REMME / REBEAN

Most metagenomic functional annotation starts by assembling reads into contigs and genes, then matching those genes to a reference database. Both steps discard information. Assembly fails in low-coverage or highly diverse communities, and reference matching cannot describe anything without an annotated relative — a severe constraint in extreme environments, where alignment pipelines label as little as 0.2% of reads. REMME and REBEAN, from Yana Bromberg's group at Emory University, skip both steps and work directly on the raw sequencing read.

REMME (Read EMbedder for Metagenomic Exploration) is a pretrained DNA language model whose unit of analysis is a single short read rather than a chromosome or a gene. Where genome-scale DNA models such as Evo and the Nucleotide Transformer learn from long assembled sequence, REMME learns from 60–300 bp fragments that carry no surrounding context, and is trained with auxiliary objectives that force it to recover biological structure — what fraction of the read is protein-coding, and in which reading frame — rather than settling for nucleotide composition statistics.

REBEAN (Read Embedding-Based Enzyme ANnotator) is REMME fine-tuned for function. Given one read, it predicts which of the seven top-level Enzyme Commission classes the gene that read came from encodes, or that the gene is not an enzyme at all. The framing is deliberate: recognize the function, do not identify the gene. The work was posted to bioRxiv in December 2024 and published in Nucleic Acids Research in September 2025.

#Key Features

• Read-level, assembly-free annotation: Operates on individual unassembled reads, so low-abundance organisms and fragmentary samples that never assemble are still annotated.
• Reference-free function calls: Because prediction runs on learned sequence representations rather than database matches, orphan genes with no annotated homolog receive labels.
• Biology-forcing pretraining objectives: Alongside masked-token prediction, REMME is trained to estimate a read's coding fraction and reading frame, which pushes it beyond surface nucleotide statistics.
• Function-relevant region highlighting: Although never trained for it, REBEAN's predictions concentrate on the parts of a read that correspond to functionally important gene regions.
• Free hosted inference: A public web service at services.bromberglab.org runs REBEAN without local installation, alongside the downloadable models and scripts.

#Technical Details

REMME is an encoder-only transformer with six layers and eight attention heads, embedding overlapping nucleotide triplets (stride one) into 128 dimensions alongside a 128-dimensional positional embedding — 1,662,177 trainable parameters in total. It was pretrained on 72.9 million reads sampled from 1,496 marine prokaryotic assemblies in MGnify, clustered at 80% identity to 53.6 million representatives. Three heads share a summed loss: a masked-token decoder, a regression head for coding fraction, and a four-way reading-frame classifier. REMME reached over 98.5% masked-token accuracy, correlated with ground-truth coding fraction at Pearson r = 0.73, and separated coding from non-transcribed reads at 88.5% accuracy.

REBEAN adds a three-layer dense classifier over REMME's encoder and was fine-tuned on 19 million non-redundant reads drawn from 19,316 metagenomic samples across nine environments, with training labels generated by the high-precision alignment method mi-faser. It attained 80.6% test accuracy, a macro AUC of 0.969, recall of 80.1%, and precision of 83.6% across eight classes. Training the same architecture from random initialization plateaued at 30% accuracy, showing that pretraining does the heavy lifting. On 525,775 synthetic reads from 4,295 experimentally annotated prokaryotic enzymes, REBEAN reached 88.1% accuracy with precision of 71.5% and recall of 33.4% at a 0.5 threshold.

#Applications

The intended use is functional profiling of metagenomes that alignment pipelines handle poorly. On 8 million reads from hydrothermal vents, hypersaline lakes, and salt crystallizer ponds, REBEAN annotated 6.1% of reads as enzymatic at a 90% precision threshold — three- to six-fold more than alignment-based tools, and thirty-fold more than HUMAnN3 — while agreeing closely with those tools where they did produce calls (Cohen's kappa 0.945 against mi-faser, 0.880 against DIAMOND). Applied to 3,820 marine MAGs, it flagged 4,901 proteins as putative oxidoreductases at 90% precision, enriched for oxidoreductase-associated Pfams at an odds ratio of 3.66, including candidates with no close structural match in the PDB.

#Impact

REMME and REBEAN extend language modeling into a regime that had largely been skipped: the short, context-free sequencing read. By training on reads rather than genomes and labeling function rather than identity, they make reference-free annotation possible for the metagenomic data that assembly and homology search leave untouched — which, in extreme environments, is nearly all of it. The models are small enough to run on modest hardware, and the release includes weights, scripts, Figshare archives, and a hosted service. Limitations are acknowledged: REBEAN's training labels come from mi-faser rather than experiment, so its accuracy partly measures agreement with a precise but low-recall aligner; recall on non-enzymatic reads is only 50.6%; and prediction stops at the seven top-level EC classes rather than a specific reaction.