Gene expression signal in colorimetric ABA ISH data is classified using an automated image segmentation algorithm (described in detail in 217) that identifies contiguous groups of pixels corresponding to higher visual concentrations of BCIP/NBT (5-bromo,4-chloro,3-indolylphosphate/nitroblue tetrazolium), a precipitate deposited at riboprobe binding sites in the cell. Working from a 10× magnification (1.05 μm/pixel) high-resolution ISH image (Figure 1b), the algorithm segments and labels putative expressing cells using a statistical classifier 47 that exploits image intensity distribution and cell shape (morphology) characteristics 17. Typical non-radioactive ISH images can have rather high levels of background intensity contaminated with non-specific hybridization products that resemble low-level expressing cells. The ABA ISH expression detection algorithm is designed to focus on achieving a high recognition probability for medium/high-level expressing cells while maintaining high specificity for rejecting non-specific labeling and artifacts among low expressing genes 17. The algorithm output is an expression segmentation heat mask with measurable characteristics (Figure 1c). High concordance between automated and manual segmentation using these methods has been previously demonstrated 17.

The traditional approach to quantifying radioactive isotopic ISH has been to use optical density 1148 of the underlying tissue section. If the probe specific binding activity is known, then the number of probe molecules bound to a certain tissue area can be calculated 1518. In both radioactive and non-radioactive ISH, most rigorous measures of quantification have employed optical density of the image area relative to background as a measure of signal intensity, and in each case increased optical density correlates with either the number of mRNA molecules (radioactive) or increased mRNA content (non-radioactive) 1549.

In order to correlate gene expression across-platforms and between distinct brain regions with different underlying cell densities, a new variant of integrated optical density that is normalized to the cell density in each region was generated. For any given cell, ISH signal is a function of both intensity and fractional area occupied within the cell. To generalize this to a region R, expression level L is defined as:

L R = a R a max ⁡   I ¯ R MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8GiVeY=Pipec8Eeeu0xXdbba9frFj0xb9Lqpepeea0xd9q8qiYRWxGi6xij=hbbc9s8aq0=yqpe0xbbG8A8frFve9Fve9Fj0dmeaabaqaciaacaGaaeqabaqabeGadaaakeaacaWGmbWaaSbaaSqaaiaadkfaaeqaaOGaeyypa0tcfa4aaSaaaeaacaWGHbWaaSbaaeaacaWGsbaabeaaaeaacaWGHbWaaSbaaeaaciGGTbGaaiyyaiaacIhaaeqaaaaakiaaykW7ceWGjbGbaebadaWgaaWcbaGaamOuaaqabaaaaa@3CFC@

that is, the product of I¯ MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8viVeY=Nipec8Eeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGacaGaaiaabeqaaeqabiWaaaGcbaGabmysayaaraaaaa@3087@ the average pixel intensity for all expressing cells in R, times the relative fractional area occupied by expressing cells in R. This latter ratio is obtained by dividing the area of expressing cells a_R in R by the effective maximum possible expression area a_max in R, computed by applying the detection algorithm to a set of ubiquitously expressed genes 2. The ubiquitously expressed genes (H2-T3, Ube2c, 0610008A10Rik, and 1810037K07Rik), genes expressed in nearly every cell, were identified through manual curation and possess the maximum observable expression throughout the brain. Expression level L is, therefore, a fraction of the average 8-bit integer pixel intensity and ranges from [0,255], and all quantities are computed from masks similar to that shown in Figure 1c. Expression level as defined is, therefore, a cell density normalized ISH analog to microarray level that models transcript copy number in a given region R, assuming both average intensity and fractional area of expressing cells scale with transcript count.

