During T₄ induced metamorphosis, Mexican axolotls progress through developmental stages (0–IV) 19 that are defined by the resorption of the upper and lower tailfins, dorsal ridge, and gills. We staged all axolotls after 0, 2, 12, and 28 days of T₄ treatment. No metamorphic changes were observed after two days of T₄ treatment and thus axolotls from both T₄ treatments were assigned to Stage 0. At Day 12, morphological changes were observed in 50 nM T₄ treated axolotls (Stages I and II) but 5 nM T₄ treated axolotls were indistinguishable from control animals (Stage 0). At Day 28, axolotls reared in 50 nM T₄ had fully resorbed tailfins and gills, and thus had completed morphological metamorphosis (Stage IV). Between Days 12 and 28, 5 nM T₄ treated axolotls initiated metamorphosis but did not complete all morphological changes by Day 28 (Stage III). On average, individuals complete metamorphosis after 35 days in 5 nM T₄ (unpublished data). Thus, a low concentration of T₄ delays the initiation timing of morphological metamorphosis but not the length of the metamorphic period.

Our first set of statistical analyses tested control axolotls for temporal changes in mRNA abundance that were independent of T₄ treatment [see Additional files 1, 2]. Temporal changes are expected if patterns of transcription (gene expression) change significantly over time as salamanders mature, or if there are uncontrolled sources of experimental variation. After adjusting the false discovery rate (FDR) to 0.05, none of the probe-sets (genes) on the custom Ambystoma GeneChip were identified as significantly differentially abundant as a function of time. Thus, we found no statistical support for differential gene expression among control animals.

At each of the times that we estimated mRNA abundances during T₄ induced metamorphosis, the number and diversity of genes that were differentially expressed between the 5 and 50 nM T₄ treatments differed (Figure 1A) [see Additional files 3, 4]. A total of 402 DEGs that differed by ≥ two-fold at one or more of our sampling times were identified among all day by T₄ treatment contrasts (Figure 1A) [see Additional files 3, 4]. We identified 30 DEGs as early as Day 2 (Table 1), and eighty percent of these DEGs were up regulated in 50 nM T₄ relative to 5 nM T₄. This small group of early response genes was statistically associated with the amino acid transport and amine transport ontology terms. Additional gene functions of these early responding genes include epithelial differentiation, ion transport, RNA processing, signal transduction, and apoptosis/growth arrest. We note that no differentially expressed transcription factors were identified as DEGs at Day 2, and neither thyroid hormone receptor (TR; alpha and beta) was identified as differentially expressed at any time point in our study. At Day 12, when axolotls in 50 nM T₄ were undergoing dramatic tissue resorption events and axolotls in 5 nM T₄ were indistinguishable from controls, we identified the greatest number of DEGs between the T₄ treatments (n = 319; Figure 1A). An approximately equivalent number of up and down regulated DEGs were identified [see Additional files 3, 4]. These genes were enriched for functions associated with epidermis development, carbohydrate metabolism, ectoderm development, response to chemical stimulus, negative regulation of cell proliferation, response to abiotic stimulus, negative regulation of biological process, development, and organ development. By Day 28, when axolotls in 50 nM T₄ had completed metamorphosis and axolotls in 5 nM T₄ were continuing to show morphological restructuring, we identified 216 DEGs that differed between the T₄ treatments (Figure 1A). Of the identified DEGs, 76% were down regulated in 50 nM T₄ relative to 5 nM T₄ [see Additional files 3, 4]. DEGs identified at Day 28 were associated with response to pest, pathogen, or parasite, negative regulation of cellular process, positive regulation of physiological process, response to stress, response to stimulus, transition metal ion transport, di-, tri-valent inorganic cation transport, response to other organism, immune response, development, organismal physiological process, muscle contraction, response to biotic stimulus, and response to bacteria. The functional categories that we identified show that the axolotl epidermal transcriptional response to T₄ is complex, involving hundreds of DEGs.

To further explore the effect of T₄ on gene expression and morphological metamorphosis we conducted a principal component analysis (PCA). This analysis shows that global gene expression and morphological metamorphosis are strongly correlated, but there is little or no correlation between gene expression and T₄ treatment (Figure 2). This suggests that after metamorphosis was initiated within T₄ treatments, molecular and morphological events were coordinately regulated. T₄ concentration affected the onset timing of metamorphosis in axolotls, but not the sequence of transcriptional and morphological events that define this process.

