microarray technology provides an unprecedented opportunity to rapidly examine large-scale gene expression of human tissues (7, 21). This powerful approach can facilitate understanding the molecular pathogenesis of biologically complex disorders such as the neurodegenerative diseases. Amyotrophic lateral sclerosis (ALS) is a neurodegenerative human disease that is almost invariably fatal, usually within 5 years. Motor neurons in ALS undergo degeneration, causing secondary muscle atrophy and weakness. Studies of ALS in humans and mouse models have identified multiple processes potentially involved in the pathogenesis of ALS including mitochondrial dysfunction, enhanced apoptosis, glutamate-mediated excitotoxicity, free radical injury, and autoimmunity (2, 31, 34, 36).

To investigate the molecular phenotype that underlies ALS, we used oligonucleotide microarrays to analyze expression of ∼6,800 genes from ALS and normal human spinal cords obtained at autopsy. The two objectives of this study were, first, to use advanced informatics tools to evaluate the overall pattern of gene expression and deduce a molecular signature of ALS, and, second, to identify functionally important genes whose altered expression profiles may help delineate underlying disease mechanisms. The results indicate that a distinct mRNA expression signature can be discerned in ALS spinal cord; this signature entails changes in expression of many specific genes potentially involved in human ALS pathogenesis.

MATERIALS AND METHODS

Tissue selection and histological examination.

All samples were obtained from local or national tissue banks and were processed under full compliance with Institutional Review Board requirements. For DNA microarray analysis, total RNA was obtained from postmortem gray matter of lumbar spinal cord from 11 individuals including 7 with ALS and 4 normal controls. The seven ALS spinal cords were obtained from five individuals with sporadic ALS (SALS) and two with familial ALS (FALS). Of the two familial specimens, one was from an individual with a superoxide dismutase 1 mutation (SOD1-A4V); the gene defect in the other was not defined (samples were not tested for the recently described Alsin mutations since the phenotype did not correspond to that described for familial juvenile ALS). The average age of death recorded for the study subjects was 56 yr for ALS and 60.5 yr for the non-ALS samples. Most ALS samples were subjected to histopathological examination prior to inclusion in the study. Findings included loss of anterior horn cells or chromatolysis accompanied by corticospinal tract degeneration. Recorded tissue collection times ranged from 2–18 h postmortem, with the averages being 10 h for ALS and 10 h for normals.

RNA isolation.

All samples were frozen at −80°C prior to RNA isolation. Frozen tissues were dissected to separate gray and white matter, followed by mortar fragmentation in the presence of liquid nitrogen. Total RNA was isolated using the Qiagen RNeasy kit with the modification that the upper lipid layer was entirely removed from all samples after resuspension and spinning with buffer RLT. Although this strategy led us to discard ∼20–40% of the RNA sample, it preserved the overall integrity of our RNA, because the lipid layer is known to compromise RNA stability and quality. Only RNA samples that met our quality/control standards (i.e., 260/280 OD ratios >1.8 in all RNA samples, and ratio of 3′ to 5′ signal >1.7 for GAPDH and β-actin on a Test1 or Test2 GeneChip for RNA samples used in microarray experiments) were used for microarray hybridization and/or real-time RT-PCR analysis. Although it is recognized that postmortem changes in levels of individual mRNAs presumably occurred, these were likely similar among the specimens.

GeneChip probe array analysis.

