Background

Recent advance in molecular biology have introduced new high-throughput tools for the analysis of differential gene expression in complex diseases, such as Alzheimer (AD), providing simultaneous overviews of the genes or proteins associated with multiple cellular pathways. The most commonly used technology for the assessment of gene expression changes in postmortem brain is the DNA microarray [1–5] This approach has not only confirmed the involvement of genes implicated in AD by conventional methods, but also revealed changes in additional genes, not previously associated with AD [6, 7]. However, as this method requires a priori knowledge of gene sequences, it cannot be applied as a discovery tool for novel transcripts. Furthermore, the expression levels of low abundance genes cannot be readily assessed by DNA microarray hybridization, since reliable results are usually obtained only for genes that are expressed in high or moderate levels. This is a significant limitation since many transcripts expressed preferentially in brain (e.g., neurotransmitter receptors and their regulatory factors) are present at very low levels [8, 9].

Differential display and conventional subtractive hybridization approaches are capable of detecting expression changes in both known and novel genes. Differential display uses arbitrarily primed PCR to fingerprint differences (from first strand cDNA) in gene expression between two samples, with the results being determined by the intensities of bands on a polyacrylamide gel [10]. The major disadvantages of this method include its lack of sensitivity for the detection of rare RNA species, the high number of false positives generated during PCR and cloning of the differentially expressed products from low resolution polyacrylamide gels, where an apparent single band may contain multiple cDNA species. Consequently, differential display is labor intensive and unreliable for this application.

Subtractive hybridization, on the other hand, permits the isolation of target sequences from one single-stranded DNA population, referred to as "tester", from another DNA population, referred to as "driver" by using an excess of sequences. The two populations are mixed and put through iterative rounds of subtraction of cross-hybridized products. Earlier subtractive methods required physical removal of hybridized driver and tester sequences, which contributed to the loss of low abundance tester sequences. Suppression subtractive hybridization (SSH) is a newer method [11] which couples hybridization-based de-selection of common cDNAs to PCR amplification which enriches differentially expressed transcripts from two mRNA sources. In contrast to differential display, the primers for PCR amplification are clearly defined, thus avoiding problems associated with random primers. The main disadvantage of this procedure is its higher detection threshold. According to the kit manufacturer's recommendation (Clontech Palo Alto, CA), the difference in mRNA levels needs to be at least 5 fold to allow reliable detection.

Here, we have developed a novel approach to the identification of differentially expressed rare sequences through a combination of subtractive hybridization and RNA amplification, termed a S ubtractive T ranscription-based A mplification of mR NA (STAR). In our method, the expressed RNAs from two source are used for the preparation of specialized cDNA libraries, from which single stranded (+) sense tester RNA and single stranded (-) sense driver DNA are generated. Subtraction is accomplished by the hybridization of single-stranded driver DNA to the complementary single-stranded tester RNA, followed by RNase H digestion. This process not only eliminates the necessity for physical removal of hybridized common sequence, but also eliminates the self-annealing step of the tester nucleic acids that is required by the SSH method for the amplification of target sequences. The self-annealing step usually imposes kinetic limitations and requires lengthy hybridization in order to recover rare cDNA sequences. Furthermore, in the STAR method, the tester RNA is designed with a defined terminal sequence which allows the unhybridized tester RNA to be amplified in a linear RNA amplification process, rather than exponential PCR amplification, thereby minimizing biased sequence amplification [12, 13] Since the tester RNA remains unchanged in the process, the products of one round of STAR are used directly in subsequent rounds to provide further enrichment of the unique sequences.

In this study, we applied the STAR method to an investigation of changes in gene expression in port-mortem AD brain. We show that, indeed, the STAR process significantly increases the levels of unique and rare sequences relative to abundant housekeeping genes and identified changes in the expression of genes not previously linked to the AD. We also performed extensive literature mining to provide a basis for considering their likely involvement in neurodegenerative processes.

