BBERT Uses AI to Recognize Bacterial DNA Hidden in Soil Samples

Researchers at the Hebrew University of Jerusalem and partner institutions have developed BBERT, an AI tool that recognizes bacterial DNA in fragments as short as 100 bases. The work addresses a persistent problem in environmental genomics: scientists can sequence the genetic material in a soil sample, yet much of it remains difficult to interpret because the organisms are poorly represented in reference databases.

The study, “Fast and accurate taxonomic domain assignment of short metagenomic reads using BBERT,” was published in Nucleic Acids Research. It was led by Dr. Amir Erez of the university’s Racah Institute of Physics and Dr. Omri M. Finkel of its Department of Plant and Environmental Sciences. Dmitry Alekhin, Moshe Alon, and Tomer Sidi contributed equally, alongside colleagues from the University of Haifa and the Agricultural Research Organization’s Volcani Institute.

Why Soil DNA Is Difficult to Sort

Shotgun metagenomics sequences DNA collected from an entire community rather than from a single isolated organism. In soil, that means mixing genetic fragments from bacteria, fungi, plants, and other life forms. Researchers then need to determine where those fragments came from and what biological information they carry.

Reference-based tools compare reads against known sequences. That is useful when a close match exists, but incomplete databases leave many reads unclassified. The resulting microbial “dark matter” can obscure both the composition of a community and the functions it performs.

BBERT tackles an earlier sorting question: does a fragment look bacterial? That is a broader classification than naming a species. Recognizing bacterial sequence patterns can help researchers decide which reads to carry into subsequent analyses, even when a close reference match is unavailable.

A Language Model Trained on DNA Patterns

The paper describes a BERT-based model that learns context within nucleotide sequences. Its tokenizer represents individual DNA bases, including an ambiguous-base symbol. During pretraining, the model learns to recover masked bases from surrounding sequence information, an adaptation of the masked-language-modeling approach used in language AI.

The researchers then fine-tuned the model for bacterial versus non-bacterial classification. The goal is to learn patterns characteristic of bacterial sequences, allowing inference without searching for a matching reference sequence each time a read is assessed. Reference-free inference still depends on what the model learned during training; it does not mean the system was developed without genomic data.

BBERT also estimates whether a fragment is protein-coding and predicts its reading frame. Reading frames determine how DNA bases are grouped into the three-letter units used to encode amino acids. Choosing the correct frame helps downstream tools interpret a coding fragment meaningfully.

What the Benchmarks Found

According to the supplied research announcement, BBERT achieved approximately 90% accuracy distinguishing bacterial from non-bacterial DNA in benchmark tests. It identified protein-coding fragments with approximately 88% accuracy and their reading frames with approximately 96% accuracy. These are separate tasks, rather than a single measure of overall genomic understanding.

The team applied the tool to 1,969 publicly available soil DNA datasets. It recognized bacterial sequences among a substantial share of reads that an established comparison tool had left unclassified. The findings also suggested that unexplained genetic material includes non-bacterial sources, rather than consisting solely of unfamiliar bacteria.

Filtering reads before genome reconstruction also reduced computational work. In simulated organism mixtures, average assembly time fell from about 66 hours to 25 hours per dataset. Reconstructed sequences had fewer mismatches and gaps while preserving approximately 96% of the average bacterial-genome coverage obtained without filtering.

Tests on wild legume root nodules provided a more complex biological example. These structures contain plant and microbial DNA together. Filtering roughly halved assembly time and lowered peak memory use from 188 GB to 82 GB, with reduced contamination in several reconstructed bacterial genomes. Some bacterial information was also lost.

Open Software Makes the Approach Usable

The team’s public BBERT repository provides inference software and model-download instructions. Its documented outputs include bacterial classification probabilities, coding probabilities, and probabilities for six possible reading frames. Researchers can also export sequence embeddings for additional analysis.

The repository supports standard FASTA and FASTQ inputs and documents GPU acceleration, Apple Silicon support, and slower CPU-only operation. It also provides utilities for translating predicted coding sequences into amino acids. These features make the model a component of an analysis pipeline, with outputs researchers can filter and inspect.

Useful Filtering Comes With Biological Limits

BBERT can incorrectly classify some fragments. Chloroplast DNA can resemble bacterial DNA because of its evolutionary origins, and distinguishing bacteria from archaea remains challenging. Filtering can therefore improve efficiency while also discarding genuine bacterial reads. Researchers need to weigh that trade-off against the goals of a particular analysis.

The ecological motivation extends beyond faster computing. Finkel’s university research profile describes work on plant microbiomes, microbial interactions, and desert plant and soil metagenomics. Understanding which organisms and functions are present is a foundation for studying how those communities assemble and influence plants.

BBERT offers a way to revisit existing environmental sequencing collections as well as process new samples. Its contribution is a faster, pattern-based first pass through short reads, helping researchers decide which genetic fragments warrant deeper investigation.