Transcriptional profiling identifies extensive downregulation of extracellular matrix gene expression in sarcopenic rat soleus muscle

J. Scott Pattison1, Lillian C. Folk2, Richard W. Madsen3, Thomas E. Childs1 and Frank W. Booth1

1 Departments of Biomedical Sciences and of Medical Pharmacology and Physiology, and the Dalton Cardiovascular Institute
2 Department of Veterinary Pathobiology, University of Missouri at Columbia, Columbia, Missouri 65211
3 Department of Statistics, University of Missouri at Columbia, Columbia, Missouri 65211


    ABSTRACT
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 DISCLOSURES
 References
 
The direction of change in skeletal muscle mass differs between young and old individuals, growing in young animals and atrophying in old animals. The purpose of the experiment was to develop a statistically conservative list of genes whose expression differed significantly between young growing and old atrophying (sarcopenic) skeletal muscles, which may be contributing to physical frailty. Gene expression levels of >24,000 transcripts were determined in soleus muscle samples from young (3–4 mo) and old (30–31 mo) rats. Age-related differences were determined using a Student’s t-test ({alpha} of 0.05) with a Bonferroni adjustment, which yielded 682 probe sets that differed significantly between young (n = 25) and old (n = 20) animals. Of 347 total decreases in aged/sarcopenic muscle relative to young muscles, 199 were functionally identified; the major theme being that 24% had a biological role in the extracellular matrix and cell adhesion. Three themes were observed from 213 of the 335 total increases in sarcopenic muscles whose functions were documented in databases: 1) 14% are involved in immune response; 2) 9% play a role in proteolysis, ubiquitin-dependent degradation, and proteasome components; and 3) 7% act in stress/antioxidant responses. A total of 270 differentially expressed genes and ESTs had unknown/unclear functions. By decreasing the sample sizes of young and old animals from 25 x 20 to 15 x 15, 10 x 10, and 5 x 5 observations, we observed 682, 331, 73, and 3 statistically different mRNAs, respectively. Use of large sample size and a Bonferroni multiple testing adjustment in combination yielded increased statistical power, while protecting against false positives. Finally, multiple mRNAs that differ between young growing and old, sarcopenic muscles were identified and may highlight new candidate mechanisms that regulate skeletal muscle mass during sarcopenia.

microarray; aged; atrophy; mRNA; statistics


    INTRODUCTION
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 DISCLOSURES
 References
 
THE CONTRIBUTION of skeletal muscle strength and mass to health is under-recognized, where losses in association with advanced age result in an increased incidence of death (32). The peak mass of male skeletal muscle occurs by the age of 25 yr in humans (30). Thereafter, ~10% of muscle mass is lost by the age of 50 yr, and another 30% is lost by 80 yr of age (30). Thus skeletal muscle transitions from a growth to an adult "steady state" to an atrophy phase with increasing age in both humans and rats (31). Because the molecular causes of sarcopenic skeletal muscle are poorly defined, the current experiment was designed to tease out contrasting mRNA levels between growing muscle in young rats and aged-atrophying (sarcopenic) muscle. The rationale for this study was to screen for gene targets in order to develop scientifically based strategies to induce growth in sarcopenic muscle. The genes identified with differential expression for muscle growth that are present during normal adolescent muscle growth, but missing in old sarcopenic muscles, or vice versa, i.e., inhibitory factors that may be present in sarcopenia, but low in growth, were targeted as candidates for regulating muscle mass. The experimental strategy also included the decision to focus on a large sample size foregoing an adult group whose muscle mass is in a steady state. Greater importance was placed on a large sample size to obtain a larger number of significant differences with fewer false positives. Thus the aim of the design was to identify factors promoting growth, rather than maturation, as potentially effective clinical interventions are needed when muscle mass decreases to the level associated with increased mortality. Use of microarrays allows a global, unbiased determination of mRNA expression and could provide an insight into the status of gene expression in skeletal muscle growth and sarcopenia. Previous studies using Affymetrix microarrays to compare young and old skeletal muscles reported 113, 70, and 449 differentially expressed mRNAs with independent observations of 3, 8, and 3, respectively (24, 27, 40). However, none of these reports provided valid statistical analyses to identify significant differences nor did any of these reports estimate the percentage of false positives in their lists of differentially expressed transcripts. Thus for the current experiment, it was reasoned that the use of larger sample sizes would allow greater statistical power so that conservative statistical adjustments could be made to minimize the presence of false positives. The purpose of the current experiment was to develop a statistically conservative list of genes whose expression differed significantly between young growing and old-sarcopenic muscles, which may be contributing to physical frailty. The hypothesis of the current study was that a subpopulation of growth factor mRNAs would be downregulated in old skeletal muscle, associated with old muscle no longer growing.


    MATERIALS AND METHODS
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 DISCLOSURES
 References
 

Animals.
Fischer 344 x Brown Norway Fl male rats obtained from the National Institute on Aging (Harlan Laboratories, Indianapolis, IN) and were killed at the ages of 3–4 mo (young, n = 25) and 30–31 mo (old, n = 20). These ages were selected on the basis of previous data (8). They received regular rat chow and water ad libitum, were housed 2–3 per cage, and maintained on a 12:12-h light/dark cycle. Prior to muscle extraction, animals were anesthetized with an intraperitoneal injection of a cocktail containing ketamine (49 mg/ml), xylazine (6.2 mg/ml), and acepromazine (2.0 mg/ml) at a concentration of 0.123 ml/100 mg body wt. Soleus muscles were excised, weighed, snap-frozen in liquid nitrogen, and subsequently powdered using a mortar and pestle cooled by liquid nitrogen. Both soleus muscles from a single rat formed one observation, where muscle RNA from a single animal was applied to an individual array. The University of Missouri Animal Care and Use Committee approved animal protocols.

