A total of 12 cDNA libraries were generated from leaf-derived embryogenic callus (EC), non-embryogenic callus (NEC) and embryoid (EMB). In order to produce useful sequencing reads, all 12 libraries were size-fractionated and directionally cloned prior to single-pass sequencing. A small-scale quality assessment was performed for all the libraries prior to commencement of large-scale sequencing (Table 1). The titers of all 12 amplified libraries ranged from 1.22 × 10⁹ to 6.10 × 10⁹ pfu/ml and the percentage of recombinant clones observed were between 54.8% and 97.0%. Insert size of these libraries ranged from 250 bp to 2,300 bp, with an average of about 1 kb.

A total of 20,949 5'-end reads were generated from the 12 cDNA libraries, yielding a total of 17,599 high-quality sequences with an average edited length of 465 bases. In order to facilitate further analysis, the sequences generated from the 12 cDNA libraries were grouped accordingly into three different developmental stages, which are embryogenic callus (EC), non-embryogenic callus (NEC), and embryoid (EMB) (Table 2). The overall sequencing success rate was approximately 84%. The most number of ESTs were generated from the EMB libraries followed by NEC libraries and the EC libraries. The EST sequences generated in this study were deposited in GenBank under the accession nos. EY396120–EY413718.

Only 57.6% of the 17,599 ESTs were assigned putative identification using the blastx analysis. Matches with an E-value ≤ 10^-10 were assumed to have significant similarity to known sequences in the GenBank non-redundant protein database 2223. There was a large percentage of ESTs that shared no significant similarity to sequences in GenBank. This was probably due to the shorter average length of these sequences (389 bases) compared to the ESTs showing significant similarities with the GenBank entries (521 bases).

A total of 3.1% (311/10,138) of the ESTs with significant amino acid sequence similarities were found to be similar to previously identified and characterized oil palm genes. These genes, which include metallothionein-like protein 24, translationally controlled tumor protein and glyceraldehyde 3-phosphate dehydrogenase are listed in Table 3. Based on the comparison with known oil palm sequences, it is clear that genes encoding metallothionein-like protein were the most abundant, representing 37.6% (117/311) of the total ESTs with significant matches to oil palm genes. Metallothionein-like proteins are genes that have been isolated from higher plants including monocot (wheat, maize) and dicot (Arabidopsis) 252627. These genes play a role in heavy metal detoxification, especially in respect to cadmium, copper and zinc 28. In oil palm, it was reported that the class I, type 3 metallothionein-like genes, MT3-A and MT3-B, when expressed as a fusion protein of glutathione-S-transferase/MT3-A, possess a strong binding affinity for zinc 24.

In order to assess the rate of gene discovery in each library, StackPACK analysis 29 was carried out and the unigenes (total number of consensi and singletons) within each library were calculated. The proportion of sequences that were unique within each library was not similar and ranged from 57% to 78% (Table 2). The wide range observed was probably due to the fact that the EC library was not deeply sampled. With the exception of the EC library, the number of ESTs that could be clustered into consensus were quite similar, which is 61% and 52% in NEC and EMB libraries respectively. This suggests that the transcripts were distributed almost equally within these libraries.

To further identify unique sequences and remove redundancies between libraries, all ESTs were clustered. Cluster analysis of the 17,599 ESTs revealed 9,584 unigenes. The analysis formed 3,268 consensus sequences (representing 11,283 or 64% of the total ESTs), with an average consensus length of 666 bases. The remaining 6,316 sequences were singletons, with an average length of 449 bases. Of these 9,584 unigenes, 5,299 showed significant similarity to known sequences in GenBank non-redundant protein database at an E-value cut off of 10^-10. The remaining 4,285 (44.7%) did not show significant similarity to any known sequences in the database. The assembly results of the libraries are provided in Additional File 1.

