BBERT uses artificial intelligence to identify hidden bacteria in soil

Reading DNA patterns like natural language

A single scoop of earth contains millions of mingled genetic fragments from bacteria, fungi, and surrounding vegetation. Standard environmental sequencing tools identify these fragments by matching them against reference databases. Because many wild soil organisms have never been isolated or cataloged, their genetic material remains unclassified, forming what biologists call microbial dark matter. According to reporting by Phys.org, the new BBERT tool harnesses artificial intelligence to approach the sorting problem by analyzing recurring statistical patterns in bacterial sequences rather than looking for recorded matches.

The deep-learning model can evaluate genetic fragments as short as 100 DNA "letters" long. Once trained, the system assesses three distinct characteristics: whether the sequence originated from a bacterium, whether it carries instructions to synthesize part of a protein, and where the cellular machinery should begin reading those instructions.

Benchmark accuracy across soil databases

In controlled benchmark tests reported by the research team, BBERT separated bacterial from nonbacterial DNA with approximately 90% accuracy, exceeding the performance of other deep-learning classification programs evaluated in the study. The model also identified protein-coding instructions with approximately 88% accuracy and located the correct genetic reading frame at a 96% accuracy rate.

The investigators applied the software to 1,969 publicly available soil DNA data sets, successfully classifying fragments that existing sequence-matching tools had left unidentified. The analysis showed that fragments lacking clear functional links were more likely to contain nonbacterial DNA rather than uncharacterized bacterial genes, providing evidence that microbial dark matter consists of diverse biological sources alongside unknown bacteria.

Reconstructing genomes with less memory

Beyond classifying loose fragments, the system assists in the computational assembly of fragmented sequences into full bacterial genomes. In benchmark trials using simulated mixtures of organisms, filtering raw data with BBERT reduced average genome assembly time from 66 hours to 25 hours per data set. The resulting sequences maintained approximately 96% of the bacterial genome coverage achieved by standard methods while producing fewer gaps and alignment errors.

The team also tested the pipeline on real biological samples taken from wild legume root nodules, which naturally blend plant tissue and bacterial symbionts. Pre-filtering the raw reads cut computational assembly time in half and reduced peak system memory usage from 188 to 82 gigabytes. Because the algorithm runs efficiently on a laptop or desktop fitted with a standard graphics card, smaller laboratories can analyze environmental metagenomic data without high-performance computing clusters.

Classification boundaries and shared origins

The researchers note several practical constraints in how BBERT sorts sequences. Because plant chloroplasts originated from ancient symbiotic bacteria, the model can mistake chloroplast DNA for bacterial fragments. The program also has difficulty distinguishing bacteria from archaea, a separate domain of single-celled organisms that share structural features.

The study was led by Dr. Amir Erez of the Racah Institute of Physics and Dr. Omri M. Finkel of the Department of Plant and Environmental Sciences at the Hebrew University of Jerusalem. Co-authors Dmitry Alekhin, Dr. Moshe Alon, and Dr. Tomer Sidi contributed equally to the work alongside collaborators at the University of Haifa and the Agricultural Research Organization Volcani Institute, representing a significant innovation in genomic analysis.

Read also on The Patent: OpenAI Models Generate Proofs in Mathematics, Computing and Quantum Physics