The correlation between the expression level L calculated by the ABA expression detection algorithm and integrated optical density 48 was tested. Measurements from expressing cells were made from the caudoputamen from 30 ISH images for six genes, five sections per gene (Table S1 in Additional data file 1). Figure S1 in Additional data file 1 shows a high correlation (R = 0.99), as expected for a fixed region, between expression level and optical density measured at a calibrated microscope and indicates that ISH expression level scales with relative mRNA signal and transcript count to the extent that optical density is a realistic measurement of that signal.

The dynamic range of the method was then investigated. This range remains limited compared to other platform expression metrics, with truncation in the higher ranges due to signal saturation. Figure 2 plots on log scale 1,270 of the highest sorted expression values for Teragenomics, GNF, and ABA common genes in the striatum and hypothalamus. (A SAGE library for the striatum is included for reference 50). Genes on each curve are sorted independently so that only the relative range of values is preserved. The profiles are nearly identical across structures and show that the dynamic range of the highest 15% of values is comparatively flat for colorimetric ISH while decreasing by a factor of 10 for microarray data and by at least a factor of 20 for SAGE. While some rescaling of the ISH dynamic range is possible to some extent, this essentially would have to be implemented on a gene by gene basis 51.

A particular advantage of colorimetric ISH is that other statistics, such as the proportion of labeled expressing cells, can be calculated. Expression density D in a region R is defined as:

D R = n R n max ⁡ MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8GiVeY=Pipec8Eeeu0xXdbba9frFj0xb9Lqpepeea0xd9q8qiYRWxGi6xij=hbbc9s8aq0=yqpe0xbbG8A8frFve9Fve9Fj0dmeaabaqaciaacaGaaeqabaqabeGadaaakeaacaWGebWaaSbaaSqaaiaadkfaaeqaaOGaeyypa0tcfa4aaSaaaeaacaWGUbWaaSbaaeaacaWGsbaabeaaaeaacaWGUbWaaSbaaeaaciGGTbGaaiyyaiaacIhaaeqaaaaaaaa@3990@

where n_R is the number of detected expressing cells in R and n_max is the approximate theoretical maximum computed as above. If all cells were spherical with identical volumes and ISH signal uniformly filled those cells, then expression level would scale linearly with density. However, in real tissue the relationship between level and density can be complex 2, as illustrated in the final section. It is also possible to develop quantitative measures of ISH signal uniformity across a structure using methods of spatial statistics, such as the k-estimator of Ripley 52.

To perform genomic scale signal mapping, quantification, and replicate analysis in the ABA requires utilizing the spatial mapping platform previously described in 217. Briefly, in this pipeline, image series are white balanced and cropped, then reconstructed and registered to a three-dimensional informatics reference atlas for the C57BL/6J mouse brain 53. Expression is quantified according to level and density metrics on a per section basis, and the results for three-dimensional anatomic regions can be compiled using the registration reconstruction and the virtual reference atlas. Details can be found in 17 and on the web site 3.

To examine the fidelity of this high-throughput platform to produce consistent quantification of regional ISH signal, a set of experiments were performed to assess variation in hapten incorporation into riboprobes, day-to-day variability associated with reagent preparation, and accuracy of mapping signal to brain anatomy. To assess the first two factors, an experiment was designed to compare ISH quantification results for nine genes (Calb1, Calb2, Cst3, Dkk3, Gad1, Man1a, Plp1, Pvalb, and Nov) processed independently on four different days. On each day, four independent riboprobes were also synthesized for each gene by in vitro transcription. The data generated for each gene consisted of 16 complete sagittal series (20 tissue sections per riboprobe) through a single brain hemisphere. Figure 3 shows a scatter plot of the computed level and density measurements at the whole brain structure level, computed hemisphere wide for 9 genes × 4 riboprobes × 4 days. Clustering of these data reveals the fidelity of this approach, in that the nearest neighbor in (level, density) space to any given gene is a replicate of the same gene in 81.8% of all cases, or a second nearest neighbor in 90.3% of cases. These results demonstrate the consistency of the ABA ISH platform day to day, treatment to treatment, and sample to sample and the reproducibility of the values level and density at the brain-wide level. It is notable that level and density values even at the level of the entire brain essentially uniquely identify this set of genes.

