To generate a diverse set of sequences involved in organizing and regenerating the primary body axis, two independent cDNA libraries were used for sequencing. One was derived from dorsal regions of stage 18-22 embryos containing neural tube, somite and notochord - called the 'neural tube' library - the other from 6-day post-amputation regenerating tail blastema. From 18,432 sequencing attempts 17,522 high-quality sequences were obtained after Phred analysis 11. All sequences are 5' reads of the inserts. Of 17,522 high-quality, single-pass sequencing runs, 32 clones contained no insert and 137 sequences were below 32 base pairs (bp). These sequences were excluded from further analysis (32 bp representing the lower limit for assembly of a sequence using TIGR-assembler), yielding 17,352 clones for final analysis. The neural tube library was the origin of 7,469 sequences and the blastema library of 9,883 sequences (Table 1, and see Materials and methods). As shown in Figure 1a, the average sequence read length peaked between 500 and 600 nucleotides with an average length of 510 nucleotides and a maximum of 871.

The blastema and neural tube libraries were unnormalized and unamplified. We assessed library quality and diversity on the basis of the number of redundant clones in the library. Redundancy was estimated by performing BLASTN searches 12 against all clones sequenced. After sequencing 10,752 clones of the blastema library 42% of the sequences were still unique, and 50% of clones were still singlets after sequencing 7,680 clones from neural tube, indicating that both libraries display high diversity.

To identify ESTs belonging to the same open reading frames (ORFs), sequences were assembled into contigs using TIGR-Assembler version 2 13. The 17,353 sequences assembled into 6,594 contigs, of which 217 were less than 100 nucleotides long and excluded from further analysis. A total of 6,377 contigs was therefore left for final analysis (Table 1). Of these, 4,561 contigs contained a single clone. The average contig length of the remaining dataset was 616 nucleotides (Figure 1b). Other than singlets, most of the contigs consisted of two ESTs (884 contigs, Figure 1c). The largest contigs included cytochrome c oxidase subunit I (469 ESTs), 12S rRNA (445 ESTs), nuclear factor 7 Zn-binding protein A33 (332 ESTs), type II keratin (274 ESTs), keratin (211 ESTs) and cytoplasmic beta-actin (206 ESTs) (Table 2).

A total of 1,134 ESTs were available from A. mexicanum in the National Center for Biological Information (NCBI) EST databases prior to this work, most of which originate from a sequencing effort of the Voss laboratory (14 and S.R. Voss, D. King, N. Maness, J.J. Smith, M. Rondet, S.V. Bryant, D.M. Gardiner, and D.M. Parichy, unpublished work (NCBI-accession numbers BI817205-BI818091); see also 15). We examined to what extent our EST dataset overlapped with the sequences available to date. Only 600 of the ESTs in the public database identified one of our contigs in a BLASTN search as a homolog; in 85% of cases, the E-value was below 1E-50 and the sequences can be considered as potentially identical. Existing ESTs in the database largely originate from regenerating limb (S.R. Voss, D. King, N. Maness, J.J. Smith, M. Rondet, S.V. Bryant, D.M. Gardiner and D.M. Parichy, unpublished work). There was, however, only a slight bias of matching contigs to regenerating blastema (49%) as compared to neural tube (44%). Seven percent of identified contigs were found in both libraries. These results mean that our EST data enriches the existing sequence resource of A. mexicanum with approximately 6,000 new gene sequences.

To identify putative homologies to known proteins, we subjected the contigs to BLASTX searches against the non-redundant protein database (NR, NCBI) where a cutoff E-value of 1e-05 was used for parsing output files. In our annotation, we used an E-value of 1e-20 as an upper limit to assign significant homology. We note that this does not imply that such sequences are true orthologs. In addition, in cases where no significant homology was found, we used an E-value limit of 1e-05 to designate weak homology. We find this additional category of 'weak homology' useful for data mining. As most contigs do not represent full-length sequences, it is possible that only a highly divergent region of a gene sequence is available in our collection. The category of weak homology allows us to find potential homologs in such situations. For example, the BLAST search for contig Am_4671 yielded the GenBank entry NP_004055, cyclin-dependent kinase inhibitor 1B (Homo sapiens), as the top hit with an E-value of 4e-07. This assignment was based on the carboxy-terminal 120 amino acids of the protein, which represents the less conserved region. When we isolated a full-length clone for Am_4671 from our library, we could confirm that it is indeed the axolotl ortholog of cyclin-dependent kinase inhibitor 1B (p27^Kip1), as discussed later.