Sample processing for microarray analysis.
Total RNA was isolated from an aliquot of muscle powder that was put directly into a TRIzol solution (Invitrogen) and homogenized on ice using a Polytron homogenizer (Kinematica) on setting 7 for three pulses of 15 s each. The total RNA was further purified using RNeasy columns (Qiagen). Methods for sample preparation are described in detail in the Affymetrix Expression Analysis Technical Manual (Santa Clara, CA) and are briefly described next. Ten micrograms of purified total RNA was put into the cDNA synthesis reactions with a T7-(dT)24 primer (100 pmol/µl). First- and second-strand cDNA syntheses were carried out using components of the Superscript Choice kit (Invitrogen) with all incubations done in a Mastercycler Gradient thermocycler (Eppendorf). The amount of resulting double-stranded cDNA was quantified using a PicoGreen kit (Molecular Probes). One microgram of cDNA was added to the in vitro transcription reaction utilizing biotinylated nucleotides provided in the BioArray HighYield RNA Transcript Labeling Kits (Enzo Diagnostics). The resulting cRNA was further purified using RNeasy columns (Qiagen). The purified biotinylated cRNA was then fragmented and subsequently hybridized to Affymetrix rat genome U34A, B, and C arrays and analyzed by fluorescent intensity scanning according to Affymetrix protocols (Affymetrix Expression Analysis Technical Manual). The hybridization and scanning of the arrays was performed in the University of Missouri DNA Core Facility (Columbia, MO).

U34 microarrays.
The U34 arrays were created in 1998. At that time, the array set was estimated to contain ~7,000 full-length/annotated genes based on UniGene Build 34, which were all located on the RG U34A microarray. Similarly, the remaining probe sets assayed for >17,000 expressed sequence tags (ESTs), predominantly located on the B and C arrays. However, since that time, many advances have been made to allow a considerable number of the ESTs to be identified. The human genome has been completed. The mouse genome is now available as a draft that is 95% complete. The rat genome project has begun and is projected to be finished before the year’s end. Now in 2002, many of the ESTs previously unassociated with known genes now are sufficiently homologous, such that "as of UniGene build 99 over 28,000 well substantiated genes exist" (Affymetrix technical datasheet rat230). Thus many of the ESTs available on the B and C arrays can now be substantiated as "true genes" based on their significant homologies to the known genes identified in the mouse and human genomes. The U34 array set assays ~24,000 genes and ESTs, represented by 26,379 probe sets. Thus some mRNAs are assayed by multiple probe sets. Also, some ESTs have since been identified as portions of full-length/known genes, where the ESTs are known to be part of same UniGene cluster as the full-length mRNAs, causing multiple probe sets to assay the abundance of a single mRNA.

GeneChip analysis.
Each probe set consisted of 16 perfectly matched (complementary) 25-mers, corresponding to different regions along the length of a transcript. Likewise, 16 mismatched pairs (containing a single mutated base) that do not perfectly complement an mRNA’s sequence were used as a measure of nonspecific background binding. The Microarray Suite 5.0 software (Affymetrix) was employed, which uses a one-sided Wilcoxon’s signed rank test to calculate a P value reflecting the significance of differences between the perfectly matched and mismatched probe pairs, based on their fluorescent intensities. The resulting P values were used as a qualitative assessment of the ability to detect a given transcript, where P values <0.04 were called "present", P values between 0.04–0.06 were called "marginal" and P values >0.06 were called "absent". Only probe sets that were "present or marginal" in >=60% of the samples for an experimental group were analyzed statistically, with 11,165 probe sets sufficiently detected in at least one of the two experimental groups, i.e., young or old soleus muscles. Approximately 20% of the 682 significantly different probe sets contained probe sets with more than one "absent" call within the 20–25 muscle samples composing an experimental group (see Supplemental Table 1, available at the Physiological Genomics web site).1 Approximately 4–9% of the 682 probe sets were consistently called absent in only one of the two experimental groups, implying turning "on" or "off" of mRNA expression. Microarray Suite 5.0 software (Affymetrix) uses statistically based algorithms to determine transcript abundance based on fluorescent intensities (termed "signal"). The "signal" for each probe set was calculated as the one-step bi-weight estimate of the combined differences of all of the probe pairs in the probe set. We used the calculated signal value for all subsequent statistical analyses. The fold changes of 347 probe sets that increased in the older group, relative to the young, were calculated as the mean old signal intensity/mean young signal intensity. The fold changes of the 335 probe sets that decreased in sarcopenic muscle were determined as the mean young signal intensity/mean old signal intensity. A fold change of a certain magnitude can be converted to a percentage decrease by [(µY - µO)/µY] x 100%. Microarray data analyses have been criticized as being "quite elusive about measurement reproducibility" (9). However, Bakay et al. (2) have reported that experimental error among Affymetrix microarrays is not a significant source of unwanted variability in expression profiling experiments (r2 = 0.979). In our hands, duplicate arrays also had small, unwanted inter-array variability (r2 = 0.981). The samples included in the current analyses were not run as replicates (i.e., assayed only once by single array). In addition, a recent publication of a single human patient, where RNA was prepared from two distinct breast tumors and placed on duplicate U95A GeneChips (four arrays total) found a very low degree of experimental variability between microarrays (r2 = 0.995) and between the two tumors (r2 = 0.987) (38).


View this table:
[in this window]
[in a new window]
 
Table 1. Comparison of microarray and RT-PCR analyses on the same young and old muscle samples

 
Statistical methods.
The Shapiro-Wilk test for normality was employed, to determine the normality distribution of the current data set. Deviations from normality were observed ~15% of the time, suggesting it was more appropriate to use parametric statistics. Thus an unequal variance Student’s t-test ({alpha} = 0.05) was employed to compare the signal values of young and old soleus groups. Furthermore, a Bonferroni adjustment was applied to correct for the multiple Student’s t-tests performed on 11,165 probe sets that had been detected as present, i.e., having sufficient hybridization as identified by the Affymetrix Microarray Suite 5.0 software. Power analysis was done to determine what statistical power could be obtained with sample sizes of 20. As rats were killed in groups of five over a 40-day period, a one-way ANOVA was performed to determine whether any one group differed from the other groups.

Groups of young and old rats were analyzed separately. A recursive analysis was performed on 20 data sets of 5 x 5, 10 x 10, and 15 x 15 that had been randomly selected from the 25 x 20 samples. An additional multiple-testing adjustment called the false discovery rate (FDR) was employed to control the chance of making a type I error. The FDR is a simple, sequential Bonferroni-type procedure that has been proven to control the FDR for independent-test statistics with a substantial gain in power over the Bonferroni technique at the expense of increasing acceptance false positives (4). Thus, to maximize the number of significant findings, an FDR-adjusted P < 0.05 was applied in the supplemental data to correct for the multiple t-tests performed.