DNA microarray analysis was performed as previously described (21), using Affymetrix HuFL GeneChip probe arrays. These microarrays contain 7,070 distinct probe sets, representing ∼6,800 human genes. Briefly, using 7–10 μg of total RNA, double-stranded cDNA was synthesized using the SuperScript Choice System (Life Technologies) with the following modifications. In the first-strand synthesis, the reverse transcription reaction contained a T7-(dT)₂₄ primer plus 0.1 M DTT and 10 mM dNTP mix. For second-strand synthesis, Escherichia coli DNA ligase (10 U/μl) and T4 DNA polymerase I (10 U/μl), 10 mM dNTP mix, and RNase H (2 U/μl) were used. Phenol-chloroform extraction was followed by in vitro transcription (IVT) (Ambion T7 Megascript System) with biotin labeling. IVT was performed with (1:3) biotinylated:unlabeled CTP and UTP. The Ambion T7 enzyme mix and T7 transcription buffer were added to the double-stranded cDNA and NTP labeling mix (ATP, CTP, UTP, GTP, Bio-11-CTP, and Bio-16-UTP). The NTP labeling mix was incubated for 5 h at 37°C, and cleaned using RNeasy columns (Qiagen). A quantity of 13–20 μg of fluorescently labeled and chemically fragmented cRNA was used for array hybridization. Fragmented cRNA and herring sperm DNA were added to the hybridization buffer containing 1.0 M NaCl, 10 mM Tris·HCl, pH 7.6, and 0.01% Triton X-100.

The hybridization mixture was heated to 99°C for 5 min, spun, and incubated at 45°C for 5 min, then injected into the probe array cartridge. Hybridizations were carried out at 45°C for 16 h with mixing at 60 rpm. Following hybridization, solutions were removed, and arrays were rinsed and incubated with 0.1× ST-T (100 mM NaCl, 10 mM Tris·HCl, pH 8.0, and 0.01% Triton X-100) at 50°C for 20 min. Hybridized arrays were stained with 5.0 μg/ml streptavidin-phycoerythrin (Molecular Probes) and 2.0 mg/ml acetylated BSA (Sigma) in 1× ST-T at 40°C for 15 min. The streptavidin-phycoerythrin step was repeated after an intermediate amplification step in which anti-streptavidin rabbit IgG antibodies and secondary biotinylated goat anti-rabbit antibodies are added to the samples. Following washes, probe arrays were scanned twice at 6 μm resolution using the GeneChip system confocal scanner. All samples were hybridized to Test1 or Test2 microarrays to assess cRNA quality control, prior to hybridization to HuGeneFL arrays.

Microarray data collection and analysis.

Scanned image files were converted to mRNA expression levels using Affymetrix GeneChip 3.1 software. This software assesses presence or absence of transcripts for each probe set, taking into account metrics such as background, noise, and comparison of intensities between “perfect match” (PM) and their control “mismatch” (MM) probe cells. The average intensity of each microarray was scaled to a target intensity of 1,500. Files containing the average difference intensity values (i.e., expression levels) for each probe set were imported into an Access database. For generation of discriminatory gene lists, all genes with maximum intensities below a value of 1,500 across all samples were excluded from the analysis. Our aim was to find the most meaningful data points, and thus our analysis was designed to obtain the genes with highest statistical significance (i.e., a limited false discovery error of <5%). The reduction in false positives was achieved using leave-one-out cross-validation (LOOCV), which allowed us, by doing random permutations, to reduce false positives from the initial list obtained by t-test (13). We assigned an arbitrary minimal expression level (i.e., average difference) value of 20 to any negative or zero values prior to performing the statistical analysis. For fold change calculations, all positive expression levels below a value of 300 were reset to 300 to purposely provide a very conservative estimate of fold increase or decrease for low-abundance transcripts.

Identification of differentially expressed genes.

Various parametric and nonparametric statistical tests have been proposed to identify genes that are informative and discriminatory between two conditions (23, 26, 39). We propose an alternative test based on the Wilks lambda score, which is readily extendable to multi-class cases and is based on group variance similar to ANOVA (8). Wilks lambda score (Λ_i) is defined by the ratio of the within-group variance (W_i) to the total variance (T_i) of gene i (8, 13).

