[논문 리뷰] Iterative Refinement Graph Neural Network for Antibody Sequence-Structure Co-design

Wengong Jin, Jeremy Wohlwend외 2명|arXiv (Cornell University)|2021. 10. 09.

Monoclonal and Polyclonal Antibodies Research참고 문헌 41인용 수 66

한 줄 요약

요약: 이 논문은 RefineGNN을 소개한다. RefineGNN은 antibody 서열과 3D 구조를 공동 설계하는 반복적 정제 그래프 신경망으로, 언어 모델링, 항원 결합 설계, SARS-CoV-2 중화 작업에서 기준 모델보다 우수하다.

ABSTRACT

Antibodies are versatile proteins that bind to pathogens like viruses and stimulate the adaptive immune system. The specificity of antibody binding is determined by complementarity-determining regions (CDRs) at the tips of these Y-shaped proteins. In this paper, we propose a generative model to automatically design the CDRs of antibodies with enhanced binding specificity or neutralization capabilities. Previous generative approaches formulate protein design as a structure-conditioned sequence generation task, assuming the desired 3D structure is given a priori. In contrast, we propose to co-design the sequence and 3D structure of CDRs as graphs. Our model unravels a sequence autoregressively while iteratively refining its predicted global structure. The inferred structure in turn guides subsequent residue choices. For efficiency, we model the conditional dependence between residues inside and outside of a CDR in a coarse-grained manner. Our method achieves superior log-likelihood on the test set and outperforms previous baselines in designing antibodies capable of neutralizing the SARS-CoV-2 virus.

연구 동기 및 목표

• Motivate automated design of antibody CDRs by jointly modeling sequence and 3D structure.
• Develop a graph-generation approach that iteratively refines both sequence and structure.
• Enable conditional generation with fixed framework regions and multi-resolution context.
• Evaluate on language modeling, antigen-binding design, and SARS-CoV-2 neutralization optimization.

제안 방법

• Represent an antibody as a sequence-structure graph with node features for residues and edge features capturing spatial relations.
• Propose RefineGNN that alternates between predicting the next residue and refining the global 3D structure via iterative graph refinement.
• Use two separate MPNs to predict sequence labels and coordinates, enabling structured, rotation/translation-invariant losses.
• Incorporate coarse-grained context blocks to efficiently propagate information from long contexts.
• Extend to conditional generation with a fixed framework via attention and multi-resolution context (block-coarsened context).
• Apply ITA-based fine-tuning to optimize generated antibodies for desired properties (e.g., neutralization).

실험 결과

연구 질문

• RQ1 Can a graph-based iterative refinement approach jointly generate antibody sequences and structures more effectively than sequence-only or one-shot graph methods?
• RQ2 Does conditioning on a fixed framework region and multi-resolution context improve CDR generation quality and structural realism?
• RQ3 Can the model improve property-guided outcomes such as antigen binding and SARS-CoV-2 neutralization predictions?
• RQ4 Is iterative refinement capable of reducing errors cascading in autoregressive graph generation for antibodies?

주요 결과

모델	CDR-H1 PPL	CDR-H1 RMSD	CDR-H2 PPL	CDR-H2 RMSD	CDR-H3 PPL	CDR-H3 RMSD	AAR
LSTM	6.79	-	7.21	-	9.70	-	22.53%
AR-GNN	6.44	2.97	6.86	2.27	9.44	3.63	23.86%
RefineGNN	6.09	1.18	6.58	0.87	8.38	2.50	35.37%
RAbD	-	-	-	-	-	-	28.53%

• RefineGNN achieves lower perplexity than LSTM and AR-GNN on CDR-H1/H2/H3 (CDR-H3 perplexity: 8.38 for RefineGNN vs 9.70 LSTM and 9.44 AR-GNN).
• RefineGNN achieves substantially lower RMSD for CDR-H3 structure prediction (1.18 Å for CDR-H1, 0.87 Å for CDR-H2, 2.50 Å for CDR-H3) compared to AR-GNN (2.97, 2.27, 3.63 respectively).
• On antigen-binding design, RefineGNN attains the highest amino acid recovery (AAR) of 35.37% versus 22.53% (LSTM) and 23.86% (AR-GNN).
• In SARS-CoV-2 neutralization optimization, RefineGNN achieves a 75.2% average neutralization score after ITA finetuning, higher than LSTM (72.0%) and AR-GNN (70.4%).
• RefineGNN pretrained on SAbDab and finetuned with ITA on CoVAbDab shows better perplexity (7.86) and higher neutralization scores than baselines.