Taken together, a total of 3,718 (58%) sequences shared homology with a protein from selected model organisms in the non-redundant database and could be assigned a putative identity. The E-value distribution of the top hits in the non-redundant database is shown in Figure 2a. Of the contigs, 11% matched a protein with an E-value below 1e-99 and are therefore likely to be true orthologs. Seventy percent of the contigs found a hit with an E-value between 1e-20 and 1e-99 and were assigned significant homology. Finally, 19% of contigs had a first hit with an E-value between 1e-19 and 1e-05 and were assigned weak homology to a protein from the non-redundant database. For annotating our database, these top hits from human, mouse (Mus musculus), rat (Rattus norvegicus), frog (X. laevis), zebrafish (Danio rerio), fugu (Takifugu rubripes), fruitfly (Drosophila melanogaster), mosquito (Anopheles gambiae), worm (Caenorhabditis elegans), newts and the yeast species Saccharomyces cerevisiae, Schizosaccharomyces pombe and Candida albicans were collected and the closest homolog from the above species was used to assign a putative identity.

To estimate how many of the clones are full length we examined the BLAST alignments for the position of the alignment in respect to the database sequence. Of the 3,718 sequences with homologs, 1,107 (29.8%) could be aligned in the amino terminus (with the alignment starting before position 10). As the library was poly(dT) primed, many of these clones are likely to represent full-length inserts. Of these 199 (5.4%) could be aligned from the amino terminus to the carboxy terminus and are potential full-length sequences.

Forty percent of our EST sequences did not generate a significant hit in the non-redundant protein database. The availability of additional sequence databases including complete genome sequences from several organisms allowed us to expand our BLAST searches to identify all possible homologs to the A. mexicanum contigs. With the remaining set of contigs, we first performed BLASTN searches against the nucleotide non-redundant (NT) database and BLASTX searches against the EST database. Finally, we performed BLASTX searches against the fugu and human proteomes. In all cases, an E-value of 1e-05 was used to assign potentially homologous sequences. Sequences in the NT database identified an additional 134 contigs and a further 220 contigs found a hit in the EST databases. A homolog was found for 3,340 (52%) contigs in the fugu proteome and 3,698 (58%) contigs shared homology with a protein from the human proteome. In total, an additional 468 contigs identified a homolog in the selected databases beyond the original assignment from the non-redundant protein database (Figure 2b).

No homologous sequence could be found for 2,191 (34%) contigs in any of the databases searched. Because the library was poly(dT) primed, many of these sequences could represent 3' untranslated regions (3' UTRs). We determined that 953 sequences (43% of non-homologous contigs) contained no ORF and were therefore potential untranslated regions. Thirty of the sequences shared homology to an existing A. mexicanum clone from the EST database (Table 3). The complete list of unique ESTs can be downloaded from 16.

From the homologous proteins found, contigs were assigned a biological process, molecular function and cellular component from the Gene Ontology (GO) database 17. The closest annotated homolog in the GO database was used, using an E-value of 1e-20 as a cutoff, for assigning these categories. A biological process could be assigned to 2,156 contigs (34% of all contigs and 58% of those sharing a homolog in the non-redundant database); 2,186 contigs (34% and 59%, respectively) were assigned a molecular function; and 2,198 contigs (34% and 59%, respectively) could be assigned a cellular component. The most abundant molecular function assigned was 'death receptor interacting protein', followed by 'peptidase', the highest-ranking biological process were 'biological process unknown' and 'proteolysis/peptidolysis' and the most abundant cellular components assigned were the 'actin cytoskeleton' and 'transcriptional repressor complex'.

