A major challenge confronting EST-based gene discovery programs is differential mRNA abundance. Usually, a few hundred highly and moderately expressed genes produce more than half of the cellular mRNA molecules, whereas several thousand genes account for the remaining mRNA mass 15. Sequencing randomly selected clones from standard cDNA libraries therefore inefficiently identifies rare transcripts, owing to the repeated occurrence of moderately and highly abundant cDNA species. We employed virtual subtraction and direct subtraction to enhance the number of unique genes identified. The virtual subtraction method 16 classifies cDNA clones according to the abundance of the mRNAs they represent (Figure 1A). The direct subtraction method removes previously identified cDNA clones from the gene discovery pipeline. We initiated this EST-based gene discovery program by sequencing 2,000 randomly selected clones. Next, we sequenced 2,304 of the low intensity clones identified by virtual subtraction. Finally, we sequenced 10,738 clones that gave very low hybridization signals when subjected to both virtual and direct subtraction.

Figure 1B presents the gene discovery rates obtained while sequencing the randomly selected clones, the clones selected following virtual subtraction, and the clones selected following virtual and direct subtraction. We obtained 5,202 singleton and contig sequences after processing 12,820 high quality EST sequences (Table 1). This means that we identified roughly one gene for every 2.5 EST sequences. This result compares favorably with the results obtained by some other large-scale EST projects of lower eukaryotes. For instance, a Neurospora crassa project produced 20,019 ESTs and identified 1,431 genes 17 for a gene discovery rate of one gene for every 14 EST sequences, and a Dictyostelium discoideum gene discovery project that generated 26,954 ESTs identified 5,381 unigenes for a gene discovery rate of one gene for every 5 EST sequences 18.

We submitted the 12,820 high quality ESTs to GenBank [GenBank: DR697868 – GenBank: DR710686]. Table 1 shows that the individual sequencing reads contained 400–800 nucleotides of high-quality sequence. The EST assembly produced by phrap 19 yielded 5,202 unisequences that included 2,183 singletonsand 3,019 contigs. Following assembly, we used BLASTN to cluster the closely related singletons and contigs. Clustering assembled 168 of the 5,202 phrap unisequences into 74 clusters, each containing 2–4 sequences. Manually confirmed ClustalW alignments showed that 56 clusters were generated by assembling alternatively spliced derivatives of 117 phrap unisequences. Taking into account the 74 clusters assembled from multiple unisequences, the 12,820 ESTs generated 5,108 clusters. The clusters predicted to have arisen through alternative splicing are available in Additional file 1. Prior to submission of our EST sequences, we found 784 A. niger genomic DNA and cDNA-derived sequence entries in the GenBank database (June 22, 2005 release). These entries formed 379 unique genes. BLASTN analysis showed that 252 of the phrap unisequences aligned with at least one of the A. niger GenBank entries (alignment length >50, identity >95%). Therefore, this study identified about 4,856 new A. niger genes. The results from our EST sequencing, contig assembly and clustering analysis are summarized in Table 1.

We attempted to determine the putative function of the set of 5,202 phrap unisequences by searching for homologs in the GenBank non-redundant protein database using BLASTX (Table 2). Setting the BLASTX cutoff value at E = 1e ^-5, about 83% of these sequences display similarity to at least one GenBank entry, 44.5% to genes of known function and 38% to genes of unknown function. The remaining sequences, 17 %, code for proteins that lack similarity with any GenBank entry.

We also compared the proteins encoded by these sequences with the proteins predicted from the completely sequenced genomes of three Ascomycetes, Saccharomyces cerevisiae 20, Aspergillus nidulans and Neurospora crassa 21, and one Basidiomycete, the white rot fungus Phanerochaete chrysosporium 22. As expected, the highest degree of similarity (BLASTX alignments with E values ≤ e^-30) is with A. nidulans, where 64% of these A. niger unisequences encode proteins that have A. nidulans homologs (Table 2). Nonetheless, almost 20% of the A. niger genes did not have a homolog (E > e^-5) in A. nidulans.

