GoldPolish-target: targeted long-read genome assembly polishing

Abstract

Background

Advanced long-read sequencing technologies, such as those from Oxford Nanopore Technologies and Pacific Biosciences, are finding a wide use in de novo genome sequencing projects. However, long reads typically have higher error rates relative to short reads. If left unaddressed, subsequent genome assemblies may exhibit high base error rates that compromise the reliability of downstream analysis. Several specialized error correction tools for genome assemblies have since emerged, employing a range of algorithms and strategies to improve base quality. However, despite these efforts, many genome assembly workflows still produce regions with elevated error rates, such as gaps filled with unpolished or ambiguous bases. To address this, we introduce GoldPolish-Target, a modular targeted sequence polishing pipeline. Coupled with GoldPolish, a linear-time genome assembly algorithm, GoldPolish-Target isolates and polishes user-specified assembly loci, offering a resource-efficient means for polishing targeted regions of draft genomes.

Results

Experiments using Drosophila melanogaster and Homo sapiens datasets demonstrate that GoldPolish-Target can reduce insertion/deletion (indel) and mismatch errors by up to 49.2% and 55.4% respectively, achieving base accuracy values upwards of 99.9% (Phred score Q > 30). This polishing accuracy is comparable to the current state-of-the-art, Medaka, while exhibiting up to 27-fold shorter run times and consuming 95% less memory, on average.

Conclusion

GoldPolish-Target, in contrast to most other polishing tools, offers the ability to target specific regions of a genome assembly for polishing, providing a computationally light-weight and highly scalable solution for base error correction.

Background

The genomics revolution has led to a surge in the development of bioinformatic technologies aimed at generating high-quality de novo genome assemblies that are crucial for extracting biological insights from sequencing data [1]. Long-read sequencing, notably using Oxford Nanopore Technologies Plc. (ONT, Oxford, UK) and Pacific Biosciences Inc. (PacBio, Menlo Park, USA) instruments, has gained prominence due to its ability to capture long-range genomic information within single reads, enabling the resolution of complex structural variants and regions with high homology [2]. Unlike Illumina, Inc. short-read sequencing that can generate reads up to 600 bp [3], ONT produces reads with average lengths of 10–100 kbp [4] and PacBio HiFi sequencing produces reads lengths averaging 10–25 kbp long [5]. While long-read technology facilitates the generation of highly contiguous genome assemblies, its associated error rates tend to be higher (1–15%) [4, 6] compared to Illumina short-read sequencing (~ 0.1–1%) [7]. These erroneous reads can negatively impact the base quality of subsequent genome assemblies, hindering accurate downstream analysis, such as the identification of true variants [8] or comparative genomic studies [1].

In response, specialized error correction tools, such as Racon [9], Medaka [10], and GoldPolish [11], have emerged, using long reads to identify and correct base errors in a genome assembly [12]. While these tools ultimately share the same goal, the approaches of these polishing algorithms vary widely [13]. Racon uses information from mapped reads to construct a partial-order alignment graph to generate a final consensus sequence [9]. Medaka generates a consensus sequence by applying neural networks to sequencing reads that are aligned to a draft assembly [10). GoldPolish uses a long-read adaptation of the ntEdit+Sealer protocol [14] implemented in the GoldRush genome assembly pipeline [11]. The method described in this work builds on the GoldPolish algorithm [11], which all of the authors of this study contributed to.

