BOAT: Navigating the Sea of In Silico Predictors for Antibody Design via Multi-Objective Bayesian Optimization

The Thesis

Designing therapeutic antibodies requires balancing dozens of competing properties — binding affinity, stability, low immunogenicity — simultaneously, and the standard approach of optimizing one property at a time through sequential filtering is slow and wasteful. BOAT addresses this by coupling a surrogate model (a cheaper, uncertainty-aware stand-in for expensive lab or computational assays) with a genetic algorithm to explore antibody sequence space more efficiently. The framework is 'plug-and-play' in that any existing in silico predictor can be slotted in as an objective, which is practically useful given the proliferation of specialized antibody property tools. The catch is that the authors honestly identify hard limits: the approach degrades as sequence dimensionality grows and as oracle (predictor) call budgets shrink, meaning it is not a universal replacement for generative deep-learning methods. Where it shines is in low-to-moderate budget regimes where running many expensive generative model queries is not feasible.

Catalyst

The last three years have produced a flood of specialized in silico antibody property predictors — for developability, immunogenicity, polyspecificity, and more — but no principled framework for jointly optimizing across them. Simultaneously, Bayesian optimization methods have matured enough to handle the discrete, high-dimensional sequence spaces common in protein engineering. The combination of abundant predictors and mature surrogate modeling makes this the right moment to build an orchestration layer on top of them.

What's New

Prior antibody optimization pipelines typically ran predictors in sequence: filter for binding, then filter survivors for stability, then for immunogenicity, discarding candidates that fail any single stage. Genetic algorithm-only approaches optimize heuristically but ignore prediction uncertainty. Newer generative learning methods like protein language model fine-tuning can propose novel sequences but are expensive to query and require large compute budgets. BOAT instead builds a probabilistic surrogate for each objective — modeling both the predicted value and its uncertainty — and uses that to guide a genetic algorithm toward the Pareto front (the set of sequences where no single property can be improved without hurting another), achieving competitive results at lower oracle-call cost in the regimes the paper tests.

The Counter

The honest version of this paper is that it benchmarks a sensible engineering combination — Bayesian optimization plus genetic algorithms — against other purely computational methods, and finds it is competitive in some regimes. That is not surprising; Bayesian optimization has been the standard-of-care for expensive black-box optimization for a decade, and applying it to antibody sequences is a reasonable extension, not a conceptual leap. The paper's own results show that generative deep-learning methods outperform BOAT when compute budget is unconstrained, which is increasingly the actual situation in well-funded drug discovery labs. There is no wet-lab data, so the entire claim rests on the quality of the in silico predictors — and every practitioner in computational drug discovery knows those predictors have significant error rates in out-of-distribution sequence regions, exactly where optimization tends to push you. The 'plug-and-play' framing is appealing but understates the integration work: each oracle needs calibrated uncertainty estimates for the surrogate to be trustworthy, and most published antibody property tools were not designed with that in mind. Finally, several large pharma companies and well-capitalized startups already have internal multi-objective optimization platforms; BOAT would need to demonstrably outperform those on real drug programs, not just published benchmarks, to change workflows.

The Paper