Although the Sordariomycetes, which include N. crassa, and the Eurotiomycetes, which include the Aspergilli, diverged about 670 million years (Myr) ago 23, over 43% of the predicted A. niger proteins are highly similar (E ≤ e^-30) to N. crassa predicted proteins. For the more distantly related Saccharomycotinna S. cerevisiae and Hymenomycete P. chrysosporium, which diverged from the Eurotiomycetes lineage about 1,090 and 1,210 Myr ago, respectively 23, only 21% and 25% of the A. niger predicted proteins had highly similar homologs (E ≤ e^-30).

The predicted A. niger protein products were assigned Gene Ontology (GO) classifiers based on BLASTX alignments (expected values of E ≤ e^-5) generated by searching the GO annotated Swiss-Prot and TrEMBL databases. GO categories were assigned to 2,549 of the 5,202 predicted protein products. Figure 2 summarizes the resulting GO assignments, which are available in Additional file 2. More detailed annotations, including the BLAST alignments, Expect Values and BLAST Scores generated by searching the GenBank nr database are available online 14 and can be used to assess the reliability of functional predictions on a gene by gene basis.

We compared the distribution of GO classifiers obtained for the A. niger unisequences and the predicted genes of six fungal species (Table 3). The gene distribution in the main ontology categories was very similar across all seven species. However, the fission and budding yeasts have a higher proportion of genes in the "cell growth and/or maintenance" categories, 45.2% and 48.5%, than did the filamentous fungi, where the proportion ranged from 29.4% to 36.2%. Since we found no correlation between evolutionary distance and these differences, it seems likely that they reflect differences in gene number. The genomes of the five filamentous fungi encode 9,000–12,000 genes 2425 whereas the fission and baker's yeast genomes have about 4,824 26 and 6,335 27 protein-coding genes, respectively. The much smaller number of genes present in these two yeast species suggests that they may have close to the minimum number of genes needed by a free-living eukaryotic cell 28.

Aspergillus niger is the source of a number of secreted proteins produced for various industrial applications. Gene Ontology mapping categorized only 15 of the predicted proteins as "extracellular" (Additional file 2 ). However, we were able to assign a GO component classifier to only 1,195 (23.4%) of the encoded proteins. To identify potential secreted proteins we used SignalP 3 29 to search for proteins with a secretion signal. SignalP predicted that about 400 of the predicted proteins had a signal peptide (Additional file 3 ). Blast searches showed that 293 of these proteins were similar (E ≤ e^-5)to at least one GenBank entry. The 27% of predicted proteins with a signal peptide that do not have a GenBank homolog is significantly higher that the 17.5% of predicted orphan proteins. The reason for these differences remains unknown although they may suggest that the fungal secretome is subject to rapid evolution.

Recent strategies for improving the efficiency of heterologous protein expression in A. niger have focused on molecular genetic manipulation of the secretory pathway. In some cases, these approaches have significantly increased the expression of selected heterologous proteins 3031. Using GO mappings and BLAST analysis we identified 118 genes that apparently participate in various steps of the protein secretion pathway (Additional file 4 ). Fifteen genes encode secretion-related ER chaperones, foldases and proteases; 77 encode putative proteins involved in protein transport, protein targeting and vesicle-mediated transport; and 26 code for proteins that are involved in secretion-related post-translational modifications. The A. niger genes identified in this study included all the previously identified secretion-related ER chaperones, foldases and quality control proteins: bipA (Asp84), pdiA (Asp734, Asp1902), prpA (Asp4188), tigA (Asp1020), cybB (Asp662), clxA (calnexin) (Asp1882), and kexB (kexin) (Asp177) 30313233.