Many assembly algorithms leave regions of the genome with highly concentrated error rates, such as the terminal overlapping regions of fragmented contigs, in the final assembly output [15]. These regions can negatively impact the reliability of downstream analyses of genome assemblies and should be addressed prior to analysis. For example, in a study conducted in 2019, human ONT long reads assembled with Canu [16], a single molecule sequence assembler, exhibited regions with higher error rates that substantially impacted the base quality of protein-coding genes, ultimately hindering protein prediction [17]. Similarly, the GoldRush genome assembly pipeline may produce regions with relatively higher error rates. During the scaffolding stage, ntLink joins contigs based on evidence from supporting reads, often leaving gaps, or ambiguous nucleotide bases (“N”s), between them [18]. Alternatively, when run with the gap-fill feature, ntLink fills these gaps with unpolished (raw) bases from representative long reads, which are soft-masked (i.e. represented in lower-case) for downstream targeting [18]. These filled-in gaps are never polished within the GoldRush workflow, introducing regions with relatively lower base quality. To address these potential challenges that many assembly algorithms face, we developed GoldPolish-Target (GP-Target), a long-read targeted polishing pipeline based on GoldPolish.

GP-Target enables the correction of specific genomic regions without having to polish entire assemblies. With experiments using Drosophila melanogaster (fruit fly) and Homo sapiens (human) datasets, we demonstrate that GP-Target produces high-quality genome assemblies by substantially reducing indel and mismatch errors, improving consensus quality, and increasing gene completeness. The tool also achieves these with lower computational costs compared to a current state-of-the-art utility.

Implementation

Algorithm overview

The GP-Target workflow is implemented using SnakeMake [19] and consists of five stages (Fig. 1). A draft sequence assembly and matching long-read sequences are required inputs. First, the long reads are mapped to the draft assembly with either ntLink (default) [18] or minimap2 [20], with the mappings output in PAF format (Pairwise mApping Format) [20]. The PAF file format is a tab-delineated 12-column file that describes the mapping positions between two sets of sequences [21]. Then, by default, GP-Target searches for soft-masked bases and excises them, providing a convenient means to polish raw bases left by the GoldRush pipeline. Alternatively, users can submit a BED file with sequence coordinates for any number of target regions, specified using the chromosome, start position, and end position columns of the standard BED format [22]. The sequences of the target regions are written to an intermediate FASTA file and the PAF mapping file generated by the mapping step is adjusted to match the naming and coordinate system of these target regions. Then, the updated PAF file and the intermediate FASTA files are passed to GoldPolish for polishing. Finally, the polished target regions are reinserted into the original draft assembly to yield the final polished assembly. The workflow is modular and can be adapted for use with any polishing algorithm that accepts a draft assembly and a corresponding PAF file as inputs.

Fig. 1

Schematic of GP-Target framework and steps. (1) Long-reads are first mapped to the draft assembly. (2) Then, by default, soft masked regions of the assembly are excised; alternatively, a BED file can be used to specify coordinates for the regions of interest to be excised. (3) The mapping file is modified to match the coordinate system of target regions and the resulting PAF file is passed to GoldPolish. (4) GoldPolish is run on the target regions to correct base errors. (5) The polished target sequences are re-inserted into the original assembly to yield the final polished assembly output

Read mapping

By default, the long-read sequences are mapped to the draft assembly using the ‘pair’ function of ntLink [18]. Users can optionally set the k-mer size (k_ntLink) and the window size (w_ntLink) for generating minimizers for the mapping; otherwise, these parameters will default to 88 bp and 1000 bp, respectively. Alternatively, minimap2 [20] with default parameters can be used for mapping.

Target sequence extraction

The target regions are stretches of soft-masked bases in the draft assembly in the default mode. If a BED file is included as an input, GP-Target excises target sequences corresponding to the specified start and end positions. It extracts these target regions, plus flanking sequences 5’ and 3’ of the region (length; default 64 bp) and saves them to an intermediate FASTA file. When flanking sequences from multiple target regions overlap in the draft assembly, they are joined into one larger sequence in the resulting FASTA file. When added to the intermediate FASTA file, the names of the target regions are updated to include some key points of information. Each target region name begins with the name of the original contig from which it is extracted, followed by a numerical identifier denoting its relative position in the contig sequence (e.g., numbered 1, 2, …, n, where 1 denotes the first target region from the 5’ end of the sequence and ‘n’ represents the total count of target regions on that contig). Lastly, the start and end coordinates of these targets on the contig are appended to the FASTA header.

