TL;DR: A team led by Duke University reported a machine learning-guided framework that combines sparse lab measurements with evolutionary and structural information to improve the ketoreductase Gre2 in catalytic performance, protein yield, and thermal stability at once.
What was announced on September 17
A preprint titled Integrating complementary biological information for multi-objective enzyme engineering was published on September 17, 2026, in bioRxiv, according to Cold Spring Harbor Laboratory records summarized by Exa.
The work was authored by Nathaniel Blalock, Yvett Sosa, James Heuschkel, Ruihao Li, Xu Ma, Suttipol Radomkit, Hao Wu, Frédéric G. Buono, Jeff Song, Noah B. Pefaur, Laura J. Kingsley, and Philip A. Romero.
Philip A. Romero of Duke University was listed as the corresponding author.
Nathaniel Blalock was affiliated with the University of Wisconsin–Madison.
Yvett Sosa, James Heuschkel, Ruihao Li, Xu Ma, Suttipol Radomkit, Hao Wu, Frédéric G. Buono, Jeff Song, Noah B. Pefaur, and Laura J. Kingsley were affiliated with Boehringer Ingelheim in the United States.
The bioRxiv record classified the paper under Enzyme Catalysis and Immobilization.
The same record also classified the paper under Protein Structure and Dynamics.
The same record also classified the paper under Microbial Metabolic Engineering and Bioproduction.
The record listed zero citations at the time of indexing.
The DOI record identified the paper as a preprint through the identifier 10.64898/2026.09.15.751856.
Why the team focused on industrial biocatalysts
Enzyme catalysts are increasingly used for sustainable pharmaceutical manufacturing, according to the abstract of Integrating complementary biological information for multi-objective enzyme engineering.
Engineering industrial biocatalysts remains challenging because multiple catalytic and developability properties must be optimized simultaneously from limited experimental data, according to the same abstract.
Ketoreductases enable stereoselective alcohol synthesis and are widely used industrially, according to the abstract of Iterative and data-driven ortholog mining enables reliable discovery of stereoselective ketoreductases.
Identifying process-compatible and selective enzymes for non-native substrates remains challenging and requires extensive screening, according to the same Nature Communications abstract.
That Nature Communications paper was published on September 15, 2026.
Its authors were Peter Stockinger, Michael Niklaus, Kevin M. Yar, Nicolas Imstepf, Ivana Pastieriková, Jasmin Küng, Steven Paul Hanlon, Hans Iding, Sumire Honda Malca, and Rebecca Buller, with Rebecca Buller listed as corresponding author.
The project combined the Department of Chemistry, Biochemistry and Pharmaceutical Sciences at the University of Bern with the Institute of Chemistry and Biotechnology at Zurich University of Applied Sciences.
The project also included Process Chemistry and Catalysis at F. Hoffmann-La Roche Ltd. in Basel, Switzerland.
Peter Stockinger and Michael Niklaus contributed equally to that ketoreductase study.
The study was created as part of NCCR Catalysis, grant number 225147 to R.B., a National Center of Competence in Research funded by the Swiss National Science Foundation.
Open-access funding for that study was provided by the University of Bern.
Biocatalysis is becoming a strategic manufacturing capability at AstraZeneca, enabling selective transformations under milder conditions and with the potential to reduce waste and resource intensity relative to conventional routes, according to the abstract of Next-Generation Manufacturing: The Evolving Role of Biocatalysis in AstraZeneca.
That review was first published on September 14, 2026.
Its authors were Alba Diaz-Rodriguez, Martin A. Hayes, Keith R. Mulholland, Fernando Tur, and Samantha Staniland.
AstraZeneca has pledged to achieve science-based, net-zero greenhouse gas emissions across all operations and its entire value chain by 2045 through its Ambition Zero Carbon initiative, according to the full text of that review.
Value-chain-associated emissions including purchased materials, transport and product use exceeded six million tons of carbon dioxide-equivalent annually, according to the same review text.
The framework tested on Gre2
The September 17 preprint developed a machine learning-guided multi-objective design framework, according to its abstract.
The framework integrates sparse functional measurements with complementary evolutionary and structural information, according to the same abstract.
The team applied the framework to engineer the ketoreductase Gre2, according to the same abstract.
Proteins are often optimized for single functions during design and engineering without consideration of other functionalities that may interfere with the intended outcome, according to the summary of Multiobjective learning and design of bacteriophage specificity.
