Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies

TL;DR

This paper addresses the challenge of designing nature-like antibodies with high binding affinity under conflicting biophysical objectives. It proposes a three-stage framework that first pre-trains on large antibody sequence data, then transfers to a diffusion-based designer conditioned on antigen context, and finally applies Pareto-Optimal Energy Alignment (POEA) with online, temperature-scaled exploration to balance multiple energy-based objectives. By extending Direct Preference Optimization to multi-objective energy rewards and incorporating online exploration, the method generates a Pareto front of antibody designs that exhibit lower repulsion and higher attraction at the antigen interface. The results show improved energy rationality and binding potential compared to state-of-the-art baselines, highlighting the approach's potential to accelerate rational, energy-aware antibody design in practical settings.

Abstract

We present a three-stage framework for training deep learning models specializing in antibody sequence-structure co-design. We first pre-train a language model using millions of antibody sequence data. Then, we employ the learned representations to guide the training of a diffusion model for joint optimization over both sequence and structure of antibodies. During the final alignment stage, we optimize the model to favor antibodies with low repulsion and high attraction to the antigen binding site, enhancing the rationality and functionality of the designs. To mitigate conflicting energy preferences, we extend AbDPO (Antibody Direct Preference Optimization) to guide the model toward Pareto optimality under multiple energy-based alignment objectives. Furthermore, we adopt an iterative learning paradigm with temperature scaling, enabling the model to benefit from diverse online datasets without requiring additional data. In practice, our proposed methods achieve high stability and efficiency in producing a better Pareto front of antibody designs compared to top samples generated by baselines and previous alignment techniques. Through extensive experiments, we showcase the superior performance of our methods in generating nature-like antibodies with high binding affinity.

Figures (4)

• Figure 1: Illustration of an antibody binding to an antigen. The antibody's light and heavy chains are shown with their variable (V) and constant (C) regions. The third CDR in the heavy chain (CDR-H3), colored in orange, is critical for determining the binding affinity to the antigen.
• Figure 2: Frontiers of CDR-Ag $E_\text{att}$ and CDR-Ag $E_\text{rep}$ alignment and typical samples produced by different reward weightings in POEA. (A) is the reference CDR-H3 (colored in orange) from PDB ID 5nuz. (B) is the best CDR-H3 design generated by AlignAb with low overall energy and high similarity with the reference structure. (C) is the typical type of design when $E_\text{att}$ reward dominates, and often consists of large side chains and contains structural collisions. (D) is the typical type of design when $E_\text{rep}$ reward dominates, and often lack of side chains with weak binding with the antigen.
• Figure 3: Frontiers of CDR-Ag $E_\text{att}$ and CDR-Ag $E_\text{rep}$ alignment produced by different reward weightings in POEA with four PDB examples.
• Figure 5: Visualization of reference antibody for antigen (PDB ID 5nuz) and different antibodies designed by our method and other baselines. The designed CDR-H3 structures are colored in orange and the designed CDR-H3 sequences are recorded accordingly.

Theorems & Definitions (3)

• Definition A.1
• Lemma A.1
• Lemma A.2