The vector, x_i (N × 1), contains the expression level of gene i in N samples, and x̄_i is the mean expression of gene i in all N samples. The superscript j represents class j among c classes. Then, an F-test, defined by Λ_i (13), can be applied to determine discriminatory genes with a significance level (α = 0.01 was used in this study)

A discriminatory gene has a large between-group variance B_i (i.e., T_i − W_i) and a relatively small within-group variance (W_i), resulting in a small Wilks lambda score, but a large F statistic value. This F-test based on Wilks lambda score might produce a number of false positives and even negatives because of violations of the underlying normality and the equal covariance assumptions as well as artifacts from significant sample variability (see results).

To reduce bias from the sample variability, we incorporate an error rate calculation through an LOOCV procedure into the determination of discriminatory genes using the Wilks lambda score (35). For each LOOCV, the following procedures are applied: 1) one sample from N samples is withheld in a test set; 2) using the rest of samples (training set; N − 1 samples), a certain number of discriminatory genes is determined based on the F-test above; 3) these selected genes are ordered by their F values; 4) a series of Fisher discriminant analysis (FDA) classifiers is then built, using the number of discriminatory genes according to the priority determined by their F values, to estimate a set of error rates. That is, the most discriminatory gene is used to construct the first classifier using the training set, and the samples withheld in the LOOCV are then tested using this classifier to estimate the error rate. The same process is repeated as one gene is sequentially added to the classifier according to its F value until all selected genes are used. Thus, for a given LOOCV, a set of error rates is generated as a function of the number of discriminatory genes used in classifiers. Then, the steps 1-4 are repeated using a different test and training sets until all samples have been withheld in the test set once.

This LOOCV procedure produces N sets of discriminatory genes and N sets of error rates. Then, the N sets of error rates are averaged to obtain the average error rate as a function of the number of genes used in the classifier. Using the error rate curve, the number of discriminatory genes that are informative can be determined at the point where the averaged error rates show an asymptotic behavior. For the comparison of SALS and normal samples, the asymptote was observed at 93 genes, and the corresponding error rate was zero (Fig. 1A). Then, a final set of informative discriminatory genes is determined from a list of selected genes sorted by frequencies by which they appeared in the N sets of discriminatory genes and then by their averaged P values (Tables 1–3). This methodology mainly removes falsely identified genes due to artifacts coming from sample variability but also affects false positive/negative rates because the LOOCV changes sample distributions and covariance structures with a certain extent. Whether these removed genes are false positives or negatives is an open question. However, removing these genes improves robustness in our further analysis such as biological interpretation of these discriminatory genes. Further details of determination of informative discriminatory genes based on Wilks lambda and the LOOCV are described in Hwang et al. (13).

Fisher discriminant analysis.

FDA is similar to principal component analysis (PCA) because they both use linear reduction. However, while PCA summarizes the data distribution in a reduced dimensional space, FDA summarizes separation of samples among groups into a reduced space. The reduced space in FDA is defined by a few linear combinations of genes, and coefficients in the linear combination (the loadings of the genes, V) are shown to be the “eigenvectors” of W⁻¹B, the ratio of between-group variance (B) to within-group variance (W)

The eigenvalues (diagonal elements of Λ) indicate the discrimination power for the corresponding linear combinations. The loadings effectively summarize how gene expression values achieve the maximum separation of sample groups in the FDA space (called expression signature; see Fig. 1B). Further details of FDA, such as discrimination rule, and its application in classification of microarray data are described in Stephanopoulos et al. (38).

Real-time quantitative RT-PCR.

Relative quantitation with real-time, one-step quantitative reverse-transcriptase PCR (QRT-PCR) was performed multiple times in triplicate with the SYBR Green PCR reagents and an ABI Prism 7700 Sequence Detection System (PE Applied Biosystems, Foster City, CA) on RNA isolated from whole spinal cord postmortem tissue.