Update PAF mapping file

The PAF file that is generated during the mapping step is updated to match the names of the target regions and their genomic coordinates; the start and end of the alignment block will be updated to map to the target region rather than the full, unpolished contig. Reads that do not align to any target region are discarded.

GoldPolish code execution

Using the updated PAF file generated in step 3 (Fig. 1) and the corresponding long-read sequences, GoldPolish is used to polish the intermediate FASTA file generated in step 2 (Fig. 1). Using reads mapped to the draft assembly, GoldPolish generates Bloom filters, a succinct and probabilistic data structure [23], and populates them with words of length k (i.e. k-mers) from the long reads. These Bloom filters are used to correct mismatches and indels on each individual ‘goldtig’, a raw sequence in a 1X representation of the genome of interest, and close gaps between ‘goldtigs’ with uncorrected bases from a representative sequence read [11]. A FASTA file with the polished target regions is generated as the final output of GoldPolish.

Re-inserting polished sequence targets into the draft assembly

The polished sequence targets are re-inserted into the draft assembly using the coordinate information stored in their FASTA headers. This produces the final output FASTA file.

Experimental data

ONT long reads from Drosophila melanogaster, sequenced using R10.4.1 pore chemistry, and the Homo sapiens NA24385 cell line, generated for the 1000 Genomes Project [24] sequenced using R9.4.1 pore chemistry, were obtained from the Sequence Read Archive (SRA) (Supplementary Table 1). For each set of reads, draft assemblies were generated using the standard GoldRush (v1.2.1) pipeline [11]. The ntLink k-mer size and window size parameters for the draft assemblies were chosen based on N50 length produced. For the NA24385 cell lines, PacBio circular consensus sequencing high-fidelity reads sequenced using Sequel sequencing kit 3.0 chemistry were also used to polish the base assembly; these were also obtained from the SRA [25] (Supplementary Table 4). To evaluate polishing with QUAST, we used a D. melanogaster reference genome retrieved from NCBI Reference Sequence Database (RefSeq) and the human genome reference GRCh38 from the Genome Reference Consortium (Supplementary Table 2). For evaluation with Merqury [26], we used Illumina HiSeq and NovaSeq reads sourced from the SRA for the D. melanogaster and H. sapiens assemblies, respectively (Supplementary Table 3).

Comparison to the state-of-the-art

GP-Target (v1.0.0; default parameters) was compared to Medaka (v1.11.1; default parameters) [10]. Identical BED files specifying the coordinates of the target regions in the draft assembly were passed to both polishing tools. Here, we targeted the soft-masked regions left by ntLink during the initial GoldRush assembly process; these are gaps that are filled with raw long reads that remain unpolished in the final draft assembly [18]. The target regions were polished by Medaka in three separate steps: alignment of reads to the input assembly, running a consensus algorithm across assembly regions, and the aggregation of the subsequent results to create consensus sequences. Additionally, we ran Medaka and GoldPolish as global polishing tools as a comparator for GP-Target. All benchmarking runs were performed with 48 threads. We assessed polishing results of both GP-Target and Medaka on the D. melanogaster and the H. sapiens assemblies.

Machine specifications

All runs were benchmarked on a DELL server with 128 Intel(R) Xeon(R) CPU E7-8867 v3, 2.50 GHz with 2.6 TB RAM.

Assessment of performance

The polished assemblies were assessed with QUAST (v5.0.2; -fast -large -scaffold-gap-max-size 100,000 -min-identity 80 -split-scaffold) [27] and their corresponding reference genome (Supplementary Table 2). We used the number of mismatches and indels per 100 kbp to assess the base quality of the assemblies. Additionally, with the matching short-read paired-end sequences (Supplementary Table 3), Merqury was used (v1.3; default parameters) [26] to provide a reference-free, k-mer based assessment of the polished assemblies. We examined consensus quality value (QV) as a measure of the base quality changes introduced by polishing. Lastly, BUSCO (v5.3.2; default parameters) [28] was used to determine the number of protein-coding marker genes that were recovered by polishing to further assess the quality and completeness of the polished assemblies. To report average computational resource usage, time and peak random-access memory usage from triplicate runs were recorded.