To further investigate the effect of T₄ concentration on induced metamorphosis, we modeled mRNA abundance estimates from the 5 and 50 nM T₄ treatments using quadratic and linear regression. The regression analyses identified 542 and 709 DEGs that changed by ≥ two-fold relative to Day 0 controls, in the 5 and 50 nM T₄ treatments respectively [see Additional files 5, 6, 7, 8]. Given our previous analyses, we expected to observe different regression patterns (expression profiles; Figure 3) for DEGs from each treatment because metamorphic initiation timing was delayed in 5 nM T₄ and metamorphosis was only completed in 50 nM T₄. Indeed, most DEGs identified by the 5 nM regression analysis exhibited linear expression profiles during metamorphosis (e.g., linear down, LD; linear up, LU; Figure 3) while the majority of the DEGs identified by the 50 nM regression analysis exhibited curvilinear and parabolic expression profiles (e.g., quadratic linear convex down, QLVD; quadratic linear concave up QLCU; quadratic convex QV; quadratic concave, QC; Figure 3; see the methods for a summary of the biological interpretations of these expression profiles in the context of our experiment). Thus, biological processes known to be fundamental to tissue remodeling and/or development were identified from both T₄ treatments, however they were statistically associated with different regression patterns (Figure 3). For example, four collagen-degrading matrix metallopeptidase (MMP) genes (MMP13, MMP9, MMP1) exhibited linear up regulated responses in 5 nM T₄ and were categorized as LU. However, under 50 nM T₄, these genes were categorized among the QC and QLCU profiles. Several genes associated with organ development (transgelin, mitogen-activated protein kinase 12, distal-less homeo box 3, actin binding lim protein 1, collagen type VI alpha 3, and msh homeo box homolog 2) were up regulated in a linear fashion in 50 nM T₄ and were categorized as LU. In 5 nM T₄, several of these genes (mitogen-activated protein kinase 12, actin binding lim protein 1, and msh homeo box homolog 2) were statistically significantly up regulated (LU and QLVU) but failed to eclipse our two-fold change criteria. A single gene (collagen type VI alpha 3) was not statistically significant and categorized as "Flat" in 5 nM T₄. However, this gene did not appreciably deviate from base-line expression levels until Day 28 in 50 nM T₄ (Figure 4A). We attribute these differences between the T₄treatments to the delayed onset timing of metamorphosis in the 5 nM T₄ treatment. Overall, the same generalized direction of expression was observed for 457 of 463 (99%) DEGs that were commonly identified from both T₄ treatments (Figure 1B). The six genes (calponin 2, ethylmalonic encephalopathy 1, 3' repair exonuclease 2, SRV_05658_a_at, SRV_09880_at, and N-myc downstream regulated gene 1) that seemingly exhibited opposite directions of expression in 5 and 50 nM T₄ were all categorized the same way (LU in 5 nM T₄ and QLCD in 50 nM T₄). Closer inspection of these expression profiles revealed that genes classified as QLCD trend upward at the early time points before decreasing at later time points (Figure 4B). These results show that T₄ concentration affected the shape of temporal gene expression profiles but essentially all epidermal DEGs that were identified in both T₄ treatments were regulated in the same direction by 5 and 50 nM T₄.

The regression analyses identified a number of DEGs that only met both of our criteria (statistically significant and ≥ two-fold change) in one of the T₄ concentrations (5 nM, n = 79; 50 nM, n = 246; Figure 1B). Of the 79 genes unique to 5 nM T₄, 36 (46%) were statistically significant in 50 nM T₄ but did not eclipse our two-fold change criteria. Of the remaining 43 genes unique to 5 nM T₄, 25 exhibited ≥ two-fold change in 50 nM T₄ at one or more sampling times but were not statistically significant. Inspection of the 50 nM T₄ regression patterns and fold change data associated with the 79 genes unique to 5 nM T₄ revealed that all of these genes exhibited similar directional trends (up versus down regulation) in 5 and 50 nM T₄ [see Additional files 9, 10]. Of the 246 genes unique to 50 nM T₄, 96 (39%) were statistically significant in 5 nM T₄, but failed to eclipse our two-fold change criteria. An additional 48 genes exhibited ≥ two-fold change for at least one sampling time in 5 nM T₄, but were not statistically significant. Inspection of the 5 nM T₄ regression patterns and fold change data associated with the 246 genes unique to 50 nM T₄ demonstrated that 209 (85%) of these genes exhibited similar directional trends in 5 and 50 nM T₄ [see Additional files 11, 12]. An additional 22 genes unique to 50 nM T₄ did not exhibit > 1.5 fold changes relative to Day 0 controls until Day 28 (at which time they were differentially regulated by ≥ two-fold), suggesting that they are expressed during the terminal stages of metamorphosis [see Additional files 11, 12]. Presumably, these genes were not detected in 5 nM T₄ because we did not sample latter time points for this concentration. These results reiterate the point that essentially all genes identified by our study were similarly, directionally expressed in the 5 and 50 nM T₄ treatments.

