To characterize the gene expression changes occurring during positive and negative selection in a non-autoimmune strain, gene expression profiles were measured in three relatively homogeneous thymocyte populations from the B10^k strain 10. Pre-selection thymocytes ('PreS') that had not yet received positive or negative selection signals were identified as CD4⁺8⁺ DP cells that are negative for CD69 and 1G12 TCR clonotype and sorted from TCR transgenic and TCR insHEL double transgenic mice. Early single positive cells that were beginning to undergo positive selection (S+) were sorted from TCR transgenic mice as CD4⁺CD8^lowCD69⁺1G12⁺ cells. Early single positive cells with an identical cell surface phenotype but beginning to undergo negative selection (S-) were sorted from TCR insHEL double transgenic animals. Three independent pools of mRNA from sorted cells were analyzed by labeling and hybridization to Affymetrix 430A microarrays and normalized using MAS 5.0. MAS 5.0 was used as the normalization method in preference to more precise project-based model fitting systems, such as RMA, GCRMA and PLM, due to advantages in simple reduction of the false discovery rate. Use of the MAS 5.0 normalization method produces 'present' and 'absent' calls and, when used as a filtering device, this has been demonstrated to reduce the false discovery rate with little true cost to the true positive rate 16. Furthermore, by taking the mismatch probe into consideration, MAS 5.0 reduces the level of false positives based on cross-hybridization 16. It should be noted that all normalization methods involve a trade-off of accuracy or precision at various levels and, while MAS 5.0 performs strongly for background correction and medium and high intensity signals 17, it is less accurate than some alternatives for probesets in the low intensity range 1819. The complete dataset, with raw values and statistical analysis is given in Additional data file 1 and has been deposited in the NCBI Gene Expression Omnibus (GEO) 20 accessible through GEO Series accession number GSE3997.

We first performed a global analysis of the differences in gene expression between negative and positive selection and pre-selection by measuring the Euclidian distance between conditions (Figure 1). Measuring Euclidian distance involves treating each condition as a single point in n-dimensional space, where each dimension is the expression of a single probeset. The distance between two conditions can then be calculated as the straight line ('ordinary') distance between the two points (conditions), so that, for example, a low value between two conditions indicates that they have similar values for gene expression, while a high value between two conditions indicates that they have different values for gene expression 21. The advantage of Euclidian distance is that it is an approximation of the closeness of global gene expression profiles under different conditions, independent of classification of gene changes as significant or non-significant by including the number of genes that are unchanged, and the degree of change in differentially expressed genes (thus allowing subtle changes in large numbers of genes, which would not necessarily be detected if a statistical threshold was required, to impact global distance). Using Euclidean distance as the measure of global 'closeness', the largest difference was between pre-selection thymocytes and selecting thymocytes (both positive and negative selection). Negative selection was closer to the pre-selection condition than positive selection.

Individual gene expression changes were then analyzed in an independent method of assessment, by assigning probesets into categories based on statistical significance of gene expression pattern in the 3 replicates, allowing categorization into 12 possible patterns. Assignment to the 12 patterns was based only on the direction of significant gene expression change rather than degree of change (Figure 2). Global expression 'closeness' could then be estimated by comparing the number of probesets that fit each category.

This measure of global expression 'closeness' between conditions gave a result consistent with the Euclidean analysis, where the pattern categories with the largest number of probesets were those (patterns A and G in Figure 2) displaying an expression difference between pre-selection and selection (both positive and negative selection) but no difference between positive and negative selection. These are likely to represent genes that are developmentally regulated as part of the differentiation of immature DP cells into more mature early SP cells. Some may be induced or repressed by TCR signaling mechanisms that are unable to distinguish between TCR engagement by weak positively selecting agonists and the strong negatively selecting agonist. Pattern A comprised 531 probesets that were induced during maturation/selection, and pattern G comprised 692 probesets that were repressed. The induced genes included well established markers and mediators of maturity, Tcr, Ccr7, S1P1 and IL7R, and genes that are known to be targets of positive and negative selection TCR signals, such as the calcineurin-response genes Ian1, Egr2 and CD52 and ERK-response genes Nab2 and Zfp36l1. The repressed genes included Rag1, preTαβ, CD8α and CD8β, known to be involved in thymocyte maturation changes, and cell cycle genes Cdc7, Cdk2, Cdk2ap1, and Cdk4, which are consistent with the exit from cell cycle that accompanies maturation of early DP cells, which are cycling, into later DP and SP cells, which are non-cycling. The identification of these expected expression changes acts as an internal validation of the dataset.

