Main

Salamanders boast an illustrious history in biological research as the animal in which the Spemann organizer1 and Sperry’s chemoaffinity theory of axonal guidance2 were discovered. Since 1768, when Spallanzani discovered tail and limb regeneration, researchers have probed this animal’s remarkable regenerative capabilities with increasing molecular resolution. A. mexicanum (Fig. 1a) was first collected by von Humboldt, and has been cultivated in the laboratory since 1864 as a model for investigating phenomena such as nuclear reprogramming, the embryology of germ-cell induction, retinal neuron processing and regeneration3. Owing to the ease with which A. mexicanum can be bred in the laboratory, a sophisticated molecular toolkit has been developed for this species, including germline transgenesis and CRISPR-mediated gene mutation as well as viral and other transfection methods. These tools have enabled discoveries such as the identification of the source cells of regeneration and molecular pathways that control regeneration4,5. A full exploitation of the axolotl model, including understanding regeneration and why it is limited in other tetrapods, requires analysis of its genome regulation and evolution. However, efforts towards comprehensive assembly of salamander genomes have been challenging owing to their large genome sizes (14–120?Gb) and the large number of repetitive regions they contain; the 32-Gb axolotl genome is ten times the size of the human genome. Here we report the sequencing, assembly and analysis of the axolotl genome.

Figure 1: Contiguity and completeness of the axolotl genome assembly.
figure 1

a, A wild-type A. mexicanum and the sequenced d/d A. mexicanum strain. b, The assembly strategy combines long-read sequencing, a novel assembler (MARVEL), error correction and scaffolding. c, A 57,385-bp PacBio read (red line) spans a large repetitive region (repeats are shown in orange; the longest repeat is 34?kb) and, together with the other long reads shown below the long PacBio read, allows assembly of the locus (green-to-red colouring indicates alignment quality; repeat-induced alignments of reads belonging to other loci have been removed). d, N(x) plot shows the percentage of the genome (x axis) that consists of contigs of at least x?kb (y axis).

PowerPoint slide

Full size image

A long-read assembler for large genomes

Our aim was to generate a genome sequence assembly for the d/d axolotl strain (Fig. 1a), which is commonly used in laboratory regeneration studies owing to its compatibility with live imaging. Taking into consideration the expected challenge of assembling the complex 32-Gb genome6, we sequenced 110?million long reads (32× coverage, N50 read length 14.2?kb) using Pacific Biosciences (PacBio) instruments (Supplementary Information section 1) to avoid the read sampling bias that is often found when using other technologies and to span repeat-rich genomic regions that cause breaks in short-read assemblies (Fig. 1b, c).

We developed an assembly algorithm (MARVEL) that integrates a two-phase read-correction procedure that keeps long PacBio reads intact for assembly (Supplementary Information section 2). MARVEL produced a contig assembly with an N50 of 218?kb. Next, we used 7× Illumina-based sequencing to correct sequence errors in 1% of the contig bases (Fig. 1b), which yielded a sequence accuracy of more than 99.2%. On the basis of the Illumina data, we estimated a heterozygosity of 0.47% (Supplementary Information section 2.2).

To provide a scaffold for the contig assembly, we generated de novo optical maps using the Bionano Saphyr system (Supplementary Information section 2.3). The Bionano map resolved contig chimaeras, which were found in 1.7% of contigs, slightly reducing N50 contig length to 216?kb (Fig. 1d). The final hybrid assembly yielded an N50 scaffold length of 3?Mb. Compared to the short-read assembly of the 20-Gb spruce genome7 or the 22-Gb loblolly pine genome8, which involved 12× long-read coverage, the axolotl assembly showed 56- and 29-fold improvements in contiguity, respectively (Table 1).

Table 1 Comparison of assembly contiguity statistics in axolotl, spruce and pine genomes
Full size table

To assess the completeness of the assembly (Supplementary Information section 4.1), we first determined the number of aligning non-exonic ultraconserved elements9 (UCEs). We found that 194 (98.5%) of 197 non-exonic UCEs that are conserved across vertebrates align to the axolotl assembly. By comparison, 189 and 192 UCEs align to the Tibetan frog and Xenopus genomes, respectively, and 195 UCEs align to the coelacanth genome, indicating that the completeness of the axolotl genome assembly is comparable to or better than the two other amphibian genomes, which are smaller than 2.3?Gb10.

