TL;DR: A review published October 7, 2026 concludes machine learning now supports the full enzyme pipeline — discovery, structure prediction, engineering, and design — but still requires close coupling to experiments.

Review ties four ML tasks into one workflow

An October 7, 2026 Review in ACS Catalysis, titled “Machine Learning Empowers Enzyme Discovery, Structure Prediction, Engineering, and Design,” presents the integration as its central result, according to ACS Publications.

The authors are Wang Muqiang, Jiang Ling, Yu Haoran, Yu Jiachen, Zheng Wenlong and Wu Jianping, with contact emails listed for Zheng and Wu at Zhejiang University. Wang Muqiang and Jiang Ling contributed equally, according to the same ACS Publications listing.

The publisher is listed as the American Chemical Society. The timeline listed for the paper is Received August 30, 2026, Revision Received September 24, 2026, Accepted September 25, 2026, Published Online October 7, 2026, and Published in Issue October 16, 2026, according to ACS Publications.

Funding is listed from the Science and Technology Department of Zhejiang Province under award identifiers 2024C03013 and 2025C01097, and from the National Science and Technology Major Project under award identifier 2025ZD0122804, according to ACS Publications. The authors declare no competing financial interest, according to the same listing.

The abstract frames enzymes as efficient and selective biocatalysts, and states that obtaining enzymes with desired properties remains challenging due to vast sequence spaces, incomplete functional annotation, limited structural information, and costly experimental screening, according to ACS Publications. Machine learning is described as a powerful complement to address these limitations, according to the same source.

The abstract organizes the field into four linked functions. For enzyme discovery, sequence and structural representations enhance candidate retrieval and functional annotation, while property predictors support further prioritization, according to ACS Publications. For structure prediction, methods progressed from evolutionary information to protein language model-based prediction and all-atom biomolecular modeling, broadening access from individual protein structures to multicomponent assemblies, according to the same abstract.

For enzyme engineering, ML-guided strategies leverage experimental measurements and pretrained representations to predict and prioritize variants, supporting iterative optimization of fitness landscapes, according to ACS Publications. Beyond optimizing existing enzymes, ML-enabled enzyme design enables construction of sequences and structures whose functional, structural, or catalytic properties are specified by design constraints, according to the same abstract.

The same abstract lists persistent challenges in data quality, model generalization, conformational dynamics, and correspondence between computational predictions and catalytic function. It concludes that closer integration of ML with mechanistic analysis and experimental validation is essential for developing robust and applicable workflows, according to ACS Publications.

Discovery: retrieval, annotation, then prioritization

The discovery section of the October 7 Review distinguishes finding a candidate from deciding it is worth testing, according to ACS Publications.

Sequence and structural representations are described as enhancing candidate retrieval and functional annotation, according to ACS Publications. Property predictors are described as supporting further prioritization, according to the same source.

That two-step framing — retrieval and annotation followed by predictor-based prioritization — mirrors a nitrilase study published online October 3, 2026 in ACS Synthetic Biology, titled “Designing a DeepPCD Strategy for Efficient Engineering Nitrilases toward Industrial Bioconversion,” according to ACS Publications.

In that study, a novel nitrilase named Bh NIT was identified from Bradyrhizobium huanghuaihaiense through a computationally assisted screening strategy, according to ACS Publications. The authors are listed as Xiao-Ting Zhou, Qian Guo, Yao-Fang Ding, Qi Shen, Ya-Ping Xue and Yu-Guo Zheng, with Xue as corresponding author, according to the same listing.

The nitrilase paper lists Received January 15, 2026, Revision Received September 22, 2026, Accepted September 23, 2026, Published Online October 3, 2026, and Published in Issue October 16, 2026, according to ACS Publications. Its citation is given as ACS Synth. Biol. (2026) 15 (10): 4007–4022, according to the same source.

Funding for the nitrilase study is listed from the Taishan Industrial Experts Program, the Zhejiang Provincial Science and Technology Plan Project under award identifier 2024C03011, and the National Natural Science Foundation of China under award identifier 32571726, according to ACS Publications. The authors declare no competing financial interest, according to the same listing.