Fisher’s exact test was employed to determine the cumulative hypergeometric probability distribution, which examines the chance probability of observing a certain number of genes in a given functional category (1). Use of Fisher’s exact test allowed for the determination of statistical significance within a given functional category (35).

Database searching.
The target sequences for the significantly different probe sets were analyzed with nucleotide BLAST analysis to identify known genes and to determine significant gene homologies with other species (http://www.ncbi.nlm.nih.gov/BLAST/). The target sequence is the region of a given gene or EST which was probed by the RG-U34 arrays. Further information about a given sequence and its homologs/orthologs was obtained from the LocusLink, HomoloGene, OMIM, Mouse Genome Informatics, Rat Genome Database, NetAffx, and Proteome BioKnowledge Library databases (http://www.ncbi.nlm.nih.gov/LocusLink/, http://www.ncbi.nlm.nih.gov/HomoloGene/, http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=OMIM, http://www.informatics.jax.org/, http://rgd.mcw.edu/, http://www.affymetrix.com/analysis/index.affx, http://www.incyte.com/control/researchproducts/insilico/proteome). A gene’s biological processes and molecular functions were determined based on the defined gene ontologies given in the aforementioned databases. Raw data from the current experiment are available at the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/), in a MIAME-compliant format ("minimum information about microarray experiments"). The accession numbers to retrieve these data are listed in Supplemental Table 2.

Real-time PCR sample preparation.
Total RNA was DNase treated using RNase-free DNase set (Qiagen), which was then purified over an RNeasy column (Qiagen). The absence of contaminating DNA was confirmed by PCR and TaqMan (real-time PCR). Two-step RT-PCR was performed according to the TaqMan Universal PCR Master Mix protocol [Applied Biosystems (ABI)], utilizing TaqMan Reverse Transcription Reagents with random hexamers (ABI) to reverse-transcribe RNA. Two micrograms of RNA was used to synthesize cDNA. Probe/primer combinations were designed using PrimerExpress 2.0. (ABI) (Primer/probe sequences are in Supplemental Table 3). All 5' nuclease assays consisted of reactions containing 25 ng of cDNA, 250 nM minor groove bender (MGB) probe, 900 nM primers and TaqMan Universal PCR Master Mix (ABI). All samples were assayed in a 25-µl total volume in triplicate with an ABI Prism 7000 Sequence Detection System. If a triplicate contained a range of >0.4 cycle time (CT), it was re-assayed. All target samples were analyzed relative to 18S rRNA. 18S was assayed using TaqMan Ribosomal RNA control reagents (ABI).

Real-time PCR analysis.
Differences in gene expression were calculated using relative quantitation to 18S rRNA, via the comparative CT method, according the User Bulletin no. 2 ABI PRISM 7700 Sequence Detection System. 18S rRNA was confirmed as an appropriate normalizer by comparing the differences in raw CT values. 18S rRNA expression did not differ with age (P = 0.55). Relative efficiency plots were run to validate use of the {Delta}{Delta}CT method, where all slopes were <0.1. The differences in {Delta}CT values for young and old were analyzed with the Student’s t-test with P < 0.05 set as significant. Data are expressed as the calculated-fold differences between young and old, with actual P values. Assay order for real-time PCR began at a twofold difference from microarray analysis and then progressively lowered the fold differences (starting with ~2.0 for atrogin, then ~1.8 for IGFBP3, then ~1.7 for homer2, then ~1.6 for elfin, then ~1.4 for prolyl-4-hydroxylase, and finally ~1.4 for agrin).


    RESULTS
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 DISCLOSURES
 References
 
Fischer 344 x Brown Norway F1 rats have previously been shown to be growing through 3–4 mo of age, which corresponds to the ages of the young rats used in the current experiment (36). Rats were killed in sets of n = 5 over a 30-day period on days 0, 3, 6, 10, and 30. The soleus muscle masses of rats killed on days 0 and 3 were compared with those killed on day 30. There was a trend toward increasing soleus muscle mass, with a 15% increase (0.129 ± 0.005 to 0.149 ± 0.008) in the young rats over the course of the experiment (P = 0.062). Conversely, over a 30-day period, the soleus muscle mass decreased 10% (0.174 ± 0.003 to 0.156 ± 0.003) in the old rats (P = 0.002). Using an {alpha} = 0.05, with a Bonferroni adjustment on 11,165 statistical tests, we obtained P <= 4.478 x 10-6 for significant differences. We found that 682 probe sets had significant differences between young growing and old, sarcopenic groups. Of the identified probe sets, 347 decreased and 335 increased in the old relative to the young (Supplemental Tables 4 and 5). Statistical power analysis showed that even with a significance level as low as 5 x 10-6, a sample size of 20 independent observations per group was sufficient (power 80%) to detect a 1.4-fold difference, assuming a 20% coefficient of variation, i.e., assuming {sigma}/µ = 0.20.

As an exercise to demonstrate the importance of sample size in finding significant differences, subsamples of data were examined. Subsets of 5 x 5, 10 x 10, and 15 x 15 that had been randomly selected from the 25 x 20 sample were evaluated with a recursive analysis, where each sample size was randomly assembled and analyzed 20 different times to better estimate the expected number of significant changes that occur with the smaller sample sizes (Supplemental Table 6). For one comparison, 15 samples were selected with a random number generator from both the young and old groups, and the same procedures as described above were used to determine how many probe sets would have significant differences. In this case, 331 probe sets were identified as significantly different compared with the 682 found using the full set of data (20 x 25). When sets of 10 samples were randomly selected, the number of probe sets showing significant differences dropped to 73, and the number dropped to only 3 significant probe set when sets of 5 samples were randomly selected (Supplemental Table 6). Note, that if other subsamples were randomly selected for the smaller sized groups, e.g., 5, 10, and 15, then the number of observed probe sets with significant differences would not be exactly the same, but would be of similar order of magnitude. Furthermore, there is a decreased probability of accepting lower fold changes as significant when sample size is decreased (Supplemental Table 7A). Similarly, there is a decreased probability of accepting larger variations at a given fold change as significant when the sample size is decreased (Supplemental Table 7B).

