Euktect

Metagenomic sequencing captures the DNA of whole microbial communities, but its analysis has been overwhelmingly prokaryote-centric. Eukaryotes are abundant in soil, ocean, food, and gut samples, yet remain poorly characterized because reference databases cover only a sliver of them: NCBI held 18,631 eukaryotic reference genomes as of May 2025, against an estimated 8.74 million eukaryotic species on Earth. Alignment-based classifiers such as BLAST, MMseqs2, and Kraken2 therefore miss sequences with no close relative on file, and the alignment-free alternatives that followed — EukRep, Tiara, and BERTax — lose accuracy on the short contigs that dominate real assemblies.

Euktect, developed at the State Key Laboratory of Microbial Metabolism at Shanghai Jiao Tong University with the BioInnovation Institute in Copenhagen and posted to bioRxiv in June 2025, is a taxonomic classifier built on the HyenaDNA genomic language model. Because HyenaDNA encodes DNA at single-nucleotide resolution rather than through fixed k-mers or 3-mer tokens, Euktect retains more of the information in a sequence, and it is most competitive exactly where earlier tools are weakest — on fragments of 500 to 1,000 base pairs.

Euktect is distributed as a toolbox rather than a single checkpoint. A superkingdom classifier separates eukaryotic, bacterial, archaeal, and viral sequences; a refinement algorithm uses its per-contig eukaryotic probability to strip prokaryotic contamination from metagenome-assembled genomes (MAGs); and two further models extend classification into the fungal kingdom.

#Key Features

• Alignment-free superkingdom classification: Assigns contigs to eukaryotes, bacteria, archaea, or viruses without any reference database, so species absent from NCBI are still labeled.
• Length-matched model ensemble: Separate checkpoints trained at 500 bp, 1,000 bp, and 1,500 bp input lengths are dispatched by contig length, with long contigs chunked and their per-chunk scores averaged.
• Eukaryotic MAG refinement: Filtering binned contigs by predicted eukaryotic probability and re-scoring with EukCC converts low-quality bins into medium- and high-quality MAGs under the MIMAG completeness and contamination criteria.
• Fungal phylum and Candida models: A ten-class fungal phylum classifier and a hierarchical cascade of binary classifiers running from kingdom down to the Candida genus extend the framework to finer taxonomy.
• Fixed-checkpoint deployment: Released checkpoints and configs install through a bioconda package, and euktect-predict runs inference on new metagenomes with no retraining step.

#Technical Details

Euktect stacks four Hyena blocks with a hidden width of 2,048 and a feed-forward expansion of 512, treating each nucleotide (plus N and special tokens) as its own token. Training follows a two-stage recipe: next-nucleotide prediction on unlabeled sequence, then supervised fine-tuning with a multi-label loss. Pretraining measurably accelerated convergence relative to training from scratch. The superkingdom corpus was drawn from NCBI reference genomes released before 14 April 2024 for bacteria, archaea, and microeukaryotes (fungi, protozoa, Chlorophyta), plus redundancy-clustered viral sequences longer than 5,000 bp, split 8:1:1. Training used a single NVIDIA A100 40 GB GPU at a learning rate of 6e-5.

On a leakage-controlled benchmark built from genomes released between April and November 2024, Euktect reached eukaryote-class F1 scores of 0.871 at 500 bp (a 5.1% gain over BERTax at 0.820), 0.952 at 1,000 bp, and 0.977 at 1,500 bp. Virus identification was comparable to PRP-Meta and ahead of other virus predictors, while retaining the ability to label the remaining three superkingdoms. The fungal phylum model averages an F1 above 0.80 at 1,000 bp across ten phyla, and the Candida cascade exceeds 0.95 F1 at each level for 5,000 bp sequences.

#Applications

Euktect targets the analysis steps that follow metagenome assembly. Applied to a public food metagenome collection, its refinement algorithm converted 407 low-quality and 20 medium-quality eukaryotic bins into 310 medium-quality and 259 high-quality MAGs, of which 239 and 15 respectively came from refinement. Direct sequence extraction supports gene-function and community-composition analysis without prior binning, and the Candida cascade identifies pathogenic yeast reads in samples where reference genomes are scarce. In a reanalysis of 193 inflammatory bowel disease metagenomes, eukaryotic species abundances derived from Euktect supported a random forest that separated IBD from non-IBD samples with an AUC of 0.74, up from 0.68 when only contigs of at least 1,000 bp were used.

#Impact

Euktect demonstrates that a single-nucleotide-resolution genomic language model can be adapted into a practical metagenomics classifier that outperforms k-mer and BERT-based predecessors on short sequence, a regime that has limited eukaryotic recovery from environmental data. Its released checkpoints, Figshare archive, and bioconda package under an Apache-2.0 license lower the barrier for routine use. The authors are candid about limits: reference genomes still represent roughly 2% of eukaryotic species, accuracy degrades when the model is pushed directly to class or genus resolution — hence the multi-level cascade for Candida — and the Hyena architecture offers little interpretability, leaving the classifier a black box. The work remains a preprint and has not yet been peer reviewed.