To further assess the completeness of the assembly, we generated a comprehensive gene catalogue by sequencing mRNA from 22 tissues (Supplementary Information section 3). Tissue-specific transcriptome assemblies and a composite assembly of all 1.5 billion transcript reads resulted in 180,649 transcript contigs (Supplementary Table 6) that contained 99% of the conserved core eukaryotic genes11 and achieved the highest BUSCO score (http://busco.ezlab.org/) of an axolotl transcriptome reported to date (Supplementary Information section 3.4). More than 85% of the transcripts aligned to the genome along at least 95% of their length (Supplementary Information section 3.5), confirming the high completeness of the assembly. Furthermore, 71% of transcript contigs in which more than 95% of the sequence aligned with the genome were located on single scaffolds, demonstrating the high contiguity of the assembly. Using this comprehensive transcript set, we annotated a total of 23,251 protein-coding genes in the axolotl genome, a similar number to those found in other vertebrate genomes (Supplementary Information section 4.2).

Expansion of long terminal repeat retroelements

Given the similar number of genes in the A. mexicanum genome in comparison to other smaller vertebrate genomes, we analysed repetitive sequences (Supplementary Information sections 4.2.2, 4.2.3). Repetitive sequences made up 65.6% of the contig assembly, representing a total of 18.6?Gb. Distinct long terminal repeat (LTR) retroelement classes and endogenous retroviruses made up the largest portion of the repetitive sequences (Fig. 2a, b, Supplementary Table 13) and included elements of more than 10?kb in length (Fig. 2c, Extended Data Fig. 1). Such long elements pose challenges for assembly, and indeed 97% of contigs ended in LTR elements. The number of substitutions to the repeat consensus, which is an estimate of the relative age of the LTR retroelement, indicates that the axolotl genome has undergone a long period of transposon activity followed by a recent and apparently continuing burst of expansion (Fig. 2d). This profile is consistent with previous small-scale characterizations of other salamander genomes12.

Figure 2: The axolotl genome contains an expansion of LTR retroelements.
figure 2

a, Pie charts of major repeat classes (LINE, long interspersed nuclear elements; SINE, short interspersed nuclear elements) show an abundance of LTR elements. b, Phylogenetic tree of axolotl LTR-element clusters (black) and all LTR elements from GyDB2.037. Annotated clusters are indicated by colour, non-annotated clusters are in grey. Note that Errantiviridae (blue), Caulimoviridae (red) and Athila/Tat (orange) families are not found. c, Box plots show the length distribution of LTR families (ERVL, endogenous retrovirus-like). Boxes indicate the first quartile, the median and the third quartile with whiskers extending up to 1.5 times the interquartile distance. Outliers are defined as data points outside the whiskers and are shown as dots. Quantitative data and sample sizes are shown in Source Data. d, Relative age (Kimura distance) suggests prolonged transposition activity followed by a recent activity burst.

PowerPoint slide

Source data

Full size image

The presence of many repeated elements contributes to a median intron size (22,759 bp) 13, 16 and 25 times that observed in human (1,750 bp), mouse (1,469 bp) and frog (906 bp), respectively (Fig. 3a, Supplementary Information section 4.3), a trend that was previously observed in five genes obtained from selective bacterial artificial chromosome sequencing of the axolotl genome13. Figure 3b shows a typical gene organization in axolotl compared to its human orthologue. Consistent with intron expansion, a distance comparison of pairs of highly conserved non-exonic elements shows that intergenic regions in the axolotl genome are 12 to 17 times larger than those in human, mouse and frog (Supplementary Information section 4.4).

Figure 3: Genome organization and loss of Pax3.
figure 3

a, Intron size of developmental genes appears to be under constraint in A. mexicanum. Violin plots represent the full distribution of intron sizes (thick bar, first to third quartile; white dot, median). ***P?<?10−11; two-sided Wilcoxon rank-sum tests. Quantitative data are shown in Source Data. b, Organization of the Agr2 exon–intron structure shows a consistent expansion of axolotl intron sizes compared to those in the human orthologue, resulting in a gene that is 4.3 times larger. c, Comparison of genes and repetitive elements in the HoxA cluster. CNEs that align to the axolotl HoxA cluster are shown in red. d, Axolotl lacks the Pax3 locus. Analysis of tetrapod-conserved genes and CNEs associated with Pax7 and Pax3 genomic loci. Red, gene sequences and CNEs that are absent in axolotl; *CNEs that overlap well-characterized mouse Pax3 enhancers38,39.

