Beneath every handful of soil lies a crowded hidden world of bacteria, fungi and other life, and for decades the sheer difficulty of naming what its genetic traces belong to has kept much of that world unread. Researchers at the Hebrew University of Jerusalem and partner institutions have now offered a new instrument for this old puzzle: BBERT, an artificial intelligence that learns to recognize bacterial DNA on its own terms rather than by matching it against catalogues that are often incomplete. It is a modest step, yet it suggests that the unnamed fragments of the soil may slowly give up mo
BBERT, an AI Tool, Helps Sort Bacterial DNA From Soil's Unidentified Genetic Material
Instead of matching sequences, it learns what bacterial DNA looks like.
Mark
Why does it matter that BBERT doesn't use reference databases?
Mimi
Because soil is full of organisms nobody has catalogued. If a tool can only recognize what's already in a database, those fragments stay unidentified or get mislabeled.
Mark
So how does it decide what's bacterial?
Mimi
It learns patterns in bacterial DNA, much as language AI learns patterns in sentences. It asks whether a fragment looks bacterial, whether it codes for part of a protein, and where the reading frame starts.
Luke
The 90% figure is from benchmark tests, though. What does that tell us about messy real soil?
Mimi
Fair. The 90%, 88% and 96% are benchmark numbers. The real-world evidence is that it classified many fragments a widely used tool left unclassified in 1,969 soil data sets, but that's not the same as confirmed correct answers.
Mark
What about the dark matter clue?
Mimi
Unexplained fragments were more likely to be nonbacterial than interpretable ones. So the mystery isn't only unfamiliar bacteria; it involves several sources.
Luke
That's a correlation in the fragments, not a full accounting of what dark matter is. The source doesn't say how much is what.
Mark
And the speed gains?
Mimi
In simulations, 66 hours dropped to 25 per data set, with about 96% of coverage kept. In root nodules, time halved and memory fell from 188 to 82 gigabytes, though some bacterial information was lost.
Luke
Plus the known weak spots: chloroplast DNA and telling bacteria from archaea. How often those errors happen isn't given here.
Le Pouls
- Soil sequencing produces millions of anonymous genetic fragments, and most of them cannot be matched against reference libraries that leave whole branches of life unrepresented.
- Existing tools leave large portions of this 'microbial dark matter' unclassified or mislabeled, stalling researchers who want to know what the soil is actually doing.
- BBERT sidesteps database matching by judging fragments as short as 100 letters, asking whether they look bacterial, whether they encode protein, and where their reading frame begins.
- On 1,969 public soil datasets it classified fragments that a widely used tool had missed, and it cut genome reconstruction time from about 66 hours to 25 while running on a laptop.
- Its limits are clear: it can confuse chloroplast DNA with bacterial sequences and struggles to separate bacteria from archaea, so its results will need careful handling as it spreads.
Beneath every handful of soil lies a crowded hidden world of bacteria, fungi and other life, and for decades the sheer difficulty of naming what its genetic traces belong to has kept much of that world unread. Researchers at the Hebrew University of Jerusalem and partner institutions have now offered a new instrument for this old puzzle: BBERT, an artificial intelligence that learns to recognize bacterial DNA on its own terms rather than by matching it against catalogues that are often incomplete. It is a modest step, yet it suggests that the unnamed fragments of the soil may slowly give up more of their stories.
A single handful of soil can hold bacteria, fungi and countless other organisms, and modern sequencing can pull out all of their DNA at once. The difficulty comes afterward, when scientists must work out which fragments belong to which life and what they do. A team led by Dr. Amir Erez of the Hebrew University's Racah Institute of Physics and Dr. Omri M. Finkel of its Department of Plant and Environmental Sciences, with colleagues from the University of Haifa and the Volcani Institute, has built a new AI tool to help with that sorting. They report it in Nucleic Acids Research.
The sorting has long been hampered by reference databases, which miss many soil organisms, leaving their DNA unidentified or wrongly labeled. The tool, named BBERT, sets those databases aside. Trained on bacterial DNA, it judges even short fragments of about 100 letters by asking whether they look bacterial, whether they encode part of a protein, and where the reading should begin. In benchmarks it separated bacterial from nonbacterial DNA with about 90 percent accuracy, and when applied to nearly two thousand public soil datasets it classified many fragments that a common comparison tool had left behind.
