MutexaGPT Turns Plain-English Enzyme Hunches Into Tested Designs
TL;DR: Researchers published an open-access, multi-agent language-model platform called MutexaGPT that turns plain-English enzyme-engineering intuition into physics-based simulations and smart mutation libraries.
On 10 September 2026, a team led by Zhongyue J. Yang published a Resource called MutexaGPT: an intuition-to-design translator for physics-based enzyme engineering in Nature Computational Science as open access, according to the journal page. The author list names Qianzhen Shao, Yinjie Zhong, Sebastian Stull, Xinchun Ran, Ning Ding, Kieran Nehil-Puleo, Ruizhe Yao, Han Xu and Zhongyue J. Yang, according to the same journal page. The author-affiliation snapshot identifies Vanderbilt University for the listed contributors and identifies Zhongyue John Yang as corresponding author, according to the publication record. The journal's research-articles index summarizes the contribution as turning plain-language enzyme-engineering intuitions into lead mutation designs through AI agent-orchestrated high-throughput molecular modeling, with experimentally validated improvements in enzyme specificity and cold activity, according to Research articles | Nature Computational Science.
What the team actually built
MutexaGPT is described as an open-access, multi-agent large language model platform that translates enzyme-engineering intuition to physics-based simulations and thus variant designs, according to the journal page. Through a web interface, MutexaGPT takes users' intuition-driven requests expressed in plain English as input, according to the publication record. It then leverages its language-model agents to comprehend the request and elicit missing information, according to the publication record. It constructs physics-based models, according to the journal page. It configures and executes high-throughput molecular-modeling workflows, according to the journal page. It eventually converts the molecular-modeling results into actionable design proposals such as a smart mutation library, according to the publication record.
The three named agents are QuestionAnalyzer, WorkPlanningBoard and ResultExplainer, according to the publication record. QuestionAnalyzer, WorkPlanningBoard and ResultExplainer are used to comprehend the request and elicit missing information, construct physics-based models, configure and execute high-throughput molecular-modeling workflows, and convert results into actionable proposals, according to the publication record. An automated evaluation framework was established to systematically benchmark the prompt-engineering strategies that allow each individual agent to achieve an optimal performance, according to the publication record.
The problem: good intuition, no translation manual
Physical intuition about how enzyme structure and dynamics shape function has guided successful engineering efforts, yet a systematic approach is still lacking for translating these qualitative and abstract thoughts into quantitative, actionable principles for enzyme design, according to the journal page. Physical intuition has been a cornerstone of enzyme engineering, according to the journal page. It arises from accumulated knowledge and experience in rationalizing how enzymes realize their catalytic functions through three-dimensional structural topology, conformational dynamics, electronic structure and interactions with the surrounding environment, according to the journal page.
The structural-topology examples listed are active-site cavities and tunnels, according to the journal page. The conformational-dynamics examples listed are flexibility and domain motion, according to the journal page. The electronic-structure examples listed are electrostatics and electric fields, according to the journal page. The environment example listed is hydration shell, according to the journal page. Such intuition helps protein engineers infer or anticipate the impact of mutations on enzyme activity, temperature adaptation, selectivity, promiscuity and stability of enzymes, according to the journal page.
The translation difficulty is that most intuitive catalytic principles in enzymology are qualitative rather than quantitative, according to the journal page. The paper cites Gordon Hammes' notion that for catalysis of an efficient enzyme, flexible but not too flexible, as well as rigid but not too rigid, is essential, according to the journal page. That conceptual statement captures a truth yet provides no guidance for identifying the optimal degree of flexibility, according to the journal page. Even when an intuition has a quantitative foundation, such as the correlation between electric field strength and catalytic rate enhancement, it remains unknown how specific mutations should be introduced to achieve the desired electrostatic environment, according to the journal page.
As a result, the unique intuition or stroke-of-genius inspiration from protein engineers always fails to get translated into a smart mutational library or a concrete protein-engineering strategy, according to the journal page. That failure makes it inevitable to adopt a hypothesis-agnostic approach by screening against a larger library for mutant discovery, according to the journal page. Screening a larger library leads to greater labor and resource cost and longer enzyme-development cycle in most enzyme-engineering scenarios, according to the journal page.
How engineers have used intuition before
Steric intuition in amine transaminases has led to a design rule for tuning stereoselectivity, according to the journal page. The rule involves adjusting the volumes of the large and small binding pockets within the active site through site-directed mutagenesis to better match the sizes of the corresponding ketone substituents, according to the journal page. That matching can invert selectivity or expand substrate scope, according to the journal page. Such intuition also strengthens the screening-based, directed-evolution approaches for effective function-enhancing mutant discovery, according to the journal page.
Reetz and coworkers exemplified this through developing the combinatorial active-site saturation test method, known as CASTing, according to the journal page. The CASTing method applies targeted saturation mutagenesis to small residue sets and recombines beneficial variants to capture cooperative effects for enhancing activity and selectivity toward non-native substrates, according to the journal page. In essence, physical intuition guides the construction of smart mutational libraries and facilitates effective enzyme engineering when high-throughput screening assays are not available, according to the journal page.