First, we tested three ALS samples for the expression of creatine kinase, inhibitor of apoptosis 1, Oxa1Hs, and phospholipid scramblase and compared with four independent normal subjects. These four normal subject RNA samples had not been used for the microarray hybridization experiments. These genes had been shown to have significant fold changes in ALS tissues but, except for Oxa1Hs, did not reach the stringent statistical criteria used by LOOCV coupled with Wilks lambda method in the microarray analysis. mRNA levels were normalized to levels of β-actin. The PCR primers used were as follows. Mitochondrial creatine kinase: forward primer, 5′-AGGCCTCAAAGAGGTTGAGAGA-3′; reverse primer, 5′-TTAGATGGACAGGTCAAGATGTATCC-3′. Inhibitor of apoptosis 1: forward primer, 5′-GGAACCTGGAGAAGACCATTCA-3′; reverse primer, 5′-GA-ACTGTCTGTTTTACCAGGCTTCTA-3′. Phospholipid scramblase: forward primer, 5′-AGAGATGTACTAAAAATAAGTGGTCCAT-3′; reverse primer, 5′-TGCCAACCACACACTGTTCAT-3′. OXA1Hs: forward primer, 5′-CCTCTGGTGGTTCCAGGATCT-3′; reverse primer, 5′-TCTCAGCACCTAGCTCAAGAACA-3′. β-Actin: forward primer, 5′-GCCCTGAGGCACTCTTCCA-3′; reverse primer, 5′-GCGGATGTCCACGTCACA-3′.

Next, we tested whole spinal cord ALS samples for the expression of the following genes, which were altered in ALS gray matter under the stringent statistical LOOCV-FDA criteria. Anaplastic lymphoma kinase receptor: forward primer, 5′-AATCGGGCGTCCAGACAAC-3′; reverse primer, 5′-AAGAAGTCCACTGCAGACAAGCT-3′. Transferrin: forward primer, 5′-TCAGCAGAGACCACCGAAGA-3′; reverse primer, 5′-CATCCAAGCTCATGGCATCA-3′. Neuronal membrane glycoprotein M6b: forward primer, 5′-CGTGGCGATTCTTGAGCAA-3′; reverse primer, 5′-TTGTATCACCTCGCTCAGCAA-3′.

Primer sets were selected to have a GC content that would support a 58–60°C melting temperature (Tm) and a content of no more than two Gs or Cs within the five last nucleotides at the 3′ end of the primer. Primer sets were designed, when feasible, to span one or more introns. All amplicons ranged between 91 and 121 base pairs. Electrophoretic analysis of expected product sizes was performed for all primer sets to confirm the fidelity of the reaction.

RESULTS

Identification of discriminatory genes and FDA.

To maximize the power of global transcript profiling to discriminate gray matter tissues of ALS from gray matter tissues of normal spinal cord, the LOOCV coupled with Wilks lambda was used to identify a robust set of discriminatory genes in seven ALS and four normal gray matter samples (see the methods). The LOOCV was done 11 times, and the asymptote was observed at 93 genes as shown in Fig. 1A. Thus a final set of informative discriminatory genes was identified as the top 93 genes from a list of genes sorted by frequency and then P values (Table 1). The corresponding error rate was estimated to be zero (see Fig. 1A). Interestingly, the first four genes, which correspond to the first four smallest P values in Table 1, also produced the “zero” averaged error rate in Fig. 1A. Thus further research is being performed for the potential use of these genes (e.g., diagnostic markers). When Wilks lambda score only had been applied to the 11 samples, 157 discriminatory genes were identified, and it can be thus concluded that the LOOCV removes a significant number of genes that show inconsistent discriminatory expression patterns.

Figure 1B shows one of the prediction results for an ALS withheld sample during a LOOCV when we used 93 discriminatory genes in a classifier. Figure 1C also shows the FDA loadings of the 93 genes to maximally separate ALS samples from normal samples shown in Fig. 1B. This FDA loading pattern represents how gene regulation ensures the maximal separation: genes with positive FDA loadings have to be upregulated in ALS, whereas genes with negative FDA loadings have to be downregulated in ALS. Thus we can consider this pattern an ALS-specific expression signature. These observations indicate that microarray data combined with the statistical analyses can provide a characteristic signature of the gray matter components of ALS pathology.