The largest fraction of the contigs was assigned a cellular process in the GO category biological process (87% of annotated contigs) (Figure 3a). We split the biological processes further into different categories: the most abundant categories were 'protein metabolism/modification' (18% of assigned contigs); 'housekeeping functions/metabolism' (17%); 'intracellular transport' (15%); 'cell cycle/proliferation' (13%); 'RNA metabolism' (13%); 'intracellular signaling' (8%); and 'DNA metabolism/repair' (5%) (Figure 3a, Table 4). A list of annotated contigs is downloadable from 16.

To identify potential domains in the axolotl contigs, we performed RPS-BLAST searches against the conserved domain database (CDD, NCBI) 1218 using the default cutoff E-value of 0.01. A total of 2,199 (34.5%) contigs had a known protein domain in either the CDD or the SMART or PFAM databases. A detailed list of common protein domains identified in our dataset is given in Table 5. Among the protein domains identified were homeobox domains such as HOX, PAX and Prox1, eight helix-loop-helix (HLH) domains, RNA-binding domains such as KH and RRM, 69 kinase domains, metal- and lipid binding domains and domains involved in cell-cycle control and ubiquitination (RING fingers, HECT domains, three cullin domains and 12 cyclin domains). Many of these domains were annotated for the first time in a sequence from A. mexicanum. We also compared the occurrence of those domains in other vertebrate species. For most of the common protein domains, only a fraction were found in our dataset; many of these are quite abundant compared to X. laevis or Gallus gallus. The RNA-binding domains KH and RRM especially showed high abundance in our contigs. A complete list of domains is downloadable from 16.

We assigned cellular functions to the identified domains and analyzed the output according to the functional distribution of contigs (Figure 3b). The most abundant domains were found in the category 'intracellular transport'; this is due to redundant annotations of small GTPases. The second largest fraction belonged to 'RNA-binding and metabolism', followed by 'DNA-binding and transcriptional control'.

As mentioned earlier, the Mexican axolotl is an important model organism for a number of reasons. First, it is the premier vertebrate model for studying regeneration. Second some aspects of caudate development, for instance mesoderm involution and notochord formation, more closely resemble those found in higher vertebrates than do those in other amphibian embryological models such as X. laevis 19. Finally, the axolotl has interesting developmental features, particularly in relation to metamorphosis. The axolotl undergoes 'cryptic metamorphosis', which is defined by its existence in a perrenibranchiate state and retaining some larval features into adulthood (for instance gills, larval skin morphology, caudal fins). The animals become sexually mature in this state, and develop only small rudimentary lungs. So far, very few markers are available to study these processes in this organism.

We examined our dataset for genes that are potentially useful for studying regeneration features or developmental processes. To this end, we analyzed our data for genes that are either involved in regulating the cell cycle - as would be expected for the highly proliferative tissue of a regenerating body structure - or could play an essential role during development and metamorphosis from the larval to the adult stage. A list of genes that could be assigned to either cell-cycle regulation or development is shown in Table 6. Among the genes involved in cell-cycle regulation were A-, B- and E-type cyclins, cyclin-dependent kinase 4 (Cdk4), Polo kinase, the kinase inhibitor p27^Kip1, the protein phosphatase Cdc25A, as well as the anaphase-promoting complex (APC) activator proteins Cdc20 and Cdh1. Representing genes involved in developmental processes, we found transcription factors such as HoxA2, B12, C4 and C8, Pax6, as well as Cdx1 and Cdx2. Furthermore we found several genes for proteins that are part of the transforming growth factor-beta (TGF-β) signaling pathway, such as TGF-β, bone morphogenetic protein 1 (BMP-1), BMP and activin membrane-bound inhibitor, activin receptor type II, as well as the transcription factors Smad5 and Smad8. Genes for proteins such as Smad8 and BMPs might be of especial interest to the research field of embryonic development, as they have been associated with mesoderm involution 20. Other important developmental genes that could be found in our dataset include those for Wnt5 and Wnt8, Sonic hedgehog, retinoblastoma binding protein 2, beta-catenin, as well as Frizzled 2, 5 and 7. Finally, it has been shown that the thyroid hormone receptor pathway has an essential role in the timing of metamorphosis in A. mexicanum 212223. We identified the protein TRIP12 (thyroid hormone receptor interacting protein 12), which is a HECT-domain-containing ubiquitin ligase and could have an essential role in regulating thyroid hormone response during development and/or metamorphosis.