To address similarity in terms of magnitude, we compared maximum fold level values for genes that exhibited QLVD and QLCU expression profiles in both T₄ treatments (Figure 3; Table 2). We assumed that genes exhibiting these profiles had achieved maximum/minimum mRNA levels during the experiment, and thus could be reliably compared between treatments. No statistical differences were observed for fold level values of 13 QLVD and QLCU genes between the 5 and 50 nM T₄ treatments (Wilcoxon signed rank test, Z = -1.293, P = 0.1961) 20, and the fold level values were highly correlated (Table 2; Spearman's rho = 1.00, P < 0.0001) 20. Although this analysis was performed on a small subset of genes, the results suggest that mRNA abundances are similar for genes that are differentially expressed in 5 and 50 nM T₄.

Salamanders and anurans may express similar genes during amphibian metamorphosis. To test this idea, we compared a list of 'core' up regulated metamorphic genes from Xenopus 10 to DEGs identified from our study of axolotl. Of the 59 genes that were reported as differentially up regulated by ≥ 1.5 fold in limb, brain, tail, and intestine from metamorphosing Xenopus, 23 (39%) are represented by at least one of the 3688 probe-sets analyzed in our study. Of these, only two (FK506 binding protein 2 and glutamate-cysteine ligase modifier subunit) were identified as statistically significant and differentially expressed by ≥ two-fold in our study [see Additional files 13, 14]. FK506 binding protein 2 was up regulated in axolotl and Xenopus. However, glutamate-cysteine ligase modifier subunit was down regulated in axolotl and up regulated in Xenopus. Thus, < 5% of the 'core' DEGs that are commonly expressed among Xenopus tissues during metamorphosis, were identified as DEGs in our study using axolotl.

We also conducted a comprehensive bioinformatics comparison between DEGs from axolotl epidermis and DEGs from T₃ induced Xenopus intestine 10. Gene expression similarities may exist between the Xenopus intestine and axolotl epidermis because both organs undergo extensive extracellular remodeling that is associated with apoptosis of larval epithelial cells and the proliferation and differentiation of adult cell types. The presumptive orthologs of 111 of the 820 non-redundant DEGs from our study correspond to DEGs from Xenopus intestine [see Additional files 13, 14]. Of these 111 genes, 50 (45%) exhibited the same direction of differential expression in axolotl epidermis and Xenopus intestine. This list includes genes that are known to be associated with metamorphic developmental processes in amphibians. For example, two MMPs (MMP9 and MMP13) that are associated with extra cellular matrix turnover were up regulated in Xenopus intestine and axolotl epidermis. However, other genes were regulated in opposite directions. For example, keratin 12 and keratin 15 were down regulated in axolotl epidermis but up regulated in Xenopus intestine. These results show that there are similarities and differences in gene expression between Xenopus and axolotls when comparing tissues that undergo similar remodeling processes.

In order to validate a subset of genes that were identified as DEGs in our microarray experiment, we conducted a second experiment in which we used quantitative real-time reverse transcriptase polymerase chain reaction (Q-RT-PCR) to generate expression profiles for five candidate genes (Table 3). These genes were chosen because they are involved in a variety of biological processes including cytoskeleton organization (desmin), cell-cell adhesion (desmocollin 1), tissue remodeling (matrix metallopeptidase 13), and ion transport (solute carrier family 31 member 1). In addition, we investigated SRV_10216_s_at in order to verify results from a gene with unknown function. Results of the regression analyses performed on the Q-RT-PCR data are presented in Figure 5 alongside plots of the analogous microarray data. Residuals from the models fit for desmocollin 1 and solute carrier family 31 member 1 exhibited significant departures from normality (Shapiro-Wilk test, P < 0.05). Overall, there was very good agreement between the expression profiles obtained from microarray and Q-RT-PCR analyses. The fact that the Q-RT-PCR results are biologically and technically independent of the microarray data strongly suggests that these patterns are repeatable and unlikely to be experimental or technical artifacts.