To examine the fidelity of the automated structural registration with respect to signal measurement, expression levels in different brain regions for the same data set described above were compared. As shown in Figure 4a, a unique reproducible profile emerges for each gene's expression pattern, highly consistent with the original ISH shown in Figure 4b and corresponding expression segmentation map in Figure 4c. The mean variation in signal across genes from laboratory through informatics processing is only 5.09%, with day-to-day variability 0.26% greater than probe synthesis variability (Additional data file 2). These control experiments establish key criteria for making global cross-gene and cross-structure comparisons, namely reproducibility of high throughput, semi-automated ISH data generation, mapping and quantification.

It should be noted that these control experiments are somewhat idealized. First, ABA data were generated over approximately a three-year period, with consequently larger process variation than the experiment described above. Second, the Allen Reference Atlas model used for structural mapping was produced in the coronal plane, while ABA ISH data for about 78% of genes are available only in the sagittal plane. Cross-axis registration is somewhat less reliable than same-plane registration. To investigate these effects further, 2,263 genes with more than one replicate were selected: either genes with two image series in the sagittal plane (1,828) or genes with one replicate in the sagittal plane and the other in the coronal plane (425). The correlations of ABA ISH expression level between the replicates retain high repeatability (ρ > 0.8) for all six structures with the correlation for the sagittal and coronal plane data ranging from 0.790 to 0.873 (lowest correlation in hippocampus and highest correlation in striatum). The correlation between the sagittal data sets ranged from 0.835 to 0.950, with the lowest and highest correlation in the same respective structures. These results are based on applying the automated spatial mapping process and the correlation results should, therefore, be understood as lower bounds for reproducibility of the replicate signals.

Several studies to date have investigated the relative internal consistency of microarray platforms under a variety of experimental conditions 282930313233545556, and high degrees of within platform consistency have been reported for Affymetrix oligonucleotide microarrays 283941. In this study, several publicly available Affymetrix-based microarray data sets were chosen for analysis. The selected data sets profile different brain regions in mouse strains and ages similar to those used for the ABA (C57BL/6J, P56), and had replicate data. Both Teragenomics 57, using the Affymetrix U74Av2 platform, and GNF 58, using a custom Affymetrix array design, present two replicates for each of the six brain structures examined (with the exception of cortex, for which only one replicate was available from GNF) (see Materials and methods; Table 1, and Additional data files 3 and 4). The condensation algorithm used for GNF data was MAS5 and a close variant was developed by Teragenomics (Matthew Zapala, personal communication). Correlation between replicates was calculated to examine the variability within each microarray platform. To avoid probe to gene mapping issues, the comparison data set was limited to genes that had a one-to-one mapping between probe and gene (see Materials and methods; Table 2).

It is well known that many factors contribute to variation between microarray experiments even when using the same platform, including probe preparation and experimental protocols 283342. Table 3 shows a very strong internal correlation between replicates (average ρ > 0.98 for Teragenomics and ρ > 0.96 for GNF) throughout all six structures. The correlation decreases to an average of 0.724 when comparison is made between the two Affymetrix platforms for a given brain structure. Excluding genes with no overlap between probe sequences from each platform (42%), the correlation shows slight but not statistically significant improvement at the α = 0.05 level in all six structures (Table 3). Scatter plots for correlation in the striatum are shown in Figure 5. All six structures have similar results (Figure S2 in Additional data file 1).

