TL;DR: A new automated pipeline hunts and tunes bottleneck enzymes in silico before any bench work
TL;DR: Researchers described GDEE, an automated structure-based platform that sources or mutates enzyme sequences, predicts atomistic structures, docks ligands, and ranks variants for experimental validation, with machine-learning rescoring to better predict mutation effects on binding.
According to the Journal of Chemical Information and Modeling record for GDEE: A Structure-Based Platform for Gene Discovery and Enzyme Engineering, the work was published on 2026-09-22 by Caio S. Souza, João P. G. Correia, Isabel Rocha, Diana Lousa and Cláudio M. Soares of Universidade Nova de Lisboa. According to that same record, the project was funded by Fundação para a Ciência e a Tecnologia and indexed under Enzyme Catalysis and Immobilization, Microbial Metabolic Engineering and Bioproduction, and Enzyme Structure and Function. The paper presents GDEE as a tool that finds and optimizes enzymes responsible for rate-limiting steps in biosynthetic pathways, with stated objectives ranging from improving catalytic efficiency to enabling novel transformations.
What was published on September 22
A Journal of Chemical Information and Modeling paper introduces GDEE
The publication identified in the Exa library as GDEE: A Structure-Based Platform for Gene Discovery and Enzyme Engineering appeared in the Journal of Chemical Information and Modeling with a publication date of 2026-09-22. The author list given in that record is Caio S. Souza, João P. G. Correia, Isabel Rocha, Diana Lousa and Cláudio M. Soares, all associated in the record with Universidade Nova de Lisboa. The ACS Publications listing for GDEE: A Structure-Based Platform for Gene Discovery and Enzyme Engineering | Journal of Chemical Information and Modeling | ACS Publications carries the same author names and the same 2026-09-22 publication timestamp.
The record describes the contribution as an automated platform for gene discovery and enzyme engineering aimed at overcoming bottlenecks in biocatalysis. The abstract text indexed for the paper states that enzymes are extensively utilized for catalyzing the production of value-added compounds in diverse sectors with significant economic impact. That framing places GDEE in an industrial biotechnology context rather than as a purely theoretical modeling exercise.
The supporting file listing for the paper, identified as GDEE supporting information, carries the same title, author set and 2026-09-22 date stamp in the ACS system. The presence of a supporting-information file indicates additional documentation beyond the main article record, although the archive snapshot available by the cutoff describes only its bibliographic entry and not its internal figures or tables.
Who built it and where
According to the bibliographic record, the first two authors are Caio S. Souza and João P. G. Correia, followed by Isabel Rocha, Diana Lousa and Cláudio M. Soares. The affiliation supplied in the indexed record for all five authors is Universidade Nova de Lisboa. The citation metrics reproduced in that record list Isabel Rocha with an h-index of 38 and 5,772 citations, Diana Lousa with an h-index of 19 and 961 citations, and Cláudio M. Soares with an h-index of 46 and 7,381 citations, while the two first authors were listed without an h-index or citation count in that snapshot.
The funding entry in the same record names Fundação para a Ciência e a Tecnologia as the funder. The topical tags attached to the record are Enzyme Catalysis and Immobilization, Microbial Metabolic Engineering and Bioproduction, and Enzyme Structure and Function. Those three tags together signal that the editors indexed the work at the intersection of catalysis, bioproduction and structural biology.
Why bottleneck enzymes matter for bioproduction
Expression, activity, cofactors and toxicity limit biocatalysts
The abstract for the GDEE paper lists four practical challenges that limit the efficiency of biocatalysts: poor protein expression, low catalytic activity, substrate and co-factor limitations, and toxicity of final products. Each of those items describes a different failure mode in moving an enzyme from a natural context into a production pathway. Poor expression means insufficient soluble protein is made, low activity means slow conversion, co-factor or substrate limits mean the reaction cannot proceed at the needed rate, and product toxicity means accumulation harms the host or process.
The abstract also states that protein engineering offers a solution by redesigning enzyme catalytic properties to enhance biosynthetic pathways. In the GDEE framing, redesign is not limited to small tweaks to speed. The stated objectives range from improving catalytic efficiency to enabling novel transformations. That range covers both optimizing a known rate-limiting step and opening chemistry that the starting enzyme did not previously carry out.
The emphasis on rate-limiting steps is central to the platform description. According to the abstract, GDEE finds and optimizes enzymes responsible for rate-limiting steps in biosynthetic pathways. In metabolic engineering, a slow step can constrain flux through an entire route, so prioritizing that step for discovery or engineering is presented as the leverage point for pathway improvement.
How GDEE works in four steps
Step 1: Start from natural sequences or generate mutants
According to the abstract, the platform operates through four key steps, beginning with an initial sequence step that either sources natural enzyme sequences or generates mutant variants. That dual entry point supports two distinct uses described in the title itself: gene discovery and enzyme engineering. Sourcing natural sequences corresponds to discovery across existing diversity, while generating mutant variants corresponds to engineering new sequence options around a starting point.