The EST collection will provide rich data for the phylogenetic comparison of particular genes. Cell cycle and cell differentiation are cellular functions that have been modified in various organisms through evolution and it will be interesting to understand the evolutionary basis of such changes. Here we analyze a particularly interesting gene family, the CDKN1 family of cell-cycle regulators which inhibit cell-cycle progression by binding to and inactivating CDKs. As a starting point for phylogenetic analysis, the mitochondrial 12S ribosomal RNA gene from our collection resulted in the expected tree, with the anuran amphibian X. laevis and the caudate A. mexicanum grouping together compared to other vertebrates such as fish, birds and mammals (Figure 4a). Next, we constructed an unrooted phylogenetic tree to compare members of the cyclin B family - cyclins B1, B2 and B3. The sequences of each family member formed strictly separate groups, with the A. mexicanum and X. laevis cyclin B1, B2 and B3 genes grouping with their vertebrate orthologs (Figure 4b).

In contrast, we obtained a quite different picture when we examined the CDKN1 family. In most vertebrates, this family consists of three members: p21 (CDKN1A), p27^Kip1 (CDKN1B) and p57 (CDKN1C). In X. laevis, however, only a single family member called p28^Kix1 (also called p27^Xic1), which shows unusual sequence features compared to the p27 sequences from any other vertebrate species, had been described in the literature 2425. We wondered whether A. mexicanum harbored the 'canonical' p27^Kip1 or a p28^Kix1 similar to that of Xenopus. We initially searched our A. mexicanum data for CDKN1 orthologs and, in contrast to Xenopus, we found a bona fide p27^Kip1 sequence that clusters closer to vertebrate p27^Kip1 sequences compared to the Xenopus p28^Kix1 (Figure 4c,d). Considering this interesting finding, we then undertook a more complete analysis of the CDKN1 family in vertebrates by searching for CDKN1 family members in several databases: the sequenced genomes from human, mouse, rat, fugu or zebrafish, the recently released genome sequence of X. tropicalis, the X. laevis EST collection, the zebrafish and fugu genomes, and a complementary A. mexicanum and A. tigrinum EST set generated by Putta et al. 26.

This data mining revealed two striking features about the distribution of CDKN1 family members among vertebrates (Table 7). First, the p28^Kix1 orthologs were only found in amphibians (X. tropicalis, X. laevis, A. mexicanum, A. tigrinum tigrinum). We were not able to identify a p28^Kix1-like gene in any other database. These p28 orthologs group as a distinct branch in an unrooted phylogenetic tree (Figure 4c,d). These data so far suggest that the p28 family is a CDK inhibitor that is specific for amphibians. With new genome sequence data being released, it will be interesting to see whether the most closely related lineage of birds contains a p28-like gene or whether this gene family is found solely in amphibians.

Second, CDKN1B (p27^Kip1) and CDKN1C (p57) were present in the A. mexicanum databases but were not found in either X. laevis or X. tropicalis, which have far more EST and genome sequence information (Table 7, Figure 4c,d). While it is not possible to conclude definitively that Xenopus species lack these genes, the current data are highly suggestive of such a scenario.

We examined in depth the phylogenetic relationships of the CDKN1 family members among vertebrates by constructing unrooted phylogenetic trees, either using the most conserved, amino-terminal 88-amino-acid domain, which includes the functionally important Cdk2-interaction region, or the entire coding sequence. Analysis of the amino terminus showed that while A. mexicanum p27 and p57 clearly grouped with their respective orthologs from other vertebrates, the p28^Kix1 proteins from axolotl and the two Xenopus species clustered as a group distinct from any of the other CDKN1 families (Figure 4c). The p28^Kix1 family showed a closer relationship to p57 than to other CDKN1 members, branching off close to the p57 family. Phylogenetic analysis using the entire coding sequence of the CDKN1 genes, which includes the Cdk2- and PCNA-binding site, resulted in a closer grouping of p28 with the p27 branch (Figure 4d). In both cases, however, the p28 family clearly formed a separate group from the other CDKN1 families.