As well as these previously identified markers of positive and negative selection, several important T cell regulatory genes were also found here to be upregulated in selection: those encoding nucleic acid binding proteins, such as Bcl3 22, Dicer123, Fosb 24, Hivep1 14, Irf1 25, Irf4 14, Irf7 26, JunB 27, Klf2 28, Nfat5 29, Rora 30, Stat1 13, and Stat6 31; signaling-associated genes Ccnd2 14, Evl 32, Hspa5 33, Ly9 34, Mcl1 35, Ndfip1 33, Psen2 36, Rassf5 37, Upf2 38 and Zfpn1a 39; and those encoding receptors, such as Crry 40, Gpr83 41, 2-K1 42, Ms4a6b 43, Sema4a 43, Slamf1 43, Tlr1 44, Tnfsf11 45, and Tslpr 46.

Previously unassociated genes identified as part of this analysis are, therefore, excellent candidates for novel maturation markers, and include those encoding nucleic acid binding proteins, such as 1810007M14Rik, AA408868, Aptx, Arts1, D11Lgp2e, D1Ertd161e, Ddef1, Ddx19b (2810457M08Rik), Dedd2, Dnmt3a, Elk3, Hnrpa1, Hrb, Ifih1, Isgf3g, Mxd4, Nab1, Rab8b, Rbms2, Rnaset2, Rpl12, Rpo1-1, Skil (9130011J04Rik), Sp100, Tcf3, Tef, Trim21, Trim30, Wasl, Zbp1, Zcchc7, Zfp260, Zfp313, and Zfp445 (AW610627); signaling-associated genes Bin1, Dlgh1, Myd88, Nedd9, Pacsin1, Pscdbp, Sdc3, Shkbp1, Sytl2, Traf1, Vps28, Xpo6; and those encoding receptors, such as 1810011E08Rik, Brd8, Cd9, Folr4, Ptger2, Ptgir, Sorl1, and Tlr2. The complete set of genes identified in this category is listed in Additional data file 2.

The next largest pattern categories comprised probesets that were unique to positive selection, consistent with this condition being the most differentiated based on the Euclidean analysis. These categories (Figure 2) included genes that were induced (patterns B and C) or repressed (patterns H and I) during positive selection, but were either not altered at all during negative selection (C and I) or underwent a change of lesser magnitude but in the same direction (B and H). Genes in these categories are candidates for translating TCR engagement by weak agonists into survival and maturation rather than negative selection. However, this category will also include genes that are developmentally regulated at later stages of SP cell maturation, after CD69 induction and increased TCR surface expression, since negative selection will remove SP cells before reaching this stage. Patterns B and C comprised 278 probesets that were preferentially induced during positive selection, including the well established functional genes CD3ζ and the calcineurin-response gene Itgb7. Patterns H and I comprised 394 probesets that were preferentially repressed during positive selection, including developmental markers Cxcr4, CD8α, CD24a, CD25, and cell cycle genes Cdc2, Cdc6, Cdc20, Cdc25b, Cdka2c and Myb.

As well as these previously identified markers of positive selection, a number of important T cell regulatory genes were also found here to be upregulated in positive selection: those encoding nucleic acid binding proteins, such as Dbp 33,Foxo1 47 and Zfp67 48; the signaling-associated gene Stam2 49; and those encoding receptors, such as Il6ra 50 and Itgb2 51.

Previously unassociated genes identified as part of this analysis are, therefore, excellent candidates for novel markers of positive selection, and include those encoding nucleic acid binding proteins, such as Bhc80, Ddb2, Ezh1, Gata1, Pcbp4, Rbms1, Smarca2, and Trim26, Ddit3, Foxp1, Oas2, Tgif2, and Zfp467; signaling-associated genes Arrb1, Bcap31, Emid1, L1cam, Numb, Pea15, Rabip4, Rcbtb1, Rsn, Selpl, and Sh3gl1; and those encoding receptors, such as AA691260, D7Ertd458e, Frag1, Gpr97 Grina, Il6st, Paqr7, Robo3 and Sh2d3c. The complete list (Additional data file 2) dramatically expands the set of candidate mediators and markers for understanding positive selection and SP cell maturation.