The output of this first stage is sequence input for downstream structural work. The abstract does not specify particular databases, organism sources or mutagenesis schemes in the indexed text, so the archive report treats the step as defined only at that functional level: provide natural candidates, provide designed variants, or provide both for comparison.
Step 2: Predict atomistic protein structures
The second stage described in the abstract is an atomistic protein structure prediction step. That step converts the amino-acid sequences from stage one into three-dimensional structural models at atomic detail. The designation as atomistic indicates that the downstream docking and energy evaluation operate on structural coordinates rather than on sequence alone.
The reference list indexed with the paper includes comparative protein modelling by satisfaction of spatial restraints by Šali and Blundell in the Journal of Molecular Biology in 1993 and a statistical potential for assessment and prediction of protein structures by Shen and Šali in Protein Science in 2006. Those cited entries document that structure prediction and assessment literature forms part of the paper's background, although the abstract itself does not name the specific structure-prediction software used in GDEE.
Step 3: Dock the relevant ligand into each model
The third stage described in the abstract is a docking step. Docking places a small-molecule ligand — substrate, co-factor, product or analogue — into the predicted protein structure to generate binding poses. Those poses become the structural hypotheses for how sequence changes might alter recognition.
The indexed reference list includes several docking and scoring publications that situate this choice in established literature. They include AutoDock4 and AutoDockTools4 with automated docking with selective receptor flexibility by Morris and colleagues in the Journal of Computational Chemistry in 2009, AutoDock Vina with a new scoring function, efficient optimization and multithreading by Trott and Olson in the Journal of Computational Chemistry in 2009, and lessons learned in empirical scoring with smina from the CSAR 2011 benchmarking exercise by Koes, Baumgartner and Camacho in the Journal of Chemical Information and Modeling in 2013. The presence of those citations shows that the paper connects its docking stage to widely used docking programs and benchmarking studies.
Step 4: Evaluate binding energy and prioritize variants for testing
The fourth stage described in the abstract is a binding energy evaluator, obtaining a set of variant enzyme sequences that are prioritized for experimental validation. In other words, the computational pipeline ends with a ranked shortlist, not with a claim that laboratory validation has already been completed in the pipeline description itself. The language of prioritization preserves the status as a computational triage step that informs which constructs should move to bench testing.
The reference list also includes the FoldX web server as an online force field by Schymkowitz and colleagues in Nucleic Acids Research in 2005 and a computational approach to enzyme design predicting omega-aminotransferase catalytic activity using docking and MM-GBSA scoring by Sirin and colleagues in the Journal of Chemical Information and Modeling in 2014. Those entries further place energy evaluation and docking-based activity prediction in the paper's cited background for enzyme design.
What machine-learning rescoring adds
Better prediction of mutation effects on binding
Beyond the four core stages, the abstract highlights an additional element: machine learning-based re-scoring of binding poses to improve binding affinities. According to that text, this re-scoring significantly enhances the accuracy of predicting the effect of mutations on binding. The claim is comparative within the platform evaluation — rescoring improves mutation-effect prediction relative to scoring without that step — and the abstract presents it as highlighting potential in metabolic engineering applications.
The citation background for scoring improvement includes a machine-learning approach to predicting protein-ligand binding affinity with applications to molecular docking by Ballester and Mitchell in Bioinformatics in 2010, a follow-up on whether a more precise chemical description of protein-ligand complexes leads to more accurate prediction of binding affinity by Ballester, Schreyer and Blundell in the Journal of Chemical Information and Modeling in 2014, and improving AutoDock Vina using random forest and effective exploitation of larger data sets by Li, Leung, Wong and Ballester in Molecular Informatics in 2015. The indexed record also lists NNScore as a neural-network-based scoring function for characterization of protein-ligand complexes by Durrant and McCammon in the Journal of Chemical Information and Modeling in 2010 and NNScore 2.0 as a neural-network receptor-ligand scoring function by the same authors in the same journal in 2011.
The record also cites stacked generalization by Wolpert in Neural Networks in 1992, a foundational ensemble-learning publication. Taken together, those cited works show that the paper draws on neural-network scoring, random-forest scoring and ensemble concepts in its discussion of affinity prediction, even though the abstract summarizes the GDEE implementation only as machine learning-based re-scoring of binding poses.
Automation from input to shortlist
The abstract describes GDEE as an automated platform. In the four-step description, automation connects sequence sourcing or mutant generation through structure prediction, docking and energy evaluation without presenting manual intervention as part of the defined workflow. The endpoint is framed as obtaining a set of variant enzyme sequences prioritized for experimental validation. That endpoint definition matters for interpretation: the platform narrows candidates computationally so that laboratory effort can focus on the most promising variants.
The broader engineering literature cited in the record includes engineering microbial factories for synthesis of value-added products by Du, Shao and Zhao in the Journal of Industrial Microbiology and Biotechnology in 2011, protein design in metabolic engineering and synthetic biology by Pleiss in Current Opinion in Biotechnology in 2011, evolving strategies for enzyme engineering by Bloom and colleagues in Current Opinion in Structural Biology in 2005, and a synthetic recursive plus-one pathway for carbon chain elongation by Marcheschi and colleagues in ACS Chemical Biology in 2012. Those citations anchor the automation goal in microbial factory engineering and pathway design rather than isolated protein biophysics.