Results and discussion

We assembled the genomes of D. melanogaster (Supplementary Table 1) and H. sapiens (Supplementary Table 1) using the GoldRush assembly pipeline. We then polished each of the draft assemblies with GP-Target (using ntLink and minimap2 mapping) and Medaka using ONT long reads. The NA24385 cell line H. sapiens assembly was also polished with PacBio high-fidelity (HiFi) reads. Currently, other long read genome sequence polishers including Racon [9] or NeuralPolish [29], do not allow users to target specific regions. As such, Medaka was chosen as the sole comparator tool for GP-Target, The resulting polished assemblies were assessed using a series of quality metrics reported by QUAST [27], Merqury [26] and BUSCO [28], and their resource usage was also compared.

Assessment of base error with QUAST

We compared the performance of GP-Target with Medaka in correcting insertion, deletion (indel), and mismatch errors in the draft assemblies using QUAST. Notably, even though the targeted regions in the D. melanogaster assembly constituted just 15.4% of the overall sequence by length of selected targets, GP-Target yielded substantial improvements in base accuracy. Specifically, it reduced indel errors by 49.2% (ntLink) and 45.2% (minimap2), and mismatch errors by 55.4% (ntLink) and 42.1% (minimap2) (Fig. 2A and Supplementary Table 7). Polishing the same regions, Medaka produced 51.6% reductions in indels and 54.6% reductions in mismatches (Fig. 2A and Supplementary Table 7). For the H. sapiens assembly, where only 7.2% of the assembly was targeted with ONT long read polishing, GP-Target (ntLink) reduced indels by 22.8%, GP-Target (minimap2) by 22.9%, and Medaka by 25.2% (Fig. 2B and Supplementary Table 8). Mismatch reductions were 20.0% (GP-Target ntLink), 20.0% (GP-Target minimap2), and 20.1% (Medaka) (Fig. 2B and Supplementary Table 8). Here, both tools demonstrate comparable performance, showing a similar efficacy in fixing indels and base mismatch errors in the draft assemblies.

Fig. 2

Indels and base mismatches before and after polishing GoldRush assemblies with GP-Target and Medaka. The draft assemblies were polished with GP-Target with minimap2 mapping, denoted as GP-Target (minimap2), GP-Target with ntLink mapping, denoted as GP-Target (ntLink), and Medaka with minimap2 mapping, denoted as Medaka. The numbers on the x- and y-axes, determined by QUAST, represent A the number of indels per 100 kbp and mismatches per 100 kbp before and after polishing the D. melanogaster draft GoldRush assembly and B the number of indels per 100 kbp and mismatches per 100 kbp before and after polishing the H. sapiens NA24385 assembly with ONT long reads

Although GP-Target was originally designed for ONT long reads, we assessed its compatibility with PacBio HiFi reads. Using GP-Target with ntLink mapping, we polished the NA24385 H. sapiens base assembly with PacBio reads from the same cell line. This resulted in a 16.0% and 20.1% reduction in indels and mismatches, respectively (Supplementary Fig. 3 and Supplementary Table 8). While ONT long reads yielded marginally more pronounced improvements in base quality, these results highlight GP-Target’s versatility, especially given that only 7.2% of the assembly was polished; this ultimately illustrates its ability to effectively enhance assembly quality even with a targeted approach.