Expression level and density are computed in the ABA pipeline for each anatomic structure (Additional data file 5) and can be correlated with other data modalities. The standard correlation for colorimetric ISH expression level (as defined above) versus microarray level was first considered. A single replicate is used in computing correlation as the internal consistency of both GNF and Teragenomics is high, and approximately 80% of ABA data is single replicate. Representative scatter plots for ABA versus Teragenomics for expression level are shown in Figure 6 (data for additional structures are in Figure S3 in Additional data file 1) and values are presented in Table 4 for ABA versus GNF and ABA versus Teragenomics with Pearson and Spearman correlations. The table shows weaker agreement of ISH data with either of the microarray data sets than the cross-microarray GNF versus Teragenomics correlation. While the difference is obviously significant, which of GNF or Teragenomics is closer to ABA can be tested by applying a Fisher r-to-z transformation 59 for observed difference in correlation. An asterisk in Table 4 indicates that the ISH Pearson correlations in the first two columns are distinct at significance level α = 0.05, and the only structures whose correlations are significantly different are cerebellum and olfactory bulb. In terms of Pearson correlation, the weighted average ABA-GNF correlation (0.479) is quite similar to the ABA-Teragenomics correlation (0.471), while the average ABA-GNF Spearman correlation (0.401) is somewhat weaker than the corresponding ABA-Teragenomics value (0.526).

To investigate the effect of increasing probe homology on platform correlation, Pearson correlation between expression levels as a function of increasing probe base-pair agreement after BLAST alignment 60 of probe sequences to the target gene was computed. Figure 7 shows ABA-Teragenomics correlation as probe overlap ranges from disjoint to 50%. A regression fit to the data with R² shown in Figure 7 illustrates strong increasing correlation with probe homology for cortex (R² = 0.861) and hippocampus (R² = 0.716) and weaker, although positive, improvement for other structures. Thus, the results are expectedly strengthened by considering more homologous targets.

To more fully account for significant data modality and platform differences, a technique originally proposed by van Ruissen et al. 30 was modified in which microarray and SAGE cross-platform comparisons are made based on correlation of ratios of expression values for transcripts. Conceptually, cross-platform variability should be accounted for by considering expression ratios rather than correlation of absolute levels. This technique was generalized to enable structural comparisons as follows: for each of six structures S, each data point input to the correlation is the log ratio of gene expression level of S to any of five other structures. (Using the cortex as an example, cortex/striatum, cortex/olfactory bulb, cortex/hypothalamus, cortex/hippocampus, and cortex/cerebellum). Similar ratios were obtained for all structures, provided the denominator structure exhibits positive expression density D and provided the gene is called 'present' in at least one of the two structures forming the ratio. In addition to reflecting the notion that the ratio of transcript abundance should be independent of inherent platform differences (particularly probe sequence differences), the ratio test is potentially a better comparison of relative abundance and is more consistent with colorimetric ISH expression level as a relative or qualitative measure.

After normalizing the histogram of ratio values for each platform to a common scale following 30, Pearson and Spearman correlation was calculated over the log expression level ratios. Representative structure ratio scatter plots for ABA versus Teragenomics for expression level are shown in Figure 8 (the full set is available in Figure S4 in Additional data file 1). Figure 9 shows the summary results of this technique with mean (median) Pearson correlations: ABA-GNF, 0.404 (0.4); ABA-Teragenomics, 0.438 (0.46); and GNF-Teragenomics, 0.624 (0.67). On the whole, these results surprisingly do not illustrate a stronger correlation and are consistent overall with the standard Pearson approach (this result holds for Spearman as well (Figure S5 in Additional data file 1)) with a slightly weaker result for GNF-Teragenomics. The results can be considerably improved, however, by restricting the analysis to genes exhibiting higher expression fold changes. For each structure ratio pair S₁/S₂, those genes that exhibited at least a two-fold expression difference between the structures S₁, S₂ (for either structure) in the Teragenomics data set were chosen. For these gene subsets the correlation over all pairs improved to: ABA-Teragenomics mean, 0.73 (mean N = 159; range, 0.63 (hippocampus/olfactory bulb) to 0.79 (cerebellum/hippocampus)); ABA-GNF mean, 0.71 (mean N = 110; range, 0.51 (striatum/cortex) to 0.79 (striatum/cerebellum)); and GNF-Teragenomics mean, 0.84 (mean N = 97; range, 0.54 (striatum/cortex) to 0.91 (cerebellum/hypothalamus)). These results indicate that both ISH and microarrays do well at identifying genes with the highest relative expression differences between brain structures. The combined results also suggest that comparing colorimetric ISH with microarray data, as with any platform, may be more problematic for lower expressing genes.

