AI for Genomes: Rethinking de novo Assembly

Accurately resolving genomic paths in assembly graphs is a key challenge in de novo genome assembly, especially in the presence of repeats that create tangles and fragmentation. We present a geometric deep learning framework that learns directly from graph structure, bypassing conventional heuristics and exploiting problem symmetries to achieve PacBio HiFi reconstructions with state-of-the-art quality and contiguity. The same approach can be implemented for other sequencing technologies.  Here, we will present results for haploid and diploid genomes. 

Our method performs robustly on both simulated and real datasets and will be able to utilise telomere-to-telomere reference expansion. By decoupling path inference from hard-coded strategies and generalising across species and genomic architectures, this framework opens the door to reconstructing highly complex genomes, including those with high ploidy or extensive structural variation.