In the comparison of 20 x 25, 412 probe sets were identified which corresponded to known genes or had homologies to known genes that were significantly altered between young growing and old sarcopenic muscles. Some of the known genes could be assembled into functional groups with common themes. For example, 20 (9%) of the 213 identifiable genes that were increased in the old soleus muscle had functional commonalities to transcripts with roles in protein degradative processes including proteolysis and ubiquitin-based degradation; 29 (14%) to immune functions; and 15 (7%) to stress/antioxidant (Fig. 1). Of 347 total decreases in mRNA in the old soleus muscle, 199 were functionally identified, with the major theme being that 48 (24%) had a biological role in the extracellular matrix (ECM) and cell adhesion (Fig. 1). Fisher’s exact test was used to determine whether the large absolute number of significant collagen genes merely reflected a large number of probe sets for collagen genes on the microarray or reflected a true shift in collagen expression. The number of collagen, ECM, and cell adhesion probe sets changed were significantly greater (P < 0.0001) than would have been expected for any functional group by chance (Supplemental Table 8). The use of Fisher’s Exact test on the current significant data extend the previous observations of increases in proteolytic genes, as well as genes related to stress, defense, and immune responses (24, 27, 40), to show that the aforementioned functional categories were disproportionately regulated in soleus muscles of 30-mo-old rats (P < 0.0001) on the microarrays. Probe sets in other functional categories (growth factors, ubiquitin-dependent degradation, signal transduction, metabolism, and phosphorylation) were not significantly altered with age (P > 0.05).



View larger version (51K):
[in this window]
[in a new window]
 
Fig. 1. Functional categories of differentially regulated genes. The relative abundance of functional assignments are shown in the pie charts. Increases and decreases have been separated into fold changes >=2.0 or <2.0 in the charts. Each color represents a particular functional category. The color and number in the legend starts at the 12 o’clock position and cycles clockwise. Not all functional categories (colors) exist in every chart. For fold changes >=2-fold, 43 probe sets decreased and 69 probe sets increased; fold changes <2-fold consisted of 304 probe sets that decreased and 266 that increased.

 
A comparison of microarray and real-time PCR analyses of the same RNA samples was performed on a subset of differentially expressed targets (Table 1). With microarray analysis, all fold changes greater than or equal to -1.58 were confirmed with real-time PCR (Table 1). However, statistical significance [-1.36 fold (a 26% decrease)] was not reached for agrin mRNA with real-time PCR analysis on the same RNA samples (Table 1). Interestingly, prolyl-4-hydroxylase mRNA, which had a significant -1.34-fold (a 25% decrease) with microarray analysis, was validated as a significant difference with real-time PCR. This demonstrates that low fold changes can be validated as significantly different. One potential reason for failing to reproduce a significant change for agrin, might be due to the fact that the levels of some genes differ between mRNA and total RNA, where a significant difference is apparent from an mRNA population, as employed on the microarray, but not in total RNA used for real-time PCR. Strikingly, the proportions of functionally related probe sets differed between mRNAs showing <2-fold and >2-fold significant changes (Fig. 1). Many of the fold changes >=2.0 corresponded to structural constituents, such as the ECM, whereas the fold changes <2.0 had a greater predominance of molecules associated with regulatory functions such as regulation of transcription and signaling. The histogram of fold change distributions (Fig. 2) shows that a twofold cutoff for mRNA identification would have missed 88% of the 347 significant decreases and 80% of the 335 significant increases with fold changes <2.0 in the current experiment.



View larger version (29K):
[in this window]
[in a new window]
 
Fig. 2. The skewed distributions in mRNA levels either downregulated (gray bars) or upregulated (black bars). The greatest percentage of mRNAs show a 25–50% change, while the >2.0-fold changes make up a small proportion of the total number of significant differences. Fold changes are in reference to the young group.

 

    DISCUSSION
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 DISCLOSURES
 References
 
Independent living in the elderly depends upon the ability to perform activities of daily living such as bathing, dressing, and walking, but 7 million Americans 65 yr of age or older are unable to perform these activities in part due to an insufficient amount of muscle mass (See Ref. 15 for references). Physical frailty is associated with increased mortality and increased rates of hospitalization, admission to nursing homes, and use of formal and informal home services (32). Knowledge of the genes contributing to physical frailty is needed to generate a scientifically based attack to counter this disability. The interpretation of the data in the current experiment is limited by the lack of an adult, steady-state set of muscles to reference, which would more precisely isolate genes involved in sarcopenia-induced muscle wasting or those unique to adolescent growth. The choice to compare only young and old animals was made to maximize the changes in gene expression that would differ between states of muscle growth in young animals and muscle wasting in old animals. Use of microarrays allowed a global, unbiased means of determining mRNA expression and may provide insights into the status of gene expression in skeletal muscle growth and sarcopenia.


Statistical rationale.
To be statistically conservative, when thousands of statistical tests are performed on all the genes represented on a microarray, the {alpha} value selected for significance must be adjusted downward to control for false positives which accumulate as the number of tests increases. The reason for this adjustment was explained by Tusher et al. (37), who wrote that although P <= 0.01 is significant in the context of experiments designed to evaluate small numbers of genes, a microarray experiment for 10,000 genes would identify 100 genes by chance. If, as in some reports, only 100 genes are identified, then all identified probe sets could be false positives. Uncertainty currently exists as to what statistical method should be used for multiple testing. Jin et al. (23) had previously concluded that the Bonferroni correction applied for almost 16,000 t-tests carried out for microarray experiments resulted in an experiment-wise significance value of 3 x 10-6, which they remarked was almost certainly too conservative of a correction, and consequently they did not employ the Bonferroni adjustment. Thus, in reference to the current study, identification of 682 differentially expressed probe sets between young (n = 25) and old (n = 20) soleus muscles was an unexpected outcome when a Bonferroni adjustment was applied. A significance level or {alpha} value of 0.05 results in a Bonferroni-adjusted significance level of <=4.478 x 10-6 for the current muscle data, approximately as stringent as the 3 x 10-6 value that was considered too conservative by Jin et al. (23). The increased sample size used in the current experiment yielded hundreds of significant differences, confirming an established mathematical principle of power; that is, as the number of observations increases, the ability to detect differences of a given level will be increased. Conservatism in minimizing false-positive differences is an advantage of using the Bonferroni adjustment. If the current experiment were to be repeated with new animals 99 more times, then use of the Bonferroni adjustment should result in no false positives in any of the significantly identified genes in ~95 experiments out of the 100. In only about five experiments (of the 100) would one or more of the probe sets be expected to contain a false positive(s) identified as significant. Thus one solution to a lack of confidence in genes being identified as differentially expressed by microarray experiments appears to be the use of a large number of observations per experimental group, and usage of the Bonferroni adjustment. An alternative to using large sample sizes to gain power is to use a more liberal level of significance ({alpha}). For example, use of an FDR P < 0.05 was sufficient to detect most of the >1.80-fold changes using a sample size of n = 10, but may accept some false positives. Similarly, if one is only interested in large fold changes (>1.80), then a sample size of n = 15 with a Bonferroni adjustment is a sufficient sample size for detecting that effect size. However, to detect smaller effect sizes while maintaining a conservative level of significance, a larger sample size is necessary (illustrated in Supplemental Tables 7A and 7B). Use of smaller sample sizes was associated with fewer significant differences between young and old due to a failure to find probe sets with larger variation or smaller fold changes significant (Supplemental Tables 7A and 7B). Determination of appropriate sample size will be dependent upon a given data set or treatment, so the information on sample size and fold change is specific only for the current data set and would only be directly applicable for other aging muscle studies. Other treatments will have different effect sizes and will cause differing numbers of genes to be differentially expressed.