To what extent do colorimetric ISH and microarray data yield reconcilable expression results? To address this question more fundamentally, correspondence at a binary present/absent level is considered. Utilizing binary present/absent calls for expression comparison has been noted as an oversimplification of true correlation due to class threshold effects 30, and alternative techniques have been developed 303361. However, this type of analysis may be appropriate where there are substantial differences in platform technology as is the case between colorimetric ISH and microarrays and to gain a benchmark.

A present/absent call for ABA ISH expression in each assayed structure was made by applying a threshold either to the expression level L or density D. With respect to density, when the threshold value for D is very low, numerous artifacts within ISH images may be detected as expression. On the other hand, if the threshold is too high, expression sensitivity is compromised. To determine an optimal threshold value a receiver operating characteristic (ROC) curve was drawn based on calling a gene expressed if its density attained a given value and by using ground truth determined from independent expert visual inspection. By examination of image series for 14,000 genes from the ABA, an optimal threshold for calling a gene expressed in a given structure based on expression density D was chosen by requiring that 1.0% of 8 × 8 pixel zones in that structure in the ISH image have expressing cells. A similar procedure can be used to threshold the ABA expression level, but thresholding on density D resulted in a closer call to visual inspection. This method yielded a sensitivity of 0.82 and specificity of 0.88 (Figure S6 in Additional data file 1).

Assessment of expression level for microarray data in terms of present/absent calls is generally made by algorithms counting transcript abundance 62. For Teragenomics and GNF data, Affymetrix software 27 is used, and probe set detection p values are calculated relative to the intensity of all probes for a given gene (Additional data files 3 and 4). Since the Teragenomics and GNF data sets have strong repeatability (ρ > 0.98, and ρ > 0.96, respectively), only one replicate from each platform was used in this analysis.

Present/absent call results between ISH and the two microarray platforms for all structures are shown in Table 5. The present/absent call agreement ranges from 0.581 to 0.717 for ISH versus microarray, with the highest correlations between ISH and Teragenomics (0.687-0.717). The higher correlation for ISH and Teragenomics may be due to the fact the biological samples were more closely matched in age and sex, whereas the GNF data included male and female samples that were 2-4 weeks older. To test these results for significance, a two-tailed binomial proportions test 59 for observed difference in values was applied. An asterisk in Table 5 indicates that a given ABA correlation in either of the first two columns is statistically significant at level α = 0.05 when compared to the GNF replicate 1 (GNF1) versus Teragenomics replicate 1 (TERA1) proportion in the final column (see Materials and methods for the data sets comprising these replicates). Diagrams of the intersections of genes with present calls in the ABA, GNF, and Teragenomics platforms in six structures are displayed in Venn diagrams in Figure S7 in Additional data file 1.

These results indicate that the ABA versus TERA1 binary correlation is slightly weaker than the GNF1 versus TERA1 binary correlation at 95% confidence, while it is as strong as in the striatum. The ABA versus GNF1 correlation is somewhat weaker than the GNF1 versus TERA1 binary correlation at much stronger confidence level. This result is consistent with a smaller scale manual comparison for embryonic ISH versus microarray presented in 14.

Variations on present/absent correlation are possible, including binary correlation of log ratios of expression levels for structures S₁, S₂ (that is, sgn log(S₁/S₂); Table S2 in Additional data file 1) and extensions to quartiles (low, low-medium, medium-high, and high; Figure S8 and Table S3 in Additional data file 1). These comparisons (Additional data file 6) show that these results are expectedly intermediate between binary and numerical correlation values.