Previous studies with A. niger identified five secretion-related GTPases belonging to the Ras super-family, SrgA, SrgB, SrgC, SrgD, and SrgE, and one member of the ARF/SAR subfamily, SarA 3134. Our A. niger sequences included the earlier identified SarA (Asp4377), SrgA (Asp5114, Asp4222), SrgB (Asp3374, Asp70) and SrgE (Asp1610) genes. We also identified contigs Asp1708, which encodes a protein with 47% similarity to the S. cerevisiae GTP-binding protein YPT52 35, and Asp1824 and Asp1217 that code for proteins with 87% and 94% identity with Aspergillus nidulans members of the Rab subfamily of small GTPases 36.

Post-translational modifications such as glycosylations are often important for the production of biologically active secreted proteins. For instance, introducing an N-glycosylation site into bovine chymosin increased the amount of secreted chymosin expressed by A. niger 10-fold 37. Identification of the various genes involved in O- and N- linked glycosylations 38 would facilitate efforts to engineer the A. niger glycosylation pathway. We identified several putative members of the N- and O-linked protein glycosylation pathways, including; six PTM related O-mannosyltransferases, contigs Asp370, Asp4472, Asp170, Asp1044, Asp1344, and Asp3205 3940 and genes that are involved in N-linked protein glycosylation such as two contigs, Asp1340, and Asp458, that encode homologs of oligosaccharyl transferases 41.

Aspergillus niger strain N402, FGSC #4732 was grown at 30°C in Minimal Medium 42 containing 1% w/v of various carbon sources with shaking at 150 RPM. The carbon sources used were: glucose, bran, maltose, xylan, xylose, sorbitol, and lactose. Mycelial samples harvested by filtration and pressed between layers of filter paper to remove excess liquid, were stored at -80°C.

Total RNA was extracted from each mycelial sample. For this, 1.5 g of each frozen mycelial sample was ground to a fine powder in liquid nitrogen. Total RNA was extracted from the powdered mycelial masses using TRIzol^® reagent following the manufacturer's recommendations (Invitrogen, Burlington, ON). Total RNA (200 μg) from each culture condition was pooled and the poly(A)⁺ RNA was purified using oligo-dT cellulose column chromatography (Amersham Biosciences Corp, Piscataway, NJ). Quality and quantification of the RNA were analyzed by running the RNA samples on an Agilent 2100 bioanalyzer (Agilent Technologies, Palo Alto, CA).

The cDNA library was constructed using a Zap-cDNA^® Synthesis Kit according to the manufacturer's instructions (Stratagene, La Jolla, CA). Double-stranded cDNA was directionally cloned into the pBluescript^® KS ⁺ vector (Stratagene, La Jolla, CA) between its EcoRI (5'-end) and XhoI (3'-end) sites and transformed into E. coli strain DH5α.

The cDNA library was plated onto LB-ampicillin agar containing X-GAL and IPTG. White colonies were picked and inoculated into 384-well plates containing LB-ampicillin medium using a VersArray robotic colony picker and arrayer system (Bio-Rad, Laboratories, Canada), grown overnight and stored at -70°C after the addition of glycerol (10% v/v). To prepare plasmid DNA from each sample, bacterial inoculates were transferred from the 384 well storage plates to 96-well growth blocks containing 1 ml of 2YT-ampicillin medium per well (Corning, Acton, MA) and grown overnight. Recombinant plasmids were extracted using alkaline lysis 43 and subjected to single-pass sequencing from the T7 universal primer site (5'-end) using an ABI 3730 XL automated sequencing machine (Applied Biosystems, Foster City, CA) at the Génome Québec Innovation Centre (Montreal, PQ).