Assessment of consensus quality with merqury

Next, we used Merqury, a reference-free, k-mer based assembly assessment tool, to assess consensus quality scores (QV) before and after the targeted polishing of each dataset. Alignment-based approaches, such as QUAST, can erroneously identify genuine variants as mismatches or indels, particularly in the absence of a high-quality or strain-specific reference genome to compare against [26]. Merqury mitigates this reference bias by using k-mers derived from high-accuracy sequencing reads to assess assembly quality, rather than a reference genome [26]. However, Merqury, while providing total error rates, does not differentiate indel and mismatch errors. As such, we used both QUAST and Merqury in conjunction to determine the impact of targeted polishing on assembly base quality. Merqury uses meryl, a k-mer counting tool, to estimate the frequency of consensus errors in the assembly [26]. Through this process, Merqury estimates the consensus quality value, where a higher QV signifies a more accurate consensus—Q30 indicates 99.9% accuracy, Q40 denotes 99.99%, and so forth [26].

Polishing the D. melanogaster assembly with GP-Target using ntLink and minimap2 mapping produced an 8.3% and 7.4% increase in assembly QV, respectively, compared to the 8.3% increase obtained with Medaka polishing (Fig. 3A). This resulted in output assembly QV scores of 30.2, 30.0, and 30.2 for GP-Target (ntLink), GP-Target (minimap2), and Medaka, respectively (Fig. 3A). Polishing the H. sapiens assembly, GP-Target demonstrated a 6.9% increase in assembly QV with ntLink mapping and 7.0% with minimap2 mapping, compared to a 7.5% increase with Medaka polishing (Fig. 3B), corresponding to QV scores of 30.9 for both GP-Target polished assemblies and 31.1 for Medaka (Fig. 3B). High QV scores represent an improvement over the appreciable error rates associated with ONT sequencing, with all polished assemblies achieving base accuracy exceeding 99.9%. QV increases are observed. These high QV scores are achieved after polishing just 15.4% and 7.2% of the D. melanogaster and H. sapiens genome assemblies based on the length of selected targets, respectively.

Fig. 3

Merqury base error rate assessment of polished D. melanogaster and H. sapiens draft assemblies. The consensus quality (QV), as determined by Merqury, after polishing the A D. melanogaster and B H. sapiens draft GoldRush assemblies

Assessment of gene completeness with BUSCO

We used the Benchmarking Universal Single-Copy Orthologue (BUSCO) tool to assess the completeness of genome assemblies in the gene space [28]. BUSCO detects the presence of evolutionarily-conserved single-copy genes within a lineage to assess the completeness of gene content within the genome [28]. This evaluation method allows us to determine the capability of these polishing tools to fix mismatch and frameshift errors that could cause fragmented reconstruction in predicted gene ortholog products [28].

Assessing the D. melanogaster assembly, GP-Target’s polishing recovered 6 complete BUSCOs (0.19% of all BUSCO groups surveyed) with both mapping options (Supplementary Table 5). Genes are classified as complete when their lengths are within two standard deviations of the BUSCO group mean length [28]. In comparison, Medaka’s targeted polishing of the D. melanogaster draft assembly recovered 8 complete BUSCOs (0.25% of all BUSCO groups surveyed) (Supplementary Table 5). For the H. sapiens dataset, the GP-Target polished assembly recovered 116 additional complete BUSCOs (0.8% of all BUSCO groups surveyed) with ntLink mapping and 106 (0.8% of all BUSCO groups surveyed) with minimap2 (Supplementary Table 6). Meanwhile, polishing with Medaka recovered 132 complete BUSCOs (1% of all BUSCO groups surveyed) (Supplementary Table 6). Despite targeting only a fraction of the D. melanogaster and H. sapiens assemblies, GP-Target’s correction of indel, mismatch, and frameshift errors aids in the recovery of additional of complete BUSCOs, ultimately producing more complete genome assemblies.