The ability to examine high-resolution cellular morphology and spatial expression patterns is a notable advantage of colorimetric ISH in the analysis of gene expression. In addition to the visual component of viewing signal in situ, the relationship between quantifiable variables such as expression level and density (as defined above) yields more insight into relative and regional expression characteristics 2. To explore this effect more thoroughly, the group of 1,030 genes defining the highest quartile of expression in the Teragenomics data set (shown in Figure 10) was considered, and a narrow microarray level range (550,650) between the median (480.7) and mean (766.7) level of this set, consisting of 84 genes, was selected. Figure 10 shows the ABA log expression level versus log density plotted in black for the full set of 1,030 genes with the 84 genes in the narrow microarray level range indicated by an asterisk. The dotted lines of Figure 10 are the median and third quartile values for level and density from the entire ABA data set.

From the distribution in Figure 10, the ABA level/density relationship for Teragenomics high expressors is essentially linear in log space (R² = 0.976), and all 1,030 genes are contained in the second to fourth quartile in both level and density (the first quartile is shown at the axes origin.). The quartile distribution of the 84 Teragenomics genes is (0, 7, 23, 54) for level and (0, 9, 22, 53) for density, and is consistent with the present/absent and categorical results presented earlier (see Additional data file 7 for expression values). Although there is some variation in the lower values of ABA level and density, the 84 Teragenomics high expressors are generally not among the outliers in Figure 10, with the possible exception of the gene Wfs1. In fact, small variations in the level/density profile of a gene can account for complexity in expression cell type and spatial pattern as shown below. To sample the variety and complexity that colorimetric ISH data may have within a limited microarray level range, four of the 84 genes were selected for consideration in detail.

Figure 11a shows the annotated image region from the caudoputamen of the striatum (Allen Reference Atlas 3) and the corresponding Nissl stained section from the Allen Reference Atlas (Figure 11b). The remaining images are ISH images for plane matched sections of the caudoputamen for genes Kcnab1, Wfs1, Gsn, and H3f3b, all having similar microarray expression levels. The potassium voltage-gated channel, shaker-related subfamily, beta member 1 (Kcnab1; Figure 11c) is widely expressed in striatal neurons, and highly enriched in the striatum relative to other brain regions. Figure 11d shows expression of Wolfram syndrome 1 (Wfs1), a gene associated with Wolfram syndrome 63, a rare autosomal recessive neurodegenerative disorder characterized by simultaneous presentation of type I diabetes mellitus and optic atrophy in youth 64. Unlike Kcnab1, Wfs1 is expressed in a subset of neurons in the caudal and ventral caudoputamen. The spatial expression pattern differences between Kcnab1 and Wfs1 are not apparent from their similar microarray expression levels but can be understood from Figure 10 as Wfs1 having a moderately reduced density for a given level compared with the more uniformly expressing Kcnab1.

Colorimetric ISH data also help to resolve the cell class by discriminating between non-neuronal and neuronal gene expression, and this cannot be distinguished by microarray levels alone. For example, Gelsolin (Gsn; Figure 11e), a Ca(2+) dependent actin regulatory protein 65, is expressed highly in oligodendrocytes 66, clearly visible by the high expression levels in white matter tracts surrounding the striatum and in scattered cells in the caudoputamen. A particularly compelling example of the value of cellular resolution is shown for the histone family member H3f3b (Figure 11f). Striatal detection of H3f3b by microarrays is apparently false and due to high levels of expression in the subventricular zone adjacent to the striatum that contains migrating cells in the rostral migratory stream. Even though Hsfb3 is reported positive for expression in the striatum by microarrays, the ISH data clearly show that H3f3b is not expressed in the striatum, but rather in the adjacent subventricular zone.