PowerPoint slide

Source data

Full size image

HoxA cluster and intron size constraints

To examine gene cluster organization within this large genomic context, we focused on the HoxA locus, which has an important role in proximal-to-distal limb development and is reactivated during limb regeneration14,15. The entire HoxA locus is contained on a single contig (Fig. 3c), and the conserved neighbouring gene Evx1 is contained on the same 3.34-Mb scaffold. Compared to the orthologous human and frog clusters, the A. mexicanum HoxA cluster has a substantially increased repeat content and is 3.5 times larger, mostly owing to a 170-kb expansion between HoxA3 and HoxA4 (Fig. 3c). Notably, highly conserved non-exonic elements that putatively overlap cis-regulatory elements are not interspersed in this 170-kb region, but remain in proximity to HoxA3 and HoxA4. The axolotl has a typical HoxA gene structure, with two coding exons separated by an intron. Notably, in contrast to the overall expansion of intron sizes, the intron sizes in the axolotl HoxA locus are very similar to those in other vertebrates, with the exception of AmHoxA3, which is also the longest of the HoxA genes in other tetrapods (Supplementary Table 17). Selected HoxC and HoxD genes examined in the red spotted newt exhibited similar properties16.

On the basis of these observations, we examined the intron size distribution among a larger set of orthologous genes involved in developmental processes. While introns of non-developmental genes in axolotl show a median size expansion of 13- to 25-fold compared to human, mouse and frog, the expansion of introns of developmental genes is significantly lower (6- to 11-fold, P?<?10−11) (Fig. 3a, Supplementary Information section 4.3). In contrast to human, mouse and frog, introns of developmental genes in axolotl are shorter than introns of non-developmental genes. Furthermore, axolotl multi-exon genes that contain only short introns exhibit gene ontology enrichments related to developmental patterning that are not enriched in multi-exon genes with larger introns (Supplementary Table 16). These results suggest that intron size in developmental genes is under constraint in the axolotl, possibly because smaller gene sizes facilitate rapid transcription and thus upregulation of these genes in specific developmental contexts.

A reduced Pax-family complement

Next, we interrogated the genome for families of canonical developmental signalling molecules (Supplementary Information section 5). All three hedgehog paralogues as well as a full set of vertebrate Wnt genes were present (Extended Data Fig. 2a, b). However, we noted that certain members of the paired box family of transcription factors, which have diverse roles in tissue formation, were not found in the assembly. Consistent with the absence of Pax4 in amphibians and other vertebrate lineages17, the axolotl genome does not contain Pax4 but does contain Pax10. Notably, despite the presence of the Pax3 and Pax7 paralogues in all other known vertebrate lineages, we were able to identify Pax7 but not Pax3 in the axolotl genome assembly (Extended Data Fig. 2c). No Pax3 sequence was found in either the raw PacBio sequencing reads or the transcriptome. To confirm the loss of Pax3, we further examined the genomic region that would be syntenic for Pax3 for the presence of neighbouring genes and highly-conserved non-exonic elements (CNEs). The orthologues of genes surrounding mouse Pax3 (Sgpp2 and Epha4) were present in the A. mexicanum genome assembly; however, neither the Pax3 gene nor any of the Pax3-associated CNEs were found (Fig. 3d). By contrast, several CNEs that overlap the Pax7 gene were identified in the assembly. Together, this evidence strongly suggests that Pax3 and several of its cis-regulatory elements are absent in the axolotl genome, probably owing to a deletion.

Axolotl Pax7 has similar functions to Pax3

To functionally assess the consequence of the absence of Pax3 in the axolotl, we used TALEN- and CRISPR-mediated gene editing18 to mutate Pax7. In other vertebrates, Pax3 and Pax7 play key roles in muscle, neural tube and neural crest-derived tissue development19. Although these two genes share some common functions, deletion of Pax3 or Pax7 causes distinct phenotypes in mice20,21,22. We investigated whether frameshift deletions introduced into the AmPax7 gene would yield a comparable Pax7 phenotype, or whether AmPax7 may have taken on functions that are carried out by Pax3 in other vertebrates. Two different TALEN-mutant alleles (7-nt and 20-nt deletions) of AmPax7 were bred through two generations (Fig. 4a, Supplementary Information section 6). In the F2 generation, the developmental phenotype described below was observed in 83 out of 345 (24%) progeny from the Pax7Δ20nt/+ intercrossing and 57 of 232 (24.6%) progeny from the Pax7Δ7nt/+ intercrossing (Fig. 4b, Extended Data Fig. 3). The phenotype was evident in homozygous mutants, as analysed by PCR and loss of protein (Supplementary Information sections 6.1, 6.3). This information, combined with the CRISPR-mediated gene mutation results (Supplementary Information sections 6.2), shows that the homozygous Pax7Δ20nt/Δ20nt and Pax7Δ7nt/Δ7nt mutants represent recessive, complete or partial loss of Pax7 function.

