Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization

Lirong Wu, Haitao Lin, Yufei Huang, Zhangyang Gao, Cheng Tan, Yunfan Liu, Tailin Wu, Stan Z. Li · 2025

Read Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization by Lirong Wu, Haitao Lin, Yufei Huang, Zhangyang Gao, Cheng Tan, Yunfan Liu, Tailin Wu, Stan Z. Li (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-2501-00013.txt.

Introduction

Lirong Wu and seven co-authors present RAAD, a setup for designing antibodies without knowing the target epitope in advance, in a two thousand twenty-four computing-and-biology paper. Antibody binding is governed by the complementarity finding regions, whose design remains hard to compute.

Earlier methods handle long regions poorly, assume a given epitope with static interactions, and ignore specificity. RAAD enriches node, edge, and relation features with context and shape information, then models antigen-antibody interaction dynamically to co-design sequences and structures. A new specificity metric and a contrastive constraint push designs toward certain binders. Tests across region types and lengths show strong modeling, making, and tuning results.

Catalog record

CREATORLirong Wu, Haitao Lin, Yufei Huang, Zhangyang Gao, Cheng Tan, Yunfan Liu, Tailin Wu, Stan Z. Li
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-00013

Browse the full library · About the Tower of Babel

Full text

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

Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization Lirong Wu, Haitao Lin, Yufei Huang, Zhangyang Gao, Cheng Tan, Yunfan Liu, Tailin Wu, Stan Z. Li Abstract Antibodies are Y-shaped proteins that protect the host by binding to specific antigens, and their binding is mainly determined by the Complementary Determining Regions (CDRs) in the antibody. Despite the great progress made in CDR design, existing computational methods still encounter several challenges: 1) poor capability of modeling complex CDRs with long sequences due to insufficient contextual information; 2) conditioned on pre-given antigenic epitopes and their static interaction with the target antibody; 3) neglect of specificity during antibody optimization leads to non-specific antibodies. In this paper, we take into account a variety of node features, edge features, and edge relations to include more contextual and geometric information. We propose a novel R elation- A ware A ntibody D esign (RAAD) framework, which dynamically models antigen-antibody interactions for co-designing the sequences and structures of antigen-specific CDRs. Furthermore, we propose a new evaluation metric to better measure antibody specificity and develop a contrasting specificity-enhancing constraint to optimize the specificity of antibodies. Extensive experiments have demonstrated the superior capability of RAAD in terms of antibody modeling, generation, and optimization across different CDR types, sequence lengths, pre-training strategies, and input contexts. 1 Zhejiang University, Hangzhou, China, 310058 2 AI Lab, Research Center for Industries of the Future, Westlake University, Hangzhou, China, 310030 {wulirong, linhaitao, huangyufei, gaozhangyang, tancheng, liuyunfan, wutailin, stan.zq.li}@westlake.edu.cn Introduction Antibodies are a family of Y-shaped proteins produced by the immune system to recognize, bind, and neutralize pathogens (e.g., antigens) ( Basu et al. 2019 ) . Since the specificity of antibody-antigen binding is mainly determined by the Complementary Determining Regions (CDRs) in the antibody, how to design CDRs that bind to specific regions (epitopes) of an antigen becomes an important problem. Considering the enormous search space of over 20^{60} potential CDR sequence candidates and the high variability of CDR structures, it is not feasible to test all potential CDRs by experimental assays in the web laboratory, which calls for computational methods for antigen-binding antibody CDR design ( Tiller and Tessier 2015 ; Kuroda et al. 2012 ) . The computational antibody design has undergone a paradigm shift from energy-base models ( Adolf-Bryfogle et al. 2018 ; Lapidoth et al. 2015 ) to deep generative models ( Lin et al. 2024 ; Kong, Huang, and Liu 2022 ; Wu et al. 2024a ; Wu et al. 2024b ; Tan et al. 2024 ; Jin, Barzilay, and Jaakkola 2022 ) . RefineGNN ( Jin et al. 2021 ) is the first work to co-design CDR sequences and structures in an autoregressive manner, and its follow-up HSRN ( Jin, Barzilay, and Jaakkola 2022 ) unifies antigen-binding antibody design and docking in a hierarchical refinement framework. DiffAb ( Luo et al. 2022 ) adopts diffusion models ( Hoogeboom et al. 2021 ) to design the CDR sequences, \mathrm{C}_{\alpha} coordinates, and orientations