Coevolution-informed Bayesian optimization for sample-efficient protein design

Evolution-inspired multi-objective Bayesian optimization for protein engineering

Sample Efficient Generative Optimization for Molecular Design

Molecular optimization in drug discovery, materials design, and catalysis requires searching vast chemical spaces under tight evaluation budgets, since high-fidelity oracles and experimental measurements are costly. The practical impact of an optimization method therefore hinges on its sample efficiency: how few evaluations it needs to find strong candidates. We introduce Sample Efficient Generative Optimization (SEGO), a framework for Bayesian optimization on adaptively generated molecules. In SEGO, a probabilistic surrogate model forms a hypothesis about where hits lie in chemical space, a generative model is steered to propose candidates in that region, the most promising candidate is selected via an acquisition function, and the resulting oracle call is used both to sharpen the surrogate and to anchor the generator in real reward. SEGO attains state-of-the-art performance on the practical molecular optimization (PMO) benchmark using only one tenth of the oracle calls consumed by other methods, and on a multiparameter docking task it reaches ten hits in roughly half the oracle calls of existing approaches. These gains move molecular optimization closer to campaigns driven by direct experimental feedback.

Combining Stability-Centered Atomistic Design with Machine Learning for Targeted Enzyme Optimization

FuncLib and high-throughput FuncLib (htFuncLib) generate diverse, functional protein libraries using a stability-centered design; however, this substrate-independent approach lacks target-specific functional constraints. We developed a machine-learning-assisted enzyme-engineering (MLEE) workflow that adds substrate-specific functional information to htFuncLib through an initial screening and sequencing round. The system was benchmarked using previously published four-position fitness landscapes of three different proteins. The MLEE workflow successfully generated compact libraries enriched in globally high-fitness variants. After the initial training phase, an MLEE-enriched library of just 12 variants increased the hit rate for the global top-0.05% variants by 5- to 12-fold relative to the htFuncLib baseline. Screening a larger set of 96 variants recovered at least one of these top-performing enzymes in 61.3–99.4% of the simulations. We then applied MLEE to MthUPO-catalyzed β-damascone hydroxylation. Across two rounds, 506 distinct variants were screened and sequenced. While the initial substrate-independent htFuncLib library yielded 14% of variants with activity above the wild type, the MLEE-enriched library increased this hit rate to 90% (97 of 108 variants) with activity above the wild type. The best variant increased the turnover number for 4-hydroxy-β-damascone by 11.8-fold and achieved >99% regioisomeric excess. MLEE may bypass the need for transition-state models and reduce the effort required for obtaining high-activity variants. TABLE OF CONTENT

Evolutionary‑Driven Bayesian Optimization for Automated Molecular Docking with AutoDock Vina

Defining the docking search space is a critical yet often overlooked step in molecular docking, especially for receptors that lack clear or well-structured binding pockets. We address this challenge by formulating grid box placement as a global, expensive black-box optimization problem and introduce Evolutionary Driven Bayesian Optimization (EA-BO), a surrogate-based framework designed for efficient exploration under strict evaluation budgets. EA-BO integrates Gaussian Process models with a Matérn 5/2 kernel, LBFGS-B hyperparameter tuning, and CMA-ES-driven acquisition maximization to balance exploration and exploitation in a computationally demanding setting. We evaluate EA-BO on the interleukin-6 receptor, where manual grid selection and standard heuristics frequently fail. Across a panel of resveratrol-like ligands selected through ECFP4-based similarity screening, EA-BO consistently identifies interaction hotspots and converges faster than Optuna, Gaussian Process Bayesian Optimization, and Scikit-Optimize, while also outperforming grid centers reported in previous IL-6R docking studies. As a second contribution, we leverage the optimized docking region to rank and select the best-performing ligand from the resveratrol analogue set based on binding affinity with the receptor, demonstrating how automated grid optimization directly facilitates ligand prioritization. These results demonstrate that EA-BO provides data-efficient strategy for locating promising docking regions when computational cost limit traditional approaches.

Rationalized All-Atom Protein Design with Unified Multi-Modal Bayesian Flow

Designing functional proteins is a critical yet challenging problem due to the intricate interplay between backbone structures, sequences, and side-chains. Current approaches often decompose protein design into separate tasks, which can lead to accumulated errors, while recent efforts increasingly focus on all-atom protein design. However, we observe that existing all-atom generation approaches suffering from an information shortcut issue, where models inadvertently infer sequences from side-chain information, compromising their ability to accurately learn sequence distributions. To address this, we introduce a novel rationalized information flow strategy to eliminate the information shortcut. Furthermore, motivated by the advantages of Bayesian flows over differential equation–based methods, we propose the first Bayesian flow formulation for protein backbone orientations by recasting orientation modeling as an equivalent hyperspherical generation problem with antipodal symmetry. To validate, our method delivers consistently exceptional performance in both peptide and antibody design tasks. Our code, checkpoint, and designed PDBs can be found in https://github.com/GenSI-THUAIR/ProBayes .

Accelerating protein engineering: an integrated framework combines protein language models and epistatic landscape modeling

In a recent study published in Science , Tran et al. introduced MULTI-evolve (model guided, universal, targeted installation of multimutants), a machine learning-guided directed evolution (MLDE) framework that rapidly designs hyperactive multimutant proteins. 1 By integrating protein language models (PLMs), epistasis-aware modeling, and a high-ef fi ciency multisite mutagenesis platform, the authors address the major combinatorial challenge that has long bottlenecked protein engineering and therapeutic development. Protein function is encoded by amino acid sequence, yet discovering productive mutational combinations remains dif fi cult because protein fi tness landscapes are high-dimensional and strongly shaped by epistatic interactions. Directed evolution has long provided a powerful route for protein optimization through repeated mutagenesis and screening. 2 More recently, MLDE has expanded accessible sequence space through in silico screening, but its performance still depends on accurately capturing epistatic interactions, the context-dependent effects between mutations that shape protein fi tness. 3 In particular, many existing approaches require large training datasets, multiple experimental rounds, or labor-intensive synthesis, and remain limited in extrapolating to high-order multimutants because of incomplete epistatic modeling. Tran et al. address these limitations with an end-to-end framework for direct multimutant exploration. MULTI-evolve is built on three conceptual components. First, the framework deploys a PLM zero-shot ensemble approach to nominate function-enhancing single mutations. By combining structure-informed and sequence-based models with z-score normalization, the authors improved the identi fi cation of productive mutations compared with individual PLMs alone. Second, they trained fully connected neural networks (FCNNs) on a compact dataset of experimentally characterized single and double mutants to learn epistatic interactions and predict higher-order variants. This data-ef