Figure 4: Pax7 mutation in A. mexicanum yields a phenotype similar to that of Pax3−/−Pax7−/− mice.
figure 4

a, Deletion of AmPax7 coding sequences using TALEN and CRISPR. Deletions were made in exon 1 or exon 2. The first three AmPax7 exons (Ex) are shown. Red rectangles, TALEN targets; arrows, CRISPR-guide RNA (gRNA)-binding sites. b, Images of 6-month-old Pax7Δ20nt/Δ20nt mutants compared to controls show loss of body elongation. Scale bars, 1?cm. c, Reduced body wall muscle in Pax7 mutants. Immunofluorescence images of myosin heavy chain (MHC, red) and DAPI (blue) in trunk cross-sections from a 6-month-old Pax7Δ20nt/Δ20nt mutant (Pax7-TALEN#2) and a control animal. Scale bar, 500?μm. d, Limbs of Pax7 mutants lack muscle. Forelimb (left; scale bars, 500?μm) and immunofluorescence images of MHC (red) and DAPI (blue) in limb cross-sections (right; scale bars, 100?μm) of a 3-month-old Pax7Δ20nt/Δ20nt mutant and a control animal. e, Loss of prefrontal bone in Pax7 mutants. Dorsal and lateral views of Alcian blue and Alizarin red-stained Pax7Δ20nt/Δ20nt and Pax7-Ex1-CRISPR#3 mutants and controls (right). Red arrows, prefrontal bone; blue arrows, maxillary bone. The yellow arrowhead points to a small remnant of the prefrontal bone. Scale bar, 1?mm. f, g, Reduced melanophores, xanthophores (f) and iridophores (g) in Pax7 mutants. Images of 17-day-old Pax7-Ex2-CRISPR#1 (f), 2-month-old Pax7Δ20nt/Δ20nt (g) mutants and controls (Ctr). Right panels in f show a magnified view of the outlined area; red arrows in g point to the eyes that are magnified in the insets; black arrows indicate the belly iridophores in the control. Scale bars, 1?mm. h, Neural tube closure defect in Pax7 mutants. Images of a 31-day-old Pax7-Ex2-CRISPR#1mutant and control (Ctr). Right, magnified view of the outlined area. Scale bar, 1?mm. Quantitative data and sample sizes are provided in the Life Sciences Reporting Summary and Source Data.

PowerPoint slide

Source data

Full size image

The Pax7Δ20nt/Δ20nt and strong F0 Pax7-CRISPR mutants exhibited a curved body, were unable to maintain an erect posture and exhibited a delay in growth compared to controls. Immunohistochemical analysis of trunk or tail cross-sections of early stage, 20-day-old Pax7Δ20nt/Δ20nt or 17-day-old F0 Pax7-CRISPR axolotls showed normal muscle mass. However, at later stages, consistent with the mouse Pax7 deletion phenotype, tail and trunk muscles were greatly decreased (Fig. 4c, Extended Data Figs 4, 5, 6). Remarkably, the Pax7 mutant axolotls also completely lacked limb muscle (Figs 4d, Extended Data Fig. 7). In mice, Pax3, but not Pax7, is required for limb muscle formation21,22,23 (Supplementary Table 18). These results demonstrate that AmPax7 has comparable functions to MmPax3 in the control of limb muscle genesis.

In mice, Pax7 deletion affects craniofacial neural crest derivatives, including the facial bones20, whereas in zebrafish, pax7 mutants show loss of xanthophores and reduction of melanophores, but no loss of iridophores24. The AmPax7 mutants lacked a prefrontal bone, had a reduced number of melanophores and were deficient in xanthophores and iridophores except in the eyes (Fig. 4e–g, Extended Data Fig. 8). Pax3 deletion in mice is associated with neural tube closure defects22,23 (Supplementary Table 18). Similarly, Pax7Δ20nt/Δ20nt-TALEN and Pax7-CRISPR axolotls displayed failed closure of the neural tube in the midbrain (Fig. 4h, Extended Data Fig. 9). In summary, mutation of AmPax7 yields a combination of the Pax3- and Pax7-mutant phenotypes that are observed in other vertebrates (Supplementary Information section 6). It will be interesting to understand how the regulation of Pax7 has changed in axolotl to enable the loss of Pax3, which is essential in other vertebrates.