The computationally assisted identification of Bh NIT provides a concrete illustration of the Review’s point that sequence-based screening can expand the starting pool before engineering begins, while still requiring experimental follow-up, according to ACS Publications and ACS Publications.

Structure prediction: from evolution to language models to assemblies

The Review describes a progression in structure prediction from use of evolutionary information to protein language model-based prediction and all-atom biomolecular modeling, according to ACS Publications.

The practical consequence stated in the abstract is broadened computational access from individual protein structures to multicomponent assemblies, according to ACS Publications.

A systematic review published October 5, 2026 in Discover Artificial Intelligence provides independent context for that progression, titled “A systematic review of protein language models for protein structure prediction and a proposed biological readiness taxonomy,” according to Springer Nature Link.

The authors are listed as Tahsien Al-Quraishi, Waseem Alwan Zaboon, Jemal Abawajy and Raza Nowrozy, with additional authors indicated, according to Springer Nature Link. The article is listed as open access and as a Review, according to the same source.

The systematic review states that it synthesises 38 eligible studies published between 2020 and 2025 using a PRISMA-guided methodology, according to Springer Nature Link. It introduces a Biological-Readiness Taxonomy for evaluating large language model- and protein language model-based approaches to protein structure prediction, according to the same source.

In that taxonomy, protein language models are defined as transformer-based models pretrained on protein sequence or structure-augmented sequence data, with ESM, ProtBERT and ProtTrans given as examples, according to Springer Nature Link. Large language models in the general-purpose sense are defined as large transformers trained primarily on natural-language corpora, applied in the corpus only to tasks such as annotation or literature mining, according to the same source.

The synthesis reports a shift from sequence-only protein language models toward hybrid and multimodal frameworks that integrate structural, evolutionary and biological information, according to Springer Nature Link. That finding aligns with the ACS Catalysis Review’s description of movement from evolutionary information toward language-model prediction and all-atom modeling, according to Springer Nature Link and ACS Publications.

The October 5 systematic review also reports that benchmark performance improved substantially, while most representative models continue to rely primarily on computational evaluation with comparatively limited biological validation, interpretability and experimental verification, according to Springer Nature Link. That caution parallels the October 7 Review’s listed challenges in model generalization and correspondence between predictions and catalytic function, according to ACS Publications.

Engineering: measurements plus pretrained representations

For engineering, the October 7 Review states that ML-guided strategies leverage experimental measurements and pretrained representations to predict and prioritize variants, supporting iterative optimization of fitness landscapes, according to ACS Publications.

Two early-October experimental reports show what that combination looks like when measurements are scarce and when they are used to guide a small set of changes.

A Bioinformatics Advances paper published October 5, 2026, titled “PRIZM: Combining Low-N Data and Zero-shot Models to Design Enhanced Protein Variants,” introduces Protein Ranking using Informed Zero-shot Modelling, described as a two-phase workflow, according to the Exa library record for the paper.

The first phase uses a high-quality low-N dataset to identify the most suitable pretrained zero-shot model for a target property, and the second phase applies that model to rank and prioritize an in silico variant library for experimental testing, according to the same record. Across diverse benchmark datasets spanning multiple protein properties, the workflow reliably separated low- from high-performing models using datasets of approximately 20 labelled variants, according to the same source.

The authors are listed as David Harding-Larsen, Brianna M. Lax, Martina Escorial García, Catarina Mendonça, Felipe Mejia-Otalvaro, Ditte Hededam Welner and Stanislav Mazurenko, with Welner and Mazurenko listed as corresponding authors, according to the same record. Affiliations listed include the Novo Nordisk Foundation Biotechnology Research Institute for the Green Transition and Masaryk University, according to the same source.