Another major advantage of sequencing ESTs from multiple libraries is the ability to identify genes that are putatively transcribed specifically within a certain tissue or during a particular developmental phase. This study revealed that 10,602 sequences (60%) were unique to one of the three tissues sampled (Table 4). As expected, several transcripts were widely expressed in all three tissues sampled. A total of 268 different consensi containing ESTs derived from all three tissues were identified. These transcripts may represent genes, such as those involved in housekeeping, which are common in these tissues (Figure 1 and Table 5). The most abundant transcript detected was putatively identified as ribosomal protein S3 (336 ESTs; cn0037), followed by metallothionein-like protein (117 ESTs; cn3069) and a hypothetical protein (109 ESTs; cn0752). The high percentage of ribosomal proteins identified is expected, as ribosomal proteins are fundamental proteins for living systems and function as intermediary for protein translation.

The sequences were also analyzed to identify the base preference of Elaeis guineensis, which can be useful for predicting the coding regions of the oil palm genome. This was achieved by identifying full-length open reading frames (ORFs) in the unigene dataset. Homology search results of the 9,584 unigenes using blastx were used to identify unigenes with relatively high similarity (score > 200) to known genes and having an in-frame start and stop codon position similar to the GenBank sequence. Selecting oil palm sequences that had a start and stop codon at a position similar to the protein sequence in GenBank resulted in a more stringent and accurate dataset of full-length ORFs.

A total of 272 putative ORFs were identified, and subsequently translated into amino acid sequences. The amino acid sequences were resubmitted for homology search to confirm the full-length ORFs. Based on the resubmission results, 211 amino acid sequences were identified as putative full-length ORFs. These sequences had a start codon similar to the GenBank sequences, in-frame stop codon and 3 untranslated region (UTR), indicating that the amino acid sequence was translated in the right reading frame. The full-length ORFs were also deposited in GenBank under the accession nos. EU284816–EU285026. The codon usage table for full-length Elaeis guineensis ORFs was subsequently generated using CODONW. The oil palm (Elaeis guineensis) codon usage table containing 44,372 codons is shown in Table 6. The codon usage table shows that the GC content of the predicted coding region (48.4%) was higher than the predicted 3 UTR (37.8%) and that the GC frequency at the third position is 51.5%. One of the prominent features is that TGA is the preferred stop codon, as it appears in 42.2% of the sequences.

Data mining of 9,584 unigenes assembled from 17,599 ESTs identified 648 non-redundant (NR) SSRs in 586 unigenes. The unigenes represented about 5.0 Mb of genic sequences. A total of 56 sequences contained more than one SSR. The NR EST-derived SSRs were represented by mono-, di-, tri-, tetra- and pentanucleotide repeat motifs. This corresponds to an overall SSR density of approximately one SSR per 7.7 kb or an SSR-containing sequence in 6.1% of the NR EST sequences. About 3.4% of the SSRs identified were compound SSRs, which are defined as two neighboring repeats that are located less than 10 nucleotides apart in a single sequence. The frequencies of the SSR motifs identified from 9,584 unigenes are summarized in Table 7.

Based on the distribution of SSR motifs, AG/CT motifs represented the most abundant dinucleotide repeat motif. These motifs corresponded to about 67% of the dinucleotide repeat motifs, whereas AT (21%), AC/GT (11%) and CG (0.3%) occurred at relatively low abundance. Among the trinucleotide repeats, AAG/CTT (23%) was the most common motif, followed by AGG/CCT (13%), CCG/CGG (11%) and AAT/ATT (11%). The remaining trinucleotide repeat motifs were less abundant (<10%) with the AGT/ATC (4%) repeat motifs being the least abundant. The most abundant tetranucleotide repeat motifs are the ACAT/ATGT (36%) and AAAT/ATTT (29%) motifs. The distribution of the repeat motifs is shown in Figure 2.

Analysis was also carried out on the 211 putative full-length ORFs to mine for SSR sequences and determine the location (UTR or ORF) of the SSRs identified. A total of 18 SSRs were identified in 17 sequences. Two mononucleotide, seven dinucleotide and nine trinucleotide repeats were identified. The mono- and dinucleotide repeats were identified only in the UTR while the trinucleotide repeats were identified in both the ORF and UTR. A higher percentage (66.7%) of the trinucleotide repeats were identified in the ORF region. The distribution of these SSRs in the putative full-length ORFs is shown in Table 8.

