10x Genomics
Chromium De Novo Assembly

Supernova2.1, printed on 12/26/2024

Supernova version 2.x performance

The Supernova 2.x releases include a number of significant changes to the code, with corresponding changes in performance. We compared the performance of versions 1.2 and 2.1 with respect to 20 different datasets, including ones that had been run previously, customer datasets, and novel datasets created just for this purpose. Most of these, along with their assemblies, are available for download.

Synopsis of Supernova's current performance

• Supernova has been tested across a broad range of sample types, including vertebrates, plants, and insects; performance on all three categories is markedly improved.

• Assembly quality is greatly improved. For example, for human samples, the typical contig length is now 160 kb and the typical scaffold length is now 40 Mb.

• We demonstrate end-to-end, practical assembly of single insects, avoiding the problems of standard approaches that inbreed or mix wild individuals.

• Streamlined user experience, with many critical metrics added and improved, including accurate measurement of molecule length and genome size.

• Assemblies are produced by a turnkey laboratory and computational process based on one library and at markedly low cost. A new simplified workflow requires only standard depth sequencing, and is compatible with the NovaSeq platform.

• Supernova runs on a single server, uses 256 GB memory for 18 of the 20 genomes tested, and 512 GB for the others. Run time varies from a few hours for small genomes to a few days for human and other similarly sized genomes. The longest observed run time was 8 days, for maize. Several computational bottlenecks have been removed, however genomes larger than human are expected to have longer run times and memory usage, especially if repeat content is high. Any genome above approximately 4 GB should be considered experimental and is not supported.

Samples

We tested Supernova on an extensive range of samples, ranging from controls to wild-caught specimens. For each sample, we created a single Chromium Linked-Read library, which we sequenced and then assembled using both Supernova 1.2 and 2.1.0, without any tuning or specification of parameters, except varying the number of input reads in a few cases (see below).

Genome and data characteristics

Supernova calculates various metrics on different aspects of the input data and the genome it represents, such as genome size, repetitivity, heterozygosity, and molecule length. For these datasets, they vary widely, as shown below; brief descriptions of the metrics follow the table.

Sample

A nickname for the sample, used in these charts.

gsize (est_genome_size)

The estimated genome size, in megabases (Mb).

%rep (repfrac)

Repeat Content Index (%): the percent of read kmers having depth ≥ twice the expected depth.

het (hetdist)

The estimated mean separation between heterozygous sites, in kilobases (kb).

%gc (gc_percent)

Estimated gc content of genome (%).

%hat (high_AT_index)

The percent of kmers in the reads having ≥ 90% AT content. Locally extreme AT content is correlated with assembly gaps.

%di (dinucleotide_percent)

Estimated dinucleotide content of genome (%).

mol_len (lw_mean_mol_len)

The length-weighted mean molecule length, in kilobases (kb).

p10 (p10)

For an average point on the genome, the estimated number of molecules that extend 10 kb in both directions from that point, counting both alleles.

seq (likely_sequencers)

Abbreviated name of the sequencing instrument used.

raw_cov (raw_coverage)

Total number of bases in all of the sequence reads, before trimming off barcodes, divided by the estimated genome size.

Contigs, phase blocks, and scaffolds are all longer

The following table adds N50 contig, phase block, and scaffold sizes. (In the metrics description table, these metrics are contig_N50, phase_block_N50 and scaffold_N50, respectively.) These all show a notable improvement in almost every case. For example, for the hgp sample, the N50 contig size rose from 120.9 to 160.7 kb, the N50 phase block size rose from 4.30 to 6.07 Mb, and the N50 scaffold size rose from 17.18 to 45.63 Mb, a greater than two-fold improvement.

Assembly accuracy and organization are also improved

The following table adds assembly accuracy and organization measures. Because the perfect stretch and misassembly estimate rely on a reference sequence, these two metrics are not generally available; we have calculated them here where we can. All three metrics are described below the table.

perf

This column provides the N50 perfect stretch, in kb. This can only be computed and is only shown for samples having a reference sequence from the same sample. It measures the N50 size of sequences in the reference that are perfectly mirrored in the assembly. Such ‘perfect stretches’ are terminated either by errors or gaps. In this context, transitioning from one allele to the other is an error. The N50 perfect stretch increased for all samples where it could be measured.

mis

This column shows the percent of the assembly that is misassembled. This can only be computed for assemblies having a reference sequence. As an example of the accounting, if a scaffold connects 5 Mb of one chromosome to 10 Mb of another chromosome, that counts as a 5 Mb error, which gets converted into a fraction by dividing by the assembly size. Errors of order and orientation are also included. The misassembly rate declined for all assemblies for which it could be measured.

m10 (m10)

This metric estimates the percent of genomic kmers that are either missing from the assembly entirely or present only in scaffolds shorter than 10 kb. Each kmer counts once regardless of its multiplicity in the genome and thus this measure discounts repeats. It measures assembly disorganization. Assembly disorganization is less in all cases and markedly less in some.

Computational performance

The following table adds computational performance statistics. As shown, all but two assemblies were run on 256 GB servers. Memory use for Supernova has increased about 10% since version 1.2, and run times have increased on average by 60%. However, because of targeted optimizations, the likelihood of the extreme run times experienced by some users of 1.2 should now be much lower.

mem (mem_peak)

The amount of memory (RAM) in GB on the server used for the assembly.

days (etime_h)

The total number of days elapsed during the assembly.

All assemblies were carried out on 28 core servers at 10x Genomics, having processor “Intel Xeon CPU E5-2697 v3 @ 2.6GHz”.

All assemblies were run twice to confirm exact reproducibility of results.