Others have emphasized the importance of using statistical approaches in microarray experiments. For example, Kerr and Churchill (25) wrote that one cannot consider just the genes with large fold changes in expression that are obvious to detect without the aid of statistics because this practice overlooks important genes that may have small, but reproducible, changes in expression. Also many genes with large fold changes in expression may not be significantly different (see below). Wolfinger et al. (41) contend that simple rules that eliminate genes with less than two- or threefold expression changes will miss very biologically important genes that have a small fold change, but which are highly significant because they can be measured with high precision as a result of replication. The histogram of fold change distributions (Fig. 2) shows that use of a twofold cutoff for mRNA identification in the current experiment would have only identified 12% of the 347 significant decreases and 20% of the 335 significant increases (Fig. 2). Although there is some preconception against low fold changes, the lowest fold change that is functionally relevant has yet to be determined and likely varies by molecule. Because sarcopenia-induced atrophy occurs over months and years, small fold changes in critical molecules could be responsible for physical frailty. Moreover, ~20% (30 separate probe sets) with detectable changes >2.0-fold did not surpass the statistically significant level for the Bonferroni-adjusted {alpha} value of 0.05, that is to say 4.478 x 10-6; examples are cardiac-responsive adriamycin (fold change = 2.35 and P = 2.0 x 10-3), skeletal fast troponin-I (fold change = 1.95 and P = 5.7 x 10-4), and activating transcription factor-3 (fold change = 2.81 and P = 1.8 x 10-4) mRNAs. Thus dependence on large fold changes alone to determine differential expression in the current microarray experiment could have resulted in numerous false positives.

The current data, consisting of 682 differentially expressed probe sets obtained with a large sample size, suggests that the expression patterns in sarcopenic skeletal muscle are the result of widespread alterations in gene expression. Usage of smaller sample sizes led Weindruch et al. (39) to state that gene expression patterns in skeletal muscle seem to be remarkably stable during the adult mammalian lifespan, a finding that they interpreted as in contrast with the hypothesis that aging is due to large and widespread alterations in gene expression. Employing a large sample size yields a more global set of results, better suiting a global technology, like microarrays.

Next, real-time PCR was employed to validate a subset of differentially expressed targets generated from microarray analysis on the same RNA samples. Usage of real-time PCR verified a fold change down to -1.34 (25% decrease of old compared with young), but a -1.36 fold change for agrin mRNA was not confirmed, suggesting that lower fold changes found with microarrays have a lower probability of being corroborated. The failure to reproduce significant differences found by microarray analysis with real-time PCR does not necessarily invalidate the original finding that a significant difference actually exists. Differences between microarray and real-time PCR results may also be explained by the following: usage of mRNA (microarray) vs. total RNA (for 18S normalization in real-time PCR) for cDNA synthesis, varying efficiencies in different methods of cDNA synthesis, different methods of normalization, varied skill levels for each technique, etc.

Kerr and Churchill (25) have previously contended that the purpose of microarray experiments is not to confirm known properties of well-studied genes, but rather the purpose is to acquire information about unknown genes and unknown functions. Use of a large sample size with a Bonferroni adjustment in the current data set found 210 ESTs of unknown/unclear functions (129 decreases and 81 increases; Supplemental Tables 4 and 5) with no significant homologies to known genes in the databases searched. Furthermore, only 17 of the 193 ESTs of unknown function which differed significantly between young growing and old atrophying skeletal muscles had fold changes greater than twofold. Thus with the statistical approach taken, the number of ESTs identified increased by greater than 11-fold. The larger sample size allowed application of the Bonferroni adjustment, providing a statistical basis for identifying and having confidence in the truthfulness of fold changes less than 2.0. As some consider a purpose of microarrays to be to identify unknown genes with potentially important functions, the current analysis provided a larger list of candidates for future testing of the hypothesis that some of them will play a role in the determination of skeletal muscle mass. As the lower limit of biological relevance is unknown and likely varies for each gene, all significant changes are reported, and the choice of a minimum fold change to test for biological relevance is left up to the reader. It should be noted that each experimental treatment will cause a varying number of genes to be differentially expressed to varying extents, such that some studies will have sufficient power with sample sizes less than 20, whereas other studies where treatments affect fewer changes in gene expression will likely require sample sizes larger than 20.

Gene expression changes in young and old skeletal muscle.
One unexpected advantage of using a large sample size was that in addition to more genes being identified with significant differences was that it allowed for the identification of differentially expressed functional families. Aged skeletal muscle has been previously shown have a selective upregulation of transcripts involved in inflammation and oxidative stress and a downregulation of genes involved in mitochondrial electron transport and oxidative phosphorylation (24, 27). Since ~30% of the increases in mRNA and ~1% of the decreases that occurred in the old soleus muscle were related to immune, defense, and stress functions (Fig. 1), a question for future research was raised as to whether these processes have some sequential order of appearance in order to establish cause and effect. For example, does degradation precede or follow immune, defense, and oxidative stress increases? One EST of interest was homologous to mouse F-box protein 32 (atrogin) mRNA. Atrogin-1, an ubiquitin-protein ligase (E3), appears to be a critical component in the enhanced proteolysis leading to muscle atrophy (18). The probe set was 1.96-fold higher in the old soleus muscle; however, the magnitude of this increase was less than that found in other examples of muscle atrophy. A 9- to 10-fold increase in atrogin mRNA occurred in the gastrocnemius muscle of young rats experiencing one of the following treatments: fasting, streptozotocin-induced diabetes, peritoneal Yoshida hepatoma, or experimentally induced uremia (18). Multi-fold increases in atrogin mRNA were also observed when the muscle was in a denervated, immobilized, or unloaded limb (5). The smaller fold rise observed in atrogin mRNA in old skeletal muscle than in the aforementioned conditions may be associated with the slower rate of muscular atrophy that occurs in sarcopenia.