Gene ontology (GO) annotations of the 17,599 oil palm ESTs were performed using Blast2GO 30. The software performed blastx similarity search against the GenBank non-redundant protein database, retrieved GO terms for the top 20 BLAST results and annotated the sequences based on defined criteria. A total of 8,436 ESTs were successfully annotated with GO terms. In order to help improve the number of sequences annotated with GO terms, additional information was obtained for the ESTs using InterProScan. An additional 755 sequences were then annotated. Overall, a total of 9,191 ESTs were annotated with 33,742 GO terms distributed among the three main GO categories, which are Biological Process (9,253), Molecular Function (13,140) and Cellular Component (11,178). The number of ESTs that were represented with GO terms is probably an underestimate, as 47.8% of the ESTs were not annotated. Generally, the sequences with no BLAST hit (3,100) could not be annotated. However 2,205 of the ESTs with BLAST hit also could not be annotated, while an additional 3,103 sequences did not fulfil the selected criteria for annotation. The majority of the 2,205 genes with BLAST hit could not be annotated with GO terms because the sequences were similar to hypothetical or unknown proteins.

The representations of genes across the various GO terms were examined with WEGO 31. The distribution and percentage of ESTs in each of the GO terms were calculated. A percentage of 100 was defined as the total number of ESTs that were assigned GO terms in a tissue type, i.e. NEC, EC or EMB. However, it must be noted that the percentages of the subcategories do not add up to 100% as many of the ESTs are involved in different classes of function and annotated with multiple GO terms. Generally the distribution of the genes based on the GO terms was quite similar in all three tissues.

Approximately 59.8% of the annotations for the ESTs were grouped into the "cellular process" category in the GO main category Biological Process (Table 9). The category includes processes that are carried out at the cellular level such as cell cycle, cell communication and cell development. The second highest category in Biological Process is the "metabolic process" category, which represents 57.1% of the annotations. The category has subcategories that are involved in photosynthesis, metabolism and regulation of metabolic processes. However, both categories have overlapping subcategories, especially those involved in the cellular metabolic process. It is interesting to note that 5.1% of the annotations were grouped into the "response to stimulus" subcategory. In this category, 307 (3.3%) and 248 (2.7%) ESTs were involved in response function towards stress and chemical stimulants respectively. This is not surprising since the tissue culture environment has been suggested to induce the stress-response mechanism 32.

Based on the distribution of the annotations for this sub-category in the individual tissues, the results also showed that the number of ESTs was lowest in NEC (3.6%) and gradually increased in EC (4.5%) and EMB (6.2%). It is possible that the expression of stress-response genes is necessary to cope and acclimatise to the stress conditions associated with tissue culture, such as mechanical wounding, osmotic shock and hormonal imbalances. The ability to endure stress can inevitably help the proliferation of culture lines into embryoids.

In the Molecular Function main category, 48.9% of the EST annotations were grouped in the binding category (Table 9). They were represented by a number of GO terms involved in binding, with 1,777 and 1,430 ESTs annotated as "nucleic acid binding" and "ion binding" respectively. Another category that has high levels of representation is in "catalytic activity". Specific catalytic activities that have at least 10% of the ESTs associated with the category are hydrolase, transferase and oxidoreductase subcategories, which are represented by 1,173, 1,118 and 982 ESTs respectively. These genes are involved in processes such as in signal transduction, metabolism and post-translational modification.

Digital northern or in silico subtraction was performed to identify candidate genes or markers that are specific in either NEC or embryogenic cultures (EC/EMB). The analysis is derived based on the relationship of the frequencies of the ESTs in a particular tissue with the expression of the genes in that tissue 163334. The analysis allowed the identification of transcripts that could represent genes that trigger the embryogenesis process.

For this purpose, the 3,268 consensi were compiled into a matrix file containing the frequency of ESTs in each consensus sequence corresponding to each of the three tissues (EC, NEC and EMB). Using the Stekel and Falciani statistical test (R test) in IDEG6 35 with a significance threshold of 0.001, 52 (1.6%) unigenes were identified. The threshold (0.001) allows the selection of differentially expressed genes with almost no false positives 35. The Stekel and Falciani statistical test was developed to compare gene expression from multiple cDNA libraries 36. A total of 30 and 22 unigenes were found to be differentially expressed in the NEC and EC/EMB respectively (Table 10). Cluster analysis to determine the relatedness between the tissues based on the abundance of transcript in each consensus, revealed that the NEC samples were distinctly different from EC and EMB cultures (Figure 3).