The pressing challenge facing the entire field of enzyme engineering is how to establish an enzyme-engineering paradigm that translates physical intuitions into quantitative, actionable design principles, according to the journal page. One promising approach to address this challenge requires high-throughput molecular simulations, abbreviated as htMS, according to the journal page. Those simulations quantify how mutations alter enzyme structure, dynamics and energetics to predict their effects on catalytic activity and selectivity, according to the journal page.
Where MutexaGPT sits beside SubTuner and CASCO
SubTuner leveraged htMS, specifically quantum mechanics and molecular dynamics simulations, to translate physical hypotheses into quantitative metrics in the in silico screening workflow, according to the journal page. The physical hypotheses included optimal transition-state binding and electric-field stabilization, according to the journal page. That workflow eventually yielded smart libraries with approximately ten mutants for each task, according to the journal page. Those libraries constantly achieved about 30% hit rate across three transferase-engineering tasks, according to the journal page.
Similarly, CASCO leveraged high-throughput molecular dynamics simulations to quantify substrate-positioning dynamics as near-attack-conformation frequencies in the engineering of enantio-selectivity, according to the journal page. The journal page introduces MutexaGPT as the next step in that line of work: an open-access, multi-agent platform that starts from plain-English intuition rather than from a manually configured simulation setup, according to the journal page. The publication record frames the advance as integrating human creativity with high-throughput molecular modeling to democratize physics-guided, intuition-driven enzyme engineering, according to the publication record.
Inside the workflow: from sentence to smart library
The input is a plain-English, intuition-driven request entered through a web interface, according to the journal page. The agents elicit missing information, according to the journal page. The system constructs physics-based models, according to the journal page. The system configures and executes high-throughput molecular-modeling workflows, according to the journal page. The system converts the results into actionable design proposals, such as smart mutation libraries, according to the journal page.
The publication record describes the same sequence as comprehending the request, eliciting missing information, constructing physics-based models, configuring and executing high-throughput molecular-modeling workflows, and converting molecular-modeling results into actionable design proposals, according to the publication record. The actionable proposal example given in both summaries is a smart mutation library, according to the journal page and the publication record. The automated evaluation framework systematically benchmarked the prompt-engineering strategies that allow each individual agent to achieve an optimal performance, according to the publication record. The named agents covered by that benchmarking were QuestionAnalyzer, WorkPlanningBoard and ResultExplainer, according to the publication record.
Test one: making room for bulkier substrates
The team demonstrated the utility of MutexaGPT in two protein-engineering tasks, according to the journal page. The first task was engineering halide methyltransferase toward bulkier substrates, according to the journal page. The second task was engineering bidomain amylase for enhanced activity at lower temperature, according to the journal page.
For the halide methyltransferase task, MutexaGPT converted a cavity-engineering intuition into a smart library design, according to the publication record. That smart library design shows a 40% hit rate, according to the publication record. It shows around 4-fold activity improvement over a baseline strategy, according to the publication record. The cavity-engineering framing connects to the broader structural-topology intuition about active-site cavities and tunnels described in the paper, according to the journal page. The publication record presents the 40% hit rate and around 4-fold improvement as the reported outcome for that cavity-to-library translation, according to the publication record.
Test two: cold-active amylase validated in the lab
For the task of engineering bidomain amylase for enhanced activity at lower temperature, MutexaGPT translated a statistics-based intuition into cold-adapted amylase variants, according to the publication record. Those variants show 1.7-fold and 3.7-fold activity enhancement at 0 °C, according to the publication record. The amylase outcome is described as experimentally validated, according to the publication record. The temperature-adaptation framing connects to the paper's discussion of intuition about temperature adaptation, conformational dynamics including flexibility and domain motion, and the difficulty of choosing an optimal degree of flexibility, according to the journal page.
The journal's index likewise describes the overall study as yielding experimentally validated improvements in enzyme specificity and cold activity, according to Research articles | Nature Computational Science. The two reported demonstrations together are presented as establishing MutexaGPT as an intuition-to-design translator that integrates human creativity with high-throughput molecular modeling to democratize physics-guided, intuition-driven enzyme engineering, according to the journal page.
Who did the work and how to see it
The work appeared as a Resource that is open access and was published on 10 September 2026, according to the journal page. The listed authors are Qianzhen Shao, Yinjie Zhong, Sebastian Stull, Xinchun Ran, Ning Ding, Kieran Nehil-Puleo, Ruizhe Yao, Han Xu and Zhongyue J. Yang, according to the journal page. The affiliation snapshot places Qianzhen Shao, Yinjie Zhong, Sebastian L. Stull, Xinchun Ran, Ning Ding, Kieran Nehil-Puleo, Ruizhe Yao and Han Xu at Vanderbilt University and identifies Zhongyue John Yang of Vanderbilt University as corresponding author, according to the publication record. The topic tags attached to the record are Enzyme Catalysis and Immobilization, Protein Structure and Dynamics, and Machine Learning in Bioinformatics, according to the publication record.
Funding sources listed are the National Institute of General Medical Sciences, the Office of Advanced Cyberinfrastructure, the National Science Foundation and the National Institutes of Health, according to the publication record. The full text is available through the open-access article page, according to the journal page. Readers can follow the data and method details through MutexaGPT: an intuition-to-design translator for physics-based enzyme engineering and the journal index at Research articles | Nature Computational Science.