Previously, we used stringent statistical criteria (one-way ANOVA, FDR = 0.001, and ≥ two-fold change) to identify 123 annotated genes that exhibited robust responses to 50 nM T₄ 11. In that study, we focused on the potential of several keratin loci to serve as biomarkers of early metamorphic changes that precede changes in gross morphology. In this study, we used less stringent criteria to identify DEGs and more fully explore temporal gene expression responses when T₄ concentration is varied. Of the 123 DEGs previously identified in the epidermis of metamorphosing axolotls, 116 genes were statistically significant and differentially regulated by ≥ two-fold in the 50 nM regression analysis. Of these, 91 (78%) were statistically significant and differentially expressed by ≥ two-fold in the 5 nM regression analysis. Only one of these 91 genes was expressed in opposite directions between the T₄ treatments (3' repair exonuclease 2). However this gene was classified as LU in 5 nM T₄ and QLCD in 50 nM T₄, and represents another example of a transiently up regulated gene that was categorized as QLCD [see Additional files 7, 8]. The 25 genes identified in 50 nM T₄ but not 5 nM T₄ (Table 4) may function in late stage metamorphic processes that were only attained within 28 days under 50 nM T₄. For example, keratin 17 is known to be a marker of proliferating basal epidermal stem cells in mammals 21; this gene may be expressed late during metamorphosis in terminal cell populations of axolotl epidermis that give rise to adult epithelial cells. Other genes associated with tissue stress, injury, and immune function (ferritin heavy polypeptide 1, ras-related C3 botulinum toxin substrate 2, and cathespin S) also appear to be late response genes although we can't rule out the possibility that these genes may be differentially expressed as a toxic response to 50 nM T₄. The majority of the DEGs identified previously using 50 nM T₄ and very strict statistical criteria were similarly identified using 5 nM T₄ and different statistical methods/criteria. These findings further emphasize that the metamorphic gene expression programs of A. mexicanum are similar even when TH concentration is varied by an order of magnitude.

All procedures were conducted in accordance with the University of Kentucky Animal Care and Use Guidelines (IACUC #00609L2003). Salamanders (A. mexicanum) were obtained from a single genetic cross, using adults from an inbred strain. Embryos and larvae were reared individually at 20–22 C in 40% Holtfreter's solution. After hatching, larvae were fed freshly hatched brine shrimp (Artemia sp., Brine Shrimp Direct, Ogden, UT) napuli until they were large enough (3 weeks) to eat California blackworms (Lumbriculus sp., J.F. Enterprises, Oakdale, CA). At approximately eight months of age, 30 salamanders were randomly assigned to each of 10 different treatments and reared in 40% Holtfreter's solution with or without T₄, (Sigma T2376, St. Louis, MO). One hundred μM stock T₄ solutions were made as described in Page et al. 11. Five and 50 nM T₄ were made fresh for each water change by mixing 2.5 or 25 mL of 100 μM stock with 40% Holfreter's solution to a final volume of 50 L. Water was changed every third day.

Skin tissue was collected from salamanders following 0, 2, 12, and 28 days of T₄ treatment. These time points were sampled to test for early gene expression changes that might precede morphological metamorphosis, and because 28 days is a sufficient period for complete metamorphosis of 50 nM T₄ induced A. mexicanum 11. To obtain tissue, salamanders were anesthetized in 0.01% benzocaine (Sigma, St. Louis, MO) and ≈ 1 cm² of skin tissue was removed from the top of the head.

Total RNA was extracted for each tissue sample with TRIzol (Invitrogen, Carlsbad, CA) according to the manufacturer's protocol; additionally, RNA preparations were further purified using a Qiagen RNeasy mini column (Qiagen, Valencia, CA). UV spectrophotometry and a 2100 Agilent Bioanalyzer (Agilent Technologies, Santa Clara, CA) were used to quantify and qualify RNA preparations. Three high quality RNA isolations from each treatment and sampling time combination were used to make individual-specific pools of biotin labeled cRNA probes. Each of the 30 pools was subsequently hybridized to an independent GeneChip. The University of Kentucky Microarray Core Facility generated cRNA probes and performed hybridizations according to standard Affymetrix protocols.

A custom Ambystoma Affymetrix (Santa Clara, CA) GeneChip was designed from curated expressed sequence tag assemblies for A. mexicanum and A. tigrinum 4445. The array contains 4844 probe-sets, 254 of which are controls or replicate probe-sets. Detailed descriptions of this microarray platform can be found in Page et al. 11 and Monaghan et al. 46.