Based on the genes identified, 15 unigenes were represented in NEC only while the ESTs for seven unigenes were present only in EMB. An additional two groups of 15 unigenes that showed a higher expression in NEC and EC/EMB respectively, were also identified. The results showed that lipid transfer protein (cn2473), catalase 2 (cn3157), PVR3-like protein (cn3238), defensin EGAD1 (cn1673) and dehydrin-like protein (cn2100) were among the genes found to be highly expressed in the EC/EMB. The identified transcripts might be involved in somatic embryogenesis initiation, differentiation during somatic embryo development and also in somatic embryo growth, maturation and desiccation. In the NEC libraries, cinnamyl alcohol dehydrogenase, which is involved in lipid biosynthesis, seemed to be abundantly expressed. Stress related genes, such as glutathione S-transferase, pathogenesis-related protein 4, glucanases and metallothionein-like protein were also up-regulated in NEC tissues. A large percentage of genes in the NEC libraries (41%) did not have significant similarity to sequences in GenBank.

Eight genes were selected for northern blot analysis using RNA extracted from tissue culture and vegetative tissue samples. Based on the digital northern results, five genes that had higher levels of abundance in EC/EMB compared to NEC were selected. One gene, glutathione S-transferase (cn0544) that had a higher level of abundance in NEC compared to EC/EMB in the digital northern results was also selected. The northern blot analysis included an additional two genes that were expressed at low levels in NEC tissues but not in EC/EMB. The two genes were a gibberellin-stimulated transcript (cn1254 having 519 bases) and a transcript with no hit to genes in GenBank (cn2945 having 503 bases). Northern blot analysis was performed and the results are shown in Figure 4. The results showed trends similar to what was observed in the digital northern analysis. The expression for the five EC/EMB enhanced transcripts was lower or not detected in most of the NEC samples when compared to EC and EMB (Figure 4a). However, the lipid transfer protein and dehydrin-like protein were also expressed in the spear leaf while defensin EGAD1 and PVR3-like protein were also found to be expressed in the flower tissue. Catalase 2 expressions were detected at high levels in EC/EMB and spear leaf with a lower expression in NEC and mesocarp tissues. These genes could be important markers to differentiate EC/EMB from NEC.

For the three probes selected for higher abundance in NEC, two transcripts (glutathione S-transferase [cn0544] and a transcript with no hit [cn2945]) were generally confirmed to be more highly expressed in NEC compared to EC/EMB in northern blot analysis (Figure 4b). The expression profile was to some extent genotype dependent, as the difference in expression level was more profound in Set 2 compared to Set 1. However, the expression profile of the last EST (gibberellin-stimulated transcript; GAST) did not correlate with the observation made through in silico analysis.

Two lines of embryogenic callus (EC), three lines of non-embryogenic callus (NEC) and two lines of embryoids (EMB) were used for cDNA library construction. These leaf-derived tissue culture materials were obtained from two laboratories. Data validation was carried out using two sets of tissue culture samples, each set represented by NEC, EC and EMB. In addition to the tissue culture samples, spear leaf, inflorescences frond-12 (female, 5 cm), 12 weeks after antheses (12 WAA) kernel and 15 WAA mesocarp tissues obtained from Malaysian Palm Oil Board (MPOB), Malaysia were also used for the validation experiments. All plant materials were of the Dura (D) × Pisifera (P) fruit type and stored at -80°C prior to RNA extraction.

A total of 12 cDNA libraries were constructed from different developmental stages of the tissue culture process: EC, NEC and EMB. Total RNA was isolated using the aqueous phenol extraction method as described by Rochester et al. 69. Poly (A)⁺ RNA was isolated using oligo-dT cellulose chromatography according to Singh and Cheah 70. Directional cDNA libraries were constructed using the ZAP-cDNA^® Synthesis Kit and the ZAP-cDNA^® Gigapack^® III Gold Cloning Kit (Stratagene USA) according to the manufacturer's instruction.

