Research map / Probabilistic and scientific modeling / Bayesian optimization
Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CloneBO, a Bayesian optimization method that uses CloneLM, a clonal-family language model, as a prior and twisted SMC to condition generated clonal families on lab measurements to suggest antibody mutations.
Problem
Iterative lab optimization of antibodies for binding and stability often fails on a limited experimental budget because the space of typical antibody mutations is huge and priors from antibody databases are not informed by optimization dynamics.
Approach
Treat clonal families from the immune system as draws from a fitness-function prior: train an autoregressive LM (CloneLM) on clonal families, sample clones containing the candidate via martingale posterior, condition them on measurements with twisted sequential Monte Carlo, and Thompson-sample mutations from the implied fitness.
Key results
- CloneBO reaches mean best normalized log10 K_D of 0.3713 and melting temperature 72.92 in the in silico table, above LaMBO-Ab (0.3381, 70.44) and other baselines.
- CloneLM achieves test perplexity 1.276 on heavy chain and 1.267 on light chain clonal family data.
- In wet lab, CloneBO's strongest binder is beaten by only 2 of 997 previously measured sequences, and its sequences are significantly stronger binders than LaMBO-Ab (Mann-Whitney p=0.021).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CloneBO instead of LaMBO-Ab | Antibody optimization · Clone oracle | log p(X|clone): -24.51 -> -7.29 | Helped |
| Use CloneBO instead of LaMBO-Ab | Antibody optimization · Binding oracle | Norm. log10 K_D: 0.3381 -> 0.3713 | Helped |
| Use CloneBO instead of LaMBO-Ab | Antibody optimization · Stability oracle | Melting temp.: 70.44 -> 72.92 | Helped |
| Add large clonal families (use M greater than 1 instead of M=1) (ablation) | CloneBO · In silico binding and stability optimization | best achieved objective | Helped |
| Use twisted SMC instead of naive importance sampling (ablation) | CloneBO · In silico binding and stability optimization | best achieved objective | Helped |
| Condition on previous experimental measurements (ablation) | CloneBO · In silico binding and stability optimization | best achieved objective | Helped |
| Use twisted SMC instead of unconditional importance sampling | CloneLM posterior sampling · Laboratory melting-temperature data | posterior fit to measurements | Helped |
| Correct twisted-sampling bias with D=4 particles | Twisted SMC · Laboratory melting-temperature data | posterior fit to measurements | Helped |
| Use CloneBO instead of LaMBO-Ab | Wet-lab antibody design · In vitro binding | binding strength | Helped |
| Use CloneBO instead of LaMBO-Ab | Wet-lab antibody design · In vitro synthesis | predicted synthesizability | Helped |
| Set the starting pool size K to 1 (ablation) | CloneBO · In silico binding optimization | best achieved objective | Hurt |
| Vary the maximum allowed mutations L (ablation) | CloneBO · In silico optimization | best achieved objective | No clear effect |
Limitations
- Sequences from the real iterative optimization experiment are proprietary and not released.
- Cost of sampling from the CloneBO posterior scales with the number of measurements N, requiring conditioning on a subset.
- Assumes an affine relationship between clonal fitness and lab measurements; in vitro validation covered only one wet-lab iteration.
Experimental setup
Datasets: Observed Antibody Space (OAS) lab measurements of VHH binding and melting temperature (in-house optimization experiment) CoVAbDab
Benchmarks: in silico clonal family fitness oracle in silico binding and stability oracles from lab measurements SARS CoV1/CoV2 binding oracle in vitro wet lab antibody optimization
Metrics: best objective value vs number of measurements test perplexity Mann-Whitney p-value for synthesizability and binding
Baselines: Greedy Sapiens LaMBO LaMBO-Ab Genetic AdaLead EvoBO CMA-ES Dyna-PPO CbAS DiffAb
Builds on: LaMBO Mistral-7B IgLM FastBCR
Closest papers by meaning
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.89)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.89)
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.88)
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.87)
- A Simple yet Effective $\Delta\Delta G$ Predictor is An Unsupervised Antibody Optimizer and Explainer (ICLR 2025 · similarity 0.87)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.