Results

Characterization of the subtracted AD cDNA libraries

Two subtracted cDNA libraries, using either the 3' cDNA of the control sample (C-AD) or the 3' cDNA of the AD sample (AD-C) as tester, were prepared using the STAR method (Fig 1). It was expected that they would contain genes that were down regulated (C-AD) and up regulated (AD-C) in AD brain, respectively. To isolate these genes, three rounds of STAR subtraction were performed, and the remaining 3' biased cDNA tester fragments were linearly amplified, ligated into a pUC-modified vector and transformed into E. Coli DH10B cells. Approximately 5000 colonies were obtained for each STAR library. Subsequently, PCR and DNA sequencing were performed on amplified cDNA inserts from 600 individual colonies for each subtracted library. Data analysis indicated that this number of colonies was sufficient to represent the entire subtracted library since many genes, including some of low abundance, appeared to be represented by more than one colony. The results summarizing the genes found in the C-AD and AD-C libraries are listed in Tables 1 and 2, respectively.

Figure 1
figure 1

Schematic diagram of the STAR procedure. The solid red line represents tester specific RNA (+). The solid green line indicates driver specific DNA (-). The solid black line shows newly synthesized tester specific first strand cDNA (-). Short-dashed lines show the common tester RNA (+). The long and short mixed dashed line denotes the common driver DNA (-).

Full size image
Table 1 Summary of sequenced clones from the C-AD subtracted cDNA library
Full size table
Table 2 Summary of sequenced clones from the AD-C subtracted cDNA library
Full size table

Of the 600 colonies analyzed from each library, over 500 contained inserts and produced readable sequences after PCR. Sequence annotation grouped the genes into three categories: (i) known genes, (ii) ESTs and (iii) novel genes, whose sequences either only match a fragment of genomic DNA or the database search did not hit any existing known sequence. The known genes were represented by multiple colonies (Table 1, 183/266, i.e., out of 266 colonies sequenced from the C-AD library 183 were unique sequences of known genes; Table 2, 187/277 in the AD-C library) far more often than the novel genes (113/131 and 109/119), indicating that known genes are relatively more abundant than the novel genes, explaining why they are more easily identified by traditional approaches. Sequence annotation also revealed that genes involved in maintenance of neuronal cell structure and function, the ubiquitin pathway and energy metabolism were predominant in the C-AD library, indicating that they were down regulated in AD brains. Whereas in the AD-C library, genes involved in inflammation, protein translation, cytoskeleton/cell adhesion and apoptosis/neurodegeneration were up-regulated in AD. Over 55% of the sequenced fragments in each subtracted library represented unknown genes, of which one half were not previously described as cDNA sequences. Representative data from these libraries is shown below.

Validation of changes from the subtracted libraries

More than 40 genes from the C-AD library were randomly selected for analysis by qRT-PCR using the pools of control and AD mRNAs as input. The results confirmed down-regulation of over 80% of them, consistent with the results of the subtracted library (Table 3). The data also confirmed that a majority of these genes (over 90%) were less abundant than β-actin. Furthermore, all novel transcripts were expressed at extremely low levels, some of them at only 0.0001–2% of the abundance of β-actin (Table 3). Taken together these results indicated that the STAR method permitted the isolation of differentially expressed genes, including very rare ones. This approach also revealed the existence of large number of previously unidentified genes expressed in the human brain which appear to have a disease association.

Table 3 Summary of qRT-PCR analysis of the selected genes from the C-AD library
Full size table

Nearly 40 randomly selected genes from the AD-C library were also analyzed by qRT-PCR in order to further validate the library data. In this case, less than 50% of the selected transcripts were found to be up regulated, with a lower percentage of novel genes (Table 4). This might be indicative of a higher content of false positive in this library, pointing out to the necessity of independent validation prior to gene selection for down stream functional studies. One possibility for the relatively low confirmation rate of the data from the AD-C library is that the increases in the expression of unconfirmed transcripts were small and, therefore, might also be difficult to validate by qRT-PCR, although this wasn't evident in the qRT-PCR validation of down regulated genes from the C-AD library. Despite this, the technique clearly identifies novel sequences that are up regulated in AD brain.

Table 4 Summary of qRT-PCR analysis of the selected clones from the AD-C library
Full size table

Supporting evidence for gene association with AD

An additional approach to increase confidence in the identified genes as likely disease candidates was to mine knowledge pertaining to gene function in AD in the existing literature using a custom built literature mining tool, LitMiner. Using this approach, we generated short lists of genes whose association with AD could be further rationalized (Tables 5 and 6).