For DNA sequencing, phagemids and plasmids were used as templates. Individual recombinant phages were selected randomly for sequencing. Inserts from phages were amplified using T3 and T7 primers. Enzymatic purification was performed on the PCR products by using Shrimp Alkaline Phosphatase (GE Biosciences USA) and Exonuclease 1 (GE Biosciences USA) to clean up excess dNTPs and primers. Plasmid clones were obtained either by single or mass in vivo excision using ExAssist^® Helper phage as recommended by the manufacturer (Stratagene USA). Plasmid DNA for sequencing was prepared using Wizard^® Plus Minipreps DNA Purification System kit (Promega USA). cDNA inserts were sequenced from the 5' end with SK primer using the ABI PRISM™ Ready Reaction BigDye™ Terminator Cycle Sequencing Kit (Applied Biosystems USA). Sequencing was performed on an ABI 377 automated DNA sequencer (Applied Biosystem USA).

All the sequences were quality checked before clustering and deposited into GenBank. Raw ABI-formatted chromatogram reads were base-called using PHRED (version 0.020425.c) 7172 with a threshold value of 20. Crossmatch (version 0.990329) and customized Perl scripts were used to trim vectors, adaptors, polyA-ends and low quality nucleotides. ESTs were also filtered for prokaryotic microbe such as Escherichia coli. Only high quality ESTs with a minimum of 100 bases and fewer than 4% Ns were retained. Manual processing was also performed to confirm the results.

Multiple sequence alignment, clustering, assembling and the generation of consensus were accomplished by StackPACK 29. StackPACK contains d2_cluster 73, where sequences were grouped together if there were at least 96% sequence similarity in any window of 150 bases. The loose clusters were then aligned using PHRAP 74 and subsequently, CRAW 75. The resulting set of consensi and singletons were considered as a set of putative unique genes (unigenes). Homology searches of all cleaned ESTs and assembled consensus sequences were compared to GenBank by using the blastx algorithm. Sequences with blastx E-value > 10^-10 were designated as having "no significant similarity".

Digital analysis of gene expression was performed with IDEG6 35 by using transcript abundance in each consensus gathered from the EST frequency for that consensus in all three tissue types (NEC, EC and EMB). The resultant gene list was clustered in TIGR Multiple Experiment Viewer software (version 3.0).

The unigenes assembled from 17,599 ESTs were mined for SSR markers. The polyA and polyT sequences in a 50 bp window at the terminal regions were removed. The unigenes were mined for microsatellite motifs. MISA (MicroSAtellite identification tool) search module 76 was used to identify and localize the microsatellite motifs. Sequences were deemed to contain microsatellite motifs if it contains ten, seven and five repeat units of mononucleotides, dinucleotides and higher-order repeats respectively. The Perl scripts provided by the MISA website also generate summary files with information such as the position and type of microsatellites identified, distribution of different repeat type classes and frequencies of identified SSR motifs.

The 17,599 ESTs were annotated with gene ontology terms using Blast2GO 30. The sequences were annotated using blastx search to the GenBank non-redundant database. The top 20 hits were evaluated and blastx results that had an E-value cutoff of 1e^-6 and a minimum similarity of 55% were annotated with GO terms. A weightage based on the default Evidence Code Weights was also used to determine the GO terms annotated. Additional information and GO terms were obtained by comparing the sequences to the InterPro database using the InterProScan tool to identify protein signatures 77. The GO terms were compared and visualized using WEGO 31.

Total RNAs (20 μg) were separated on a 1.0% (w/v) agarose/formaldehyde gel and transferred to positively charged nylon membranes (Hybond™-N⁺, GE Healthcare Biosciences, UK), by downward capillary transfer with NorthernMax™ kit transfer buffer (Ambion Inc., USA). Specific probes labelled with α³²P-dCTP were generated by purified PCR amplification of the relevant differentially expressed clones. After a 3-hour prehybridization step performed in ULTRAhyb™ (Ambion Inc., USA) and hybridization in the same solution for overnight at 65°C, the membrane was washed twice in 2× saline sodium phosphate EDTA (SSPE) (0.36 M sodium chloride, 20 mM sodium hydrogen phosphate, 2 mM EDTA) pH 7.4 at room temperature for 5 min and then was washed in 0.1× SSPE at 65°C for 15 min. The blot was exposed at -80°C with an intensifying screen overnight or up to 1 week.