Generative Humanization for Therapeutic Antibodies

TL;DR

Antibody immunogenicity risk complicates development and traditional humanization methods yield limited viable candidates or degrade function. The authors reframe humanization as a conditional sampling problem from a masked language model trained on human antibodies and augment sampling with product-of-experts that incorporate therapeutic attributes such as binding affinity and thermostability. They demonstrate in silico generation of large, diverse sets of highly humanized antibodies and validate lab performance in real programs, showing improved or preserved target binding when guided by affinity and structure oracles. The approach supports iterative optimization, enabling rapid exploration of humanness and developability trade-offs to accelerate antibody therapy development.

Abstract

Antibody therapies have been employed to address some of today's most challenging diseases, but must meet many criteria during drug development before reaching a patient. Humanization is a sequence optimization strategy that addresses one critical risk called immunogenicity - a patient's immune response to the drug - by making an antibody more "human-like" in the absence of a predictive lab-based test for immunogenicity. However, existing humanization strategies generally yield very few humanized candidates, which may have degraded biophysical properties or decreased drug efficacy. Here, we re-frame humanization as a conditional generative modeling task, where humanizing mutations are sampled from a language model trained on human antibody data. We describe a sampling process that incorporates models of therapeutic attributes, such as antigen binding affinity, to obtain candidate sequences that have both reduced immunogenicity risk and maintained or improved therapeutic properties, allowing this algorithm to be readily embedded into an iterative antibody optimization campaign. We demonstrate in silico and in lab validation that in real therapeutic programs our generative humanization method produces diverse sets of antibodies that are both (1) highly-human and (2) have favorable therapeutic properties, such as improved binding to target antigens.

Figures (9)

• Figure 1: Sampling-based approaches generate many candidates with increased humanness from both murine (a) and manually-humanized (b) starter antibodies. Allowing up to 2 CDR mutations but measuring resulting affinity in the lab gives improved humanness (c) and a range of affinities (d) including some equal to or better than the starter.
• Figure 2: Guided sampling generates sequences which balance MLM log likelihood ($\sim$ humanness) and predicted target binding affinity yielding higher-affinity samples compared to unguided sampling (left, contours indicate areas of high sample density, middle and right affinity and log likelihood marginals respectively).
• Figure 3: Guided sampling can flexibly capture multiple desired properties. Multioracle guidance (a) improves predicted affinity (left), thermostability (middle) and maintains log likelihood (right). Even in the absence of a trained oracle, sampling guided to minimize structural perturbation (b) yields samples which less frequently ablate target binding.
• Figure 4: In lab validation, guided generative humanization yields improved binding affinity. Affinity guidance outperforms both unguided sampling, as well as unguided samples which are ranked post-hoc by an affinity oracle.
• Figure 5: Log-likelihood under the MLM is correlated with the OASis percentile score.