Species-restricted genes in regeneration

Previous searches for mRNA and microRNA (miRNA) transcripts associated with limb regeneration relied on mapping to de novo transcriptome assemblies. We sought to re-examine these datasets using our newly acquired genomic data. Recent functional work has highlighted the role of diverged gene or protein function during regeneration25,26,27. Analysis of published tissue-enriched datasets28, combined with regeneration time courses29,30 and our own transcriptional profiling of 22 tissues, identified five transcripts that are upregulated in the limb blastema (the mass of proliferating cells involved in regenerating the limb) with orthology limited to non-amniote vertebrates (Supplementary Information section 7). One of these protein sequences shows a weak similarity to tectorin, a basement membrane component normally found in the inner ear, consistent with studies that implicate extracellular matrix components with having an important role in limb regeneration31,32. Notably, another of these transcripts encodes a Ly6 family member in the urokinase type plasminogen activator surface receptor (uPAR) class. Previous studies had identified the salamander-specific Ly6 family member Prod1 as a key factor involved in salamander limb development and regeneration25,33. Our results suggest that Ly6 family members have a broader role in limb regeneration. Finally, we also investigated the role of non-coding RNAs by mapping a dataset of small RNA sequences expressed in the limb and limb blastema34 to our genome assembly. This analysis classified 93 small RNAs as pre-miRNA sequences, of which 42 appear to be novel miRNAs (Supplementary Information section 7.2). Taken together, these data point to a potential role in limb regeneration for several coding and non-coding sequences that have been lost or diverged rapidly in amniotes. Future investigations of such sequences are likely to be a fruitful avenue for understanding the evolution of regeneration capabilities.

Discussion

We have generated a comprehensive whole-genome assembly for the salamander A. mexicanum, and analysis of this assembly has allowed us to draw conclusions about the structure of the expanded genome. Our data, together with data from plants and partial data from several other salamander species, show that LTR expansion is a major contributor to giant genome size across animals and plants6,12,35. Our assembly is sufficiently complete to reliably detect the absence of Pax3, which is present in fish and other amphibians. This analysis was confirmed using gene editing, which showed that AmPax7 has assumed functions that are carried out by Pax3 in other animals.

Functional analysis of axolotl development, physiology and regeneration is facilitated by the availability of tissue- and time-dependent gene expression profiles28,29,30,36. The axolotl genome provides a foundation for applying methods such as chromatin immunoprecipitation with sequencing (ChIP–seq) or assay for transposase-accessible chromatin using sequencing (ATAC-seq) to investigate the genomic basis of gene regulation during regeneration. Together with methods such as CRISPR-mediated gene editing, viral expression methods, transplantation and transgenesis, the axolotl is a powerful system for studying questions such as the evolutionary basis of its remarkable regeneration ability. Our approach of long-read sequencing, optical mapping and genome assembly using MARVEL also demonstrates that it is now feasible to assemble very large repeat-rich genomes.

Methods

No statistical methods were used to predetermine sample size. The experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment.Axolotl genomic DNA was prepared from freshly isolated liver and spleen of an individual three year old adult d/d male using DNAzol followed by phenol/chloroform extraction and ethanol precipitation.

A total of 50 size-selected SMRTbell libraries were prepared with a minimum fragment length cutoff between 10?kb and 20?kb. We sequenced medium and large insert libraries on the PacBio RSII instrument, making use of three different sequencing polymerases (P4, P5 and P6) and the corresponding sequencing chemistries (C2, C3 and C4). Movie times ranged from 180?min to 360?min with the majority of SMRT cells (1,414 of 1,933) at 240?min.

Sequences were assembled using the MARVEL assembler.

Optical mapping was performed using the Saphyr System (Bionano) based on NanoChannel array Technology. DNA was labelled with Nt.BspQI and Nb.BssSI enzymes in separate labelling reactions. Each enzyme reaction was run on the Saphyr System. 2.813 Tb of data were collected on three Saphyr Chips for Nt.BspQI and 2.0 Tb of data were collected on two Saphyr Chips for Nb.BssSI samples; single molecule N50 lengths were 240?kb and 184?kb, respectively. Each dataset was de novo assembled using Bionano Solve 2.1 software.

RNA was isolated from 22 tissue types using TRIzol or RNeasy reagents and sequenced using Illumina technology. The Trinity software package was used for transcriptome assembly.

Code availability

The MARVEL assembler with documentation is available at https://github.com/schloi/MARVEL.

Data availability

A browser of the axolotl genome is available at https://genome.axolotl-omics.org. The transcriptome assembly and the genome and transcriptome BLAST database can be accessed at https://www.axolotl-omics.org with no restrictions. The sequence data and both assemblies have been deposited in the NCBI BioProject database with accession numbers PRJNA378970 (genome data) and PRJNA378982 (transcriptome data). Both genome data and transcriptome data were deposited to the NCBI Nucleotide Database (nuccore) with accession numbers PGSH00000000 and GFZP00000000, respectively.