The major theme of the decreasing mRNAs in the old soleus muscle was 48 probe sets featuring a biological role in the ECM and cell adhesion (Fig. 1). On the other hand only three transcripts with an ECM or cell adhesion function increased. These data extend the earlier observations of an age-related decline of procollagen I and III mRNA expression in mice and humans (16, 27, 40), which Goldspink et al. (16) interpreted to strongly suggest that the increased fibrosis of skeletal muscle with age is not the result of increased collagen gene expression but is most likely due to an impairment of degradation. Muscle inactivity has previously been associated with a downregulation of three extracellular-related mRNAs (laminin B2 chain, pro-{alpha}1-collagen type III, and actin-binding protein 280) (10). As the old rats in the current study were less physically active (22), we speculate that decreased mechanical loading of the old soleus muscle may contribute to its downregulation of ECM transcripts. Indeed, in old muscle compared with young, repair of old skeletal muscle is slower (19) and regrowth from atrophy is often absent (8). Muscle growth might be delayed or impaired due to the loss of ECM mRNAs. ECM molecules are thought to serve as more than a mechanical link of ECM/integrin/cytoskeleton in the transduction of forces into and out of cells (21). Physical distortion of the ECM results in cytoskeletal remodeling and signals alterations in intracellular biochemistry such as cell proliferation (21). The inverse also holds: changes in cytoskeletal tension results in remodeling of structural scaffolds within the ECM and within neighboring cells, thereby altering biochemistry outside the cell (21). For example, application of biaxial strain to fibroblasts led to a transcriptional profile of a "synthetic" fibroblast phenotype characterized by induction of connective tissue synthesis while simultaneously inhibiting matrix degradation (26). As older animals have less voluntary physical activity (7), corresponding environmental signals to gene expression are diminished with aging. ECM proteins also function by interacting directly with cell surface receptors or growth factors to initiate signal transduction pathways, controlling the activity and presentation of a wide range of growth factors and their signaling pathways.

Collagen XV is a major component of various basement membranes, and two ESTs similar to mouse collagen XV were found to be 1.75- and 1.55-fold lower in the old, than the young, soleus muscle (Supplemental Table 3). Mice null for collagen XV by 3 mo of age show progressive histological changes that are characteristic of muscular diseases and are more vulnerable than controls to exercise-induced muscle injury (12). Eklund et al. (12) suggested that the prominent changes seen in the collagen type XV-deficient muscle fibers could reflect a defect in linkage between the muscle cell basement membrane and the surrounding fibrillar matrix. As old skeletal muscle is more susceptible to contraction-induced injury (6), it is reasonable to hypothesize that reductions in collagen XV in old muscle could contribute to its increased susceptibility to injury.

Aging skeletal muscle demonstrates an increased number of extrajunctional receptors (33) and spontaneously denervated muscle fibers (29). Indeed {alpha}-, ß-, {delta}-, and {gamma}-subunits of the acetylcholine receptor (AChR) increased 4.94-, 1.64-, 2.71-, and 1.56-fold, respectively, in the old soleus muscle compared with young (Supplemental Table 4), similar to increases found by others, suggestive of transient denervation (17, 33). In addition, a previous study found that in 30-mo-old mouse muscle, genes involved in neuronal growth accounted for 9% (5/58) of genes highly induced in old muscle (27). Thus the increased neuronally related mRNAs in old muscles allows the generation of a hypothesis that AChR gene expression reflects spontaneous denervation of old muscle fibers and the eventual loss of muscle fibers. Another neural mRNA, previously unidentified as differentially expressed, was {gamma}-synuclein, which was 5.28-fold higher in the old than young, soleus muscle (Supplemental Table 4). {gamma}-synuclein is expressed in the brain and spinal cord but is most abundant in the peripheral nervous system and may play a mechanistic role in the onset or progression of several neurodegenerative disorders (14). Insoluble {alpha}-synuclein filaments accumulate intracellularly in brains of patients with Parkinson’s disease and dementia with Lewy bodies (14). As motor neurons to skeletal muscle are lost with aging, resulting in the disappearance of entire motor units (13), the increased {gamma}-synuclein mRNA in the old soleus muscle provides a hint for a potential candidate to explain this defect in sarcopenic muscle.

Follistatin is an extracellular factor that binds and modulates activin and myostatin activity (28). Follistatin mRNA was threefold higher in old sarcopenic muscle. As overexpression of follistatin has been shown to double muscle mass (28), speculation can be made that follistatin is acting as an insufficient compensatory response to sarcopenia. Follistatin may fail to prevent the atrophy in old muscle because downstream regulation is defective either with follistatin translation or in subsequent follistatin-induced signaling. Interestingly, a follistatin-like (FS) domain is found in the matricellular protein, Sparc (also called osteonectin or BM-40) (20). FS domains derive their name from a cysteine-rich domain repeated three times in the follistatin protein. Sparc is one of the handful of ECM proteins that have been conserved from worm to man, but while it’s biological functions are still largely unknown, it binds with moderately high affinity to collagens I–V (See Ref. 20 for references). Sparc mRNA also decreased approximately threefold in the old soleus muscle (Supplemental Table 3).

As expected, for postmitotic muscle, few genes involved in the regulation of cell cycle were observed as significantly different with age. Also of interest, two genes involved in myogenic differentiation, MyoD and myogenin, were differentially increased in sarcopenia compared with young growing muscle (+3.35 and +1.45, respectively). These mRNA data support the previous findings that MyoD and myogenin protein levels were elevated in senile muscle. Dedkov et al. (11) suggested that the elevated MyoD and myogenin levels were the result of "partial or complete denervation and participation of satellite cells in compensatory myogenesis."

