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

Yinghua Yao, Yuangang Pan, Xixian Chen

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.

Introduction

Designing effective therapeutic antibodies is a complex task often described as finding a "needle in a haystack." The primary objective is to engineer Complementarity-Determining Regions (CDRs) that exhibit high specificity for an antigen while maintaining essential "developability" traits, such as solubility, thermal stability, and low aggregation.

Current deep learning approaches for antibody design face two significant hurdles. First, the search space problem: optimizing directly in the raw data space (amino acid sequences and 3D coordinates) is computationally inefficient due to the vast, discrete nature of the space. Second, the differentiability problem: many critical biological property evaluators function as "black boxes," providing scores without mathematical gradients to guide optimization. LEAD addresses these challenges by shifting the optimization process into a shared latent space.

The LEAD Framework

LEAD (LatEnt blAck-box Design) is built upon diffusion models that jointly model antibody sequence and structure. The framework introduces three core innovations to improve efficiency and synchronization:

  1. Shared Latent Space Optimization: Instead of tweaking raw data, LEAD identifies a shared latent code ($Z^t$) produced by the model's encoder. This code captures the intrinsic relationship between sequence and structure. By updating this single code, LEAD ensures that modifications to the sequence and structure remain perfectly synchronized.
  2. Black-Box Guidance: To navigate non-differentiable property evaluators, LEAD employs a gradient estimation technique. The latent code is perturbed with random noise, and the resulting antibodies are evaluated. These rewards are then used to estimate an "optimal perturbation" ($\delta^*$) to guide the generation process.
  3. Strategic Efficiency: Guidance is applied only during the later stages of the denoising process (e.g., starting at step 50 of 100). This avoids wasting computational resources on the highly stochastic early stages of generation.
Overview of the LEAD framework showing the co-design process and the latent-space optimization step.

Experimental Evaluation

The framework was evaluated using a benchmark of 19 antibody-antigen complexes from the Structural Antibody Database (SabDab), including targets like influenza and SARS-CoV-2. The authors tested several selection strategies for perturbations, including LEAD-H (Hard selection of the best perturbation), LEAD-S (Soft sampling), and LEAD-W (Weighted average gradient estimation).

Results and Discussion

LEAD demonstrated superior optimization performance across all six CDR loops (H1-3, L1-3) for both single and multi-property objectives. A key finding was the significant efficiency gain: LEAD achieved better results using only 50 steps of guidance compared to the 100 steps required by raw-space methods, effectively halving the computational query cost.

Qualitative comparison of antibodies generated by LEAD versus baselines, showing improved hydropathy and binding energy (ΔΔG).

In multi-property tasks requiring a balance between stability ($\Delta\Delta G$) and solubility (Hydropathy), LEAD consistently identified superior Pareto-optimal solutions. Furthermore, the model maintained high structural integrity, with low Root Mean Square Deviation (RMSD) indicating that the optimized antibodies remained physically plausible and similar to natural frameworks.

Conclusion

LEAD represents a significant advancement in antibody co-design by demonstrating that latent space optimization is more efficient and effective than raw space manipulation. While currently dependent on the quality of the underlying diffusion model and requiring future wet-lab validation, the LEAD framework provides a robust template for other AI-driven molecular design tasks, including the engineering of RNA and other proteins.

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