The ISH level and density profile in Figure 10 indicates that quantitative methods in colorimetric ISH may encode some variation in spatial pattern and cell type, even when restricted to a set of genes with a comparatively narrow range of high microarray expression level. The four genes shown in Figure 11c-f illustrate the range of expression characteristics (widespread, regional, cell class specific) that can be seen within a particular brain region with cellular resolution techniques. Expression ranges from widespread to highly restricted patterns and can include both neuronal and non-neuronal cell populations. While this heterogeneity can be resolved to some degree with methods for isolating much smaller tissue samples, such as microdissection 67 or voxellation 68, the resolution of colorimetric ISH for distinguishing intermingled cell populations remains a significant benefit of the technique.

Adult mouse brain gene expression data for C57Bl/6J based on Affymetrix MG-U74Av2 chip (Teragenomics 557, Gene Expression Omnibus (GEO) GSE3594) and Affymetrix GNF1M chip (Mouse GNF1M(MAS5-condensed; GeneAtlas 4658, GEO GSE1133) were used for the microarray platform data. Both data sets have two replicates for six brain structures: striatum, cortex, cerebellum, hippocampus, olfactory bulb, and hypothalamus, except GNF in the cortex, which has one. TERA1 and TERA2 are replicates specified by GEO IDs. Data sets for TERA1 for the striatum, cortex, cerebellum, hippocampus, olfactory bulb, and hypothalamus are GSM82978, GSM82974, GSM82988, GSM82952, GSM82986, and GSM82953, respectively. Data sets for TERA2 for the striatum, cortex, cerebellum, hippocampus, olfactory bulb, and hypothalamus are GSM82979, GSM82975, GSM82989, GSM82951, GSM82987, and GSM82954, respectively. GNF1 and GNF2 are replicates identified by the column names of the downloadable data file GNF1M, MGJZ030207007A and MGJZ030207007B, respectively. For ISH, data from the ABA was gathered for these six brain structures following 2. Microarray, SAGE and ISH data are provided in Additional data files 3-5 and 8.

For ABA ISH data, genes with corresponding Entrez IDs and having positive expression values in one of six structures were mapped in the pipeline described in 217, including pre-processing steps, image registration and signal segmentation. From each registered ISH image series the expression level and density statistics defined can be computed using the formulas from signal segmentation masks (Figure 1c) on a structure basis. The gene selection process and counts in the comparisons are shown in Table 2. For microarray data, probes that mapped to a unique Entrez ID and have one-to-one mapping between gene and probe ID were chosen. Both Teragenomics and GNF microarray data sets have two replicates. Replicates were used in all platforms when available in order to examine the consistency in present/absent calls, excluding those with inconsistent values. The final column of Table 2 indicates genes labeled as having present expression by present/absent calling algorithms for each respective platform minus those genes without consistent calls for those data available in replicate (except for the ABA, which has minimal replicates in the sagittal plane). For ABA present/absent calling, the density threshold values are 1.2582 (striatum), 1.2267 (cortex), 0.7946 (cerebellum), 0.3198 (hippocampus), 1.7140 (olfactory bulb), and 1.9723 (hypothalamus). Following preprocessing, each structure has a variable gene count for comparison (Figure S9 in Additional data file 1).

ImagePro software (MediaCybernetics, Silver Spring, MD, USA) was used for the manual annotation of the region of interest and optical density measurements. The statistical computing package R 92 was used for figure generation.

Probe sequences of Teragenomics, GNF, and ABA were downloaded from their respective websites and target gene sequences were obtained from NCBI 93. Affymetrix MG-U74Av2 chip for Teragenomics and Affymetrix GNF1M chip for GNF both have 16 probes for each target gene. In each platform for each gene, the corresponding probe sequences were mapped against the target sequence using NCBI BLAST 60 and the aligned locations were recorded, measuring sequence overlap for aligned locations. The significance of the correlation difference before and after filtering out genes with no overlap was tested since the sample sizes changed significantly.