Ariadne: synthetic long read deconvolution using assembly graphs

Abstract

Synthetic long read sequencing techniques such as UST's TELL-Seq and Loop Genomics' LoopSeq combine 3[Formula: see text] barcoding with standard short-read sequencing to expand the range of linkage resolution from hundreds to tens of thousands of base-pairs. However, the lack of a 1:1 correspondence between a long fragment and a 3[Formula: see text] unique molecular identifier confounds the assignment of linkage between short reads. We introduce Ariadne, a novel assembly graph-based synthetic long read deconvolution algorithm, that can be used to extract single-species read-clouds from synthetic long read datasets to improve the taxonomic classification and de novo assembly of complex populations, such as metagenomes.

Keywords: Assembly graphs; Barcode deconvolution; Metagenomics; Synthetic long read.

Fig. 1

The size-weighted purity of SLR read clouds increases after applying deconvolution methods. All graphs were generated from 40,000 randomly sampled clouds from the dataset. Top row: No deconvolution. Middle row: Reference deconvolution based on read alignment to species and then grouping reads in 200 kbp regions. Bottom row: Ariadne deconvolution with a search distance of 5 kbp and a minimum cloud size cutoff of 6

Fig. 2

Read cloud deconvolution improves metagenomic assembly compared to raw SLR data. We compared assemblies built from raw linked reads (no deconvolution) to assemblies built from reads deconvolved using two methods: reference deconvolution, which maps reads to reference genomes, and deconvolution using Ariadne. Top row: The NA50 of assemblies for each species in each sample between deconvolved reads and raw reads. Larger numbers indicate better performance. Middle row: The largest alignment of assemblies for each species in each sample between deconvolved reads and raw reads. Larger numbers indicate better performance. Bottom row: The proportion of misassembled bases $p_{\mathit{miss}}$ is the number of bases in misassembled contigs over the total number of assembled bases

Fig. 3

Halving and doubling the maximum fragment length does not meaningfully change the quality of de novo assembly using reference-deconvolved linked-reads. Shown here are the NA50, largest alignments, and relative misassembly rate of the MOCK5 LoopSeq reference-deconvolved assembly. As before, we compared assemblies built from raw linked reads (no deconvolution) to assemblies built from reads deconvolved using two methods: reference deconvolution with maximum fragment lengths set to 100 kbp (Ref_100), 200 kbp (Ref_200), and 400 kbp (Ref_400), and deconvolution using Ariadne

Fig. 4

Read cloud deconvolution improves the specificity of short-read taxonomic classification, especially from high ranks such as root (R), kingdom/domain (D), and phylum (P) to low ranks such as species (S). A MOCK5 LoopSeq. B MOCK20 TELL-Seq, for which the y-axis has been truncated at 120,000 promoted reads for display purposes. There were 683,635 domain-to-species promotions with reference-based deconvolution

Fig. 5

Graphical description of the Ariadne deconvolution process. A Reads with the same 3${}^{\prime }$ UMI are in a read cloud. Blue and red reads originate from different fragments. The B de Bruijn assembly graph is generated by cloudSPAdes, and C a focal read is mapped to one of its edges. From a read’s 3${}^{\prime }$-terminal vertex, D a Djikstra graph (indicated by a large black circle) is created from all edges and vertices within the maximum search distance from the 3${}^{\prime }$-terminal vertex. These vertices and edges (within the black circle) comprise read i’s search-distance-limited connected subgraph within the whole assembly graph. Reads aligning to edges in this connected subgraph are added to read i’s connected set. E Reads originating from different fragments likely coincide with non-included vertices. F Connected read-sets with at least one intersection (i.e., one read in common) are output together as an enhanced read cloud