REFinder Mines Hidden Enzymes From Metagenomes at Lower Cost

TL;DR: A new pipeline finds more enzymes for less compute by assembling only AI-flagged reads

TL;DR: Researchers at Emory University reported on Oct. 1, 2026 that the REFinder pipeline identified 1.2 to 2.1-fold more enzymes than homology-based annotation of full assemblies across 50 microbiome metagenomes while costing as much as 6.4-fold less computationally.

According to the poster titled REFinder: Mining new enzymes from metagenomes, the work was presented by Prabakaran Ramakrishnan and Yana Bromberg of the Emory University Department of Biology in Atlanta, Georgia, USA. The poster was published Oct. 1, 2026 as version 1 and noted as not peer reviewed, with no competing interests disclosed, according to the same F1000Research poster page.

The core idea is selective assembly. The authors describe REBEAN, a DNA language model that assigns each sequencing read a high-level Enzyme Commission (EC) class or a non-enzyme label without alignment, according to the abstract captured in the Exa library record for REFinder: Mining new enzymes from metagenomes. REFinder then routes REBEAN-annotated reads to assemble only the putative enzymatic reads that have identified catalytic signatures, according to that same abstract record.

That design addresses two bottlenecks described by the authors. Recovering enzymes of novel, yet-undescribed functionality typically requires whole-community assembly plus a means of identifying active proteins ab initio, according to the REFinder abstract record. The authors state that the first step is compute-intensive and misses low-abundance sequences, while the second is complicated by inability to predict novel functions, according to that same record.

In evaluation, the pipeline both recovered known-accessible enzymes and added invisible ones. Across all samples, REFinder identified at least three fourths and as many as 90% of the homology-accessible enzymes identified via full assembly of the complete metagenomes, according to the REFinder abstract record. It also identified 1.2 to 2.1-fold more enzymes than could be recovered via homology-based annotation of the proteins from the corresponding full assemblies, according to that same record.

The added enzymes included many with no match in a reference database. A fraction of the identified enzymes, 22% to 45% per EC class, carried no similarity to Swiss-Prot proteins, described as a set of enzymes homology cannot annotate, according to the REFinder abstract record. For roughly two fifths of the over thirteen thousand such novel oxidoreductases from two saliva samples, ESMFold predicted structures aligned with TM-score greater than or equal to 0.7 to a characterized enzyme structure in the PDB, described as substantial structural similarity without sequence homology, according to that same record.

The problem: full assembly is costly and misses rare sequences

Microbial metagenomes encode vast catalytic diversity, according to the REFinder abstract record. The standard route to that diversity has required whole-community assembly, according to that same record.

Whole-community assembly is described as compute-intensive and as missing low-abundance sequences, according to the REFinder abstract record. That limitation matters because metagenomic datasets are large and uneven, with rare members contributing reads that can fail to assemble when pooled with dominant community members.

Identifying active proteins ab initio is the second required step described by the authors, according to the REFinder abstract record. That step is described as complicated by inability to predict novel functions, according to that same record.

Homology-based annotation depends on similarity to already known proteins. The REFinder authors frame enzymes with no similarity to Swiss-Prot proteins as a set homology cannot annotate, according to the REFinder abstract record. The reported range was 22% to 45% per EC class, according to that same record.

The fix: label reads first with REBEAN, then assemble only enzyme candidates

REBEAN is described by the authors as a DNA language model, according to the REFinder abstract record. It sidesteps function prediction by assigning each sequencing read a high-level Enzyme Commission class or a non-enzyme label without alignment, according to that same record.

The Enzyme Commission numbering scheme provides a hierarchical way to classify enzymes according to their catalytic functions, according to the PLOS Computational Biology article titled Fast structural search for classification of gut bacterial mucin O-glycan degrading enzymes. That article was published Sept. 25, 2026, according to its page record.

REFinder builds on REBEAN by routing REBEAN-annotated reads to assemble only the putative enzymatic reads that have identified catalytic signatures, according to the REFinder abstract record. The authors describe this as a pipeline that addresses both steps of recovery and identification, according to that same record.

The practical effect is targeted, alignment-free assembly. The authors state that targeted, alignment-free assembly turns even well-mapped microbiomes into a source of thousands of previously invisible but credible novel enzymes, according to the REFinder abstract record. The evaluation covered 50 microbiome metagenomes, according to that same record.

What was tested: 50 microbiome metagenomes against full assembly

The evaluation compared REFinder against homology-based annotation of the proteins from the corresponding full assemblies, according to the REFinder abstract record. The test set comprised 50 microbiome metagenomes, according to that same record.