Table 5 A short list of down regulated genes and supporting evidence for their involvement in AD
Full size table
Table 6 A short list of up regulated genes and supporting evidence for their involvement in AD
Full size table

The data mining analysis identified 12 genes (Table 5 items 1–12), whose down-regulation in AD has been previously reported, in agreement with our results as these genes were present in the C-AD subtractive library. Furthermore, down-regulation of synuclein alpha was also confirmed by qRT-PCR analysis (Table 5, item 12). The remaining genes on this list (items 13–37) have never been formally linked to AD, but their possible involvement could be either inferred from the literature or from the results of the present study. For example, mutations in A2M, BRI3, EEF1A1, EIF2S2, MAPT and UBB (items 13–18) have been linked to various other neurodegenerative diseases [14–19] (Similarly, ATB2B1, CDK5R1, CPLX1, NES, RAD21, ST13 and TAGLN3 (items 19–25) have been shown to be down-regulated in some brain disorders [20–26] The down-regulation of 12 genes from the C-AD library (items 26–37) was validated by qRT-PCR (Table 3) and a similar literature analysis further supports the potential involvement of RELN, PRKCE and SYT4 (items 35–37) [2, 27, 28] Therefore, it is likely that the down-regulation and the loss of these gene functions could play a role in AD.

Table 6 contains a short list of genes cloned from AD-C subtractive library and supporting evidence for their role in AD. Sixteen of these genes (Table 6, items 1–16) have been previously associated with AD. For example, up-regulation of CLU, GFAP, MT1F (items 1, 2 and 10, respectively) have often been reported in AD brains [29–31] This is consistent with our data from the AD-C library. Two separate groups, Takahashi et al 2000 [32] and Loring et al 2001 [3], reported up-regulation of STXBP1 (item 16) in AD and the transcript of KNS2 (item 17) was reported elevated in injured optic nerve [33]. Both genes were found in the AD-C library and KNS2 was represented by multiple clones; however, we could not validate these changes by qRT-PCR (Table 5). Seven genes from this list (Table 6, items 17–23), have been shown to be elevated in other neurological disorders such as Schizophrenia, Down syndrome and ischemic neurodegeneration. Changes in the additional 17 genes (Table 6, items 24–40) were validated by qRT-PCR, six of which (items 35–40) could be also inferred from the literature.

Possible associations between the 37 down-regulated (Fig. 2) and 40 up-regulated (Fig. 3) genes from the subtracted libraries were mapped by additional literature mining. We imported the Unigene names of each group of genes into LitMiner and used "Alzheimer" and "neurodegeneration" as knowledge terms. All articles containing a gene name, or alias and the knowledge terms, in either title, MeSH terms or abstract were reported in a graphic format. Any two given genes or a gene and one of the knowledge terms appearing in the same article were considered to constitute an association. The frequencies of these associations are represented by the number adjacent to the connecting lines. Within LitMiner, we could rapidly retrieve and manually scan articles to eliminate false association due to the misuse of gene aliases in the database. The graphs in Figs 2 and 3 contain genes that exhibited actual association to the knowledge terms or to other genes. Some genes summarized in Table 5 and 6 are not included in Figs 2 and 3 because the current software finds only associations identified in the title, abstract or MeSH terms. Nevertheless, the LitMiner output gives a high level overview of the relationships of these differentially expressed genes that is not obtainable from a simple gene list.

Figure 2
figure 2

Association of down regulated genes with AD and neurodegeneration. The associations between genes or between genes and knowledge term (Alzheimer or Neurodegeneration) are represented as a graph with the frequency indicated by numbers on the links. A red box or line indicates a knowledge term or a link between a gene and a knowledge term, respectively. A blue box and line shows a gene that is associated with both knowledge terms. A black box and line denotes a gene that is associated with only one knowledge term or another gene.

Full size image
Figure 3
figure 3

Association of up regulated genes with AD and neurodegeneration. The associations between genes or between genes and knowledge term (Alzheimer or Neurodegeneration) are represented as a graph with the frequency indicated by numbers on the links. A red box or line indicates a knowledge term or a link between a gene and a knowledge term, respectively. A blue box and line shows a gene that is associated with both knowledge terms. A black box and line denotes a gene that is associated with only one knowledge term or another gene.

