-based optimization frameworks Pantazes & Maranas (2010) ; Lapidoth et al. (2015) ; Adolf-Bryfogle et al. (2018) , provided initial solutions but fell short of fully capturing the structural and functional interactions between antibodies and antigens Ingraham et al. (2019) . More recently, learning-based approaches, particularly equivariant graph neural networks, deep generative models, and diffusion models Saka et al. (2021) ; Jin et al. (2022b) ; Watson et al. (2022b) ; Luo et al. (2022) ; Kong et al. (2023a) ; Martinkus et al. (2023) ; Wang et al. (2024a) , have emerged as powerful tools for simultaneously co-designing CDR 1D sequences and 3D structures. These models represent a significant advancement over traditional approaches by leveraging deep learning to model complex sequential and structural relationships. However, many of these methods simplify the problem by focusing exclusively on backbone modeling or specific CDR regions, overlooking side-chain interactions and the full-atom geometry that are crucial to accurately modeling antibody-antigen binding Ruffolo et al. (2023b) ; Martinkus et al. (2023) . Traditional computational antibody design relies on multi-stage pipelines that separately handle structure prediction, docking, and CDR generation. A recent work, dyMEAN Kong et al. (2023a) , integrates these tasks into a unified end-to-end framework, providing a more streamlined approach to antibody co-design. However, despite these advances, dyMEAN does not fully take advantage of intricate antibody-antigen interactions, leaving opportunities for further improvement in capturing the detailed molecular interplay at binding interfaces. These limitations highlight the need for a more sophisticated approach that can simultaneously handle full-atom geometry and complex hierarchical interactions in antibody-antigen complexes. Further discussion of related works can be found in Appendix A . To address these limitations, in this work, we introduce Igformer 1 1 1 I mmuno g lobulin Trans former (Igformer) , a novel end-to-end framework for antibody co-design that advances antibody co-design through three key innovations. At its core, Igformer introduces a sophisticated antibody-antigen inter-graph refinement strategy that uniquely integrates personalized propagation with a global attention mechanism. The personalized propagation scheme models local chemical properties and geometric constraints, while the complementary global attention mechanism captures long-range structural dependencies. This novel dual-scale architecture enables Igformer to maintain atomic-level precision while effectively capturing intricate residue-level interactions within the antibody-antigen binding interface. Furthermore, Igformer incorporates E(3)-equivariance throughout the framework, ensuring that all structural predictions adhere to the physical and geometric constraints of biological systems. Through comprehensive experimental evaluation, Igformer demonstrates superior effectiveness across multiple challenging antibody design tasks. For epitope-binding CDR design, Igformer shows a 2.02% improvements over state-of-the-art models in terms of amino acid recovery rate. For antigen-antibody complex structure prediction, Igformer achieves remarkable accuracy with an 11.84% reduction in RMSD compared to existing approaches. These substantial improvements validate the effectiveness of our technical innovations in addressing the challenges of full-atom geometry and sequence-structure co-design. Igformer thus represents a significant advancement in computational antibody design, offering promising new directions for accelerating drug discovery and therapeutic development. Figure 2: Framework of Igformer. 2 Results 2.1 IgFormer Figure 2 illustrates the overall framework of Igformer. Specifically, the Equivariant Message Passing (EMP) modules serve as foundational blocks of Igformer, integrating both spatial and biochemical properties during the message-passing process. Our theoretical analysis confirms the E(3)-equivariant property of EMP, which lays the theoretical foundation of Igformer. The key innovation of Igformer lies in its sophisticated inter-graph refinement strategy. The process begins with the initialization of residue representations and coordinates, followed by intra-graph construction. Subsequently, based on the constructed intra-graph, Igformer captures the complex interactions within the binding range between antibody paratope and antigen epitope through two key modules: Approximate Personalized Propagation (APP) , which preserves local interaction information, and Simplified Graph Transformer (SGFormer) , which captures global binding dependencies. Next, the learned intricate binding patterns are utilized to refine the inter-graph, followed by Triangle Multiplicative (TM) module and Axial Attention (AA) module, which generate pairwise residue interactions through different mechanisms. Finally, the refined inter-graph, together with the intra-graph are processed through a dual-scale EMP modules , responsible for the final coordinate and sequence predictions. Systematical elaboration on our method and supplementary details can be found in Appendix C and appendix D . Dataset. Following previous works on antibody design Jin et al. (2022a) ; Kong et al. (2023a) ; Kong et al. (2023b) , we train all models on the Structural Antibody Database (SAbDab) ( Dunbar et al., 2013 ; Schneider et al., 2021 ) using the snapshot from November 2022. We split the dataset into training and validation sets with a ratio of 9:1, yielding 3,246 antibodies for training and 365 antibodies for validation. The antibodies are clustered based on their CDR-H3 regions using a 40\% sequence identity threshold, calculated using the BLOSUM62 substitution matrix Foote & Winter (1992) . This clustering process, performed using MMseqs2, results in 1,644 clusters in the training set and 182 clusters in the validation set. After that, we evaluate Igformer and other competitors on the RAbD benchmark Adolf-Bryfogle et al. (2018) for Tasks 1-3, which consists of 60 diverse antibody-antigen complexes. For Task 4, Igfold benchmark Ruffolo et al. (2023a) consisting of 51 antibody-antigen complexes is utilized. This test set selection prevents data leakage during the evaluation phase. Detailed introduction

Read the complete text