Structure-Aware Antibody Humanization via a Novel Black-box Optimization Algorithm

ICLR 2026 Conference Submission13319 Authors

Keywords: humanization, antibody design

TL;DR: In this work, we humanize antibodies using a novel black-box optimization algorithm based on Monte Carlo Tree Search.

Abstract: Antibodies are essential for viral neutralization and therapeutic applications. However, non-human antibodies can provoke anti-antibody immune responses, prompting the development of humanization strategies to reduce immunogenicity. Most existing approaches focus primarily on improving humanness scores from a sequence perspective, often overlooking the structural stability of humanized antibodies, which is essential for preserving complementarity-determining region (CDR) conformations. To overcome this limitation, we propose Hu-MCTs, a two-stage framework comprising (1) a pretraining phase that learns latent representations of human antibody sequences, and (2) a humanization phase that optimizes murine sequences toward human-like sequences using a novel black-box optimization algorithm based on Monte Carlo Tree Search. This algorithm jointly considers humanness and structural integrity, particularly minimizing disruption to CDR conformations. Experimental results demonstrate that Hu-MCTs outperforms baseline methods by achieving higher humanness scores while better preserving CDR structural stability. Moreover, the generated sequences exhibit the highest biological plausibility scores, closely resembling natural antibodies. These results suggest that Hu-MCTs is an effective solution for humanizing antibodies while preserving key structural features for functionality.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 13319