BBERT AI recognizes bacterial DNA in short soil fragments

Researchers at the Hebrew University of Jerusalem and partners developed BBERT, a BERT-based model that identifies bacterial DNA in fragments as short as 100 bases. It reaches about 90% accuracy on bacterial classification, 88% on coding fragments and 96% on reading frames, cutting assembly time from 66 to 25 hours per dataset.

• Bacterial DNA classification accuracy is about 90%
• Coding fragments identified at 88%, reading frames at 96%
• Assembly time fell from 66 to 25 hours per dataset
• On legume nodules peak memory dropped from 188 GB to 82 GB