Language-Model-Based Detection of Genetic Editing in Bacteria
TL;DR
A transformer-encoder-based machine-learning classifier that, instead of analyzing words in sentences, models gene families in genomes, provides a scalable method for identifying engineered sequences without relying on specific marker genes, with potential applications in biosecurity, agriculture, GMO regulation and more.
Abstract
Recent advances in genome editing allow easy genetic manipulation of bacteria, providing them with new traits, some of which could be hazardous, e.g. enhanced virulence or extended resistance to antibiotics. The ability to detect artificially modified bacteria is crucial for identifying potential bio-threats. However, malicious genome editing could be challenging to trace due to the natural exchange of genes among bacteria through horizontal transfer. After curating extensive datasets including natural genomes and simulated edited genomes, we utilized a natural language processing approach to detect edited genomes. We developed a transformer-encoder-based machine-learning classifier that, instead of analyzing words in sentences, models gene families in genomes. After training the model on our datasets, it is able to accurately detect genes artificially added to bacterial genomes due to their unnatural context. Our approach provides a scalable method for identifying engineered sequences without relying on specific marker genes, with potential applications in biosecurity, agriculture, GMO regulation and more.