Moreover, distinct signatures for FALS (Table 2) and SALS (Table 3) spinal cord gray matter gene expression could be discerned, demonstrating that global transcriptional profiling can capture the prominent molecular phenotypes of various forms of ALS, compared with normal spinal cords.

Decreased expression of a neuronal survival gene in late-stage human ALS spinal cords.

A comparison of ALS spinal cord gray matter to normal specimens revealed a reduction of expression of the survival motor neuron protein (SMN) gene, recently reported to exist in abnormal (decreased or increased) copy numbers in ALS patients (5). This observation substantiates and extends recent reports that ALS involves induction of apoptotic events in spinal cord of both humans and transgenic FALS mice expressing a mutant human Cu-Zn SOD1 (18, 25, 37, 40).

Role for oxidative stress in ALS.

ALS and other neurodegenerative diseases exhibit tissue damage reflective of oxidant stress and enhanced free radical generation (9, 20, 28). In a minority of cases, ALS is caused by mutations in the Cu-Zn SOD1 (6, 32), whose normal function is to catalyze the dismutation of superoxide radicals to hydrogen peroxide. Mutations in SOD1 result in a gain of one or more adverse, neurotoxic functions. Although the exact nature of this acquired toxic property remains controversial, data suggest that the mutant SOD1 enzyme has altered affinities both for zinc and for some substrates and may have enhanced peroxidase activity. Moreover, most of the mutant SOD1 proteins demonstrate diminished protein stability and, under appropriate circumstances, a tendency to aggregate. One consequence of both types of enzyme abnormality is increased generation of toxic oxygen and nitrogen radical species (3, 41, 43). Whatever the exact mechanisms of its toxicity in ALS, it is striking that the mutant SOD1 protein is clearly pro-apoptotic, whereas the wild-type molecule distinctly remains an anti-apoptotic molecule (29). Also striking is the observation that transgenic expression of mutant SOD1 in mice generates a remarkably accurate model of ALS (30). This model is at least partially responsive to apoptosis inhibitors (18). In this context, we were intrigued to find that our analysis of mRNA expression captured changes in genes associated with metal ion homeostasis as well as cellular stress. Of particular importance, there was a decrease in transferrin and heat shock protein hsp90. In addition, the expression of the metal ion homeostasis gene NRAMP1 was prominently elevated in FALS gray matter tissues.

An inflammatory signature in ALS.

An inflammatory signature was evident from changes in immune system-related genes. Transcripts for the antigen presentation molecules HLA class I and the Ia-associated invariant γ-chain, as well as cathepsin B, possibly originating from macrophage/microglia lineage cells, were detected at high levels in ALS. We did not detect significant evidence of a T cell or NK cell response, with pro-inflammatory cytokine mRNAs such as TNF-α, IL-6, and IL-2 being undetected in ALS. Interestingly, the IL-7 receptor was downregulated in ALS gray matter tissues, and IL-7 has been shown to have neurotrophic properties (24). In FALS gray matter tissues, complement C1q expression was elevated. Our results substantiate the long-held view (27, 36) that an inflammatory state characterizes postmortem ALS spinal cord.

Additional pathways altered in ALS.

In addition to revealing evidence of cell death, stress, and inflammation, ALS spinal cord was characterized by downregulation in transcripts associated with adhesion (e.g., integrin-β3), extracellular matrix (e.g., type IX and XVI collagen genes), and cytoskeletal organization (e.g., dynactin, α-tubulin, and synaptobrevin), consistent with observations in mutant mice with perturbed neurofilament assembly that develop an ALS-like illness (4, 14). These genes may need to be viewed as potentially related to ALS signaling pathways, in light of the recently described FALS mutation in a Rho GTPase homolog (Alsin, ALS2) (10, 42), which may participate in cell-cell adhesion, linkage of growth signals to the cytoskeleton, or intracellular trafficking.

Signal transduction, transcription, proteasome activity, and growth.