IGFBP5 mRNA was recently found to decrease in fasting muscle, which was interpreted to favor the atrophy process (18), but was found to be twice as high in old as young soleus muscle (Supplemental Table 4). In a previous experiment (34), we confirmed that IGFBP5 mRNA, assayed by RT-PCR, doubled in the old soleus muscle, but that IGFBP5 protein was decreased by one-half. These data support the postulate that a "thermostat-like" mechanism, which senses the loss of mass in old skeletal muscle, upregulates numerous mRNAs to attempt to rescue muscle from sarcopenia, but that the expression and translation of other necessary factors are deficient and/or unresponsive in the old muscle. The ESTs with unknown functions identified in the present study are candidates for both the "thermostat-like sensors" detecting mass, as well as the deficient factors limiting sarcopenic muscle’s ability to maintain its pre-atrophy size.

The hypothesis of the current study that a subpopulation of growth factor mRNAs would be downregulated in old skeletal muscle was thus not upheld (Supplemental Table 8). However, the failure to find a disproportionate expression of growth factors genes should not be interpreted to mean that these are unimportant in the old soleus muscle, because changes in only a few of these could trigger sarcopenia, and thus a disproportional change in growth factors was not significant.

Gill et al. (15) contend that interventions designed to prevent functional decline into physical frailty have the potential not only to generate large health care savings, but also to lead to important reductions in the physical, emotional, social, and financial problems attributable to disability. As a previous study showed that the overexpression of IGF-I could ameliorate muscle wasting in older mice (3), knowledge of other genes responsible for sarcopenia could become clinically significant. Although changes in growth factor and signaling mRNAs were not the most predominant alterations observed, some growth factor proteins have been previously shown to functionally interact with the ECM, a substantially decreased functional category at the mRNA level in old skeletal muscle. The current findings demonstrate that usage of large sample sizes with the conservative Bonferroni adjustment provides a more global output of gene candidates responsible for sarcopenia while minimizing false positives.


    DISCLOSURES
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 DISCLOSURES
 References
 
This research was supported by National Institutes of Health Grant AG-18881.


    ACKNOWLEDGMENTS
 
We thank Dr. Gary Allen who provided access to the Bioinformatics Consortium at the University of Missouri and Dr. Mark McIntosh for leadership in establishment of the Affymetrix Core facility at the University of Missouri. We also thank Dr. Espen Spangenburg, Tsghe Abraha, David Kump, and Aaron Wheeler for assistance.


    FOOTNOTES
 
Article published online before print. See web site for date of publication (http://physiolgenomics.physiology.org).

Address for reprint requests and other correspondence: F. W. Booth, Univ. of Missouri, Dept. of Biomedical Sciences, E102 Vet. Med. Bldg., 1600 E. Rollins, Columbia MO 65211 (E-mail: boothf{at}missouri.edu).

10.1152/physiolgenomics.00040.2003.

1 The Supplementary Material for this article (Supplemental Tables 1–8) is available online at http://physiolgenomics.physiology.org/cgi/content/full/00040.2003/DC1. Back


    References
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 DISCLOSURES
 References
 

