How to Mine Novel Enzymes from Databases: An AI-Driven Practical Guide

In the fields of biocatalysis, synthetic biology, and industrial enzyme production, discovering novel enzymes with new functions or superior performance has always been a core driver for both industry and fundamental research. Traditional enzyme mining heavily relies on experimental approaches such as microbial isolation and cultivation, and in vitro activity screening—methods that are generally time-consuming, low-throughput, and yield limited positive hits. With the widespread adoption of high-throughput sequencing and the exponential growth of public protein databases, computational biology-driven database mining has now become a mainstream and highly efficient approach in the field.

UniProtKB currently houses hundreds of millions of protein sequences. Precisely identifying target candidate enzymes from this vast sequence space requires a systematic bioinformatics workflow. This article integrates major public database resources, traditional sequence-based mining strategies, and emerging AI-driven paradigms to provide a comprehensive technical framework for mining novel enzymes from public databases.

I. Core Databases: The Data Infrastructure for Enzyme Mining

The Enzyme Mining Data Stack

The first step in enzyme mining is to identify the data sources. Four categories of databases form the foundation of modern computational enzyme mining:

Sequence and functional annotation databases. UniProt is the most critical resource in this domain. UniProtKB comprises the manually curated Swiss-Prot and automatically annotated TrEMBL datasets, integrating protein sequences, functional annotations, three-dimensional structural information, mutation sites, post-translational modifications, and many other data types. Most enzyme mining projects begin with UniProtKB as their primary data source. Supplementary databases include NCBI nr (the most comprehensive non-redundant protein database) and RefSeq (a high-quality reference sequence database).

Enzyme-specific databases. BRENDA curates experimentally validated enzymology data and is the world's most comprehensive enzyme information system, covering over 8,000 EC numbers, approximately 90,000 enzyme entries, and millions of kinetic, ligand, and literature annotations. All data are manually curated from the literature, making BRENDA a critical reference for validating computational predictions. Other commonly used enzyme databases include KEGG/MetaCyc (metabolic pathways and enzyme reactions), CAZy (carbohydrate-active enzymes, a core database for industrial enzyme research), and ExplorEnz (an authoritative resource for enzyme nomenclature and classification).

Metagenomic databases: unlocking enzyme resources from uncultured microorganisms. Over 99% of microorganisms in nature cannot be obtained through traditional pure-culture methods. Metagenomic sequencing bypasses this cultivation bottleneck by directly extracting total DNA from environmental samples. Standardized metagenomic databases such as gcMeta serve as bridges between sequence discovery and functional exploitation, providing AI-ready standardized datasets. Other major international metagenomic repositories include IMG/M (Joint Genome Institute) and MGnify (EMBL-EBI).

A domestic research team applied machine learning to analyze over 10,000 metagenome-assembled genomes (MAGs), identifying millions of enzyme sequences belonging to nearly 100,000 homologous protein clusters, representing a vast reservoir of unexplored natural enzyme resources.

Protein structure databases: AlphaFold DB fills the structural dimension. AlphaFold-predicted structures now cover over 200 million UniProt entries. Tools such as ActSeek enable searches for proteins with similar active-site spatial conformations within these predicted structure databases, opening up a structure-based screening dimension for enzyme mining that operates independently of sequence similarity. Other structural databases include the experimental structure database PDB and ESM Atlas (ESMFold-predicted structure database).

II. Traditional Enzyme Mining Methods: Sequence Similarity-Driven Classic Approaches

The Twilight Zone of Homology

BLAST homology search: the most fundamental starting point. BLASTp is the entry-level tool for novel enzyme mining. Researchers use a functionally characterized target enzyme as a query sequence to search for homologous sequences in protein databases via BLASTp. For example, in a study mining novel 4-vinylguaiacol (4-VG) oxidases, the team used CSO2 as the query sequence and retrieved numerous homologs from UniProtKB for further analysis.

When sequence identity drops below approximately 30% (the classic "twilight zone"; below approximately 20%, it enters the "midnight zone"), homology-based searches relying on sequence similarity become highly unreliable—distant homologs may still share similar functions even though their sequences have diverged substantially. A large number of distant homologs with potential novel catalytic functions can thus be missed. In practice, E-value thresholds and sequence coverage should be considered together, with specific cutoffs depending on the conservation level of the enzyme family.

Multiple sequence alignment, phylogenetic analysis, and sequence similarity networks. Using Clustal Omega or MUSCLE for multiple sequence alignment and constructing phylogenetic trees to identify evolutionarily divergent sequences that may have acquired functional differentiation is an advanced screening strategy.

Sequence Similarity Networks (SSN) are powerful tools for large-scale enzyme family analysis. Compared with traditional phylogenetic trees, SSNs visualize similarity relationships among thousands of sequences in a network format, enabling intuitive identification of functional clusters and potential new functional branches. In a BRENDA-based study on feruloyl esterases, 2,085 target sequences were analyzed; protein similarity network analysis divided them into three major clusters, guiding stratified candidate screening.

Hidden Markov Models (HMM) and conserved domain screening. The Pfam database (now integrated into InterPro) provides numerous HMM models for protein domains. Combined with hmmscan/hmmsearch, these tools can rapidly enrich candidate sequences carrying target conserved domains. Compared with BLAST, HMMs are more sensitive for detecting distant homologs and can precisely locate conserved functional domains, making them a highly efficient and accurate filtering strategy.

