Machine-learning predicts genomic determinants of meiosis-driven structural variation in a eukaryotic pathogen

Species harbor extensive structural variation underpinning recent adaptive evolution and major disease phenotypes. Most sequence rearrangements are generated non-randomly along the genome through non-allelic recombination and transposable element activity. However, the causality between genomic features and the induction of new rearrangements is poorly established. Here, we analyze a global set of telomere-to-telomere genome assemblies of a major fungal pathogen of wheat to establish a nucleotide-level map of structural variation. We show that the recent emergence of pesticide resistance has been disproportionally driven by rearrangements. We used machine-learning to train a model on structural variation events based on 30 chromosomal sequence features. We show that base composition and gene density are the major determinants of structural variation. Low-copy LINE and Gypsy retrotransposons explain most inversion, indel and duplication events. We retrain our model on Arabidopsis thaliana and show that our modelling approach can be extended to more complex genomes. Finally, we analyzed complete genomes of haploid offspring in a four-generation pedigree. Meiotic crossover locations were enriched for newly generated structural variation consistent with crossovers being mutational hotspots. The model trained on species-wide structural variation predicted the position of >74% of the newly generated variants along the pedigree. The predictive power highlights causality between specific sequence features and the induction of chromosomal rearrangements. Our work demonstrates that training sequence-derived models can accurately identify regions of intrinsic DNA instability in eukaryotic genomes.