3.6.4 Downsampling
Reads in De Novo
Assembly
While it might be assumed that the quality of a de novo assembly
will necessarily increase as the number of reads used for input
increases, the truth is, predictably, far more nuanced.
The primary case where this generally is true is when
performing long read-only assembly. In this case, having a preponderance of long reads allows increased ability to correct errors
found in long reads by averaging across more reads; however,
even this benefit plateaus and generally holds true only when
reads are sufficiently long. In the case of short-read sequencing,
too many reads (and by extension, too much coverage) can result in
amplification of sequencing errors and very low-frequency genomic
elements. The consensus is that between 100 and 200Â coverage is
desirable for accurate short-read assemblies. When considering
optimal depth of coverage for Nanopore reads in a hybrid assembly,
assemblers like Unicycler only require 5–30Â coverage of sufficiently long Nanopore reads. In the cases where samples are
“over-sequenced”, some bioinformatic approaches to reduce the
number of reads going into assembly can be applied, a process
known as downsampling.
These approaches fall into two broad categories: deterministic
downsampling and random downsampling. In deterministic
downsampling, reads are ranked by desirability going into assembly (e.g., ranked by average basecall quality score); reads are then
removed starting with the worst ranked reads and proceeding
upward until a desired average coverage threshold is achieved.
This is akin to how quality trimming is already performed on
Illumina reads. In random downsampling approaches
(as implemented primarily for Illumina sequencing in tools such
as Khmer [57] and BBNorm [58]), reads are broken down into
smaller constituent representative pieces (called k-mers) and reads
containing redundant k-mers are randomly removed until a desired
average coverage is achieved. When applied correctly, downsampling can result in better assembly N50 values and assemblies that
are better representatives of the bacterial population being
sequenced.
3.6.5 Reference-Based
Assembly Methods
Unlike de novo genome assembly methods, reference-based assembly methods start with the assumption that the genome sequenced
is highly similar to a previously assembled genome, typically one
generated via de novo sequencing. In reference-based assembly,
sequenced reads are first individually mapped to the reference
genome, typically using a tool like bowtie2 or bwa-mem. This
mapping process results in reads aligned to the reference genome
in a file called a SAM file (or a compressed version of a SAM file
called a BAM file); each line of this text file describes the position to
which each read maps relative to the reference genome. The collection of reads mapped to the reference genome is called a pile-up,
where each DNA base from the reference genome is matched to the
homologous base (or gap) contained in reads from the sequenced
Mycobacterial ID and Characterization Using WGS
417
Reads in De Novo
Assembly
While it might be assumed that the quality of a de novo assembly
will necessarily increase as the number of reads used for input
increases, the truth is, predictably, far more nuanced.
The primary case where this generally is true is when
performing long read-only assembly. In this case, having a preponderance of long reads allows increased ability to correct errors
found in long reads by averaging across more reads; however,
even this benefit plateaus and generally holds true only when
reads are sufficiently long. In the case of short-read sequencing,
too many reads (and by extension, too much coverage) can result in
amplification of sequencing errors and very low-frequency genomic
elements. The consensus is that between 100 and 200Â coverage is
desirable for accurate short-read assemblies. When considering
optimal depth of coverage for Nanopore reads in a hybrid assembly,
assemblers like Unicycler only require 5–30Â coverage of sufficiently long Nanopore reads. In the cases where samples are
“over-sequenced”, some bioinformatic approaches to reduce the
number of reads going into assembly can be applied, a process
known as downsampling.
These approaches fall into two broad categories: deterministic
downsampling and random downsampling. In deterministic
downsampling, reads are ranked by desirability going into assembly (e.g., ranked by average basecall quality score); reads are then
removed starting with the worst ranked reads and proceeding
upward until a desired average coverage threshold is achieved.
This is akin to how quality trimming is already performed on
Illumina reads. In random downsampling approaches
(as implemented primarily for Illumina sequencing in tools such
as Khmer [57] and BBNorm [58]), reads are broken down into
smaller constituent representative pieces (called k-mers) and reads
containing redundant k-mers are randomly removed until a desired
average coverage is achieved. When applied correctly, downsampling can result in better assembly N50 values and assemblies that
are better representatives of the bacterial population being
sequenced.
3.6.5 Reference-Based
Assembly Methods
Unlike de novo genome assembly methods, reference-based assembly methods start with the assumption that the genome sequenced
is highly similar to a previously assembled genome, typically one
generated via de novo sequencing. In reference-based assembly,
sequenced reads are first individually mapped to the reference
genome, typically using a tool like bowtie2 or bwa-mem. This
mapping process results in reads aligned to the reference genome
in a file called a SAM file (or a compressed version of a SAM file
called a BAM file); each line of this text file describes the position to
which each read maps relative to the reference genome. The collection of reads mapped to the reference genome is called a pile-up,
where each DNA base from the reference genome is matched to the
homologous base (or gap) contained in reads from the sequenced
Mycobacterial ID and Characterization Using WGS
417