in an iterative manner. Similarly, AbDiffuser ( Martinkus et al. 2024 ) incorporates family-specific priors into the diffusion process, which greatly improves generation efficiency and quality over DiffAb. To improve computational efficiency, MEAN ( Kong, Huang, and Liu 2022 ) replaces autoregressive decoding with progressive full-shot decoding and involves the light chain as a conditional input to generate CDRs. It is further extended as dyMEAN ( Kong, Huang, and Liu 2023 ) , an end-to-end full-atom antibody design method. Two recent works, ABGNN ( Gao et al. 2023 ) and HTP ( Wu and Li 2023 ) , focus on the benefits of antibody sequence pre-training and hierarchical pre-training, respectively. A detailed comparison of the characteristics of these methods has been summarized in Table. A1 of Appendix A . Figure 1: Correlation of the performance of sequence generation (Left) and structure generation (Right) w.r.t amino acid position in CDR-H3 for different lengths L from SAbDab. Despite the fruitful progress, existing approaches still face several challenges. The first issue is that it is hard to model complex CDRs with long sequences. Take MEAN as an example in Fig. 1 , it fails to model the types and coordinates of those amino acids in the middle of CDRs, especially for CDRs with long sequences. This is probably because the amino acids in the middle region of CDRs are far from the framework regions on either side and thus cannot acquire sufficient contextual information to make correct predictions. The second issue is that previous methods are mostly conditioned on pre-given epitopes ( Jin, Barzilay, and Jaakkola 2022 ; Kong, Huang, and Liu 2022 ) and take as input only the static interactions between known epitopes and antibody. However, antigen-antibody interactions may evolve dynamically with the generation of different CDRs, especially when epitopes are not known. Note that dynamism is a polysemous term; it does not refer to dynamic conformations in this paper, but rather to the dynamism of antigen-antibody interaction graphs. The third issue is that existing antibody optimization methods may generate non-specific antibodies that not only bind with the target antigen but also other antigens with high affinity, thereby losing the important property of antibody specificity. In this paper, we represent each antigen-antibody complex as an Attributed Heterogeneous Graph (AHG) and define E(3)-invariant node and edge features in terms of six aspects: category, position, distance, direction, angle, and orientation, respectively, in order to include more geometrical information. Moreover, we defined a variety of different edge relations to expand the sequence and structural contexts of residues. We propose a simple but efficient R elation- A ware A ntibody D esign (RAAD) framework that performs message passing for each relation separately and then jointly encodes different relational contexts into the final representations. RAAD takes the complete antigen (rather than known epitopes) as input and formulates antigen-antibody interaction dynamics modeling as an edge relation optimization problem. Furthermore, we propose a new evaluation metric for measuring antibody specificity by taking into account both the altered binding affinity of the optimized antibody to the target and non-target antigens. Moreover, a contrastive specificity-enhancing constraint is developed to improve the specificity of the optimized antibodies. Extensive experiments have demonstrated the superiority of RAAD in antibody modeling, generation, and optimization. Codes will be available at: https://github.com/LirongWu/RAAD . Figure 2: Structure of an antigen-antibody complex, where the antibody is symmetrically Y-shape, and we mainly focus on three CDRs of the variable domain in the heavy chain. Preliminary An antibody is a Y -shaped symmetric protein with two identical sets of chains, as shown in Fig. 2 . Each set contains a heavy chain and a light chain, either of which contains a variable domain ( V_{H} / V_{L} ) and several constant domains. Each variable domain can be further divided into four framework regions (FRs) and three complementarity-determining regions (CDRs), each occupying a contiguous subsequence. The CDRs, especially those in the heavy chain (i.e., CDR-H1, CDR-H2, CDR-H3), are key in determining the antigen-antibody binding. Therefore, we mainly focus on the design and optimization of 3 CDRs in heavy chains in this paper. Graph Construction We represent each antigen-antibody complex as a heterogeneous graph \mathcal{G}=(\mathcal{V},\mathcal{E}) , where \mathcal{V}=(\mathcal{V}_{H},\mathcal{V}_{L},\mathcal{V}_{A}) is the set of nodes (residues) of the heavy chain,

Read the complete text