The comparison measured both yield and computational cost. REFinder identified 1.2 to 2.1-fold more enzymes than could be recovered via homology-based annotation of the proteins from the corresponding full assemblies, according to the REFinder abstract record. It was as much as 6.4-fold cheaper computationally than full assembly, according to that same record.

The pipeline also retained most of what full assembly could already find by homology. Across all samples, REFinder identified at least three fourths and as many as 90% of the homology-accessible enzymes identified via full assembly of the complete metagenomes, according to the REFinder abstract record.

The authors include saliva samples in the detailed novel-enzyme analysis. For over thirteen thousand novel oxidoreductases from two saliva samples, roughly two fifths showed strong predicted structural matches, according to the REFinder abstract record. Those matches were defined as ESMFold predicted structures aligned with TM-score greater than or equal to 0.7 to a characterized enzyme structure in the PDB, according to that same record.

How many more enzymes: 1.2 to 2.1-fold gain and 75% to 90% retention

The headline yield figures cover total enzyme recovery versus a homology baseline. REFinder identified 1.2 to 2.1-fold more enzymes than homology-based annotation of full-assembly proteins, according to the REFinder abstract record. The range reflects variation across the 50 microbiome metagenomes, according to that same record context.

Retention of homology-accessible enzymes is reported separately. The pipeline identified at least three fourths and as many as 90% of the homology-accessible enzymes found by full assembly of the complete metagenomes, according to the REFinder abstract record. That interval describes coverage of enzymes that homology methods can already annotate.

The computational saving is reported as a maximum. The pipeline was as much as 6.4-fold cheaper computationally than full assembly, according to the REFinder abstract record. The authors present that efficiency alongside the yield gain rather than as a trade-off that reduces total recovery.

The novel fraction is reported by EC class. From 22% to 45% per EC class carried no similarity to Swiss-Prot proteins, according to the REFinder abstract record. Those proteins constitute enzymes that homology cannot annotate, according to that same record.

Why homology misses so much: no Swiss-Prot similarity for 22% to 45%

Homology methods require a match to a known sequence. Proteins without similarity to Swiss-Prot proteins fall outside that approach, described as a set homology cannot annotate, according to the REFinder abstract record.

The REFinder evaluation quantifies that blind spot per high-level catalytic group. The no-similarity fraction ranged from 22% to 45% per EC class, according to the REFinder abstract record. That means more than one fifth and up to nearly one half of the REFinder enzymes in each reported class lacked Swiss-Prot similarity.

Sequence-only comparison has known limits for distant relationships. A thesis study on embedding-based alignment states that reliable transfer of annotation by sequence comparison alone is possible only when two proteins share a high level of similarity, according to the record for Embedding-Based Alignment for Protein Function Annotation: a comparative study of Protein Language Models. That record also states that the study examined 256,769 enzymes whose function is hierarchically encoded with EC numbers.

The REFinder approach avoids requiring alignment at the read-labeling stage. REBEAN assigns each sequencing read a high-level EC class or non-enzyme label without alignment, according to the REFinder abstract record. Only reads with identified catalytic signatures are then routed to assembly, according to that same record.

Structural check: ESMFold and TM-score 0.7 for saliva oxidoreductases

The authors tested credibility of novel sequences with predicted structures. For roughly two fifths of the over thirteen thousand novel oxidoreductases from two saliva samples, ESMFold predicted structures aligned with TM-score greater than or equal to 0.7 to a characterized enzyme structure in the PDB, according to the REFinder abstract record. That outcome is described as substantial structural similarity without sequence homology, according to that same record.

ESMFold predicts a protein's three-dimensional structure from its amino acid sequence alone, with no multiple sequence alignment and no database search, according to the tutorial titled ESMFold Tutorial: Fold a Protein Sequence in Seconds. That tutorial states the method was published by Lin and colleagues in Science in 2023, volume 379, pages 1123 to 1130.

Speed is a central feature of that tool. On a single NVIDIA V100 GPU, ESMFold predicts a 384-residue protein in 14.2 seconds, about 6 times faster than a single AlphaFold2 model, rising to roughly 60 times faster on shorter sequences, according to the ESMFold Tutorial record. The homology search that AlphaFold2 runs first can take over 10 minutes on its own, according to that same record.

TM-score provides the similarity scale used in the REFinder check. A TM-score above 0.5 is used as the working criterion for the same fold, while the value for random structures is about 0.17 regardless of protein size, according to the Edgepedia article titled Structural alignment of proteins. The REFinder threshold of 0.7 therefore exceeds the commonly cited same-fold cutoff described in that structural-alignment record.