We used the Bioconductor package affy 47 that is available for the statistical programming environment R 48 to perform a variety of quality control and preprocessing procedures at the individual probe level 4950. These procedures included: 1) generating matrices of M versus A plots for all replicate GeneChips (n = 3 GeneChips for 10 treatment/sampling time combinations), 2) investigating measures of central tendency, measures of dispersion, and the distributions of all 30 GeneChips via boxplots and histograms, 3) viewing images of the log₂(intensity) values for each GeneChip to check for spatial artifacts, and 4) viewing an RNA degradation plot 50 that allows for visualization of the 3' RNA labeling bias across all GeneChips simultaneously. In addition, we used ArrayAssist Lite software (Stratagene, La Jolla, CA) to assess several quality control measures that are recommended by Affymetrix such as average background (minimum = 48.36, maximum = 59.28, n = 30) scale factors (minimum = 0.404, maximum = 0.629, n = 30), and percent present (minimum = 81.5, maximum = 86.5, n = 30). Next, we processed our data by implementing the robust multi-array average (RMA) algorithm of Irizarry et al. 51.

To obtain estimates of between GeneChip repeatability, we generated correlation matrices for the hybridization intensities across all probe-sets among replicate GeneChips. Very high and consistent mean r-values were calculated for each of the 10 treatment by sampling time combinations (range of mean r ± standard error = 0.966 ± 0.002 to 0.986 ± 0.001). These results demonstrate that we were able to obtain a high level of repeatability between replicate GeneChips. Our data are MIAME compliant and raw data files can be obtained at Sal-Site 4552.

Microarray platforms may not accurately or precisely quantify genes with low intensity values 5354. Because low intensity genes contribute to the multiple testing problem that is inherent to all microarray studies, we filtered probe-sets whose mean expression values across all GeneChips (n = 30 per gene) were smaller than or equal to the mean of the lowest quartiles (25^th percentiles) across all GeneChips (n = 30, mean = 6.53, standard deviation = 0.04; data presented on a log₂ scale). Upon performing this filtration step, 3688 probe-sets were available for significance testing. We then performed PCA on the centered and scaled data from these probe-sets. This analysis allowed us to visualize the relationships between GeneChips within and across treatments.

We investigated whether genes exhibited differential expression as a function of time in the absence of T₄ (control animals sampled at Days 0, 2, 12, and 28) via linear and quadratic regression 55. We corrected for multiple testing by evaluating α₀ according to the algorithm of Benjamini and Hochberg 56 with a FDR of 0.05. α₁ was set to 0.05.

We conducted three analyses to investigate the effect of T₄ on epidermal gene expression. For the first analysis, we used limma 5758 to identify genes that were differentially expressed as a function of T₄ concentration. The limma package couples linear models with an empirical Bayes methodology to generate moderated t-statistics for each contrast of interest. This approach has the same effect as shrinking the variance towards a pooled estimate and thus reduces the probability of large test statistics arising due to underestimations of the sample variances. Operating limma requires the specification of two matrices. The first is a design matrix in which the rows represent arrays and the columns represent coefficients in the linear model. The second is a contrast matrix in which the columns represent contrasts of interest and the rows represent coefficients in the linear model. For this analysis, the design matrix specified a coefficient for each unique treatment by sampling time combination (10 coefficients) and the contrast matrix specified the calculation of contrasts between the two T₄ concentrations (5 and 50 nM) at each of the non-zero sampling times (Days: 2, 12, and 28). In addition to moderated t-statistics, limma also generates moderated F-statistics. These moderated F-statistics test the null hypothesis that no differences exist among any of the contrasts specified by a given contrast matrix. A FDR correction 56 of 0.05 was applied to the P-values associated with the moderated F-statistics of the contrast matrix. In order to further reduce the number of false positives, we required that all "identified" DEGs be differentially regulated by ≥ two-fold at one or more of the contrasted time points.

The last two analyses were conducted using the regression-based approach of Liu et al. 55 to detect genes that exhibit differential expression as a function of days of T₄ treatment. This approach also classifies DEGs into nine categories based on their temporal expression profiles as determined via linear and quadratic regression. In the context of our experiment, these profiles have specific biological interpretations. However, exceptions to these interpretations exist (see results for examples). In general, genes that exhibited LD, LU, QLCD, and QLVU expression profiles were still actively undergoing changes in their expression levels when the study was terminated. In contrast, genes that exhibited QLVD and QLCU expression profiles underwent down and up regulation respectively but reached steady state expression levels before the experiment was terminated. Finally, genes that exhibited QC and QV expression profiles were transiently up and down regulated respectively before returning to baseline expression levels. Null results are described by the 'Flat' category. Separate analyses were conducted for the 5 and 50 nM datasets with α₀ evaluated at a FDR of 0.05 according to the algorithm of Benjamini and Hochberg 56 and α₁ = 0.05. DEGs were required to exhibit ≥ two-fold changes relative to Day 0 controls at one or more sampling times before they were categorized as "identified".

