Introduction
The first human genome draft (1) was based on Sanger sequencing technology (2), cost $2.7 billion and lasted over a period of 10 years (3). In comparison, the sequencing of the human genome (~3 Gbp haploid genome size) in a next generation sequencing (NGS) platform where millions of reads are efficiently mapped to the reference genome, currently costs <$1,000 and it can be performed in <2 days (4). Short-read de novo genome assemblers have difficulty to produce large and reliable contigs, particularly in low complexity regions such as centromeres, telomeres and other repetitive regions (5,6). To address this issue, third generation sequencing (7) technologies have been developed. Nanopore (https://nanoporetech.com/) (8,9) and PacBio (https://www.pacb.com/) (10) sequencing platforms were launched around 2010. Third generation sequencers are sequencing single-molecules in real-time (10) without the need of PCR amplification and thus, avoid PCR bias (11,12). The main drawback of long reads is lower accuracy compared to Illumina short-reads: Typical Nanopore and PacBio Sequel I long-reads have an average accuracy of 90% (13) compared to 99.9% of typical Illumina short-reads (4). As a consequence, assemblies produced only by long-reads were more contiguous, but they also contained more errors, which made genome annotation, variant calling and other genome analyses, challenging tasks (6,12).
By following the hybrid assembly strategy (14,15), the advantages of the two generations are combined, incorporating the information contained in the two read types, overcoming their drawbacks. Recent advantages in long-read sequencing by PacBio have shown very promising results: Sequel System II was released in 2019 with an upgraded SMRT flow cell that was first introduced in 2013 (16), which was able to increase the sequencing yield up to 8-fold. However, the greatest breakthrough was the advance of circular consensus sequencing (CCS) (17) which sequences the same circular DNA molecule 10 times, to produce a highly accurate (99.9%) high-fidelity (HiFi) consensus read, while increasing unique molecular yield and insert size (up to 25 Kbp). At the same time, recent advances in Nanopore's base identification algorithm, Bonito (https://github.com/nanoporetech/bonito) (18), have led to greater than 97% base accuracy.
Usually, the primary genome assembly is very fragmented and some contigs are misassembled. For this reason, the completion of the assembly requires the construction of scaffolds (19). To this end, Hi-C sequencing method provides chromosomal conformation information necessary to assemble chromosome-level scaffolds. The general principle of this method is based on the proximity and contacts of chromosomal regions in the cell nucleus. The frequency of contacts is higher between regions of the same chromosome; thus, different chromosomes can be distinguished during the assembly (20). The result of this method is a collection of pairs of reads of chimeric fragments that can be mapped to the assembly, joining very remote areas.
Using the recent sequencing and scaffolding technologies, it is now possible to construct new reference genomes and finish the assembly of existing ones, by closing gaps in the centromeres, telomeres and other low complexity regions. For this reason, new projects have been launched and new consortia have been formed (21–23). The telomere to telomere (T2T) consortium (https://sites.google.com/ucsc.edu/t2tworkinggroup/) (24,25) aims to finish the entire human genome by producing chromosomes without gaps. Almost two decades after the first draft of the human genome by the International Human Genome Sequencing Consortium, T2T published a completed human genome with the exception of five known gaps withing the rDNA arrays (https://genomeinformatics.github.io/CHM13v1/).
The development of sequencing technologies and assembly and scaffolding algorithms, as well as the sharp increase of publicly available data (https://www.ncbi.nlm.nih.gov/genbank/statistics/), democratised de novo genome assembly projects by making them more approachable to smaller labs. The present study aimed to compare genome assembly pipelines, which use different assembly strategies, evaluating them in terms of accuracy, speed and computational power needed. Finally, the need for scaffold construction, incorporating Hi-C sequencing data was also evaluated.
Discussion
The use of a reference genome in the study of medical genetics, with the help of novel tools and methods, can help the identification of novel drug-sequence variant interactions (46) and the identification of variants which may be related to mutations with a genetic base of a variety of genetic diseases, such as cancer (47) and produce further analysis (48). By studying these variants, we are able to analyse the differences and the heterogeneity of different populations in order to understand their differences (49).
To propose an optimised de novo genome assembly workflow, in the present study, factors such as the maximum assembly contiguity, accuracy and completeness were taken into account, without ignoring other parameters crucial for the execution of the sequencing experiments and the production of the assemblies, such as financial, computational power and time limitations.
These findings suggest that the assembly exclusively based on long highly accurate PacBio Hifi reads outperforms Illumina-Nanopore hybrid and Nanopore assembly. de novo genome assemblers which use HiFi reads, require lower amounts of data compared to other strategies. It has been reported that a 30× genome coverage, using HiFi data, is sufficient in order to produce high quality assemblies (18,50). The present study revealed that even a 16× coverage of the human genome was adequate for that purpose. Thus, subsampling in Hifiasm assembly strategy allows the adaptation of sequencing data to the computational resources available as follows: Sequencing data with a coverage of no higher than 40× can be produced as the current findings and previous experience from other Hifiasm users (https://downloads.pacbcloud.com/public/dataset/redwood2020/hifiasm/v12/) suggest, and if the computational system fails to run, the data can be subsampled using the divide and conquer approach, until the computational resources are adequate for the analysis. However, if the subsampled data correspond to <30× coverage, the final assembly can be deteriorated, as we notice on Homo sapiens assembly, where chromosome 22 is missing from the primary assembly and chromosomes 16, 19, 21 and 22 from the final assembly, after the scaffolding and correction process. On the other hand, it has been reported that a hybrid assembly would need 50× Illumina short-read coverage and 30× Nanopore long-read coverage of the genome (15,51,52). In the case of the human genome, notable results with a 34× Illumina and 30× Nanopore coverage were able to be produced. Therefore, the volume of data used for HiFi assemblies is much smaller. As the volume of data decreases, so do the computational requirements for CPU power and particularly memory. In addition, the use of highly accurate long reads, bypasses several computationally demanding, time consuming steps of the assembly workflow.
