HaploCart: Human mtDNA haplogroup classification using a pangenomic reference graph

Current mitochondrial DNA (mtDNA) haplogroup classification tools map reads to a single reference genome and perform inference based on the detected mutations to this reference. This approach biases haplogroup assignments towards the reference and prohibits accurate calculations of the uncertainty in assignment. We present HaploCart, a probabilistic mtDNA haplogroup classifier which uses a pangenomic reference graph framework together with principles of Bayesian inference. We demonstrate that our approach significantly outperforms available tools by being more robust to lower coverage or incomplete consensus sequences and producing phylogenetically-aware confidence scores that are unbiased towards any haplogroup. HaploCart is available both as a command-line tool and through a user-friendly web interface. The C++ program accepts as input consensus FASTA, FASTQ, or GAM files, and outputs a text file with the haplogroup assignments of the samples along with the level of confidence in the assignments. Our work considerably reduces the amount of data required to obtain a confident mitochondrial haplogroup assignment.Author summary: Pangenome graphs are powerful and relatively nascent data structures for representing an entire collection of genomic sequences and their homology. Here we present HaploCart, a tool which leverages the power of pangenomics, in conjunction with maximum-likelihood estimation, to improve human mtDNA haplotype inference on single-source samples (i.e. the sample is not a mixture of multiple contributors, be they human or contaminant). In this context, mapping to many reference genomes at once vastly reduces the Eurocentric bias inherent in contemporary methods, and also improves haplotyping performance at low coverage depths. We show that HaploCart is far more accurate than competing programs on simulated and empirical datasets, and reports clade-level posterior probabilities that accurately reflect confidence in our phylogenetic assignments. Our work can easily be generalized to other haploid markers and suggests that pangenome-based approaches combined with Bayesian methods show promise for improving inference and mitigating ethnicity-related bias in a large class of bioinformatics problems involving sequencing data.