  1. Agresti A. Categorical Data Analysis. New York: Wiley Interscience, Wiley, 1990, p. 60.
  2. Bakay M, Chen YW, Borup R, Zhao P, Nagaraju K, and Hoffman EP. Sources of variability and effect of experimental approach on expression profiling data interpretation. BMC Bioinformatics 3: 4, 2002; Epub 2002 Jan 31.[Medline]
  3. Barton-Davis ER, Shoturma DI, Musaro A, Rosenthal N, and Sweeney HL. Viral mediated expression of insulin-like growth factor I blocks the aging-related loss of skeletal muscle function. Proc Natl Acad Sci USA 95: 15603–15607, 1998.[Abstract/Free Full Text]
  4. Benjamini Y and Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Stat Soc Ser B Methodological 57: 289–300, 1995.[ISI]
  5. Bodine SC, Latres E, Baumhueter S, Lai VK, Nunez L, Clarke BA, Poueymirou WT, Panaro FJ, Na E, Dharmarajan K, Pan ZQ, Valenzuela DM, DeChiara TM, Stitt TN, Yancopoulos GD, and Glass DJ. Identification of ubiquitin ligases required for skeletal muscle atrophy. Science 294: 1704–1708, 2001.[Abstract/Free Full Text]
  6. Brooks SV and Faulkner JA. The magnitude of the initial injury induced by stretches of maximally activated muscle fibres of mice and rats increases in old age. J Physiol 497: 573–580, 1996.[Abstract]
  7. Brown M, Ross TP, and Holloszy JO. Effects of ageing and exercise on soleus and extensor digitorum longus muscles of female rats. Mech Ageing Dev 63: 69–77, 1992.[ISI][Medline]
  8. Chakravarthy MV, Davis BS, and Booth FW. IGF-I restores satellite cell proliferative potential in immobilized old skeletal muscle. J Appl Physiol 89: 1365–1379, 2000.[Abstract/Free Full Text]
  9. Claverie JM. Computational methods for the identification of differential and coordinated gene expression. Hum Mol Genet 8: 1821–1832, 1999.[Abstract/Free Full Text]
  10. Cros N, Tkatchenko AV, Pisani DF, Leclerc L, Leger JJ, Marini JF, and Dechesne CA. Analysis of altered gene expression in rat soleus muscle atrophied by disuse. J Cell Biochem 83: 508–519, 2001.[ISI][Medline]
  11. Dedkov EI, Kostrominova TY, Borisov AB, Carlson BM. MyoD and myogenin protein expression in skeletal muscles of senile rats. Cell Tissue Res 311: 401–416, 2003.[ISI][Medline]
  12. Eklund L, Piuhola J, Komulainen J, Sormunen R, Ongvarrasopone C, Fassler R, Muona A, Ilves M, Ruskoaho H, Takala TE, and Pihlajaniemi T. Lack of type XV collagen causes a skeletal myopathy and cardiovascular defects in mice. Proc Natl Acad Sci USA 98: 1194–1199, 2001.[Abstract/Free Full Text]
  13. Flanigan KM, Lauria G, Griffin JW, and Kuncl RW. Age-related biology and diseases of muscle and nerve. Neurol Clin 16: 659–669, 1998.[ISI][Medline]
  14. Galvin JE, Lee VMY, and Trojanowski JQ. Synucleinopathies: clinical and pathological implications. Arch Neurol 58: 186–190, 2001.[Abstract/Free Full Text]
  15. Gill TM, Baker DI, Gottschalk M, Peduzzi PN, Allore H, and Byers A. A program to prevent functional decline in physically frail, elderly persons who live at home. N Engl J Med 347: 1068–1074, 2002.[Abstract/Free Full Text]
  16. Goldspink G, Fernandes K, Williams PE, Wells DJ. Age-related changes in collagen gene expression in the muscles of mdx dystrophic and normal mice. Neuromuscul Disord 4: 183–191, 1994.[ISI][Medline]
  17. Gomes RR Jr and Booth FW. Expression of acetylcholine receptor mRNAs in atrophying and nonatrophying skeletal muscles of old rats. J Appl Physiol 85: 1903–1908, 1998.[Abstract/Free Full Text]
  18. Gomes MD, Lecker SH, Jagoe RT, Navon A, and Goldberg AL. Atrogin-1, a muscle-specific F-box protein highly expressed during muscle atrophy. Proc Natl Acad Sci USA 98: 14440–14445, 2001.[Abstract/Free Full Text]
  19. Grounds MD. Reasons for the degeneration of ageing skeletal muscle: a central role for IGF-1 signalling. Biogerontology 3: 19–24, 2002.[ISI][Medline]
  20. Hohenester E and Engel J. Domain structure and organisation in extracellular matrix proteins. Matrix Biol 21: 115–128, 2002.[ISI][Medline]
  21. Ingber DE. Mechanical signaling, and the cellular response to extracellular matrix in angiogenesis and cardiovascular physiology. Circ Res 91: 877–887, 2002.[Abstract/Free Full Text]
  22. Ingram DK. Age-related decline in physical activity: generalization to nonhumans. Med Sci Sports Exerc 32: 1623–1629, 2000.[ISI][Medline]
  23. Jin W, Riley RM, Wolfinger RD, White KP, Passador-Gurgel G, and Gibson G. The contributions of sex, genotype and age to transcriptional variance in Drosophila melanogaster. Nat Genet 29: 389–395, 2001.[ISI][Medline]
  24. Kayo T, Allison DB, Weindruch R, and Prolla TA. Influences of aging and caloric restriction on the transcriptional profile of skeletal muscle from rhesus monkeys. Proc Natl Acad Sci USA 98: 5093–5098, 2001.[Abstract/Free Full Text]
  25. Kerr MK and Churchill GA. Statistical design and the analysis of gene expression microarray data. Genet Res 77: 123–128, 2001.[ISI][Medline]
  26. Kessler D, Dethlefsen S, Haase I, Plomann M, Hirche F, Krieg T, and Eckes B. Fibroblasts in mechanically stressed collagen lattices assume a "synthetic" phenotype. J Biol Chem 276: 36575–36585, 2001.[Abstract/Free Full Text]
  27. Lee CK, Klopp RG, Weindruch R, and Prolla TA. Gene expression profile of aging and its retardation by caloric restriction. Science 285: 1390–1393, 1999.[Abstract/Free Full Text]
  28. Lee SJ and McPherron AC. Regulation of myostatin activity and muscle growth. Proc Natl Acad Sci USA 98: 9306–9311, 2001.[Abstract/Free Full Text]
  29. Lexell J. Human aging, muscle mass, and fiber type composition. J Gerontol A Biol Sci Med Sci 50: 11–16, 1995.[ISI][Medline]
  30. Lexell J, Taylor CC, and Sjostrom M. What is the cause of the ageing atrophy? Total number, size and proportion of different fiber types studied in whole vastus lateralis muscle from 15- to 83-year-old men. J Neurol Sci 84: 275–294, 1988.[ISI][Medline]
  31. Marsh DR, Criswell DS, Carson JA, and Booth FW. Myogenic regulatory factors during regeneration of skeletal muscle in young, adult, and old rats. J Appl Physiol 83, 1270–1275, 1997.[Abstract/Free Full Text]
  32. Metter EJ, Talbot LA, Schrager M, and Conwit R. Skeletal muscle strength as a predictor of all-cause mortality in healthy men. J Gerontol A Biol Sci Med Sci 57: B359–B365, 2002.[Abstract/Free Full Text]
  33. Pestronk A, Drachman DB, and Griffin JW. Effects of aging on nerve sprouting and regeneration. Exp Neurol 70, 65–82, 1980.[ISI][Medline]
  34. Spangenburg EE, Abraha T, Childs TE, Pattison JS, and Booth FW. Skeletal muscle IGF-binding protein-3 and -5 expressions are age, muscle, and load dependent. Am J Physiol Endocrinol Metab 284: E340–E350, 2003.[Abstract/Free Full Text]
  35. Tavazoie S, Hughes JD, Campbell MJ, Cho RJ, Church GM. Systematic determination of genetic network architecture. Nat Genet 22: 281–285, 1999.[ISI][Medline]
  36. Turturro A, Witt WW, Lewis S, Hass BS, Lipman RD, Hart RW. Growth curves and survival characteristics of the animals used in the Biomarkers of Aging Program. J Gerontol A Biol Sci Med Sci 54: B492–B501, 1999.[ISI][Medline]
  37. Tusher VG, Tibshirani R, and Chu G. Significance analysis of microarrays applied to the ionizing radiation response. Proc Natl Acad Sci USA 98: 5116–5121, 2001.[Abstract/Free Full Text]
  38. Unger MA, Rishi M, Clemmer VB, Hartman JL, Keiper EA, Greshock JD, Chodosh LA, Liebman MN, and Weber BL. Characterization of adjacent breast tumors using oligonucleotide microarrays. Breast Cancer Res 3: 336–341, 2001.[ISI][Medline]
  39. Weindruch R, Kayo T, Lee CK, and Prolla TA. Microarray profiling of gene expression in aging and its alteration by caloric restriction in mice. J Nutr 131: 918S–923S, 2001.[Abstract/Free Full Text]
  40. Welle S, Brooks A, and Thornton CA. Senescence-related changes in gene expression in muscle: similarities and differences between mice and men. Physiol Genomics 5: 67–73, 2001.[Abstract/Free Full Text]
  41. Wolfinger RD, Gibson G, Wolfinger ED, Bennett L, Hamadeh H, Bushel P, Afshari C, and Paules RS. Assessing gene significance from cDNA microarray expression data via mixed models. J Comput Biol 8: 625–637, 2001.[ISI][Medline]