Designing novel proteins with desired characteristics remains a significant challenge due to the large sequence space and the complexity of sequence-function relationships. Efficient exploration of this space to identify sequences that meet specific design criteria is crucial for advancing therapeutics and biotechnology. Here, we present BoGA (Bayesian Optimization Genetic Algorithm), a framework that combines evolutionary search with Bayesian optimization to efficiently navigate the sequence space. By integrating a genetic algorithm as a stochastic proposal generator within a surrogate modeling loop, BoGA prioritizes candidates based on prior evaluations and surrogate model predictions, enabling data-efficient optimization. We demonstrate the utility of BoGA through benchmarking on sequence and structure design tasks, followed by its application in designing peptide binders against pneumolysin, a key virulence factor of Streptococcus pneumoniae. BoGA accelerates the discovery of high-confidence binders, demonstrating the potential for efficient protein design across diverse objectives. The algorithm is implemented within the BoPep suite and is available under an MIT license at GitHub.
Figures & tables
Figure 1: BoGA couples evolutionary optimization with Bayesian selection for sequence design. a Schematic of the optimization cycle. From an initial evaluated dataset D0 , an elite set Sk is chosen from a leaderboard where each sequence is ranked by its fitness f(x) . A proposer (genetic mutations: substitutions/insertions/deletions) generates a pool of candidates X′ of size kpropose . Sequences are embedded and scored by a surrogate model f^θ ; an acquisition function α selects the mselect top candidates for explicit evaluation. The fitness function is determined by the user, and can utilize properties derived from the amino acid sequence, the folded structure, and/or interchain interactions defined by a folded complex. The newly scored sequences are added to the dataset Dt and the loop is repeated. b Conceptual search in sequence space. A standard GA (orange) evaluates many proposals, including those that decrease fitness, whereas BoGA (blue) uses the surrogate to discard low-value proposals (gray) and concentrates evaluations to proposals that are more likely to increase fitness. c Example of optimization trajectories when maximizing molecular weight, showing improved performance as kpropose increases.
Figure 2: BoGA improves optimization efficiency for sequence- and structure-level objectives. a Optimization of β -sheet fraction (fraction of E, M, A, and L residues). The left panel shows optimization trajectories for different values of kpropose , and the right panel shows the distribution of fitness values at various stages of optimization. The inset shows the surrogate model R2 over the course of optimization. b Similar layout as in a but the objective function is the normalized hydrophobic moment (uHrel). c Structure-guided optimization using AlphaFold 2-predicted secondary structure weighted by predicted TM-score (pTM). The left panel shows optimization trajectories and violinplots show the distribution of fitness values for the last 10 generations. The right panel shows a triplot of the secondary structures of samples for the run with kpropose=500 .
Figure 3: BoGA enables efficient design of peptide binders targeting pneumolysin. a Illustration of the neutralizing monoclonal antibody and PLY complex. PLY is shown in gray and the antibody in green. The model at the bottom represents domain 4. b Optimization trajectories for the binder score across 100 generations for kpropose=10 (teal) and kpropose=500 (purple). The points indicate individual evaluated candidates; solid lines are running top quartile means. Right: Kernel density estimates of the binding score distribution during early (0-10), intermediate (10-50), and late (50-100) generations, showing that larger proposal pools accelerate discovery of high-scoring binders. c Relationship between predicted interface pTM (ipTM) and peptide predicted aligned error (PAE) for evaluated sequences. Larger kpropose increases sampling of high-confidence, low-PAE binders. The dashed lines show cutoffs for ipTM=0.9 and PAE=5 Å. d Outline of the post-optimization refinement approach. The top 100 BoGA sequences were subjected to three rounds of sequence recovery using ProteinMPNN and FastRelax, resulting in a total of 400 candidates, which are subsequently re-docked to the full-length PLY and scored. e pLDDT versus ipTM for refined candidates. Points are colored by Boltz-2 interface distance score. High-confidence binders cluster at high ipTM and high pLDDT (upper right quadrant, dashed lines). f Example predicted complexes (Boltz-2) of top-performing binders (cyan) bound to PLY (gray). Filtering of the top candidates using quality, solubility and structural consistency yielded a final set of 41 high-confidence binders. g Orthogonal structure prediction using AlphaFold 3 and Boltz 2 for one of the top binders. Both models support a consistent binding pose with high interface confidence (ipTM = 0.85 for AlphaFold 3; ipTM = 0.92 for Boltz 2). Estimated binding free energy ( ΔG ) from PyRosetta [ 34 ] indicates a strongly favorable interaction.