Biological process gene ontology categories that are over represented in our lists of DEGs (statistically significant and ≥ two-fold change) were identified using the Database for Annotation Visualization and Integrated Discovery (DAVID) 59. In all analyses, the 3085 probe-sets on the Ambystoma GeneChip with established orthologies were used as the background for generating expected values. The EASE threshold was always set to 0.05, and the count threshold was always set to two.

In a recent microarray study, Buchholz et al. 10 presented "a core set of up regulated genes". These genes have been identified as up regulated in response to T₃ treatment by ≥ 1.5 fold in every tissue that has been examined in metamorphosing X. laevis via microarray analysis (limb, brain, tail, and intestine) 910. We determined the orthologies of these genes to human as described in Page et al. 11 and Monaghan et al. 46. We then identified genes listed by Buchholz et al. 10 that were also differentially expressed in our experiment. The same approach was used to compare our gene lists to the 2340 DEGs identified by Buchholz et al. 10 from the intestine of metamorphosing X. laevis.

We conducted a second experiment using Q-RT-PCR to investigate the temporal expression patterns of five genes identified as differentially expressed by microarray analysis (Table 3). Animals used in our second experiment were raised as described for the animals used in the microarray experiment, with the exception that T₄ treatment (50 nM) was initiated at 120 days post-fertilization. Tissue samples from two or three individuals were collected as described above beginning on Day 0 (prior to initiating T₄ treatment) and were collected every two days for 32 days (i.e., Day 0, Day 2, Day 4... Day 32).

Total RNA was extracted from integument as described for the microarray experiment with the exception that all samples were treated with RNase-Free DNase Sets (Qiagen, Valencia, CA) according to the manufacturer's protocol. For each sample, the Bio-Rad iScript cDNA synthesis kit (Hercules, CA) was used to synthesize cDNA from 1 μg total RNA. Primers (Table 3) were designed using Primer3 60, and design was targeted to the same gene regions that are covered by Affymetrix probe-sets. All PCRs were 25 μL reactions that contained: cDNA template corresponding to 10 ng total RNA, 41 ng forward and reverse primers, and iQ SYBR-Green real-time PCR mix (Bio-Rad, Hercules, CA). Reaction conditions were as follows: 10 minutes at 50 C, five minutes at 95 C, 45 cycles of 10 seconds at 95 C followed by 30 seconds at 55 C, one minute at 95 C, and one minute at 55 C. Melting curve analysis was used to ensure the amplification of a single product for each reaction. All reactions were run in 96 well plates and blocked by sampling time (i.e., each of the 17 sampling times was equally represented for each gene on each plate). PCRs were performed using a Bio-Rad iCycler iQ Multi-Color Real Time PCR Detection System (Hercules, CA). All plates contained template free controls 61. Primer efficiencies were estimated via linear regression and relative expression ratios (R) were calculated according to Pfaffl 62. All expression ratios are relative to the average of the Day 0 animals, and normalized to transcriptional intermediary factor 1 (probe-set ID: L_s_at). This gene was selected as a control because it had an extremely small standard deviation across all treatment regimes in the microarray experiment (n = 30, mean = 14.44, standard deviation = 0.03; data presented on a log₂ scale).

Log₂ transformed R-values for each gene were analyzed separately via linear and quadratic regression models in which days of T₄ treatment was the predictor variable. These analyses were carried out using JMP, Version 5 (SAS Institute, Cary, NC). We decided whether to use a linear or quadratic model for a given gene via forward selection 20. In short, quadratic models were accepted when the polynomial terms were significant (P < 0.05) and resulted in an increase in the proportion of variation in the data explained by the model (adjusted R²) of ≥ five percent relative to the linear model. The residuals of all models were inspected graphically. In addition, the residuals of all models were checked for normality. In cases where the assumption of normality was violated, regression analyses were run to obtain equations that describe the response of these genes to T₄ as a function of time. However, such analyses were conducted with the understanding that a strict hypothesis testing interpretation could prove problematic.