Full size image

Discussion

With the advent of high throughput genomics and proteomics technologies, the involvement of new genes in AD continues to emerge, indicating clearly that the full extent of molecular and functional aberrations responsible for the etiology of AD is not yet understood. The current high throughput methods are mainly designed to identify changes in relatively abundant genes, whereas weakly expressed genes or proteins are still overlooked. For example, cDNA microarrays, although widely used, are limited to known mRNA sequences or ESTs, thus it is not suitable for the identification and isolation of novel rare transcripts. Complementary to microarray technology, subtractive hybridization is capable of detecting changes of both known and novel genes. The existing SSH method enriches the differentially expressed genes by selectively suppressing the amplification of non-differentially expressed transcripts in the PCR reaction, resulting in the enrichment of differentially expressed transcripts. An annealing step is necessary for PCR amplification and subsequent cloning and since rare sequences re-anneal more slowly, this process reduces their chances of discovery. The STAR technology uses single strand RNA/DNA hybridization to remove non-differentially expressed genes. Rare sequences are enriched by slower hybridization and the non-hybridized RNA is recovered and amplified by a linear RNA amplification procedure, which minimizes the biased exponential amplification of subsets of genes in a mixture. These features make the STAR method attractive for examining changes of gene expression in normal and diseased brain tissues, where many mRNAs are present in low abundance. In the present study, we used the STAR method to identify genes that are not only novel in human brains, but also differentially expressed in AD. Our subtracted cDNA libraries constructed by the STAR technology contained genes whose relative differences in expression levels ranged from 25% to 100%, illustrating the fidelity of this method. Although there has been little, and in some cases, no overlap in identified genes between different microarray studies of AD samples, due to experimental variability, different sources of microarray bearing different sets of genes and different areas of brain tissues used, STAR has identified a number of known, moderately abundant genes, whose alternation in AD was in agreement with previous microarray analyses. Of these APP and SNCA were found down regulated [2], AGT [34], CLU [34], FTL [3, 34], MAP2 and STXBP1 [3] were reported up-regulated. More than 55% of clones in each subtracted library contain cDNA of unknown function and more then 25% are novel sequences only matching genomic DNA in database searches. These novel genes can now be vigorously pursued to identify their function and precise role in AD pathology.

Similar to other high throughput approaches, the changes in gene expression identified by the STAR method need to be validated by qRT-PCR. This is especially true for the rare and novel sequences, for which there is no information available in the literature. Combinations of PCR-selected cDNA subtraction and cDNA microarray analysis of the subtracted clones, or using probes generated by PCR-selected subtraction to screen Affymetrix GeneChips have been attempted [35, 36]. The clones produced by STAR can be printed on microarray slides for hybridization with labeled cDNAs from individual brains samples, however, the rare RNA species must be enriched by amplification in order to give reliable hybridization signals. This method, currently under development, could be a high throughput approach to eliminate false positives before qRT-PCR validation on individual genes. We have measured changes in nearly 40 genes from each subtracted cDNA library by qRT-PCR. However, there was a difference in the confirmation rate between the two subtracted libraries. Besides the fact that there seemed to be more down- than up- regulated genes in the AD brains where neurodegeneration was already evident [2, 3]., we cannot offer an definite explanation why some genes, such as, Syntaxin binding protein 1 and kinesin 2 found up-regulated by STAR procedure, consistent with the results obtained by others [3, 32, 33], yet not confirmed by qRT-PCR. Further investigations with different fine-tuned qRT-PCR primers and conditions and by using individual brain samples may offer insight on this issue.

While identification of novel genes functions in neurodegeneration remains to be our ongoing objective, the current study was focused on the association of some of the known genes with AD. Among the down regulated genes in Table 5, those belonging to two major functional categories are noteworthy. The first group contains genes involved in maintenance of neuronal cell function, including CAST, PFRK, SERPINE2, SNCA, SV2A and SYT4 [2, 37–41] Their roles in normal brain are to promote neurite outgrowth and to regulate synaptic vesicle transport or trafficking at the synapse. Their down regulation is consistent with the compromised synaptic transmission observed in AD brain. This undoubtedly contributes to the impairment of memory and cognitive function. Our findings support the current view that AD is a disease of synaptic failure [42].

The second group of genes consists of members of the ubiquitin pathway involved in protein degradation. Although only UBB, UBE2B and UCHL1 passed our current screening criteria, several other ubiquitin pathway-related genes also appeared in the C-AD library. A straightforward interpretation of this result would be that the decreased ability of protein degradation caused by the down regulation of ubiquitin pathway genes, resulted in the accumulation of unwanted proteins in the senile plaques and neurofibrillary tangles. However, studies of individual ubiquitin pathway genes, such as UCHL1, suggest other mechanisms. For example, UCHL1 is sensitive to redox changes and is oxidized in AD brain [43, 44]. indicating that its function under stressed conditions is more than "house keeping" [45]. Elevated expression of other ubiquitin pathway proteins in response to oxidative stress has also been documented. [46, 47] However, a better understanding of the role of individual ubiquitin pathway genes in AD pathology requires comprehensive study.