In hybrid assembly strategy, Wengan performed most effectively in terms of accuracy and speed. Wengan produced the most contiguous Drosophila virilis assemblies. Although no hybrid assembler produced a human genome assembly in BRFAA cluster, Wengan was the only assembler that managed to construct a primary assembly in AUTh computational system.
The assembler we recommend for HiFi reads is Hifiasm, as it outperformed HiCanu in a small genome and it succeeded to produce a notable assembly of a large genome whereas HiCanu failed to run. Hifiasm performed equally well in respect of insert size and coverage, while HiCanu output improves with the increase in insert size and coverage. We recommend the use of Hifiasm or HiCanu assemblers, depending on the available computational resources as well as the organism's genome size and complexity. Hifiasm produced the most contiguous assemblies and its assembly strategy is highly efficient in terms of computational power and time on a single node of the cluster. For this reason, Hifiasm is also used by the Human Pangenome Project (https://humanpangenome.org/). On the other hand, HiCanu gives the possibility to run the assembly on grid when using a computational cluster. Distributing the tasks on different nodes allows the use of more computational resources than running on a central resource and jobs can be executed in parallel speeding performance. Although running on grid, HiCanu was unable to produce a human genome assembly, as the main bottleneck of all assemblers is RAM size. Finally, by following PacBio HiFi assembly strategy for small genomes, we utilise only one sample preparation and one sequencing technology, in contrast to the Illumina/Nanopore hybrid strategy where we need to make three sample preparations (Illumina, Nanopore and Hi-C) and utilise two sequencing technologies (Illumina sequencing for short genomic and Hi-C reads and Nanopore for long genomic reads). For larger genomes, similar to the human one, PacBio HiFi assembly strategy relies on two sample preparations and two sequencing technologies (Illumina sequencing for short Hi-C reads and PacBio long genomic reads).
Our analysis suggests that the use of additional information for scaffolding is not necessary in small genomes (such as insect genomes); however, it offers a noticeable improvement in larger and more complex genomes (such as the human genome and higher plant genomes). The computational resources required for scaffolding, even for the most complex genomes, are far less than those for the assembly step. Ideally, the use of multiple types of data, seems to exploit different genome features. The successive use of 10× (https://www.10×genomics.com/) (53,54), Bionano (https://bionanogenomics.com/) (55) and Hi-C data will generate the most accurate scaffolds (25,56). Although the use of 10× and Bionano data is not imperative, Hi-C sequencing reads are highly recommended for complex genomes, in order to increase the continuity of the assembly, while improving the accuracy by reducing major misassemblies and translocations.
The development of sequencing technologies led to a great reduction on sequencing cost. The purchase of a sequencer is no longer compulsory for genome assembly projects, as different institutes provide a variety of sequencing services at affordable, by many labs, prices. Each of PacBio, Illumina and Nanopore, offers a network of certified sequencing service providers. Some of these providers are certified for more than one of those sequencing technologies. Moreover, the purchase of a computational cluster is no longer necessary, as bioinformatics infrastructures, such as ELIXIR (57), can offer researchers the computational recourses necessary for the accomplishment of demanding tasks, such as a de novo genome assembly.
The major bottlenecks in genome assembly projects were the computationally demanding assembly algorithms and the large cost of sequencing. The development of new assembly algorithms, which require much less computational power and memory, is the result of major improvements in long-read accuracy by PacBio. The future of genomics relies on long-reads in order to resolve low complexity regions of the genomes and perform telomere-to-telomere assemblies. Alongside to the advances of read accuracy, third generation sequencing led to the reduction of sequencing cost. Furthermore, the increase of genomic data availability in public databases (58), such as Sequence Read Archive (SRA) (59), allows researchers to find and use a variety of raw sequencing data from the same species of interest, already produced by others, for the primary assembly and/or the scaffolding process. Finally, it is important to note that all assembly algorithms and methods we utilised during this work, are being constantly updated in order to improve in terms of performance and computational efficiency, allowing even the reanalysis of older data and the discovery of novel information. In addition, as basecallers are also constantly updated, reusing raw signal files (for example, fast5-formatted files in Nanopore) can produce more accurate reads.
In conclusion, continuous advancements in all fields mentioned above, lead towards the democratisation of de novo genome assembly projects, by enabling scientific laboratories with limited technical and financial resources to perform a great variety of genomic studies, without the need for expensive sequencing equipment and computational infrastructure.