Research map / Probabilistic and scientific modeling / Bayesian optimization
Tree Search-Based Evolutionary Bandits for Protein Sequence Optimization
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 a tree-search-based bandit method (TreeNeuralUCB/TreeNeuralTS) for protein sequence optimization that grows a mutation/recombination tree guided by UCB or Thompson Sampling scores, with a Bayesian regret bound under GP assumptions, evaluated on simulated oracles from AAV, TEM, and AAYL49 datasets.
Problem
Efficiently exploring the vast protein sequence space in iterative directed-evolution screens while keeping the number of mutations small and retaining theoretical guarantees.
Approach
Combines local tree search (mutation and recombination from lead sequences, mimicking directed evolution) with bandit scoring (UCB or Thompson Sampling over neural network embeddings) to guide which nodes are expanded, and proves a Bayesian regret bound of order gamma_T sqrt(T) for a GP-prior simplification.
Key results
- TreeNeuralUCB achieves average max fitness of 4.3723 on AAV versus 3.9286 for PEX and 3.7697 for NeuralUCB-DE after 10 rounds of 100 queries.
- Tree search methods keep mutation counts within single digits (about 5 on AAV) while NeuralUCB-DE and NeuralTS-DE have very large mutation counts on longer TEM and AAYL49 sequences.
- Theory gives a Bayesian regret bound of order gamma_T sqrt(T), matching GP bandit results up to the local-search term.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use tree search with UCB instead of non-tree-search baselines | TreeNeuralUCB · AAV, TEM, and AAYL49 | maximum fitness | Helped |
| Use tree search with TS instead of non-tree-search baselines | TreeNeuralTS · AAV, TEM, and AAYL49 | maximum fitness | Helped |
| Add tree search to NeuralTS-DE (ablation) | NeuralTS-DE · AAV, TEM, and AAYL49 | maximum fitness | Helped |
| Add tree search to NeuralUCB-DE (ablation) | NeuralUCB-DE · AAV, TEM, and AAYL49 | maximum fitness | Helped |
| Add a small positive recombination rate (ablation) | TreeNeuralBandit · AAV | fitness | Helped |
| Increase recombination rate beyond the beneficial range (ablation) | TreeNeuralBandit · AAV | fitness | Hurt |
| Keep previous parent/root nodes in the active set (ablation) | TreeNeuralBandit · AAV | fitness | Helped |
| Add a diversity bonus (ablation) | TreeNeuralBandit · AAV | average fitness | Hurt |
| Increase the diversity bonus weight q (ablation) | TreeNeuralBandit · AAV | diversity | Helped |
| Use tree search instead of uniform mutation and recombination (ablation) | NeuralTS-DE and NeuralUCB-DE · AAV, TEM, and AAYL49 | mutation count | Helped |
Surprising findings
- Adding recombination helps only at small rates; performance declines at larger recombination rates, falling below the zero-recombination variant.
Limitations
- Oracles (and PEX/AdaLead oracles and pretrained PLMs) do not accurately predict full fitness landscapes, so results depend on simulated screens.
- Bayesian regret bound does not directly give a frequentist guarantee.
- No wet-lab validation yet; the authors note wet-lab experiments are slow.
Experimental setup
Datasets: AAV TEM-1 beta-lactamase AAYL49 anti-SARS-CoV-2 antibody
Benchmarks: AAV oracle TEM oracle AAYL49 oracle 3gfb oracle
Metrics: max fitness per round mutation count Bayesian regret
Baselines: PEX AdaLead Bayesian Optimization DbAS LaMBO NeuralUCB-DE NeuralTS-DE
Builds on: NeuralUCB Thompson Sampling GP-UCB AdaLead PEX
Closest papers by meaning
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.88)
- Evolution-Inspired Loss Functions for Protein Representation Learning (ICML 2024 · similarity 0.88)
- A Variational Perspective on Generative Protein Fitness Optimization (ICML 2025 · similarity 0.88)
- Improving protein optimization with smoothed fitness landscapes (ICLR 2024 · similarity 0.88)
- ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 2025 · similarity 0.87)
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.87)
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.