Extensive evidence suggests that inflammation plays a major role in AD [48]. It is therefore not surprising to find genes, such as MBP, NPTX1 and SCARB2 involved in neuroinflammation [49–51], to be up regulated in AD brains (Table 6). We also found several genes related to the cellular distribution of iron (FTL), transport of copper (CP) and metal-binding (MT1F) that were up-regulated in AD (Table 6). [31, 52, 53] These data strongly suggest a disruption of metal homeostasis and a potential metal neurotoxicity component in AD. The increase of these proteins may indicate an acute phase-type reaction and/or a compensatory response to stress conditions.

We did not find any typical cytoskeletal proteins, such as members of the actin or tubulin families, to be significantly up-regulated in AD brain. However, we did identify three clusters of genes related to cytoskeletal organization (Table 6). The first group contains organelle membrane skeletal proteins such as SAFA, LMNA and SPTBN2 [54–56] The second group encodes skeletal binding proteins, including FKBP2, interacting with erythrocyte membrane cytoskeletal protein [57], KNS2 and MAP2, all associated with microtubules [58, 59] The third group contains GTPase activating proteins, ARHGAP1, CENTG2 (AGAP1), SRGAP2, which regulate membrane trafficking and actin remodeling [60–62] SEPT6 can also be categorized into this group since it is a polymerizing GTPase required for cytokinesis and cortical organization [63]. It is intriguing that so many cytoskeleton related genes are up regulated in the degenerating brains. It is currently unclear whether these changes were the causes or the consequences of neurodegeneration. One possibility is that these genes might be up regulated in activated microglia or astrocytes, which triggered signals that mediate cytoskeletal reorganization and vesicular trafficking during glial cell migration.

Conclusion

We have used a proprietary subtractive hybridization technology (STAR) to identify differentially expressed genes in AD brains, extending existing gene profiling and subtraction methods, such as DNA or protein microarray analyses to identify rare sequences. 55% of the identified differentially expressed genes have no known function, of which, 25% had no matching ESTs in the databases. These sequences represented novel and newly discovered transcripts in the brain and were also differentially expressed in AD brains. Using literature mining tools we have established (Figs. 2 and 3) many new gene associations, not yet reported to be involved in AD. This information will facilitate future efforts aimed at establishing alterations in molecular pathways involved AD pathology.

Methods

Brain tissues and RNA extraction

Poly A+ RNA was isolated from the frontal cortex of frozen post mortem human brains from the same 4 AD and 5 age-matched control subjects as used in our previous study using the same the extraction procedure [34]. Equal amounts of mRNA were taken from each brain and subsequently combined to generate separate pools of AD RNA and normal RNA. Five micrograms of the mixed mRNA from each pool was converted to double-stranded cDNA (ds-cDNA) and used to prepare full-length tester and driver cDNA libraries.

Subtractive Transcription-based Amplification of mRNA (STAR)

The STAR subtraction procedure was performed as illustrated in Figure 1. Briefly, single stranded (+) sense tester RNA and single stranded (-) sense driver DNA were generated from specialized tester and driver cDNA libraries (see below), respectively. Subtraction was accomplished by hybridization of single-stranded driver cDNA to the complementary single-stranded tester RNA, followed by RNase H digestion. The unhybridized tester RNA remained active when subjected to a linear RNA amplification process comprising the steps of (i) reverse transcription, to synthesize cDNA from the tester RNA; (ii) DNA conversion, to append a promoter to the cDNA; and (iii) in vitro transcription, to synthesize additional copies of tester RNA.

Construction of full length tester and driver cDNA libraries

Double-stranded cDNA (ds-cDNA) was synthesized from 5 μg of mRNA from each pool of brains using a modified ThermoScript™ ds-cDNA synthesis kit (Invitrogen, Burlington, ON) and a locking-dT19V oligonucleotide comprising a Not I restriction enzyme site. An Asc I adaptor was then ligated to the 5' terminus of the ds-cDNA. Following digestion with Asc I and Not I enzymes (NEB, Pickering, ON), the ds-cDNA was directionally ligated into pUC18-derived vectors, p17+ for the production of driver DNA and p14 for the production of tester RNA. Both vectors contained a T7 promoter and specific oligonucleotide sequences (OGS302: 5'-GCCTGCACCAACAGTTAACA, in the case of p17+, and OGS77: 5'-CGAGAGCACCTGGATAGGTT, in the case of p14), immediately upstream of the cDNA inserts. These plasmids were then transformed into E. coli DH10B cells to generate the full length cDNA libraries.

Construction of 3'-UTR tester and driver cDNA libraries

Subtraction using STAR was performed using only the more variable 3'-UTR regions of mRNA sequences in order to minimize losses of gene family members that share homologous 5'-UTR and coding regions. Thus, 3'-UTR tester and driver libraries were subcloned from the original full-length p14 and p17+ cDNA libraries as follows. Plasmid DNA (2 μg) from each library was digested with Not I restriction enzyme and purified using Qiaquick (Qiagen, Mississauga, ON) and 1 μg of each was used to in vitro transcribe (IVT) full-length RNA copies of the cDNA inserts with T7 RNA polymerase, according to the manufacturer's instruction (USB, Cleveland, OH). The plasmid DNA template was digested with 2U RNase-free DNase I (Promega, Madison WI) and the RNA was purified with RNeasy kit (Qiagen). The newly synthesized RNA from each library now contained specific oligonucleotide sequences OGS77 and OGS302 at its 5' terminus, which were initially carried by p14 and p17+ plasmid vectors, respectively. Twenty micrograms of each IVT RNA were converted to first-strand cDNA (as described above) and purified by Qiaquick (Qiagen). Second-strand cDNA synthesis was then accomplished in a reaction containing Klenow DNA polymerase and specific oligonucleotide primers, OGS77 or OGS302, according to the manufacturer's instruction (NEB, Pickering, ON). The resulting full-length ds-DNA for each library was purified by Qiaquick (Qiagen). To prepare the 3'-UTR tester and driver libraries for STAR, 6 μg of the full-length ds-DNA for each library was divided into 1 μg aliquots and each aliquot was digested with one of six restriction enzymes (Bsh 1236 I, HinP1 I, Mse I, Msp I, Rsa I or Sau3A I) (NEB and MBI Fermentas, Burlington, ON). Following digestion, each set of 6 reactions was extracted with phenol, pooled and desalted. Each pooled DNA sample was blunt-ended using T4 DNA polymerase (NEB) and ligated to 2.5 μg Asc I linker (NEB) in a 10 μL reaction. Each linker-adapted DNA sample was then digested with Asc I and Not I enzymes (NEB) and purified using Qiaquick (Qiagen). The digested DNA samples were then ligated into Asc I-Not I digested p14 and p17+ plasmid vectors respectively and transformed into E. coli DH10B cells. The resulting transformants for each library were pooled to produce the 3'-UTR p14-tester and p17-driver libraries.

Construction of 3' STAR subtracted cDNA libraries

One microgram of plasmid DNA isolated from each 3'-UTR library was digested with Not I and purified by Qiaquick (Qiagen), and then in vitro transcribed as described above to produce RNA. The RNA copies from the tester library are now ready to be used in STAR. Twenty micrograms of the p17-3' driver RNA were further converted to single-stranded driver DNA in a first-strand cDNA synthesis reaction as described above with the exception that a oligo rU, instead of oligo dT, was used as primer for the cDNA synthesis and the rU primer attached to every first strand cDNA was then digested with RNase A. The p14-3' tester RNA (10 ng) was then hybridized with 100-fold excess p17-3' single-stranded driver DNA in a hybridization buffer containing 40 mM Tris-HCl, pH7.5, 0.1 M NaCl, 7.6% D-Trehalose and 40% DMSO. The reaction was carried out in a thermocycler in descending temperature sequence as follows: 65°C, 63°C and 61°C each for 10 min; 59°C for 30 min; 57°C, 55°C, 53°C, 51°C, 49°C, 47°C and 44.5°C each for 98 min; 42.5°C for 18 hours, followed by RNase H digestion at 40°C for 30 min. The tester RNA was then converted to cDNA and amplified by in vitro transcription as described above. After 3-rounds of STAR (Fig. 1), the remaining tester RNA was converted to double-stranded DNA, digested with Asc I and Not I and ligated into a similarly digested pUC-modified vector. The plasmids containing specific tester DNA inserts were transformed into E. coli DH10B cells to form the STAR libraries.