Antibody lead optimization is inherently a multi-objective challenge in drug discovery. Achieving a balance between different drug-like properties is crucial for the development of viable candidates, and this search becomes exponentially challenging as desired properties grow. The ever-growing zoo of sophisticated in silico tools for predicting antibody properties calls for an efficient joint optimization procedure to overcome resource-intensive sequential filtering pipelines. We present BOAT, a versatile Bayesian optimization framework for multi-property antibody engineering. Our `plug-and-play' framework couples uncertainty-aware surrogate modeling with a genetic algorithm to jointly optimize various predicted antibody traits while enabling efficient exploration of sequence space. Through systematic benchmarking against genetic algorithms and newer generative learning approaches, we demonstrate competitive performance with state-of-the-art methods for multi-objective protein optimization. We identify clear regimes where surrogate-driven optimization outperforms expensive generative approaches and establish practical limits imposed by sequence dimensionality and oracle costs.
Jackie Rao, Ferran Gonzalez Hernandez, Leon Gerard +1
MRC Biostatistics Unit, University of Cambridge, UK · Centre for AI, Data Science and Artificial Intelligence, R&D, AstraZeneca, Barcelona, Spain
We introduce RosettaSearch, an inference-time multi-objective optimization approach for backbone conditioned protein sequence design. We use large language models (LLMs) as a generative optimizer within a search algorithm capable of controlled exploration and exploitation, using rewards computed from RosettaFold3, a structure prediction model, under a strict computational budget. In a large-scale evaluation, we apply RosettaSearch to 400 suboptimal sequences generated by LigandMPNN (a state-of-the-art model trained for protein sequence design), recovering high-fidelity designs that LigandMPNN's single-pass decoding fails to produce. RosettaSearch's designs show improvements in structural fidelity metrics ranging between 18% to 68%, translating to a 2.5x improvement in design success rate. We observe that these gains in success rate are robust when RosettaSearch-designed sequences are evaluated with an independent structure prediction oracle (Chai-1) and generalize across two distinct LLM families (o4-mini and Gemini-3), with performance scaling consistently with reasoning capability. We further demonstrate that RosettaSearch improves the sequence fidelity of ProteinMPNN designs for de novo backbones from the Dayhoff atlas, showing that the approach generalizes beyond native protein structures to computationally generated backbones. We also demonstrate a multi-modal extension of RosettaSearch with vision-language models, where images of predicted protein structures are used as feedback to incorporate structural context to guide protein sequence generation. To our knowledge, this is the first large-scale demonstration that LLMs can serve as effective generative optimizers for backbone-conditioned protein sequence design, yielding systematic gains without any model retraining.
Meghana Kshirsagar, Allen Nie, Ching-An Cheng +5
AI for Good, Microsoft Redmond · Google DeepMind · Google Research +3
Protein sequence optimization under tight oracle budgets requires methods that explore vast combinatorial spaces while making each evaluation informative. Existing reinforcement learning and off-policy generative approaches often degrade under surrogate noise, and position-agnostic mutation proposals risk disrupting functionally critical residues. We introduce SILO, a trajectory-level self-improvement imitation framework for oracle-budgeted protein design. SILO uses a hierarchical edit policy that decomposes each mutation into a position choice followed by a residue choice. In each active-learning round, the policy samples candidate trajectories via incremental stochastic beam search without replacement (SBS), and a UCB-based proxy ensemble, combined with an alanine-scan fitness score (AFS), selects candidates with functionally relevant edits for in silico oracle evaluation. The policy is then updated by next-action cross-entropy imitation on the round's best oracle-labeled trajectories, avoiding value-function estimation. Across eight reproduced protein fitness landscapes and five strong baselines from prior work, SILO achieves the highest maximum and top-100 mean fitness on 8 of 8 landscapes within our evaluations, often exhibiting faster early-stage improvement. In low-data and noisy-proxy stress tests on two landscapes per setting, SILO remains competitive or best when several baselines degrade. Ablations show that SBS with AFS account for much of the gains, with iterative imitation providing additional improvement. Code is available at: https://github.com/grimmlab/SILO.git
Ashima Khanna, Dominik Grimm
Bioinformatics, Technical University of Munich, TUM Campus Straubing, 94315 Straubing, Germany · University of Applied Sciences Weihenstephan-Triesdorf, 94315 Straubing, Germany