The researchers also found hints about the dark matter itself: fragments that resisted interpretation were more likely to be nonbacterial, suggesting the mystery draws on several sources at once. BBERT further helped reconstruct bacterial genomes, cutting processing time roughly in half and reducing memory demands on root-nodule samples, though some bacterial information was lost in the process.
The tool is not flawless. It can miss real bacterial fragments, mistake others for bacterial, confuse chloroplast DNA with bacterial sequences, and struggle with archaea. Still, it runs on a personal computer with a capable graphics card, which may open these analyses to laboratories without large computing resources. The authors hope it will also let scientists return to old environmental DNA collections and extract answers earlier methods could not reach.
A handful of soil holds bacteria, fungi and other organisms, and scientists can pull their DNA out all at once. The harder part is working out what that DNA is. Researchers at the Hebrew University of Jerusalem and partner institutions have now built an AI tool, called BBERT, meant to help with that sorting job. They describe it in a study published in Nucleic Acids Research.
The work was led by Dr. Amir Erez of the Hebrew University's Racah Institute of Physics and Dr. Omri M. Finkel of its Department of Plant and Environmental Sciences. Dmitry Alekhin, Dr. Moshe Alon and Dr. Tomer Sidi contributed equally, alongside colleagues from the University of Haifa and the Agricultural Research Organization's Volcani Institute.
The problem starts with how soil is studied. Sequencing a whole sample produces millions of short genetic fragments from many different organisms, and the task is to determine which ones are bacterial and what they encode. The source compares it to sorting a huge pile of pieces from many puzzles without knowing what the finished pictures look like. Existing tools usually compare fragments against reference databases, but many soil organisms are missing from those databases. Their DNA can go unidentified or be labeled wrongly. Researchers often call this unexplained material microbial dark matter.
BBERT skips the matching step. It uses a type of AI also found in language technology, but instead of learning patterns in sentences, it learns patterns in bacterial DNA. Once trained, it can judge an unfamiliar fragment as short as 100 DNA letters without a matching reference sequence. It asks three questions: does the DNA look bacterial, does it contain instructions for making part of a protein, and if so, where should the cell's reading machinery start grouping the letters to read those instructions correctly.
In benchmark tests, BBERT separated bacterial from nonbacterial DNA with about 90% accuracy, ahead of the other deep-learning tools evaluated. It identified protein-coding fragments with about 88% accuracy and found the correct reading frame about 96% of the time. The team then applied it to 1,969 publicly available soil DNA data sets, where it classified a substantial share of fragments that a widely used comparison tool had left unclassified.
Those results also offered a clue about dark matter. Fragments that could not be tied to known functions were more likely to contain nonbacterial DNA than fragments that could be interpreted. That suggests the mystery involves genetic material from several sources, not only unfamiliar bacteria.
The tool also helped with assembling short fragments into longer stretches to reconstruct bacterial genomes. In simulated mixtures of organisms, filtering with BBERT first cut average processing time from about 66 hours to 25 hours per data set. The assembled sequences had fewer mismatches and gaps, while keeping about 96% of the average bacterial-genome coverage achieved without filtering. On DNA from wild legume root nodules, where plants house bacteria and samples mix plant and microbial DNA, filtering halved reconstruction time and cut peak memory use from 188 to 82 gigabytes. It reduced contamination in several reconstructed bacterial genomes, though some bacterial information was lost.
BBERT can run on a personal computer with a suitable graphics card, and the researchers demonstrated it on a laptop, which could help laboratories without large computing facilities analyze environmental DNA.
It is not a perfect sorter. It can miss bacterial fragments and mistake other DNA for bacterial. It can confuse chloroplast DNA with bacterial sequences, reflecting chloroplasts' ancient bacterial origins, and it can struggle to tell bacteria from archaea. Those limits help researchers decide when and how to use it. The authors say it could also let scientists return to existing collections of environmental DNA and find information earlier analyses could not resolve.
Citations marquantes
The tool can confuse chloroplast DNA with bacterial sequences and can struggle to distinguish bacteria from archaea; these limits help researchers decide when and how to use it.— Study findings as reported by Phys.org
Envie de l'histoire complète ?
Lire l'original sur Phys.org ↗
Science & Space