A relational database with a web-based front end was created to store, navigate and annotate analyzed contigs. The main object of the database is the annotated sequence contig, which contains information about its length, putative identity, computationally calculated expression profile, GO annotation, homologous proteins and identified domains, as well as number and identity of ESTs that build the contig (Figure 5a). The Gene Identifier (GI) and GO annotation can be modified by the administrator. To circumvent the problem of split contigs, we introduced a super-contig, to which related contigs can be assigned. Furthermore, the administrator can modify the relationship of EST to contig manually. All protein and domain alignments, as well as the assembly of the EST sequences of a contig are stored and can be viewed by the user. On the contig main page, three homologs at most from selected species are shown, with a full list of homologs from selected species displayed on the protein information page (Figure 5c). To make use easier, an image of the identified domains with the beginning and end base pair of the alignment is shown on the contig page. Individual ESTs can be accessed via the contig page, including their length, storage information, quality information and available trimmed EST-sequence (Figure 5b).

Some of the main advantages of this database are: first, the direct links to source databases such as the NCBI sequence database, GO database, CDD, and the Smart and Pfam databases for identified domains; second, direct visualization of source data such as sequence alignments of contigs to homologs and domains, as well as alignments of EST assemblies; third, easy retrieval of sequences for further analysis like BLAST-searching; fourth, user-specific annotation of contigs; and fifth, easy manipulation and editing of contig annotations. The database will be available from 27.

Total RNA was purified using Trizol (Invitrogen) from 6-day regenerating tail blastemas and from neural tube-somite-notochord-containing tissue dissected from stage 18-22 A. mexicanum embryos. Total RNA quality was assessed by determining the relative brightness of the 28S:18S rRNA bands (2:1). For library construction mRNA was purified and size fractionated, then poly(dT)-primed cDNA was synthesized and directionally cloned into the NotI-SalI sites of the pCMVSport6 vector. DNA was transformed by electroporation into EMDH10B-TONA bacteria (library construction performed by Invitrogen). Two separate, unnormalized libraries were produced. The blastema library contained an average insert size of 1.67 kb and 2.67 × 10⁷ independent transformants and the neural tube library had an average insert size of 1.5 kb and 1.9 × 10⁷ transformants. From each library 100,000 clones were arrayed into 384-well plates (Resource Zentrum/Primary Database, Berlin, Germany).

For sequencing, single-pass reads from the 5' end of the library inserts were performed using a custom-designed SP6 primer: GCACATTAGGCCTATTTAGGTGACA. DNA from bacterial library clones was amplified using the Templiphi reaction, based on φ29 rolling-circle replication of DNA (AP Biotech). Briefly, approximately 0.5 μl of bacterial glycerol stocks were picked up using 96-pin plastic replicators (Genetix) and centrifuged into 96-well PCR plates. Five microliters of denaturing buffer was added, and samples heated to 95°C for 3 min. After cooling, 5 μl Templiphi enzyme was added and samples incubated overnight in a 30°C incubator. The Templiphi reaction provides two advantages for large-scale sequencing projects on capillary sequencers. First, the reaction proceeds to an endpoint where all nucleotide is incorporated, yielding uniform quantities of DNA from varying amounts of starting bacteria (or DNA). Second, the rolling-circle reaction results in large pieces of DNA that, in contrast to plasmid DNA, do not enter the capillary and interfere with the sequencing run.

For sequencing reactions, the DNA preparation was diluted fivefold with distilled water. Sequencing reactions were performed using the DYEnamic ET Dye terminator kit diluted twofold with DYEnamic ET dilution buffer (AP Biotech). Five microliters of DNA was added to 5 μl of sequencing reaction mix with primer and cycled 30 times under the following conditions: 95°C 20 sec, 60°C 1 min. Sequencing was performed on a MegaBACE 1000 (AP Biotech). Runs were either performed at injection: 3 kV 60 sec, run: 8 kV 120 min, or injection: 3 kV 60 sec injection, run: 3 kV 360 min.

The redundancy of the arrayed libraries was tested by performing BLASTN searches 12 against all sequenced ESTs from the two libraries. Hits against clones other than the query with an E-value lower than 1e-50 were considered for clustering.