In enzyme case studies targeting sucrose synthase thermostability and glycosyltransferase activity, PRIZM-guided selection identified improved variants, including gains of approximately 3 degrees Celsius in apparent melting temperature and approximately 20 percent higher relative activity, according to the same record. Funding is listed from the European Commission through CETOCOEN Excellence, according to the same source.

The DeepPCD nitrilase study illustrates the complementary case where a defined computational workflow organizes multiple filters before testing. Using Bh NIT as the target for engineering, the authors established DeepPCD as an integrated computationally guided workflow that organizes sequence-design, structure-filtering, stability-assessment, evolutionary-compatibility and molecular dynamics analyses to prioritize testable mutations, according to ACS Publications.

Compared with approaches requiring extensive experimental screening of large mutant libraries, the DeepPCD workflow is described as facilitating efficient modifications, according to ACS Publications. By modifying only seven candidate residues of Bh NIT, the team obtained a mutant exhibiting a 1.3 times 10 to the 5th-fold improvement in hydrolytic efficiency, according to the same source.

The same mutant showed a 20.4 degrees Celsius increase in melting temperature and a broader operational pH range extending toward more alkaline conditions, substantially enhancing its potential for industrial applications, according to ACS Publications. Molecular dynamics simulations and structural analysis revealed that enhanced performance stems from reshaping of the substrate-binding pocket and access channel, along with a strengthened interaction network, according to the same source.

Preliminary computational application to additional nitrilases suggested potential transferability within the nitrilase family, although broader generalization will require further validation across more diverse enzyme systems, according to ACS Publications.

Design: building to constraints, not just improving natives

The October 7 Review draws a line between optimizing existing enzymes and constructing sequences and structures whose functional, structural or catalytic properties are specified by design constraints, according to ACS Publications.

The DeepPCD description reflects constraint-based thinking in a limited form: ProteinMPNN-generated sequence analysis, structural filtering, evolutionary conservation analysis, electrostatic potential analysis, substrate-binding energy calculations, pocket volume, tunnel radius and docking energy analysis are among the supporting analyses listed for Bh NIT and its variants, according to ACS Publications.

Supporting Information for the nitrilase paper is listed as available at the ACS Synthetic Biology DOI, and includes computational prediction of thermostability and catalytic parameters, sequence alignment and phylogenetic analysis, AlphaFold3 structural modeling and model quality assessment, primers for combinatorial mutagenesis, and activity tables for purified Bh NIT and mutants, according to ACS Publications.

The October 7 Review does not present design as solved. It states that persistent challenges remain in data quality, model generalization, conformational dynamics and correspondence between computational predictions and catalytic function, according to ACS Publications.

The October 5 protein language model systematic review reinforces that assessment by proposing its Biological-Readiness Taxonomy around biological knowledge source, representation mechanism, structural reasoning capability, biological prediction task, validation maturity and interpretability, according to Springer Nature Link. Priorities identified for future work centre on biological validation, explainability and translational applicability, according to the same source.

Why integration with experiments remains the test

The take-home from the October 7 Review is procedural: closer integration of machine learning with mechanistic analysis and experimental validation is described as essential for robust workflows spanning discovery, structure prediction, engineering and design, according to ACS Publications.

The nitrilase example shows validation through purified-protein measurements, including specific hydrolytic and hydration activities, temperature and pH profiles, and structural comparison of wild-type Bh NIT and engineered variants, as listed in the Supporting Information description, according to ACS Publications.

The PRIZM example shows validation with minimal data, using approximately 20 labelled variants to choose among pretrained models before ranking a larger library, with demonstrated gains in thermostability and activity case studies, according to the Exa library record.

The protein language model synthesis shows why such validation matters: substantial benchmark gains have not yet translated into routine biological validation and experimental verification for most representative models, according to Springer Nature Link.

Together, the three early-October records support the Review’s four-part map without overstating it: retrieval and annotation improve discovery, language-model and all-atom methods widen structural access, measurement-guided ranking focuses engineering, and constraint-based construction extends design — each step still judged by experiment, according to ACS Publications, Springer Nature Link and the Exa library record.

Sources referenced