Sept. 22 context: enzyme design was a busy news day
LLM agents for intuitive protein engineering
GDEE was not the only enzyme-design item dated 2026-09-22 in the archive window. According to the Nature Computational Science item titled Intuitive enzyme design with LLM agents, author Terra Sztain of the Department of Medicinal Chemistry and the Department of Biophysics at the University of Michigan, Ann Arbor, described a collaborative system of large language model agents that brings protein engineering closer to natural-language interaction. The summary text for that News & Views piece states that the system lowers the barrier between human intent and biomolecular design.
That News & Views entry, published on 22 September 2026, references work by Shao and colleagues with a 2026 Nature Computational Science identifier, agents documentation by Wiesinger, Marlow and Vuskovic from 2025, a 2025 Journal of Chemical Information and Modeling paper by Ma and colleagues, a 2022 Nature Communications paper by Ferruz, Schmidt and Höcker, a 2023 Cell Systems paper by Nijkamp and colleagues, a 2024 Digital Discovery paper by Ghafarollahi and Buehler, a 2025 ICML workshop paper on ProteinCrow by Ponnapati and colleagues, a 2026 Journal of Chemical Information and Modeling paper by Durumeric and colleagues, and a 2026 bioRxiv preprint by Dolorfino and colleagues on assessing generalizability of machine learning and physics-based methods with DNA-encoded libraries. The correspondence address given is [email protected], and the competing-interests statement declares no competing interests.
Template-free mechanism prediction with EZSolver
A second item dated 2026-09-22 provides complementary context on computational enzymology. According to the Journal of the American Chemical Society record for EZSolver: Template-Free Prediction of Polar Enzymatic Mechanisms via Bidirectional Flow Matching and Search, authors Lun-Hsin Kuo, Jason Yang and Frances H. Arnold present EZSolver for polar enzymatic mechanism prediction. The abstract states that current computational predictors rely on deterministic, rule-based dictionaries that perform well on in-distribution tasks but fail to generalize to out-of-distribution chemistry.
According to that same abstract, EZSolver is powered by a flow-matching predictor called EZFlow, described as a template-free generative framework, and navigated by an evaluator-guided bidirectional beam search. The text states that the system learns the chemistry of electron redistribution instead of memorizing rigid templates. Evaluated across diverse enzyme classes, the abstract reports 60.0% accuracy and an 84.6% chemical plausibility rate for full-mechanism prediction of unseen polar enzymatic reactions. The JACS record lists receipt on July 8, 2026, revision receipt on September 9, 2026, acceptance on September 14, 2026 and online publication on September 22, 2026, with funding from the Division of Molecular and Cellular Biosciences under award MCB-2016137, Google, and the National Science Foundation Graduate Research Fellowship Program.
Structure-guided amine dehydrogenase engineering
A third nearby data point illustrates the laboratory side that computational shortlists are intended to feed. According to the ACS Sustainable Chemistry & Engineering record for Structure-Guided Engineering of Amine Dehydrogenase for Chiral Bulky Beta-Aryl Amine Synthesis, authors Jianqiao Liu, Kesheng Fu, Jing Bai, Yunting Liu, Liya Zhou and colleagues report an amine dehydrogenase from Aquisalibacillus elongatus and an engineered variant called M2 1 developed through structure-guided mutagenesis. That record, published online on September 17, 2026 and in issue on September 28, 2026, states that mutant M2 1 efficiently converts multi-substituted aryl, heteroaryl and aliphatic ketones, achieving up to a 252-fold enhancement in catalytic efficiency for bulky substrates, with process work enabling six pharmaceutical intermediates in 62 to 98 percent yields and enantioselectivity above 99 percent ee.
That amine dehydrogenase example is separate from GDEE and does not demonstrate GDEE validation, but it documents the type of outcome — expanded substrate scope and large catalytic-efficiency gains feeding pharmaceutical-intermediate synthesis — that discovery and engineering pipelines aim to prioritize for testing.
What to watch next for GDEE users and readers
From ranked list to bench results
The actionable output defined in the GDEE abstract is a set of variant enzyme sequences prioritized for experimental validation. For a laboratory considering the platform, the next information to look for in the full paper and its supporting file would be which rate-limiting steps and ligand systems were used to exercise the four stages, how many natural sequences versus generated mutants entered the comparison, and how the machine-learning rescoring changed the ranking relative to the base binding-energy evaluator. Those details belong to the complete article and supplement rather than to the indexed abstract, so this archive report does not infer them.
For readers tracking United States science and technology news, the September 22 cluster is notable because three different computational strategies appeared together: GDEE as a structure-based discovery and engineering pipeline in the Journal of Chemical Information and Modeling, EZSolver as a template-free mechanism predictor in the Journal of the American Chemical Society, and a natural-language agent system for enzyme design discussed in Nature Computational Science. Each addresses a different bottleneck — candidate ranking, mechanistic inference and human-AI interaction — while converging on reducing experimental burden in enzyme development.