The sequences were submitted to GenBank. After quality control, individual ESTs were used to search the non-redundant protein database (release of July 2004) using the program BLASTX from the standalone NCBI-BLAST package 12. For annotation of sequenced ESTs, the top hit of the BLAST output was used, whereby an E-value of 1e-20 was used for significant similarity and an E-value of 1e-05 was used as a cutoff value for weak similarity.

Quality control of sequenced ESTs was performed using the program Phred 11 using a cutoff of 20 for trimming low-quality regions, and vector trimming was performed using the program cross-match 11. (We note here that the arbitrary Phred score reflects the likelihood of a false base. A Phred score of 20 indicates that in 1 out of 100 trials (10²), the base would be false, 30 would reflect a wrongly sequenced base in 1 of 1,000 trials (10³), and so forth.) Sequence and contig files can be downloaded at 16. The resulting high-quality sequences were assembled into sequence contigs with the program TIGR-Assembler version 2 13. Alignment of contigs was performed with the program ClustalW with the settings Gap Opening 5 and Gap Extension 85 29 or Cap3 30, when ClustalW could not correctly assemble the sequences. Assembled contigs were used to perform BLAST searches (BLASTX, BLASTN from NCBI-BLAST 12) against the non-redundant protein sequence database (release of November 2003), human and fugu protein databases and the NCBI EST database, all downloaded from the NCBI. Domain searches were done with RPS-BLAST against the conserved domain database (CDD 18) from the NCBI. BLAST and domain-search output files were parsed for homologous sequences, whereby an E-value of 1e-05 was used as a cutoff for BLASTN and BLASTX searches against the sequence databases and the default cut-off of 0.01 was considered to yield significant homology to conserved domains from CDD. A gene identifier was assigned to those contigs that showed reliable homology to a sequence in the non-redundant database (E-value cutoff of 1e-20 for significant similarity and 1e-05 for weak similarity). Potential untranslated regions were identified using the program ESTScan 31.

Based on the GO annotation of the closest annotated homolog, contigs were assigned a molecular function, biological process and cellular component from the GO database 17. To this end, the GenBank annotation files from the GO database were downloaded and parsed for the gene identifier (gi) numbers of previously identified homologs. The cutoff for annotating an A. mexicanum contig was an E-value of 1e-20.

Two EST sequences of the p27^Kip1 gene were sequenced in the EST collection but neither were full-length sequences. To isolate the full-length sequence, 200,000 clones of our arrayed blastema and neural tube libraries were screened by PCR. Briefly, the bacterial library clones in each 384-well plate were pooled, mini-prepped and arrayed into 96-well plates (RZPD, Berlin), resulting in 576 DNA pools. These DNA pools were screened by PCR using the custom SP6 primer (GCACATTAGGCCTATTTAGGTGACA) as a forward primer and a gene-specific p27 reverse primer (TGATTTCCAATGGCTGGTTT). Fifty nanograms of DNA from each pool was used for PCR reactions and PCR cycling was performed at the following conditions: 94°C 2 min, 30 cycles of 94°C 15 sec, 65.5°C 30 sec, 72°C 90 sec, followed by 72°C 7 min). The largest positive band (1.1 kb) was gel purified and sequenced on an ABI377 machine using the SP6 primer.

Multiple sequence alignments were done with the program ClustalX 32 using standard parameters. Phylogenetic analysis of mitochondrial 12S rRNA was done using the programs dnadist, phylogenetic analysis of the cyclin B family and the CDK inhibitor family (CKI family) was done using protdist, both from the Phylip package 33. Trees were calculated with the program fitch from the same software package, using 100 iterations. For the CKI family, only the amino-terminal, CDK-inhibitory domain or the full-length sequences were used for construction of a phylogenetic tree. For the cyclin-B family, only the region overlapping in A. mexicanum contigs was used for tree construction. Trees were displayed using the program nj-plot 34 for the mitochondrial 12S rRNA tree and unrooted 34 for the CKI- and cyclin B families.

A relational database was created using the open source software MySQL as the database server to store and navigate through resulting sequence contigs and annotations. Scripts connecting the web-based front end to the database were written in the programming language Python.