III. AI-Driven Enzyme Mining: From "Similar Sequence Retrieval" to "Functional Reasoning"

Artificial intelligence methods, particularly protein language models (PLMs), are redefining the boundaries of enzyme mining. Models such as the ESM series (ESM-1b/ESM-2/ESM-3) learn from vast numbers of natural protein sequences and transform amino acid sequences into high-dimensional embedding vectors. Within this embedding space, functionally similar proteins—even those with very low sequence identity—remain close in vector distance.

Embedding-based functional similarity retrieval is a typical application of this approach: researchers use a functionally characterized seed enzyme as a query to retrieve proteins with the nearest vector distances in the embedding space, thereby discovering distant homologs that share similar functions despite large sequence divergence. The 2025 study on ESM-Ezy, published in Nature Communications, exemplifies this strategy, using ESM-1b's embedding space for similarity calculations. In mining campaigns targeting multiple enzyme classes, such as multicopper oxidases and L-asparaginases, this approach consistently achieved higher distant-homolog hit rates than traditional BLAST—44% of candidate enzymes outperformed the query enzymes in at least one property, including catalytic efficiency, thermostability, organic solvent tolerance, and pH stability, with 51% excelling in environmental remediation applications. For L-asparaginases, 40% of candidates showed higher specific activity and catalytic efficiency than the query enzymes.

Machine learning classifiers and enzyme property prediction models can compress tens of thousands of candidate sequences down to a small set of high-potential targets. These models can predict optimal temperature, optimal pH, thermostability, solubility, and kinetic parameters such as Km and kcat. In the 4-VG oxidase mining work, researchers first used CataPro to predict kinetic parameters for 1,500 candidate sequences, selected the top 150, and then combined structural clustering to finalize 5 candidates for experimental validation. The best variant, SsCSO, exhibited 19.53-fold higher catalytic activity than the reference enzyme.

Integrated automated mining workflows. Gene Surfing is a bioinformatics workflow built on the Snakemake framework, integrating data quality control, genome assembly, functional annotation, and homology search modules. Published validation cases show that the workflow identified over 1.31 million potential lignocellulolytic enzyme sequences, with 127 functionally validated and an 84.25% activity-positive rate. In polyethylene-degrading enzyme mining, 38 candidates were heterologously expressed, yielding an 81.5% activity-positive rate.

IV. Practical Workflow: A Standardized Protocol for Mining Novel Enzymes from Scratch

enzyme database landscape

Step 1: Define the mining target and collect seed sequences. Clearly define the target catalytic function (e.g., thermostable lipase, PET hydrolase, oxidase). Collect 2–5 experimentally validated, high-confidence seed sequences, prioritizing Swiss-Prot manually curated entries. If a high-quality HMM model is needed, collecting 10+ representative sequences covering different evolutionary branches is recommended.

Step 2: Initial homolog expansion. Use BLASTp with the seed sequences as queries to search UniProtKB. Start with relatively loose E-value thresholds (e.g., 1e-5 to 1e-20, adjusted based on family size) and coverage (50–70%) to recover close-to-mid-range homologs. Then use InterPro/Pfam domain validation to remove false positives. Candidates with sequence identity below approximately 30% proceed to Step 3 for AI-driven enrichment.

Step 3: AI-driven candidate enrichment (core step). For distant homologs in the low-similarity range that are likely missed by BLAST, use ESM-series protein language models to generate sequence embeddings and retrieve proteins functionally similar to the seed sequences in the embedding space. Concurrently, employ machine learning models to batch-predict key physicochemical properties and kinetic parameters for the candidate enzymes.

Step 4: Structural modeling and candidate prioritization. Use AlphaFold to predict three-dimensional structures for high-priority candidates. Apply TM-align to assess overall structural similarity and combine with ActSeek/active-site alignment to further evaluate catalytic functional conservation. Rank candidates using multi-objective scoring that integrates sequence novelty, predicted catalytic performance, and structural plausibility.

Step 5: Wet-lab functional validation. Select the top 10–50 candidate enzymes for gene synthesis, heterologous expression, and in vitro activity assays. For well-defined enzyme families, after multiple rounds of AI-driven screening, the experimental positive rate for candidates can reach over 80%.

V. Summary and Outlook

Database-driven novel enzyme mining has evolved from a purely BLAST-based homology search into an integrated technological framework combining public database retrieval + AI-based function prediction + structural validation + wet-lab confirmation. While traditional sequence alignment excels at finding closely related homologs, AI models enable function prediction that transcends sequence-similarity limitations, pinpointing high-performance candidates within vast sequence spaces and reducing the validation workload from thousands of sequences to dozens.

UniProt provides sequence breadth; BRENDA provides experimental functional depth; metagenomic databases expand the dimensional scope of enzyme resources from uncultured microorganisms; and AI serves as the core engine that connects these diverse data sources to enable high-throughput, precise mining. Case after case—improved distant-homolog hit rates, automated workflows achieving over 80% experimental positive rates, and AI-guided screening enabling order-of-magnitude activity enhancements—demonstrates that computational enzyme mining has formally entered the "AI-first" era.

At the same time, the field still faces limitations: protein language models can produce false-positive functional predictions; AlphaFold's static structures cannot fully capture conformational dynamics; and large numbers of metagenome-derived sequences lack annotation, adding to the screening challenge. Going forward, natural enzymes mined from databases can not only be directly applied as biocatalysts but also serve as templates for AI-assisted directed evolution, bridging the entire R&D pipeline from "natural enzyme discovery" to "rational engineering" and accelerating the translation of biocatalysis technologies into practice.