Computational resource usage

While both GP-Target and Medaka produce polished genomes with comparable base quality and accuracy, GP-Target is more efficient in terms of total run time and peak random-access memory (RAM) usage regardless of the mapping algorithm used (ntLink or minimap2). When polishing smaller genomes, like D. melanogaster, GP-Target with minimap2 mapping is the most efficient, running in 8.2 min with 3.0 GB peak RAM, on average—significantly faster and more resource-friendly than Medaka, which used an average of 3.0 h and 33.8 GB of RAM (Fig. 4A, B). Similarly, comparing GP-Target with ntLink mapping to Medaka, GP-Target remains consistently more efficient, averaging 14.3 min and 1.4 GB of peak memory (Fig. 4A, B).

Fig. 4

Compute resource usage of targeted polishing with GP-Target and Medaka. D. melanogaster draft assembly mapping and polishing A wall-clock time (in hours) and B peak memory (RAM, in gigabytes) and H. sapiens draft assembly polishing C wall-clock time (in hours) and D peak memory (RAM, in gigabytes) associated with GP-Target and Medaka. GP-Target with minimap2 alignment is denoted as GP-Target (minimap2) and GP-Target with ntLink mapping is denoted as GP-Target (ntLink). All GP-Target and medaka benchmarking runs were performed with 48 threads specified in the command

GP-Target has a linear time complexity algorithm and scales favourably to large genomes, as shown by polishing the substantially larger H. sapiens assembly, which has a genome size 21.4-fold larger than that of D. melanogaster. With ntLink mapping, GP-Target achieves an average polishing time of 1.7 h with a peak RAM usage of 2.6 GB with the H. sapiens dataset (Fig. 4C, D). In contrast, Medaka requires 45.8 h and 38.0 GB of peak RAM to polish the same target regions (Fig. 4C, D). Comparing the two GP-Target mapping options, polishing with ntLink utilizes 11.8% of the peak RAM that minimap2 uses (Fig. 4D).

Comparison with global polishing algorithms

When polishing the entire genome with Medaka, the base quality of both test assemblies was substantially reduced. Medaka, in its non-targeted mode, was able to decrease the number of mismatch errors by 72.2% and indel errors by 80.0% in the D. melanogaster assembly (Supplementary Fig. 1 and Supplementary Table 7). Moreover, it was able to reduce the number of mismatches and indels by 48.1% and 72.9%, respectively, for the H. sapiens assembly (Supplementary Fig. 1 and Supplementary Table 8).

However, like its targeted mode, running Medaka as a global polisher required substantially more computational resources compared to GP-Target. Compared with GP-Target (ntLink), global polishing with Medaka spent 15.6X more time and 23.9X more RAM usage, on average, to polish the D. melanogaster assembly (Supplementary Fig. 2 and Supplementary Table 9). For the H. sapiens assembly, global polishing with Medaka used an average of 23.8X more time and 14.6X more memory (Supplementary Fig. 2 and Supplementary Table 10). While GP-Target does not improve the base quality of the two test genome assemblies as effectively as Medaka in its non-targeted mode, it offers a substantially more computationally lean polishing option, which has become increasingly vital to keep up with the onset of large-scale, high-throughput sequencing projects.

Next, we used GoldPolish as a global polisher to correct base errors in the draft assemblies in order demonstrate the added value of targeted polishing. When polishing the D. melanogaster draft assembly with GoldPolish using minimap2 mapping, mismatch errors increased by 4.8% and indel errors decreased by 1.9% (Supplementary Fig. 1 and Supplementary Table 7). GoldPolish with ntLink mapping decreased mismatch errors by 13.5% and indel errors by 15.5% (Supplementary Fig. 1 and Supplementary Table 7). GoldPolish performed similarly on the H. sapiens assembly, increasing mismatch errors by 1.0% and decreasing indel errors by 3.3% with minimap2 mapping and increasing mismatch errors by 0.2% and decreasing indel errors by 4.0% with ntLink mapping (Supplementary Fig. 1 and Supplementary Table 8). Within the GoldRush pipeline, GoldPolish is used to polish the “golden path”, which is a set of raw sequences that are relatively short in comparison to the scaffolded output of the pipeline [11]. GoldPolish builds a Bloom filter k-mer size per input contig, populating it with reads that are mapped to that respective contig. The relatively long length of scaffolded sequences likely results in large, noisy Bloom filters, which can impede polishing quality. GP-Target addresses this by targeting specific regions on the draft assembly and simplifying the Bloom filter-guided polishing step, improving polishing quality. Instead of generating one Bloom filter for each k-mer size per scaffolded sequence, GP-Target generates Bloom filters for each k-mer size per target region, only using reads mapped specifically to the target region, for error correction. Thus, we see substantially improved polishing results (Supplementary Fig. 1 and Supplementary Tables 7 & 8) with GP-Target compared with GoldPolish, despite GP-Target polishing less of the draft assembly (Supplementary Fig. 2). Moreover, with both the D. melanogaster and the H. sapiens datasets, GP-Target completed in less time and with lower peak RAM usage than GoldPolish as a global polisher regardless of the mapping tool used (ntLink or minimap2) (Supplementary Fig. 2 and Supplementary Tables 9 & 10).