Several signaling genes that serve as intracellular mediators of growth factors and activation signals had augmented expression in ALS gray matters, including transducin-β2, adenylyl cyclase, and protein kinase C-δ. Numerous changes in mRNA transcript levels were also detected in molecules involved in proteasomal/protein degradation activity, transcription, and growth factor pathways. Several ubiquitination-related gene transcripts were dysregulated in ALS tissues, including downregulation of ubiquitin and upregulation of the E6-AP ubiquitin protein ligase 3A mRNAs. In FALS gray matter tissues we found an increase in an ubiquitin conjugating enzyme (accession number L22005) and a decrease in expression of the ubiquitin COOH-terminal hydrolase (UHX1). Of particular note, altered levels of members of the ubiquitination system may correlate with increased ubiquitination as an early event in ALS (17), and could be related to protein aggregation or disruption of normal proteasomal function. Finally, fibroblast growth factor-4 (FGF-4), the FGF receptor flg-2, and the bone morphogenetic protein type II receptor (BMPRII) were downregulated in ALS spinal cord gray matter tissues.

Dysregulation of mitochondrial function and neuronal metabolism genes.

Impairment of mitochondrial function and neuronal energy metabolism is a characteristic of ALS (1). We detected changes suggesting altered folate metabolism, which has been previously implicated in ALS (45), with a decrease in the folate receptor-α. Downregulation of the mitochondrial ATP synthase δ-subunit was also found.

Glutamate-mediated excitotoxicity and neurotransmission profile.

Our study detected no changes in the levels of expression of the astroglial glutamate transporter EAAT2, although we did detect upregulation of the glutamate transporter EAAT1 in ALS gray matter. The finding that EAAT2 expression levels, as predicted by mRNA in spinal cord, were not decreased is of interest, as Western blotting and immunofluorescence analyses have revealed a distinct loss of EAAT2 protein in motor cortex in ALS patients and SOD1 transgenic ALS mice (33). It has been proposed that this may reflect aberrant splicing of the EAAT2 transcript (19), a finding that was not addressed by these microarray studies. The metabotropic glutamate receptor type 3 mRNA was found to be decreased in ALS gray matter tissues. Finally, the ε-subunit of the GABA receptor was downregulated in ALS, consistent with a recent report of decreased GABA receptor in G93A SOD1 mice spinal cords also using microarrays (44).

Miscellaneous genes.

Interestingly, genes involved in lipid metabolism and clearance were dysregulated in ALS. The pre-apolipoprotein C-III gene was downregulated in ALS gray matter tissues. Of note, genes involved in chromatin modification were affected in ALS, as shown by downregulation of the heterochromatin protein homolog (HP1) and the histone acetylator chromatin assembly factor 1 (CAF1). Finally, genes involved in lysosomal metabolism were found to be altered, as shown by upregulated oligosaccharyltransferase and downregulated Gal-β(1–3/1–4)GlcNAc α-2.3-sialyltransferase.

Genes prominently altered in FALS gray matter tissues.

Many genes were found to be predominantly altered in FALS gray matter (Table 2). These included the ε-Fc receptor, the HLA class II genes DC-α, F, and DQ, and the inflammation and extravasation-associated genes cathepsin D and thromboxane synthetase. Several genes associated with high intracellular calcium levels were found to be elevated, including reticulocalbin, calcyclin, and calgizzarin. The channel-like integral membrane protein CHIP28 gene had remarkable upregulation. The B6 metabolism gene pyridoxal kinase, the axonal sprouting-related axin and Kallman syndrome genes, and the DNA repair gene Rad23A homolog were also elevated in FALS gray matter tissues. Other genes involved in transcription, neurotransmission, signaling, cell cycle, and enzymatic functions were also exclusively found in FALS gray matter.

Genes predominantly altered in SALS gray matter tissues.

SALS gray matter tissues had an exclusive gene expression profile (Table 3), which included high-level expression of the nicotinic acetylcholine receptor α4-subunit, the γ-interferon-inducible protein IP-30, and an immunoglobulin light chain transcript. Downregulation of the homeotic gene NBPhox was also detected in SALS gray matter. Finally, apolipoprotein C-II was elevated only in SALS tissues.