That Cell Systems paper applied deep learning to understand and design the multifunctional host-targeting landscape of the T7 bacteriophage receptor-binding protein for enhanced infectivity, predefined specificity, and high generality toward unseen strains, according to its summary.
The Cell Systems paper was received on November 25, 2025.
The Cell Systems paper was revised on April 7, 2026.
The Cell Systems paper was accepted on August 3, 2026.
The Cell Systems paper was published online on September 1, 2026.
Its authors were Naia Novy, Phil Huss, Sarah Evert, Philip A. Romero, and Srivatsan Raman.
Philip A. Romero was affiliated with Duke University in Biomedical Engineering in Durham, North Carolina, in that study.
Naia Novy, Phil Huss, and Sarah Evert were affiliated with Biochemistry at the University of Wisconsin-Madison in Madison, Wisconsin, in that study.
Srivatsan Raman was affiliated with the Department of Bacteriology and the Department of Chemical and Biological Engineering at the University of Wisconsin-Madison.
The Cell Systems study compared four model architectures.
The Cell Systems study experimentally characterized engineered phages optimized for 26 tasks.
With multiobjective machine learning, it is possible to engineer complex specificities at success rates that enable low-throughput validation of predicted hits, according to the same Cell Systems summary.
The targeting capabilities of T7 are highly plastic, with opposite specificities occasionally separated by only a few mutations, according to the same summary.
The same framework can guide multiobjective optimization of other proteins or biological systems, offering a general strategy for modeling multifunctional landscapes, according to the same summary.
Three properties improved together
The resulting Gre2 designs achieved simultaneous improvements in catalytic performance, protein yield, and thermal stability, according to the abstract of Integrating complementary biological information for multi-objective enzyme engineering.
Selected designs retained improved performance under process-relevant conditions, according to the same abstract.
The abstract described the outcome as simultaneous improvement across catalytic and developability-related measures rather than improvement in one property alone.
Protein yield was one of the three improved properties named in the abstract.
Thermal stability was one of the three improved properties named in the abstract.
Catalytic performance was one of the three improved properties named in the abstract.
Process-relevant conditions were the setting in which selected designs retained improved performance, according to the same abstract.
The authors presented the result as evidence that integrating complementary biological information enables efficient multi-objective enzyme engineering from sparse experimental data, according to the same abstract.
They described the approach as providing a general strategy for accelerating industrial biocatalyst development, according to the same abstract.
How this fits ketoreductase discovery
The September 15 Nature Communications ketoreductase study presented an iterative ortholog mining strategy to explore functional space, according to its abstract.
Starting from a known ketoreductase, the team sampled its orthogroup across 48 evolutionarily distant variants, according to the same abstract.
That distant sampling identified five enzymes producing pharmaceutically relevant alcohols with diastereomeric excess or enantiomeric excess values from 42.8% to greater than 99%, according to the same abstract.
Refined sampling of 60 phylogenetically close orthologs yielded further improved ketoreductases with diastereomeric excess or enantiomeric excess values from 98% to greater than 99%, according to the same abstract.
Because ortholog mining is genome-dependent, it overlooks enzymes from unsequenced organisms, according to the same abstract.
To address that limitation, the authors used their dataset to explore homologous ketoreductases beyond the sampled orthogroup, expanding the search space fifteenfold, according to the same abstract.
To support that expansion, the team developed the automated pipeline HomoLogic, whose functional descriptors, fed into interpretable models, predicted ketoreductase performance accurately with R-squared values up to 0.76, according to the same abstract.
The authors concluded that evolutionarily balanced ortholog panels combined with data-driven modeling enable efficient discovery of stereoselective biocatalysts, according to the same abstract.
Sparse data as the central constraint
Limited experimental data was identified as a central difficulty for industrial biocatalyst engineering in the Gre2 preprint abstract.
Sparse functional measurements were one of the inputs integrated by the machine learning-guided framework, according to the same abstract.
Complementary evolutionary information was another input integrated by the framework, according to the same abstract.
Complementary structural information was another input integrated by the framework, according to the same abstract.
The combination was applied to Gre2 rather than to a screening panel of unrelated enzymes, according to the same abstract.
The reported benefit was efficient multi-objective enzyme engineering from sparse experimental data, according to the same abstract.
The broader claim was that complementary biological information can compensate when functional measurements alone are sparse, according to the same abstract.
Who contributed and in what roles
Nathaniel Blalock was listed with an h-index of 2 and 19 citations in the indexing record.
Yvett Sosa was listed with an h-index of 5 and 75 citations.