Two methods were used to normalize the library. For virtual normalization 16, bacterial colonies harboring independent cDNA clones were arrayed from the 384-well plates onto nitrocellulose membranes, 9,216 colonies per 492-cm² membrane. The membranes were probed using radiolabeled cDNA. The probe was prepared as follows; double-stranded cDNA was produced from the same mRNA population that was used for library construction using the SMART cDNA construction kit (BD Biosciences, Mississauga, ON) according to the manufacturer's instructions. The double-stranded cDNA was labeled with [³²P]dCTP by random priming, using the Rediprime™ II Random Prime Labeling System (Amersham Biosciences Corp, Piscataway, NJ). The labeled cDNA was used to probe six membranes, arrayed with 55,296 clones, and the clones were ranked according to the relative intensity of their hybridization signals (Figure 1A). Based on these intensity ratios the colonies were divided into three groups, high (relative intensity 50%-100%), moderate (relative intensity 10%-50%), and weak (relative intensity less than 10%).

For direct subtraction, plasmid DNAs representing each of the non-redundant genes that had already been identified was pooled. The pooled plasmid DNAs were linearized with the restriction endonuclease XhoI and radiolabeled "run-off" transcripts were generated using the Riboprobe in vitro Transcription System (Promega, Madison, WI). The probe RNA was then used to hybridize to the same membranes that had been subjected to virtual subtraction.

After hybridization, the membranes were exposed to X-ray film, and the intensity of the signal for each colony was quantified using GeneTools image software (Synoptics Limited). The intensity data for each clone was stored in our in-house database. The clones chosen for sequencing were based on the relative intensity of their hybridization signals, determined as a ratio of signal intensity of the individual clone to the maximum signal intensity present on the array.

The chromatograms obtained following single pass sequencing of the cDNA clones were processed using three software tools, phred to assign sequence quality values 4445, lucy to remove vector sequences and regions of low quality sequence 46, and phrap to assemble overlapping sequences into contigs 19. Sequence similarity searches against the NCBI non-redundant database were conducted using BLASTX 47 with default BLAST parameters. The top 5 scoring BLASTX hits with E values less than e^-5 were used to annotate each EST and EST-derived assembly using our annotation program TargetIdentifier 48. Sequences that did not return alignments with E values less than e^-5 were then used to perform BLASTN searches against the NCBI non-redundant nucleotide database. The top 5 BLASTN hits for each query, where the E value was required to be < e^-5, were then used for annotation. The resulting output files are uploaded to a local MySQL database.

Redundancy was also analyzed by means of clustering based on the BLASTN alignments. Sequences that exhibited more than 93% identity over lengths of at least 100 bases were assigned to the same cluster. Cluster assignments were confirmed by additional analysis using ClustalW 49.

For comparing E values obtained by searching databases of different sizes, we normalized the E-values using the following formula:

E_n = E specific *S nr/S specific,

E_n: the normalized E value, it is the subject/query E value that would have been obtained had the alignment been generated by searching a database having the same number of amino acids as the NCBI-nr database; E specific: E-value retuned by BLASTX when searching a user specified database other than the NCBI-nr database; S specific: number of amino acids in the user defined database; S nr: number of amino acids in the NCBI-nr database (total 617,284,665).

TargetIdentifier was used to estimate the proportion of the clones that contained complete coding sequences. The criteria used for establishing that a cDNA included the complete ORF can be found on our web site 50.

Annotation and functional binning were accomplished using tools provided by the Gene Ontology Consortium 51. Annotations were based on the Gene Ontology (GO) terms and hierarchical structure 52. Reference sequences were selected from the BLASTX results with E values less than e-5 obtained by searching the Swiss-Prot database of manually annotated proteins and the TrEMBL database of proteins with automated annotations. The GO categories associated with the BLASTX subject giving the highest score from the Swiss-Prot and TrEMBL databases were used to annotate our A. niger singletons and contigs. The GO term annotations were merged and loaded into the AmiGO browser and database 53. The resulting GO-derived annotations can be viewed with the AmiGO browser at our website 54.

The coding region of each singleton and contig was predicted and translated into protein sequences using our OrfPredictor program 55. The N-terminal 50 amino acids of each predicted polypeptide were searched for a signal peptide using SignalP version 3 29.