A smaller transcriptome of 246 probesets comprised genes that were preferentially or selectively induced or repressed during negative selection (patterns D, E, F, J, K, L in Figure 2), consistent with negative selection being closer to pre-selection in the global analysis. Genes in these categories are candidate mediators or markers of negative selection and thymocyte apoptosis. Pattern categories E and K were selectively induced or repressed during negative selection, showing no change in expression between pre-selection cells and positive selection. Pattern E comprised 87 probesets, including the TCR-induced pro-apoptotic gene Bim (Bcl2l11), which has previously been shown to be induced selectively during negative selection in this system and is an essential mediator of negative selection, and activation markers such as Ccr6.

Genes in patterns D and J were induced or repressed more strongly in negative than positive selection. Probesets in these categories are likely to include genes that report the quantitative differences in TCR signaling thought to differentiate strong TCR engagement by negative selecting agonists from weak engagement by positively selecting agonists. Pattern D comprised 48 probesets, including the gene encoding the thymocyte apoptosis-inducing transcription factor, Nur77, markers of activated T cells or regulatory T cells, Gitrd, Ox40 and 41bb, and the ERK-response gene Fos. Category F and L comprised a small number of probesets that exhibited a change in expression during negative selection that was in the opposite direction to that induced during positive selection. Pattern L includes genes such as CD25, CD24a and Annexin A4.

Overall, the negative selection transcriptome is quite small: 144 induced probesets, and 102 repressed probesets. By contrast, positive selection of early DP thymocytes into early SP thymocytes induced 809 probesets and repressed 1,086 probesets - a transcriptional program that is eight-fold larger. A complete list of the positive and negative selection transcriptomes, all the genes falling into patterns A to L, in the B10^k strain is given in Additional data file 2.

Recent studies have found that sets of genes induced by similar stimuli can co-localize in genomic clusters 52. To determine if positive or negative selection likewise work by the activation or suppression of broad clusters of genes we queried the data to identify cytogenetic bands with a significantly increased proportion of induced or repressed genes, normalized to their gene density.

Focusing first on genes associated with positive selection, there is evidence for a small degree of gene expression change by cytogenetic band. Six cytogenetic bands were broadly suppressed upon positive selection, and no cytogenetic bands were broadly activated. Of the six altered cytogenetic bands, three meet stringent false discovery rates, 12F, 3B and 2F (Table 1).

The gene expression changes induced upon negative selection, by contrast, are highly localized. While the global difference (Euclidian distance) between positive selection and negative selection is only half that of positive selection to pre-selection (Figure 1), a greater number of cytogenetic bands show enrichment of gene expression differences. That is, while fewer genes were changed, and with a lesser magnitude, in positive selection when compared to negative selection than in positive selection compared to pre-selection, the genes that were differentially expressed were highly co-localized to certain genomic regions. One region was broadly activated in response to negative selection, 2F, while seven regions were broadly suppressed upon negative selection, of which two meet stringent false discovery rates, 11E and 6B (Table 1). Region 2F is also of interest because it is a region that contains a cluster of genes that decreased in expression during positive selection (Figure 3a), and a cluster of genes that increased in expression during negative selection (Figure 3b). The induced and repressed clusters are, however, largely distinct, with little correlation (Figure 3d). Of significance, the key pro-apoptotic gene Bim is encoded within 2F. Bim is controlled at least partially by chromatin structure, as histone deacetylase inhibition allows the spontaneous induction of Bim followed by apoptosis 53. Of the 2F probesets changed by negative selection in the B10^k strain, the highest level of upregulation is observed in Dut, Cops2, Dusp2, Bub1, Bim, Mertk, Smox, 6330527O06Rik, LOC545465 and 2610101J03Rik. Similarly, the 2F probesets with the greatest defects in upregulation in the NOD^k strain are Cops2, Galk2, Bub1, Bim, 1600015H20Rik, IL1a, Smox, Bmp2, Hao1, 6330527O06Rik and 2610101J03Rik. These data generate the hypothesis that the cytogenetic band 2F contains a concentration of apoptotic initiators for negative selection that may be coregulated by chromatin structure.