GP-Target provides users with the flexibility to focus on specific regions of interest, such as critical areas for analysis or potentially erroneous regions, without requiring polishing of the entire assembly. For example, users can target protein-coding regions within a draft genome by generating BED files based on transcript or protein alignments or focus on refining regions surrounding closed scaffold gaps. Conversely, users may choose to exclude certain genomic regions, such as highly repetitive areas [30], to optimize computational resources and improve overall efficiency.

In contrast, Medaka’s targeted functionality was intended to parallelize polishing the entire genome in separate regions rather than concentrating solely on targeted sections. Using Medaka for this approach required us to develop a BASH script to execute the three main steps separately—aligning reads, running the consensus algorithm, and combining the polished sequences together [10]. GP-Target, on the other hand, requires minimal additional preparatory work.

When polishing draft assemblies with D. melanogaster and H. sapiens ONT long reads, GP-Target consistently reduces the number of indel and base mismatch errors, increases the quality value (QV), and recovers complete marker genes, as assessed by QUAST [27], Merqury [26], and BUSCO [28], respectively (Figs. 2, 3, Supplementary Tables 4 and 5). While Medaka shows similar or slightly better improvements in the same assembly metrics, GP-Target maintains a more modest computational resource usage for both test datasets (Fig. 4).

GP-Target has been integrated as the last step in the GoldRush pipeline (v1.2.1) [11], automatically identifying and polishing gaps filled in with raw, unpolished reads by ntLink that are particularly prone to errors. This approach leads to substantial improvements in the base quality of the final output without having to polish the entire assembly. When using GP-Target to polish an assembly that is not soft-masked or for more controlled targeting of regions, users can instead generate a BED file with their desired polishing coordinates. Moreover, the consistently shorter run times associated with GP-Target polishing, facilitates quicker turnaround times, making it particularly advantageous for large-scale datasets and time-sensitive analyses.

Conclusion

Here, we introduced GP-Target, a targeted long-read polishing pipeline designed to improve genome assembly base quality by targeting specific assembly regions. Our experiments, featuring fly and human datasets, demonstrate that GP-Target achieves polishing accuracy comparable to that of Medaka—a state-of-the-art polishing tool with a similar targeted functionality. However, GP-Target is consistently more computationally efficient, with shorter average run times and smaller memory footprints. Furthermore, GP-Target offers versatility, supporting mapping-based polishing through minimap2 or ntLink, while maintaining adaptability for potential integration with future tools. In summary, GP-Target emerges as a powerful, efficient, and flexible solution for targeted polishing genome assemblies using long reads.

Availability and requirements

Project name: GoldPolish-Target.

Project home page: https://github.com/bcgsc/goldpolish.

Operating system(s): Platform independent.

Programming language: Python, C++, C, GNU Make.

Other requirements: btllib v1.6.2+, ntLink v1.3.5+, minimap2, snakemake, intervaltree. License: GPL v3.

Any restrictions to use by non-academics: No.