James Heuschkel was listed with an h-index of 1 and 4 citations.
Ruihao Li was listed with an h-index of 0 and 0 citations.
Xu Ma was listed with an h-index of 10 and 352 citations.
Suttipol Radomkit was listed with an h-index of 8 and 780 citations.
Hao Wu was listed with an h-index of 0 and 0 citations.
Frédéric G. Buono was listed with an h-index of 23 and 1,615 citations.
Jeff Song was listed with an h-index of 1 and 123 citations.
Noah B. Pefaur was listed with an h-index of 6 and 171 citations.
Laura J. Kingsley was listed with an h-index of 9 and 526 citations.
Philip A. Romero was listed with an h-index of 22 and 3,710 citations.
The collaboration therefore joined University of Wisconsin–Madison, Boehringer Ingelheim in the United States, and Duke University.
Duke University was the affiliation of the corresponding author.
Boehringer Ingelheim in the United States accounted for ten of the twelve listed authors.
What remains a preprint finding
The September 17 Gre2 record was explicitly labeled as a bioRxiv preprint.
The indexing record was published by Cold Spring Harbor Laboratory through bioRxiv.
The Nature Communications ketoreductase paper was described as peer-reviewed, accepted research carrying a permanent DOI and subject to replacement by the final Version of Record.
The Cell Systems bacteriophage paper was published on September 1, 2026, with copyright held by Elsevier Inc. in 2026.
The AstraZeneca biocatalysis review was published in ChemBioChem and dedicated to the memory of Prof. Ivan Lavandera and his contributions to biocatalysis.
The Gre2 preprint did not report a journal acceptance date in the indexed abstract record.
The Gre2 preprint did not report citation counts beyond zero in the indexed record.
Why manufacturers track catalytic and developability traits together
Multiple catalytic and developability properties must be optimized simultaneously, according to the Gre2 preprint abstract.
Catalytic performance, protein yield, and thermal stability were the three properties simultaneously improved in the Gre2 designs, according to the same abstract.
The pharmaceutical relevance comes from the increasing use of enzyme catalysts for sustainable pharmaceutical manufacturing, according to the same abstract.
Ketoreductases are widely used industrially for stereoselective alcohol synthesis, according to the Nature Communications abstract.
Process compatibility and selectivity for non-native substrates remain difficult to secure without extensive screening, according to the same Nature Communications abstract.
The AstraZeneca review described capability building across people through expansion of cross-site expertise, automated high-throughput experimentation platforms, and academic partnerships that extend innovation and capacity, according to its abstract.
Case studies in that review showed how enzyme identification, optimization, and directed evolution translate into robust, scalable processes for active pharmaceutical ingredients, according to the same abstract.
The review defined priorities including end-to-end automated workflows and artificial intelligence-enabled design to shorten enzyme engineering cycles and broaden biocatalysis impact across modalities, according to the same abstract.
The pharmaceutical industry has historically carried one of the highest environmental factors of any manufacturing sector, generating tens to hundreds of kilograms of waste per kilogram of active pharmaceutical ingredient produced, according to a September 14, 2026, review in Genetics and Molecular Research.
That review consolidated enzyme classes including transaminases, ketoreductases, halohydrin dehalogenases, cytochrome P450 monooxygenases and nitrilases, alongside whole-cell microbial hydroxylation and bioreduction.
That review analyzed five case studies in depth including the transaminase-catalysed asymmetric synthesis of sitagliptin and the ketoreductase and halohydrin dehalogenase cascade used for the atorvastatin side-chain intermediate.
The same review identified machine-learning-guided enzyme engineering, synthetic biology and standardized green metrics as promising directions for sustainable transformation of pharmaceutical manufacturing.
What to watch next
The Gre2 authors framed complementary biological information as a general strategy for accelerating industrial biocatalyst development, according to the preprint abstract.
The Nature Communications authors framed evolutionarily balanced ortholog panels combined with data-driven modeling as an efficient route to stereoselective biocatalysts, according to that paper’s abstract.
The Cell Systems authors framed their multiobjective framework as a general strategy for modeling multifunctional landscapes in other proteins or biological systems, according to that paper’s summary.
Future reports building on the Gre2 preprint would need to disclose experimental scale, measurement methods, and process conditions to allow independent assessment.
For now, the verified advance reported on September 17 is a machine learning-guided design framework that used sparse measurements plus evolutionary and structural information to produce Gre2 variants improved in catalytic performance, protein yield, and thermal stability, with retained gains under process-relevant conditions.