Jovian enables direct inference of germline haplotypes from short reads via sequence-to-sequence modeling
Detection of germline variants in next-generation sequencing data is an essential component of modern genomics analysis. Variant detection tools typically rely on statistical algorithms such as de Bruijn graphs, Hidden Markov Models and regression models, often coupled with heuristic techniques and thresholds to identify variants. Here we describe a new approach that replaces these handcrafted statistical methods with a single, end-to-end deep learning model that directly infers germline haplotypes from short read pileups. Our model, called Jovian, frames variant detection as a sequence-tosequence modeling task, akin to language translation, and employs a transformer-based architecture to translate alignment columns into two predicted haplotype sequences. After training with 17 whole genome sequences from Genome-in-a-Bottle cell lines, we demonstrate that this method learns to realign complex and ambiguous read mappings to produce accurate haplotype predictions, predicts variant genotypes and phase accurately, and leverages the local read context to inform predictions about a given position. We also demonstrate that a 2-dimensional positional encoding significantly improved precision of the detected variants. Compared to other callers, sensitivity and precision is higher than GATK HaplotypeCaller, but lower than DeepVariant and Strelka2.