Broader context: metagenomes hold millions of uncharacterized families

The REFinder result arrives alongside large-scale evidence that known protein families cover only part of natural diversity. From 40.3 billion predicted open reading frames across 40,446 metagenomes, 9,540 metatranscriptomes, and 539 million proteins from 167,415 reference genomes, researchers identified 608,258 previously uncharacterized putative protein families with at least 100 members and 6.5 million families with at least 25 members, none matching known Pfam domains or reference proteins, according to the Nucleic Acids Research article titled Quadrupling the protein family space with global metagenomics. That article was published Sept. 28, 2026, according to its record.

The scale of expansion is summarized as doubling and quadrupling. The effort doubles the known repertoire of large families and quadruples that of smaller families, according to that same Nucleic Acids Research abstract record. Integration of AlphaFold2-based structural predictions with gene-neighborhood and taxonomic analyses enables characterization of previously unannotated proteins, according to that same record.

Other recent work also links metagenomes to therapeutic and functional discovery. A Microbiology Spectrum article published Oct. 1, 2026 and titled GCF-anchored and target-oriented mining of metagenomic BGC space for bioactive product discovery presents a framework using experimentally validated gene cluster families as evolutionary templates and functional beacons, according to its abstract record. That study reports application to large-scale public metagenome-assembled genome datasets and synthesis of six novel compounds with in vitro cytotoxic profiles, according to that same record.

The tumor microbiome is another active mining target. A Nature Communications article published Oct. 1, 2026 and titled Genome mining of the tumor microbiome reveals biosynthetic diversity and potential pathology-modulating metabolites reports a survey using high-quality microbial metagenome-assembled genomes from 3,526 human tumor tissue samples, identifying 528 biosynthetic gene clusters with potential to encode specialized metabolites relevant to tumor pathology, according to its abstract record.

What changes for labs: cheaper targeted assembly of catalytic reads

For laboratories working with human-associated communities, the immediate implication is efficiency. REFinder was as much as 6.4-fold cheaper computationally than full assembly in the reported evaluation, according to the REFinder abstract record. The saving comes from assembling only putative enzymatic reads with identified catalytic signatures rather than the complete metagenome, according to that same record.

For enzyme discovery, the implication is expanded yield from familiar samples. Targeted, alignment-free assembly turns even well-mapped microbiomes into a source of thousands of previously invisible but credible novel enzymes, according to the REFinder abstract record. The saliva example produced over thirteen thousand novel oxidoreductases, with roughly two fifths showing strong structural matches, according to that same record.

For annotation strategy, the result supports combining sequence-language models with structure. REBEAN provides alignment-free read-level EC or non-enzyme labels, while ESMFold provides structure predictions that can be compared with characterized PDB entries, according to the REFinder abstract record and the ESMFold Tutorial record. The TM-score threshold of 0.7 used for the saliva oxidoreductases sits above the 0.5 same-fold working criterion described in the structural-alignment record.

The work remains an early disclosure. It was published as version 1 and marked not peer reviewed on the F1000Research poster page for REFinder: Mining new enzymes from metagenomes. The authors are listed as Prabakaran Ramakrishnan and Yana Bromberg of Emory University, with no competing interests disclosed, according to that same poster page.

Limits and next questions: peer review, low-abundance recovery, environmental tests

The disclosure status sets the first limit. The F1000Research poster page lists version 1 and states not peer reviewed for REFinder: Mining new enzymes from metagenomes. Independent replication and peer review would be needed to confirm performance beyond the authors' 50-metagenome evaluation.

The second limit concerns what was measured. The abstract reports recovery relative to homology-based annotation of full-assembly proteins and reports retention of homology-accessible enzymes, according to the REFinder abstract record. It does not report experimental biochemical validation of catalytic activity in the abstract text, which instead uses ESMFold structural alignment as credibility evidence for the novel oxidoreductases.

The third question is generalization beyond well-mapped communities. The authors state that the results raise expectations for exploration of environmental microbiomes, according to the REFinder abstract record. That statement is framed as an expectation rather than a completed environmental survey.

Low-abundance sequences are a stated motivation. Full assembly misses low-abundance sequences, according to the REFinder abstract record. Whether targeted assembly recovers specific rare taxa or functions at scale remains a question for follow-up studies with detailed per-sample assembly comparisons.

Standardization will also matter. The pipeline depends on REBEAN read labels and on selection of reads with identified catalytic signatures, according to the REFinder abstract record. Public code, parameters, and benchmarks would allow other groups to test the 1.2 to 2.1-fold gain and the up to 6.4-fold cost reduction on independent datasets.