Real-time QRT-PCR verification.

To verify that alterations in expression levels detected by DNA microarrays reflect changes in mRNA levels (particularly with relatively small changes in expression), we confirmed changes in expression of multiple transcripts in SALS whole spinal cords using real-time QRT-PCR. First, we compared gene expression alterations from a group of genes that were selected solely on fold change >2. By testing against four independent normal spinal cord control RNA samples, comparable increases were detected in expression of phospholipid scramblase (4.4-fold microarray, 3.8-fold QRT-PCR) and OXA1Hs (3.2-fold microarray, 2.1-fold QRT-PCR), whereas similar decreases were seen observed with inhibitor of apoptosis 1 (IAP1) (9.0-fold microarray, 4.0-fold QRT-PCR) and mitochondrial creatine kinase (2.7-fold microarray, 1.6-fold QRT-PCR) (Fig. 2A). OXA1Hs is essential for assembly of cytochrome c oxidase and the ATP synthase complexes, and its elevation may reflect a compensatory response. Of note, OXA1Hs had been detected in our microarray analysis as a predominant FALS signature gene, and our demonstration by QRT-PCR of its upregulation in this new set of SALS spinal cords indicates that there may be some variability and overlap of expression of some genes in FALS and SALS. Next, we tested by QRT-PCR three additional genes that had been shown to be altered in DNA microarrays based on P values. As shown in Fig. 2B, downregulation of transferrin, neuronal membrane glycoprotein M6b, and anaplastic lymphoma kinase receptor (ALKR) mRNAs was confirmed in SALS whole spinal cords using QRT-PCR.

DISCUSSION

Transcriptional profiling provides a global index of mRNA expression corresponding to major patterns of biological activity in a particular specimen. This study revealed changes in many biological processes in human ALS spinal cord. These included evidence of inflammation, oxidative stress, neuronal cell death, abnormal signaling, and cytoskeletal dysfunction. Many observations substantiate prior work, whereas many previously unappreciated mRNAs will provide new hypotheses regarding the pathobiology of ALS. In addition to discovering changes in individual mRNA transcripts, FDA analysis was able to show that there are specific gene expression fingerprints that distinguish ALS from healthy specimens, as well as subtypes of ALS from one another.

Studies with postmortem tissues, however, are more challenging to interpret since heterogeneous cell populations may blur the distinctive profile originating from the most affected cell types. Furthermore, in addition to representing alterations in gene transcription, these profiles may also reflect shifts in cell populations associated with losses or changes of cell types in a tissue. In the present study, some mRNA changes may reflect the underlying loss of motor neurons, and others may represent increases in the numbers of activated microglia/macrophage lineage cells. Nonetheless, we do not believe that changes in the proportions of various cell types are likely to be the primary factor explaining the expression patterns we describe here. Most gene expression changes were detected in both gray and white matter (the latter not shown). Moreover, a number of neuronal genes were actually unchanged, such as the neuronal marker catechol-O-methyltransferase (COMT). Furthermore, most transcripts detected in spinal cord were actually unchanged in expression level across samples.

