Large language model surrogates for Bayesian optimization of antibody sequences
Year: 2026
Type: article
Abstract: Enhancing therapeutic antibody properties typically requires iterative sequence engineering and experimental validation, such as binding assays, making antibody optimization time-consuming and resource intensive. To address these challenges, we introduce AbL2BO, a Bayesian optimization (BO) framework integrating a large language model (LLM) with Bayesian fine-tuning. AbL2BO leverages prior knowledge embedded in an antibody-specific LLM and efficiently updates model parameters using Bayesian Low-Rank Adaptation (LoRA) with approximate inference, including the Laplace approximation or MC dropout. The fine-tuned LLM serves as a surrogate model within the BO framework. Our results show that AbL2BO achieves superior sample efficiency compared to conventional BO methods that rely on Gaussian process regression to model black-box functions. Across experimental and simulated benchmark datasets, AbL2BO consistently demonstrated competitive or superior optimization performance with better-calibrated uncertainty estimates than kernel-based BO approaches. Integrating LLMs into BO offers three key advantages: (1) warm-starting the optimization with pretrained knowledge; (2) easy adaptation to new domains; and (3) bypassing the complexities of kernel design choices, kernel computation, and sequence alignment to handle sequences of varying length. Overall, AbL2BO improves both optimization efficiency and uncertainty estimation, particularly for complex antibody sequence landscapes, with the potential to reduce time and resource requirements in therapeutic antibody development.
Source: Scientific Reports
Authors: Eunna Huh, Seungjin Choi, Hyunjin Shin
Institution: Seoul School of Integrated Sciences and Technologies
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7237355106
FWCI:
Cites:
Cited by:
Related to:
