Informatics-assisted Protein Profiling in a Transgenic Mouse Model of Amyotrophic Lateral Sclerosis*S

One of the causes of amyotrophic lateral sclerosis (ALS) is due to mutations in Cu,Zn-superoxide dismutase (SOD1). The mutant protein exhibits a toxic gain of function that adversely affects the function of neurons in the spinal cord, brain stem, and motor cortex. A proteomic analysis of protein expression in a widely used mouse model of ALS was undertaken to identify differences in protein expression in the spinal cords of mice expressing a mutant protein with the G93A mutation found in human ALS. Protein profiling was done on soluble and particulate fractions of spinal cord extracts using high throughput two-dimensional liquid chromatography coupled to tandem mass spectrometry. An integrated proteomics-informatics platform was used to identify relevant differences in protein expression based upon the abundance of peptides identified by database searching of mass spectrometry data. Changes in the expression of proteins associated with mitochondria were particularly prevalent in spinal cord proteins from both mutant G93A-SOD1 and wild-type SOD1 transgenic mice. G93A-SOD1 mouse spinal cord also exhibited differences in proteins associated with metabolism, protein kinase regulation, antioxidant activity, and lysosomes. Using gene ontology analysis, we found an overlap of changes in mRNA expression in presymptomatic mice (from microarray analysis) in three different gene categories. These included selected protein kinase signaling systems, ATP-driven ion transport, and neurotransmission. Therefore, alterations in selected cellular processes are detectable before symptomatic onset in ALS mouse models. However, in late stage disease, mRNA expression analysis did not reveal significant changes in mitochondrial gene expression but did reveal concordant changes in lipid metabolism, lysosomes, and the regulation of neurotransmission. Thus, concordance of proteomic and mRNA expression data within multiple categories validates the use of gene ontology analysis to compare different types of “omic” data.

Amyotrophic lateral sclerosis (ALS) 1 is a fatal disease characterized by the progressive loss of motor neurons from the spinal cord, brainstem, and motor cortex. An important discovery in the study of the disease was the existence of mutations in the gene encoding Cu,Zn-superoxide dismutase (SOD1) in patients with a familial form of ALS (1). SOD1 mutations found in ALS are predominantly missense point mutations that occur at nearly 70 individual sites along the amino acid sequence. Most of these SOD1 mutants retain enzymatic activity indicating that the effects of the mutant are not due to a loss of normal function but rather a gain in a property that is toxic to motor neurons. The nature of this toxic function is not understood, although recent work has implicated the formation of SOD1 aggregates as a causal phenomenon (2,3). Yet the mechanism of how mutant SOD1 initiates ALS is not known.
Transgenic (Tg) mice that express a mutant SOD1 gene can develop motor neuron disease and serve as a model system for human ALS (4,5). Characterization of several SOD1 mutants indicated that the severity and time of onset of disease are highly dependent upon the specific mutation and level of expression. Within the spinal cord of SOD1 mutant mice, the highest levels of expression occur in the ventral horn motor neurons (6). It is suspected that this localized expression coupled with biochemical properties of motor neurons renders them vulnerable to mutant SOD1 (7). One of the best characterized SOD1 Tg mouse is the strain that carries the G93A-SOD1 mutation (5). These mice are apparently healthy until about 90 days of age at which point they develop symptoms of motor neuron disease such as shaking limbs when suspended. After 120 days most of these mice are in the end stages of disease when they become completely paralyzed. Although symptom-free at 60 days of age, G93A-SOD1 mice exhibit spinal cord pathology characterized by fragmented Golgi (8), mitochondrial swelling, and vacuole formation (9). Also prominent at 90 days is the presence of cytoplasmic inclusions, many of which contain SOD1 (10). Indeed biophysical studies of mutant SOD1 proteins indicate that they have a greater tendency toward aggregation than wild-type SOD1 and have altered metal binding properties (7,11,12). Thus, the disease course is linked to changes in cellular biochemistry that occurs prior to the presence of mutant SOD1 aggregates and inclusion bodies (10).
Dysfunction of spinal cord motor neurons in mutant SOD1 mice has been attributed to cumulative cellular stress that appears to converge upon the mitochondria as a key target (13)(14)(15)(16)(17)(18)(19). Mitochondria subjected to oxidative and other stresses can lead to the release of apoptotic factors that trigger a caspase-mediated pathway of cell death (20,21). SOD1 mutants expressed in cultured cells lead to the activation of caspases in SOD1 mutant Tg mouse spinal cord (22)(23)(24) but only in a few cases of human ALS (24). Although they slow the disease progression, neither agents that promote neuronal survival (25,26) nor expression of antiapoptotic proteins (22,(27)(28)(29) prevents the death of SOD1 mutant mice. Although the subject of intense study over the past decade, the sequence of events triggered by mutant SOD1 that initiate cell dysfunction and ALS symptoms and mediate cell death remain to be established.
Gene expression studies (6) and signaling pathway studies (30 -32) done on mouse spinal cords from normal and SOD1 mutant mice have provided some clues to the nature of the cellular insults that have been initiated. Transcriptional profiling of G93A-SOD1 mice at 90 and 120 days revealed gene expression changes consistent with glial activation, activation of scavenger pathways for proteins and lipids, and altered expression of a few metal ion-binding proteins (6). However, these mRNA expression analyses failed to point toward a common pathway that leads to neuronal dysfunction. This is one of the problems confronting the characterization of neurodegenerative diseases because the cause and effect relationships among all of the proteins affected and genes differentially expressed in the degenerative compared with normal state are disconnected. Therefore, we sought a potential unifying approach using proteomics and informatics to investigate the functional changes that occur during the progression of neurodegenerative disease. We used a method that can be used with multiple protein profiles derived from animals expressing different transgenes or from a common transgenic animal at different ages as well as mRNA-based microarray data.
The current work is an initial step in addressing the problem of cause-effect relationships in ALS using a direct assessment of changes in protein expression in the spinal cords of transgenic mice expressing wild-type and G93A-SOD1 proteins compared with those expressing wild-type human SOD1 and normal nontransgenic mice. Protein profile data were analyzed using gene ontology relationships. These analyses revealed an enrichment of several protein groups in diseased (G93A-SOD1) compared with control animals. Previously unidentified in mutant SOD1 mouse spinal cords were changes in G-protein regulation and lysosomal enzymes. Overexpression of mutant G93A-SOD1 in Tg mice appears to accelerate the course of the disease by altering selected mitochondrial proteins in the electron transport chain necessary for ATP synthesis and unbalancing the spectrum of proteins involved in cellular communication systems, particularly those involved in the regulation of neurotransmitter secretion.

EXPERIMENTAL PROCEDURES
Fractionation of Spinal Cord Extracts-G93A-SOD1 and wt SOD1 mice have been described previously (5,33). Non-Tg mice were the background strain B6SJL. Spinal cords were carefully dissected from 60 -70-or ϳ120 (end stage)-day mice following euthanization according to Northwestern University approved animal care protocols. Dissected tissue was immediately frozen (Ϫ70°C) and stored at Ϫ80°C until used. For preparation, spinal cords were quickly thawed and homogenized using a Teflon-glass Dounce (four strokes) in lysis buffer consisting of 25 mM Tris, 0.1 M sucrose, 1 mM EDTA, pH 7.5 on ice. The extracts were centrifuged at 600 ϫ g for 20 min to sediment cellular particles (P1 pellet) and then at 10,000 ϫ g for 20 min to create a clear supernatant (S2) and pellet fractions (P2). The particulate fractions were resuspended in lysis buffer, and protein concentrations were determined for all fractions using the BCA reagent (Pierce) and bovine serum albumin as a standard.
Mitochondria were prepared by enriching the P2 fraction on a Nycodenz gradient in the 25-30% zone (34). A detergent-based fractionation method was used to prepare three mitochondrial subfractions. Extraction of mitochondria with 0.1% digitonin provided an outer membrane ϩ inner membrane space fraction followed by 1% Nonidet P-40 extraction to yield an inner membrane fraction and a pellet fraction representing the remaining mitochondrial proteins. Each fraction was then subjected to one-dimensional SDS-PAGE on a 4 -12% acrylamide gradient gel. Individual lanes were then sliced into 14 -18 bands, and each was digested with trypsin, extracted, and concentrated. The peptides obtained from each band were then analyzed by LC-MS as described below.
Preparation of Protein Digests-An aliquot of each fraction, P1, P2, and S2, containing 0.5 mg of protein was diluted with an equal volume of 6 M guanidine hydrochloride in 25 mM Tris, pH 7.5. To these solubilized fractions was added dithiothreitol to a concentration of 25 mM. After incubation at room temperature for 1 h, iodoacetamide (prepared as a 0.5 M stock solution in acetonitrile) was added to a final concentration of 55 mM, and the solutions were incubated for 1 h at room temperature. The reduced and alkylated fractions were then dialyzed (10,000 molecular weight cutoff membrane) against 50 mM ammonium bicarbonate solution at 4°C for 15 h. Fractions were digested with 0.03 mg/ml trypsin (L-1-tosylamido-2-phenylethyl chloromethyl ketone-treated) for 6 h at 37°C, and then another aliquot was added to a final concentration of 0.045 mg/ml. Digestion continued at 37°C for an additional 12-15 h. Reactions were stopped by adding glacial acetic acid to a final volume of 2%. Samples were then applied to C 18 spin columns (Nest group) that were preequilibrated with 0.1% (v/v) trifluoroacetic acid. After sample application, the columns were washed with 3 ϫ 0.6 ml of 0.1% trifluoroacetic acid. Peptides were eluted by application of 2 ϫ 0.3 ml of 60% (v/v) acetonitrile in 0.1% trifluoroacetic acid and 1 ϫ 0.3 ml of 80% (v/v) acetonitrile in 0.1% trifluoroacetic acid. The combined eluants were evaporated to dryness in a SpeedVac, and the residue was redissolved in 50 l of 10% (v/v) acetonitrile in 0.1% (v/v) formic acid. A 1-2-l aliquot of each redissolved eluate was diluted to 100 l, and the absorbance at 280 nm was recorded. This value was used to normalize the amount of each eluant to be analyzed by two-dimensional LC-MS.
Ion Trap LC-MS-Two-dimensional chromatography coupled to nanospray mass spectrometry was done in the following manner. Samples were applied to a strong cation exchange column (PolyLC, 0.3 ϫ 50 mm) at 20 l/min. Unbound peptides were trapped by a C 18 trapping column (0.3 ϫ 5 mm, Zorbax) in line with the ion exchange column. The trapping column was then switched in line with the analysis column (Zorbax C 18 300SB, 0.075 ϫ 150 mm) connected to an Agilent XCT 1100 mass spectrometer fitted with a nanospray interface eluted at a flow rate of 0.3 l/min. Peptides were eluted with a gradient of 5-65% (v/v) acetonitrile in 0.1% (v/v) formic acid over a period of 80 min. Mass spectra were collected at 26,000 m/z/s. During each spectral acquisition cycle, four to five MS/MS acquisitions were done on selected peaks in the elution spectrum. Dynamic exclusion was implemented whereby an m/z signal was excluded for 0.8 min after an analysis within the instrument data acquisition cycle. In the two-dimensional LC-MS experiments, the reversed phase analysis was repeated upon elution of the ion exchange column with steps of 25, 50, 75, 100, and 500 mM ammonium acetate. Samples from gel electrophoresis were analyzed in a similar fashion without using the ion exchange prefractionation steps. Peptides were chromatographed on a Zorbax C 18 SB300 column (0.075 ϫ 50 mm) eluted with a gradient of 25-65% acetonitrile (0.1% (v/v) formic acid) over 45 min.
Data Processing and Semiquantitative Analysis-Two-dimensional chromatography-mass spectrometry data were submitted to the Spectrum Mill program (Agilent Technologies) for processing and database searching. One potential problem with MS/MS data obtained from ion trap instruments are false positives resulting from database searching with poor quality mass spectral data. Spectrum Mill merged spectra of peaks with the same m/z and rejected spectra that did not have an MS/MS fragmentation pattern that included a string of three amino acid residues in the peptide fragmentation profile. Filtered MS/MS spectra were used to search the Swiss-Prot protein data base. A positive peptide hit required a minimum score of 6 -8 and had at least 70% of the 25 largest MS/MS peaks found in the predicted MS/MS data of a peptide identified in the database.
A semiquantitative analysis of protein profile data was facilitated by using a standard amount of protein extract for each analysis and the adjustment of the amount analyzed by LC-MS to maximize the number of proteins identified from each sample. The mean peptide intensity data were obtained from each positive hit in database scoring. These values were summed and then divided by the number of peptide hits to obtain an average intensity for the protein. Individual protein scores were then divided by the average overall proteins identified to give the relative expression values shown in Supplemental Table 1. Usually the same sets of peptides were identified for a given protein. If only a single peptide was identified from a given protein, these were included as hits based upon manual inspection of the quality of the MS/MS spectra and a maximum deviation from the identified peptide MH ϩ of Ϯ1.9 mass units. Thus, for both the commonly detected ϩ2 and ϩ3 charge peptides, the deviation is less than a mass unit. To estimate the variation in peptide signal intensities obtained from multiple experiments, the digests of soluble fractions of G93A-SOD1 mouse spinal cords and soluble fractions from non-Tg mouse spinal cords were run twice. A list containing ϳ100 normalized (signal/average) intensities of the most abundant proteins (two or more peptides) was compiled and compared between runs. The average variation in signal intensity for the same protein identified in two separate runs was less than 1.8 -2-fold. Supplemental Table 1 contains the hit lists and the normalized expression values for G93A-SOD1 and wt SOD1 mice.
Informatics-assisted Data Analysis-For gene ontology analysis normalized ratio data were transformed to single values of Ϫ1, 0, or 1 corresponding to underrepresented proteins, no change, or overrepresented proteins. We filtered the data to obtain proteins that exhibited at least 1.8 -2-fold changes in normalized ratio. This was based upon the variation observed in multiple runs of the same sample. Thus, the "changed protein" data imported to GoMiner analysis was "flattened" to be compatible with GoMiner. We generated microarray data from control and G93A-SOD1 mice (75-110 days) on the Affymetrix MG_U74A (version 2) chip starting from 10 g of total mRNA. mRNA was processed according to standard Affymetrix protocols, and normalized (.cel) data files were generated for subsequent analysis. We also used expression data for G93A-SOD1 and wt mice obtained from the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) database (35,36). These datasets were from 70-day-old presymptomatic SOD1 mice GSM75505, -6, and -7 and controls GSM75486 and -7. These data used the Affymetrix mouse 430A (version 2.2) chip. Microarray data were analyzed with BRB ArrayTools developed by Dr. Richard Simon and Amy Peng Lam (Biometric Research Branch, NCI, National Institutes of Health) within Microsoft Excel.
Single channel values were scaled and classified to disease or control groups. Gene ontology analysis on the mRNA expression data was done with the plug-in module in BRB ArrayTools and using GoMiner (37). In the GoMiner analysis, expression data were filtered to include only genes that changed by greater than 1.5-fold and were assigned Ϫ1 or 1 for GoMiner analysis. The new microarray data presented in this publication have been deposited in the NCBI GEO (www.ncbi.nlm.nih.gov/geo) and are accessible through GEO Series accession number GSE4390.
Protein Expression by Western Blot-Western blots were performed on the supernatant and pellet fractions of spinal cords extracted as done for protein profiling except that the low speed centrifugation was omitted. Polyacrylamide gradient (4 -20%) SDS gels were loaded with 25 g of soluble or particulate protein per lane. Gels were blotted to PVDF membranes and blocked with 1% nonfat dry milk in Tris-buffered saline containing 0.1% Tween 20 overnight at 4°C. Blots were probed with rabbit polyclonal antibodies to RabGDI (Zymed Laboratories Inc., 1:2000) and synapsin II (QED Bioscience, 1:5000). After incubation for 1 h at room temperature, membranes were washed and then incubated with bovine anti-rabbit horseradish peroxidase-conjugated second antibody (1:5000) for 1 h at room temperature. Blots were developed with ECLϩ reagent (Amersham Biosciences) and visualized by autoradiography.

Global Protein Expression Analyses-
In an attempt to optimize coverage of cellular proteins within the spinal cord we used a fractionation technique based upon differential centrifugation. Thus, two particulate and one soluble fractions were generated from a spinal cord extract. These fractionations were not designed to preserve organelles but rather to disrupt them and release proteins. Table I summarizes the recovery of protein in these fractions from mouse spinal cord. Tg mice containing the G93A-SOD1 mutation were from end stage animals (120 days), whereas the others were from wildtype SOD1 Tg and normal animals of similar age. We consistently found that the total extractable protein from mutant SOD1 mice was 30 -40% lower than from nontransgenic or wild-type SOD1 animals. Because the motor neurons contribute only 10 -15% of the total protein, this decrease probably results from loss of both neurons and associated support cells from the G93A-SOD1 animals and/or reduced total protein expression within the mixture of all cell types found in spinal cord.
This relatively simple fractionation method allowed the identification of 10-fold more proteins from mouse spinal cord than could be obtained without prior fractionation. Several proteins were found in multiple fractions, but the majority was represented in one pool (Fig. 1). Thus, two-dimensional peptide chromatography provided a limited profile representing several groups of proteins expressed in the spinal cord. Realizing that the methodology carries some intrinsic bias toward the peptides that are identified by mass spectrometry following multiple chromatographic steps, we focused efforts into identifying changes in groups of proteins representing various biological processes using a gene ontology database mining program.
Informatics-assisted Analysis of Protein Expression-In our analysis, coverage of the proteome using multiple step chromatography coupled to mass spectrometry was limited to less than 1000 proteins in a single experiment. Although a recent study provided identification of more than 3000 proteins in multiplex LC-MS experiments (38), the large number of chromatographic steps required for these analyses makes such schemes impractical for multiple sample analysis. Therefore, we engaged a simplified qualitative data reduction method coupled to informatics to analyze the LC-MS-derived protein profiles from spinal cords of transgenic and normal mice. We reasoned that changes in protein composition trends (up or down) may be best revealed by clustering groups of proteins associated with categories representing cellular structure or function. Thus, the goal of the analysis was to discover common trends in groups rather than the largest changes of individual proteins. In this regard, classification of proteins or genes based upon such criteria is known as gene ontology (GO) (39), and databases have been built utilizing such classification schemes (37,40). GO mapping is common in analysis of microarray-based mRNA expression data (41), but few protein profiling studies have utilized GO analysis (42)(43)(44)(45), and none of these applied GO mapping to a comparative study of transgenic animal tissues. Thus, a secondary goal of these studies was to validate the use of GO mapping for comparing proteomic and genomic data.
To investigate the utility of GO mapping, proteomic identification datasets were prepared by normalizing the MS intensity data for identified peptides (see "Experimental Procedures"). A gene list was prepared by merging all of the positive identifications from transgenic SOD1 mice expressing wildtype and mutant SOD1 proteins. The query lists were prepared from the individual transgenic mouse datasets that included 400 -540 proteins based upon the underlying peptide score (see "Experimental Procedures"). We used the ability of the gene ontology program, GoMiner, to group related proteins and determine the potential significance of changes in groups of proteins (exhibiting high or low abundance) representing biological process and functions. Because we used as base set all of the proteins identified in our mass spectrometry-based proteomics, those absent in one dataset are also included in the analysis as "unchanged." The method (Fisher two-tailed exact test) used by GoMiner in assessing the potential significance of changes in GO categories is applicable to categories with few representatives as well as many because the significance factor is based upon whether the proportion of changed proteins (regardless of direction) is greater than would be expected by random chance (37). Thus, the GO categories with low p values contain groups of related proteins that are either enriched or depleted. For the protein data we used a cutoff p value of less than 0.05; there is less than a 5% chance that the proteins in the category arise from a random distribution. Overlapping GO groups are represented with the categories/subcategories exhibiting the most significant changes within a single Tg mouse dataset. We did not consider any categories with fewer than three proteins represented regardless of the p value calculated by GoMiner. These data are summarized in Tables II and III. In the G93A-SOD1 transgenic mice, significantly changed protein categories were found in enhanced representation of  selected protein kinases, antioxidant proteins, vacuole-associated proteins, and protein kinase inhibitors (Table III). Lower relative expression was found among mitochondrial proteins, especially membrane-associated species. The opposite trend was observed in the data from wt SOD1 transgenic mice (Table II). Thus, the common link in both SOD1 Tg animals is mitochondria with low representation in the G93A spinal cord compared with enhanced representation in the wild-type SOD1 Tg mouse spinal cord. Other GO categories flagged (as potentially significant) in the analysis of mouse spinal cord data may be correlated with changes in proteins that may not have been detected in both G93A and wt SOD1 mice. We attribute this finding to the intrinsic variability in proteome expression data. Such findings have been reported even in the quantitative analyses of samples using isotopic tagging methodology (46). Mitochondria, Oxidative Stress, and Metabolism-The mitochondrial proteins whose expression varied in the G93A Tg mouse spinal cords are shown in Table IV. A limited number of proteins were identified in common with both the G93A-SOD1 and wt SOD1 mouse spinal cord so that other identified proteins in G93A-SOD1 mice are listed as high or low average abundance. Changes in the relative ratio of some proteins were found to be comparable to those found in a proteomic study of a motor neuron cell line transfected with G93A-SOD1 protein (47). The agreement between these studies supports our method of protein profiling. Inspection of the members in the ATP-driven transporters category revealed that some of these proteins are mitochondrial membrane-bound ATP synthase subunits (Table IV). These are underrepresented in the G93A-SOD1 mouse spinal cord. On the other hand, the elevated antioxidant protein in the G93A-SOD1 mouse mitochondria (Table IV), peroxiredoxin 5 (PRDX5), is complemented by overall enhanced antioxidant protein levels in the G93A-SOD1 mouse spinal cord. This GO grouping includes another peroxiredoxin (PRDX6), mouse SOD1, and a glutathione S-transferase (maleylacetoacetate isomerase). Similarly changes in mitochondrial carbohydrate-metabolizing enzymes such as aldolase were enriched in another GO category (glucose metabolism) containing enzymes such as enolase, phosphoglycerate mutase, and others. Overall these differences are consistent with decreased ATP production in G93A-SOD1 mouse spinal cord coupled to a paradoxical increase in metabolic activity. Although several reports indicate that both mutant and wt SOD1 enter the mitochondria (see references above), our proteomic results suggest that the presence of mutant SOD1 may be affecting components of the inner membrane (AT-Pases) and matrix. Therefore, we fractionated spinal cord mitochondria from transgenic mice (wt and G93A-SOD1) using a differential detergent extraction method (48) and then identified the component proteins using LC-MS. As shown in Table V wt SOD1 predominantly associates with readily extractable neutral detergent fractions that represent outer membrane and inner membrane space. On the other hand, mutant G93A-SOD1 is distributed among all of the fractions and includes a portion of the protein that is only extractable by the denaturing detergent sodium dodecyl sulfate. Also found in the same fractions are heat shock proteins of the HSP70 family. Note, however, that the cytoplasmic HSP70 and mitochondrial HSP70 are found in the same fractions as G93A-SOD1, whereas only the mitochondrial HSP70 was detected in fractions containing wt SOD1. Therefore, consistent with previous reports of the localization of cytoplasmic HSP70 with G93A-SOD1 (49), we found that both proteins are distributed in multiple mitochondrial fractions, including materials that are extractable only with harsh detergent conditions.
Protein Kinase Regulation and Apoptosis-GO analysis of the proteomic data indicates that G93A-SOD1 mouse spinal cord has an enrichment of protein kinase regulatory activities (Table III). The MAP kinase kinase kinase category includes protein kinase MAP kinase I (ERK1) and the regulatory subunit of protein phosphatase 2A as overrepresented, but globally there is an underrepresentation of protein kinases in the expression profile. On the other hand, two GO categories contain enriched kinase regulatory proteins that are associated with MAP kinase and protein kinase C (PKC) signaling. The protein kinase regulators include those directed at protein kinase C (14-3-3 proteins) and cyclin-dependent protein kinase inhibitor 1B (p27 kip1 ). Also one of the proteins in this group is a phosphatase associated with MAP kinases (MAPK8IP1). Although not a direct protein kinase regulator, S100␤, a Ca 2ϩ -and Zn 2ϩ -binding protein, inhibits the phosphorylation of p53 by PKC (50). This phosphorylation prevents p53 oligomerization and proapoptotic function. S100␤ is expressed predominantly by glial cells and can be either a trophic or neurotoxic factor for neurons (51). Low levels of S100␤ are neurotrophic, whereas higher levels are neurotoxic (51). Collectively these data implicate a change in protein kinase effector signaling involving stress response in the nonneuronal cells in the spinal cord. Although previous reports indicated an increase in PKC activities associated with G93A-SOD1 mouse spinal cord (52), an associated increase in proteins that inhibit phosphorylation of selected PKC substrates is one of the new findings in the current study. Two of these proteins (S100␤ and 14-3-3 ) were also found in another GO category, the regulation of apoptosis. The over/under groups of proteins in this category are split between positive and negative regulators of apoptosis in the G93A mouse spinal cord. Three potentially important proteins in this group include Nogo (inhibitor of neurite outgrowth), apoptosis inhibitor mIAP-2, and RIP-kinase2 (RIP2). Differential expression of various isoforms of Nogo have been observed in Tg mouse models (G85R-SOD1) and human ALS postmortem tissue (53). Apoptosis inhibitor mIAP-2 is a member of the BIR (baculoviral IAP repeat) family that binds to specific tumor necrosis factor receptor (TNF-R) complexes (54) and induces anti-apoptotic activities through NFB signaling (55). Finally RIP-kinase2 is involved in inflammatory pathways that are mediated by NFB activation (56). Like the IAP proteins, RIP2 is recruited to signaling receptor complexes such as TNF-R1 (57). Overexpression of RIP2 induces apoptosis or potentiates caspase-8-mediated cell death (56). Although the activation of proapoptotic proteins has been reported previously in G93A-SOD1 mouse spinal cord (24,58), our GO analysis of the protein profile indicates that there is altered expression of proteins involved in both proapoptotic and antiapoptotic pathways.
Membrane Organelles and GTPase Regulatory Proteins-The protein profile obtained from the current study indicates that certain lysosomal proteins are enriched in G93A-SOD1 mouse spinal cord (Table III). These include a lysosomal protease (cathepsin D), a proton-pumping ATPase (ATP6V1H), antioxidant protein peroxiredoxin 6, and an low density lipoprotein receptor-related protein (LPR2_human).
Under the GTPase regulator category, only one protein from this group, the Rab guanine nucleotide dissociation inhibitor, RabGDI, was detected in both wt SOD1 and G93A transgenic mice. To confirm enhanced expression of this protein in G93A-SOD1 mouse spinal cord, we performed Western blots of soluble and particulate fractions of tissues from G93A-SOD1 mice compared with nontransgenic mice. Shown in Fig.  2 are the results of these experiments. RabGDI was modestly increased in G93A-SOD1 mouse spinal cord at both 120 days (paralysis stage) and 60 days (presymptomatic), but it was not as evident at 60 days (presymptomatic) compared with nontransgenic mice. For comparison, in the same sample extracts we found that synapsin II was underexpressed in G93A-SOD1 mice compared with nontransgenic mice at both 60 and 120 days.

Analysis of Protein and Gene Expression in SOD1 Mice
GO Analysis of mRNA Expression in G93A-SOD1 and wt Spinal Cord-The GO analysis is one of the few "omic" analyses that can provide meaningful comparisons between protein and gene expression data. We used two independent sets of microarray data. One was from our own microarray analysis of mouse spinal cord mRNA at 70 -110 days (Table  VII), whereas the other was from the NCBI GEO database of 10-week mouse spinal cords (Table VI). In both datasets, the comparison is with age-matched nontransgenic mouse controls. Between the microarray datasets, the overlap in the data centers on lysosomal proteins, ion-binding proteins, lipid transport/metabolism, and selected transcription factors. In Table VI, cation-binding proteins, membrane-associated AT-Pases, proteins associated with stress-activated protein kinase activities, second messenger signaling, and gene transcription are also overrepresented. In Table VII microarray data revealed changes in G93A-SOD1 mouse spinal cord at a later stage in disease progression and more "changed" categories in the GO analysis. These include lipid metabolism, complement activation, lysosomal activity, and ion channel activities. The larger number of proteins (genes) represented in the GO analysis of the microarray data provides added sensitivity to the detection of potential protein changes within these selected categories. Concordance of the protein and microarray data within these categories is notable in the ATPdriven copper-transporting ATPase (ATP7A/B), phospholipidtransporting ATPase (ATP11B/8A2), and vacuolar ATPase (ATP6V1). Similarly lysosomal proteins such as cathepsin proteases are enriched in both the proteomic and mRNA expression data (cathepsins D and A, respectively). Changes in selected members of the MAP kinase cascade are also found in the microarray data concordant with the protein data (Table  III). On the other hand, the differences in mitochondrial proteins (Table IV) are not found in GO analysis of the microarray data from either study. This suggests that this disconcordance between mRNA and protein expression data is related to changes of protein turnover, possibly due to oxidative damage, or regulation of mRNA translation. Proteomics is destined to become an integral part of investigations into the causes and cures of human neurodegenerative diseases such as Alzheimer disease, Parkinson disease, and ALS. In the current work, we have shown that in the transgenic mouse model expression of mutant SOD1 genes causes distinct changes in the protein profile detected in spinal cord extracts. These include alterations in mitochondrial proteins and proteins involved in cellular signaling and lysosomal function. The elucidation of these differences was facilitated by using an informatics-based approach. The flattening of expression ratios required for GO analysis makes the semiquantitative data obtained in the current studies quite useful for a comparative approach. Thus, this strategy alleviates one of the limitations of mass spectrometry-based proteomics because comparisons are based upon the GO analysis that utilizes groups of proteins rather than specific differences of individual proteins.
Most of the specific differences in mitochondrial protein profile are consistent with observed changes in mitochondrial enzymatic activities in G93A-SOD1 mouse spinal cord (59,60) and human ALS (16). For example, increased mRNA levels of thioredoxin peroxidases were detected in other studies of spinal cords of SOD1 transgenic mice (6) and in selected human ALS spinal cord tissues (61) consistent with our finding of increased levels of peroxiredoxin. The overlap of proteins differentially expressed in the NSC34 cell line (47) expressing mutant SOD1 compared with the mouse spinal cord in our studies indicates significant agreement in the direction (up/ down) of selected mitochondrial proteins (Table IV) and that the overexpression of mutant SOD1 protein is correlated with altered mitochondrial protein composition. The association of G93A-SOD1 and HSP70 proteins with mitochondrial fractions containing the ATP biosynthesis complexes suggests that part of the "gain of function" of mutant SOD1 is to target specific mitochondrial proteins. Because disease onset in transgenic mice is later in mice expressing a lower level of G93A-SOD1 (62), the accumulation of mutant SOD1 in the mitochondria may be viewed as an accelerant for disease progression.
Up-regulation of several protein kinase activities such as protein kinase C (30) and the stress-activated kinase p38 have been reported (63) in spinal cords of SOD1 mutant mice. Our results confirm these findings and extend the alteration in protein kinase-mediated signaling to include changes in protein kinase inhibitory proteins and selected stress-activated protein kinases in the MAP kinase family.
To place the GO results in a larger context, we investigated mapping of the GO category containing the largest number of changed genes from mouse spinal cords. This category, cell communication, has nearly 2000 genes in the microarray data and Ͼ180 proteins in the G93A-SOD1 proteomic data. GO analyses of both the microarray and proteomic data gave similar maps. Fig. 3 depicts the cell communication tree and several branches as mapped from the proteomic data. Red squares at each node indicate GO categories that are enriched in changed genes. The major groups include cell adhesion, signal transduction, and cell-cell signaling. A consistent finding in both the proteomic and microarray datasets was that in the cell-cell signaling branches, there was a convergence of changed proteins/genes within the neurotransmitter secretion groups (Fig. 3).
For example, we found that synapsin II levels are low in G93A-SOD1 mouse spinal cord at both end stage (120 days) and presymptomatic mice (60 days) (Fig. 2). Synapsin II is a synaptic vesicle-associated protein that functions to direct targeting and staging of neurosecretory vesicles that carry neurotransmitters such as glutamate, ␥-aminobutyric acid, and acetylcholine (64). Both sets of microarray data exhibited decreased expression of synaptotagmins (SYT1 and SYT4) that are necessary for Ca 2ϩ -dependent synaptic vesicle fusion (65,66). From the literature, the only other indication of altered synaptic vesicle proteins comes from a postmortem immunochemical analysis of the anterior horn from ALS subjects (67). In this study, an altered distribution of immunoreactivity for synapsin and synaptophysin was observed in ALS tissue compared with control (67). These data suggest that deficiencies in synaptic vesicle-associated proteins may be one of the early indicators of motor neuron dysfunction that directly impacts upon the secretion of neurotransmitters.
Another representation of the potential influence of SOD1 on other cellular processes is shown in the interactome map (Fig. 4). Most of the branched nodes on this map represent the direct or interact action of SOD1 on ubiquitin ligases (68), oxidative inhibition of calcineurin (69), ciliary neurotrophic factor expression (70), alsin (71), and glutathione peroxidase (72). Based on the results presented here, we may add additional nodes for mitochondrial and cellular communication proteins not previously associated with the expression of mutant SOD1. Within the latter group, the early alteration of proteins associated with neurotransmission may be the underlying factor that makes motor neurons especially susceptible to the stress that mutant SOD1 inflicts upon mitochondrial function.
In summary, the integration of proteomic and genomic data using a higher level gene ontology analysis allowed the delineation of common and unique disease pathways in mouse models of ALS and their relevance to human disease. FIG. 4. Interactome map of Cu,Znsuperoxide dismutase. The circles indicate the primary interactions (direct and indirect points for SOD1 (copper chaperones, p53, ciliary neurotrophic factor (CNTF), etc.). Note that many of these interactions involve only mutant SOD1 and not wt SOD1. The branches extending from the primary points show other genes that have either direct or indirect interaction with those species. New nodes for mutant SOD1 from the results of the current study are shown with the name of the affected cellular system such as neurotransmitter regulation and mitochondrial ATP generation (from Fig. 3).