Analysis of STAR cDNA libraries from human brain samples

A STAR library, where AD 3' cDNA was used as tester (termed the AD-C library) which should contain genes up-regulated in AD brains. Conversely, when control 3' cDNA was used as tester, we produced a C-AD library, which should permit the isolation of genes down regulated in AD. Approximately 600 individual colonies from each subtracted library were picked and the cloned inserts were amplified by PCR with HotStart Taq polymerase (Qiagen) using forward and reverse flanking primers on the vector. The PCR amplicons were purified using the Corning filter polystyrene 96-well plate system (Fisher Scientific, Ottawa, ON). One microliter of the purified PCR product was used for sequencing on the ABI Prism 377 DNA sequencer or 3100 Genetic Analyzer. DNA sequences were analyzed using Sequencher and batch BLAST search.

cDNA synthesis, and qRT-PCR

cDNA was synthesized from the same RNA pools used to construct the original AD and control cDNA libraries using Superscript II Reverse Transcriptase according to the manufacturer's instruction (Invitrogen). The reaction was stopped by adding EDTA to a final concentration of 5 mM. RNA templates were subsequently hydrolyzed in 0.5 M NaOH solution at 65°C for 20 min. The cDNA was further purified using a QIAquick PCR purification kit (Qiagen) and quantified using the OliGreen ssDNA Quantitation Kit (Molecular Probe, Hornby, ON). Forward and reverse primers for sequences of interest were designed using Primer Express (Applied Biosystems, Foster City, CA). Equal amounts of cDNA (2 ng each) were used for qRT-PCR analysis using the QuantiTect SYBR Green PCR Kit (Qiagen) according to the manufacturer's instructions. Fluorescent products were detected using a GeneAmp 5700 Sequence Detection System (PE Applied Biosystems). Percentage of changes was calculated according to the manufacturer's instruction. The experiments were performed in triplicate. Only significant differences (ρ < 0.05; t-test on the qRT-PCR experiments) between AD and control samples are reported as differentially expressed genes.

Literature mining

UniGene symbols of the known genes from each subtracted library were obtained from the SOURCE database using their respective Genbank accession numbers. These symbols were imported into a literature mining prototype software, LitMiner [64], developed by the National Research Council of Canada to identify relationships among genes and their association with biological processes. The search uses the standard UniGene symbols and all possible aliases appearing in the title, abstract or MeSH terms of publications. The associations amongst genes are represented as a graph with the frequency indicated by numbers on the links. Occasionally, the numbers in the graphs may not accurately represent associations because some articles might simply mention genes or terms without specifying an association or using incorrect or incomplete gene aliases. In practice these errors are manageable, because within LitMiner, we can rapidly retrieve and manually scan articles to eliminate such false associations. A knowledge term, such as "Alzheimer" or "Neurodegeneration" was added into the search to explore possible association of these genes with AD. The LitMiner tool was used since it is a much faster than searching for gene relationships manually. In this context, relationship means a co occurrence between either gene names or gene names and a biological process.

Every step performed with LitMiner for this study could be replicated manually using Entrez Gene, PubMed, and graph drawing tools available in Microsoft Word or similar sources. The first step of this manual process would be to find all the gene name aliases available in Entrez Gene for each known UniGene symbol from each subtracted library. Second, each such list of aliases would be converted into a disjunctive PubMed query by adding the OR operator (|) after each alias. As well, any alias that is longer than a single word would be enclosed in quotation signs (" "). For example, part of the query for the gene A2M would be, "A2M" | "alpha 2 macroglobulin" | "alpha2 M". Other manual improvements could be made to this query, i.e., to remove overly general terms. Third, researchers would scan the results removing incorrect or uninformative matches. Fourth, a database or file would record all the PMIDs that matched each of these gene queries in PubMed. Fifth, queries would be created for biological processes and the resulting lists of PMIDs would also be included in the database. Using this database, the researcher could then count how many PMIDs were shared by any pair of genes or biological process. For example, there would be tens of PMIDs whose title, abstract, and MeSH terms mention both A2M and APP. Once all co occurrence counts are collected, those counts could be entered into a manually drawn graph. These steps would require a considerable amount of time, to remove incorrect matches and to produce graphs of the co occurrences between gene names and biological processes.