A Graph Auto-Encoder for Haplotype Assembly and Viral Quasispecies Reconstruction

Abstract

Reconstructing components of a genomic mixture from data obtained by means of DNA sequencing is a challenging problem encountered in a variety of applications including single individual haplotyping and studies of viral communities. Highthroughput DNA sequencing platforms oversample mixture components to provide massive amounts of reads whose relative positions can be determined by mapping the reads to a known reference genome; assembly of the components, however, requires discovery of the reads' origin -an NP-hard problem that the existing methods struggle to solve with the required level of accuracy. In this paper, we present a learning framework based on a graph auto-encoder designed to exploit structural properties of sequencing data. The algorithm is a neural network which essentially trains to ignore sequencing errors and infers the posterior probabilities of the origin of sequencing reads. Mixture components are then reconstructed by finding consensus of the reads determined to originate from the same genomic component. Results on realistic synthetic as well as experimental data demonstrate that the proposed framework reliably assembles haplotypes and reconstructs viral communities, often significantly outperforming state-ofthe-art techniques.

A Convolutional Auto-Encoder for Haplotype Assembly and Viral Quasispecies Reconstruction

“…Note that no method can correctly reconstruct the strains in 4 genes (vpu, gp120, gp41 and nef). As noted in (Ke and Vikalo, 2020), this may be due to translocations of short segments in those genes, causing mismatch between the 5 HIV-1 strains reconstructed by (Di Giallonardo et al, 2014) and the actual ground truth.…”

“…While early methods explored a variety of metrics including minimum single nucleotide polymorphism (SNP) removal (Lancia et al, 2001) and maximum fragments cut (Duitama et al, 2010), the vast majority of more recent techniques is focused on minimum error correction (MEC) optimization (Lippert et al, 2002), i.e. determining the smallest number of inconsistencies between reads and the reconstructed haplotypes (Ke and Vikalo, 2020;Edge, Bafna, and Bansal, 2017;Xie et al, 2016;Bonizzoni et al, 2016;Patterson et al, 2015;Pisanti et al, 2015;Kuleshov, 2014). Existing MEC score optimization methods can be divided into two categories: those in pursuit of exact solutions to the MEC optimization problem, and computationally efficient heuristics.…”

“…Reconstructing polyploid haplotypes is more difficult than solving the same task for diploids due to the expanded search space. Existing methods capable of handling haplotype assembly of both diploids and polyploids include HapCompass (Aguiar and Istrail, 2012); HapTree (Berger et al, 2014), a Bayesian method; SDhaP (Das and Vikalo, 2015); BP (Puljiz and Vikalo, 2016); matrix factorization frameworks including (Cai, Sanghavi, and Vikalo, 2016) and AltHap (Hashemi, Zhu, and Vikalo, 2018); and GAEseq (Ke and Vikalo, 2020), a method based on a graph auto-encoder.…”

“…Finally, prior work on viral quasispecies reconstruction includes ViSpA (Astrovskaya et al, 2011), a method based on read clustering; ShoRAH (Zagordi et al, 2011), a method based on read-graph path search; QuRe (Prosperi and Salemi, 2012), an algorithm that relies on combinatorial optimization; QuasiRecomb (Töpfer et al, 2013), a technique based on a hidden Markov mode; PredictHaplo (Prabhakaran et al, 2014), an algorithm that relies on Dirichlet process generative models; aBayesQR (Ahn and Vikalo, 2017), an approach based on hierarchical clustering and Bayesian inference; TenSQR (Ahn, Ke, and Vikalo, 2018), a successive clustering framework using tensor factorization; and GAEseq (Ke and Vikalo, 2020), a graph euto-encoder technique. Among all the existing methods, GAEseq (Ke and Vikalo, 2020) is the only one designed to handle both haplotype assembly and viral quasispecise reconstruction problems. Note, however, that due to aiming to minimize the MEC score directly, GAEseq uses full-batch gradient descent which makes it exceedingly slow and practically infeasible when dealing with large numbers of reads.…”

“…Illumina's MiSeq paired-end reads of length 2 × 250 bp are aligned to HIV-1 HXB2 reference genome. The MEC improvement rate used to estimate the number of strains is set to 0.09 following (Ahn and Vikalo, 2017;Ke and Vikalo, 2020). Reads with mapping score lower than 60 and length shorter than 150 bp are filtered out for quality control.…”