[Paper Review] Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization

Xiangxin Zhou, Dongyu Xue+4 more|arXiv (Cornell University)|Mar 25, 2024

Monoclonal and Polyclonal Antibodies Research7 citations

TL;DR

The paper proposes AbDPO, a direct energy-based preference optimization method that fine-tunes a pre-trained antibody diffusion model at the residue level to yield rational CDR-H3 structures with high antigen binding affinity, achieving state-of-the-art energy and binding performance on the RAbD benchmark.

ABSTRACT

Antibody design, a crucial task with significant implications across various disciplines such as therapeutics and biology, presents considerable challenges due to its intricate nature. In this paper, we tackle antigen-specific antibody sequence-structure co-design as an optimization problem towards specific preferences, considering both rationality and functionality. Leveraging a pre-trained conditional diffusion model that jointly models sequences and structures of antibodies with equivariant neural networks, we propose direct energy-based preference optimization to guide the generation of antibodies with both rational structures and considerable binding affinities to given antigens. Our method involves fine-tuning the pre-trained diffusion model using a residue-level decomposed energy preference. Additionally, we employ gradient surgery to address conflicts between various types of energy, such as attraction and repulsion. Experiments on RAbD benchmark show that our approach effectively optimizes the energy of generated antibodies and achieves state-of-the-art performance in designing high-quality antibodies with low total energy and high binding affinity simultaneously, demonstrating the superiority of our approach.

Motivation & Objective

• Tackle antibody design as a sequence-structure co-design problem focused on rationality and function.
• Leverage a pre-trained conditional diffusion model for CDR design conditioned on antigens.
• Introduce residue-level energy-based preference optimization to guide generation toward lower energy (better rationality) and higher binding affinity.
• Decompose and mitigate conflicts among multiple energy terms to improve optimization efficiency and outcomes.

Proposed method

• Pre-train a conditional diffusion model on real antigen–antibody data modeling CDR sequences and structures with SE(3)-equivariant networks.
• Fine-tune the diffusion model via direct energy-based preference optimization using residue-level energy signals.
• Decompose energy into multiple terms (e.g., CDR total energy, CDR–antigen attractive/repulsive energies) and apply gradient surgery to mitigate conflicts among terms.
• Formulate preference data as winning/losing antibody samples based on energy-driven rewards, and apply a residue-level DPO objective for efficient training.
• Use forward diffusion sampling to estimate KL-based divergences and compute the AbDPO loss that reweights gradient updates by residue-level rewards.
• Incorporate gradient surgery to reduce interference among energy types and enable stable multi-term optimization.

Experimental results

Research questions

• RQ1 Can direct energy-based preferences guide diffusion-based antibody design toward safer (lower energy) and more functional antibodies?
• RQ2 Does residue-level energy signaling improve rationality and binding functionality over amino acid recovery or RMSD-driven metrics?
• RQ3 How can multiple energy terms (attraction/repulsion) be decomposed and their conflicts mitigated during optimization?
• RQ4 What is the comparative performance of AbDPO against state-of-the-art sequence-structure co-design methods on the RAbD benchmark?

Key findings

• AbDPO outperforms baselines in both rationality (lower CDR total energy) and functionality (better CDR–antigen binding energy) on the RAbD benchmark.
• Energy decomposition with gradient surgery reduces interference between terms and improves stability and optimization efficiency.
• Residue-level preference signals enable fine-grained credit assignment and more effective optimization than whole-molecule rewards.
• On the evaluated 55 antigen targets, AbDPO achieves lower total energy and more favorable binding energy than baselines, indicating higher-quality antigen-specific antibodies.