Generative Co-Design of Antibody Sequences and Structures via Black-Box Guidance in a Shared Latent Space

Yinghua Yao, Yuangang Pan, Xixian Chen

Abstract

Antibody engineering requires simultaneous optimization of sequence and structure to ensure both binding affinity and developability. Traditional generative models often struggle with inefficient search in high-dimensional raw data spaces and the non-differentiable nature of property evaluators. We present LEAD (LatEnt blAck-box Design), a framework that performs optimization within a shared latent space of sequence and structure. By leveraging black-box guidance, LEAD efficiently navigates the latent manifold to produce antibody candidates with superior stability and solubility. Experimental results on the SabDab benchmark demonstrate that LEAD achieves state-of-the-art performance while requiring significantly fewer evaluator queries compared to raw-space optimization methods.

Introduction

Designing therapeutic antibodies is a complex balancing act. Beyond binding to a specific antigen, a successful candidate must possess "developability" traits, such as thermal stability, low aggregation propensity, and high solubility. Current AI-driven approaches often utilize diffusion models to generate candidates. However, optimizing these designs for specific biological properties remains challenging due to the vast, discrete nature of the raw sequence and coordinate space. Furthermore, many high-fidelity property evaluators are "black boxes" — they are non-differentiable, preventing the use of standard gradient-based optimization techniques.

Methods: The LEAD Framework

To address these challenges, the LEAD framework operates on a pre-trained diffusion model (DiffAb) but shifts the optimization process from the raw data space to a shared latent space.

Shared Latent Space

Instead of treating amino acid sequences and 3D structures as independent variables, LEAD employs a shared encoder $\mathcal{E}$ to map both modalities into a unified latent code $Z$. This synchronization ensures that any optimization step maintains physical consistency between the sequence and its corresponding structure.

Overview of the LEAD framework for latent-space black-box optimization.

Black-Box Guidance Mechanism

LEAD optimizes designs without requiring gradients from the property evaluator $\mathcal{F}$. The process involves perturbing the latent code $Z^t$ with noise to generate a set of candidate perturbations, decoding these into sequence/structure pairs, and scoring them using the property evaluator. The latent code is then moved toward regions associated with higher rewards. To maximize efficiency, optimization begins at intermediate denoising steps (e.g., step 50), focusing on structured designs rather than initial noise.

Experimental Results

The framework was evaluated using 19 antibody-antigen complexes from the Structural Antibody Database (SabDab), including targets for influenza and SARS-CoV-2. Performance was measured via Hydropathy scores (solubility/aggregation proxy) and Folding Energy ($\Delta\Delta G$, structural stability).

Performance and Efficiency

LEAD consistently outperformed GuideRaw — a baseline that optimizes directly in the raw data space — across all six Complementarity-Determining Regions (CDRs). Notably, LEAD achieved these superior results while utilizing 50% fewer queries to the property evaluators. This efficiency is critical in drug discovery pipelines where evaluations can be computationally intensive.

Comparison of antibody designs generated by LEAD versus baseline methods.

Multi-Property Optimization

LEAD demonstrated the ability to navigate trade-offs between competing properties. By finding Pareto-optimal designs, the model successfully generated antibodies that were both highly stable and highly soluble, a key requirement for therapeutic developability.

Conclusion

LEAD introduces a robust method for antibody co-design by leveraging the "semantics" of a shared latent space. By decoupling the optimization from the raw data representation and the differentiability of the evaluator, it provides a flexible, "plug-and-play" solution for protein engineering. While current results are simulated (in silico), the framework's efficiency and performance suggest significant potential for accelerating the design of developable therapeutic antibodies.

Summary updated

Abstract

Advancements in deep generative models have enabled the joint modeling of antibody sequence and structure, given the antigen-antibody complex as context. However, existing approaches for optimizing complementarity-determining regions (CDRs) to improve developability properties operate in the raw data space, leading to excessively costly evaluations due to the inefficient search process. To address this, we propose LatEnt blAck-box Design (LEAD), a sequence-structure co-design framework that optimizes both sequence and structure within their shared latent space. Optimizing shared latent codes can not only break through the limitations of existing methods, but also ensure synchronization of different modality designs. Particularly, we design a black-box guidance strategy to accommodate real-world scenarios where many property evaluators are non-differentiable. Experimental results demonstrate that our LEAD achieves superior optimization performance for both single and multi-property objectives. Notably, LEAD reduces query consumption by a half while surpassing baseline methods in property optimization. The code is available at https://github.com/EvaFlower/LatEnt-blAck-box-Design.
BibTeX
@inproceedings{yao2025generative,
  author = {Yinghua Yao and Yuangang Pan and Xixian Chen},
  title = {{Generative Co-Design of Antibody Sequences and Structures via Black-Box Guidance in a Shared Latent Space}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1050},
}
Figure