In studying human disease tissue, as in the present study, it is difficult to distinguish primary from secondary events or to recognize with certainty, in the case of drug response studies, deleterious from beneficial effects. Furthermore, gene expression changes late in a disease process may not necessarily reflect the initial transcriptional responses to the pathogenic insult. We therefore acknowledge that because this is an autopsy study in late-stage ALS spinal cords, some of our findings are likely to reflect terminal disease course or even agonal gene expression changes and thus not represent the molecular pathology that is central to the motor neuron death process in ALS. However, we believe that the gene expression profiles observed in late-stage human ALS will serve as a valuable frame of reference for investigations of potential drug targets for this fatal disease. Since early-stage human ALS postmortem spinal cord samples are exceedingly difficult to obtain, these investigations must be supplemented by transcriptional profiling of tissues at various disease stages from ALS animal models and evaluation of their transcriptional response to various therapies. A decrease in a GABA receptor transcript found in our study parallels similar findings in end-stage disease reported for nontreated SOD1 mice (44). Our broad-scale study also supports the growing evidence that the molecular events supporting the current ALS hypotheses are not mutually exclusive, and some gene signatures may underlie potential links among these pathways. Since recent studies have shown that rare cases of juvenile FALS are associated with mutations of a putative Rho GTPase regulator possibly implicated in regulating neuronal cytoskeletal integrity, we suggest that signaling and cytoskeletal genes highlighted in this report may be functionally linked and should be the focus of further investigations. Our demonstration of downregulated dynactin mRNA in ALS spinal cords is also a provocative finding since recent reports have linked dynactin dysfunction with motor neuron disease (15). Interestingly, recent reports have validated the use of DNA microarrays in revealing chronic pathology-related profiles in complex human tissues, including postmortem brains (11) and surgically obtained joint and bowel samples (12). In addition, a recent profiling study using nonredundant expressed sequence tag (EST) cDNA array filters revealed inflammation and apoptosis signatures in human ALS spinal cord, demonstrating the utility of large-scale approaches for examining postmortem ALS tissues (22).

Several aspects of our study deserve emphasis. In particular, we utilized highly sensitive, fluorescence-based DNA microarrays to identify an in vivo mRNA expression signature suggesting compromised neuronal survival in the central nervous system (CNS). Furthermore, we characterized other gene pathway signatures that reflect mechanistic processes previously implicated in neurodegeneration such as cellular stress and altered neuronal metabolism. We also identified unsuspected changes in specific genes including transcription factors, hormone receptors, signaling components, metal ion homeostasis genes, and neurotransmission molecules. Interestingly, several observations such as changes in the expression of genes involved in inflammation or neurotrophic support could guide therapies, as new modulators of these pathways are developed.

Our ability to detect gene expression profiles that help distinguish FALS vs. SALS spinal cord tissue samples may reflect the presence of distinctly altered pathways in these individuals that culminate in common outcomes (i.e., degeneration of motor nerves). For instance, the presence of reticulocalbin, calcyclin, and calgizzarin upregulation in FALS gray matter tissues suggests greater calcium metabolism dysregulation, and this merits further study. Similarly, marked upregulation of axin mRNA in FALS but not SALS suggests that the response to oxidative stress or other offending stimuli may actually be distinct in disease subtypes and lead to dynamic changes in disease progression at the molecular level, which could lead to a rational design of individualized therapies. Caution should be exercised when categorizing microarray-identified genes into disease subtypes, however, since factors such as timing of postmortem tissue collection, stage of disease, degree of inflammation, sample type (i.e., gray vs. white matter vs. both), or variability of statistical criteria used may lead to erroneous assumptions of classification. A good example is the OXA1Hs gene, which we initially categorized as a FALS “specific” gray matter gene from our microarray studies and later found it to be similarly altered in 1) a microarray comparison of grouped FALS and SALS gray and white matter samples against grouped gray and white matter control samples (not shown) and 2) an unrelated set of SALS whole spinal cord samples by QRT-PCR. Thus we interpret our gene classification as approximate estimates of disease subtype signatures and prefer the terms “prominent” or “predominant” instead of “specific” gene signatures, as we recognize that some overlap between SALS and FALS expression alterations exists.

Finally, we observed a number of parallel changes in gene expression among the degenerative samples used in this study and the profile of the ageing murine brain reported elsewhere (16). These include increases in cathepsins, MHC molecules, CD9, and transducin β2. These parallel observations reinforce the known age-dependent nature of ALS. Moreover, they offer the prospect that ALS may represent, in part, an accelerated aging process in motor neurons. This study indicates that DNA microarray analysis of postmortem CNS tissue can provide a window into the late stages of the molecular pathophysiology of neurodegeneration.