Deciphering microbial gene function using natural language processing

This paper presents a novel approach for identifying the function of uncharacterized microbial genes using deep learning techniques adapted from natural language processing (NLP). The researchers created a biological corpus of over 360 million microbial genes, modeling gene families as 'words' and genomic regions as 'sentences'. Using this corpus, they trained a word2vec model to generate gene embeddings, which were then used to predict functional categories for 56,617 genes. The method demonstrated high accuracy, particularly for recently discovered defense systems (98% accuracy). Furthermore, the researchers identified functional categories with high 'discovery potential' and showcased its application in discovering systems associated with microbial interaction and defense, highlighting the potential of this combined genomics-NLP approach.

Danielle Miller, Adi Stern, David Burstein