Scientists at the Massachusetts Institute of Technology (MIT) and the Institut Pasteur in France have developed a technique for reconstructing whole genomes, including the human genome, on a personal computer. This technique is about a hundred times faster than current state-of-the-art approaches and uses one-fifth the resources. The study, published September 14 in the journal Cell Systems, allows for a more compact representation of genome data inspired by the way in which words, rather than letters, offer condensed building blocks for language models.
“We can quickly assemble entire genomes and metagenomes, including microbial genomes, on a modest laptop computer,” says Bonnie Berger, the Simons Professor of Mathematics at the Computer Science and AI Lab at MIT and an author of the study. “This ability is essential in assessing changes in the gut microbiome linked to disease and bacterial infections, such as sepsis, so that we can more rapidly treat them and save lives.”
To approach genome assembly more efficiently than current techniques, which involve making pairwise comparisons between all possible pairs of reads, Berger and colleagues turned to language models. Building from the concept of a de Bruijn graph, a simple, efficient data structure used for genome assembly, the researchers developed a minimizer-space de Bruin graph (mdBG), which uses short sequences of nucleotides called minimizers instead of single nucleotides.
“Our minimizer-space de Bruijn graphs store only a small fraction of the total nucleotides, while preserving the overall genome structure, enabling them to be orders of magnitude more efficient than classical de Bruijn graphs,” says Berger.
Berger and colleagues used their method to construct an index for a collection of 661,406 bacterial genomes, the largest collection of its kind to date. They found that the novel technique could search the entire collection for antimicrobial resistance genes in 13 minutes—a process that took 7 hours using standard sequence alignment.
“We can also handle sequencing data with up to 4% error rates,” adds Berger. “With long-read sequencers with differing error rates rapidly dropping in price, this ability opens the door to the democratization of sequencing data analysis.”
Berger notes that while the method currently performs best when processing PacBio HiFi reads, which fall well below a 1% error rate, it may soon be compatible with ultra-long reads from Oxford Nanopore, which currently has 5-12% error rates but may soon offer reads at 4%.