In parallel with the above analyses of negative and positive selection in B10^k mice, the same cell subset markers, sorting methods, and mRNA labeling and hybridization to microarrays were applied to pre-selection (PreS), early positive selection (S+) and early negative selection (S-) thymocytes from TCR and TCR insHEL animals on the NOD.H2^k strain background. This allowed developmentally matched, homogeneous populations of T cells to be traced during positive and negative selection using the same TCR and self peptide-MHC ligands, but carrying all the NOD genomic differences from B10 with the exception of the congenically matched H2^k haplotype.

The global gene expression differences between pre-selection DP cells and early SP cells undergoing positive or negative selection were first used to compare these states by Euclidian distance (Figure 1). The difference between pre-selection and positive selection thymocytes was similar on the NOD^k background (23.4 units) to that observed on the B10^k background (26.7). On the NOD background, however, there was much less difference between positive and negative selection, with the Euclidean distance between these states decreased from 13.9 in B10^k to 8.9 in NOD.

Individual gene expression differences between pre-selection, positive selection and negative selection on the NOD background were categorized into the same 12 patterns, as conducted above for the equivalent cells from B10 strain animals (Figure 2).

Focusing first on patterns A and G, representing probesets that were induced or repressed equivalently during positive or negative selection, these categories contain the largest number of genes that were induced (787 probesets) or repressed (994 probesets), which are comparable to the numbers observed for these categories in the B10 strain (Figure 2). Again, category A includes genes that are developmentally increased during maturation from DP to SP cells, such as IL7R, and genes that are induced by TCR signals during positive and negative selection, such as calcineurin-response genes Ian1, Egr2 and CD52 and ERK-response genes Nab2 and Zfp36l1. In total, 240 of the probesets assigned to category A in NOD were also assigned to this category in B10. The stringent cut-offs used to assign probesets to pattern categories underestimated the similarity of gene expression during DP to SP maturation on the two strain backgrounds, because only 39 of the 531 pattern A probesets in B10 thymocytes have significantly different values from NOD for the corresponding cell types. Likewise, genes that were decreased during maturation (pattern G) included expected developmentally regulated genes such as Rag1, Cd8α, PreTα and cell cycle genes. Of the 692 B10 pattern G probesets, only 56 have significantly different values to NOD for both positive and negative selection. Thus, the NOD background had little effect on gene expression changes associated with early SP maturation from pre-selection DP cells.

By contrast, the NOD background has markedly reduced numbers of genes with expression patterns that differentiate negative from positive selection (patterns B to F and H to L), consistent with the smaller Euclidean distance between these two states in the global analysis. Thus, of the patterns with increased expression in both positive and negative selection (A, B, D), 79% show the same degree of regulation during positive and negative selection (assigned to pattern A) in B10^k mice, but 94% show the same degree of regulation in NOD^k mice (with NOD^k pattern A containing many probesets assigned to patterns B or C in B10^k mice). Likewise, of the patterns with decreased expression in both positive and negative selection (G, H, J), 72% show the same degree of regulation (assigned to pattern G) in B10^k mice, but 95% show the same degree of regulation in NOD^k mice. By this assessment, positive and negative selection are less distinct in NOD^k mice than in B10^k mice.

Focusing specifically on genes that were preferentially or selectively induced (patterns D, E, L) or repressed (J, K, F) during negative selection revealed a global dampening of the negative selection response in the NOD background (Figure 4). Patterns D, E, and L, comprising probesets that were induced during negative selection either selectively (E, L) or to higher levels than during positive selection (D), contained only 66 probesets in NOD mice, whereas these sets were more than twice as large (144 probesets) on the B10 background. Moreover, of the 144 B, E or L probesets that were specifically induced during negative selection in B10 mice, 137 were diminished in expression during negative selection in NOD mice, 112 by more than 20% and 82 by a significant amount (Figure 4a). Thus, as noted previously, Bim induction (category E in B10) is undetectable in NOD thymocytes, while Nur77 induction (category B in B10) is greatly diminished.

Similarly, genes that were selectively decreased in negative selection (patterns J, K, F) accounted for only 46 probesets in the NOD strain compared to 102 in B10. Of the 102 probesets decreased upon negative selection in B10 mice, 100 remained at higher levels during negative selection in NOD^k mice, 84 by more than 20% and 44 significantly so (Figure 4a). Combining both transcriptional increases and decreases, of the 246 probesets specifically changed by negative selection in the B10^k mouse, 51% were significantly less changed in the NOD^k mouse. By contrast, of the 531 pattern A probesets that were increased equivalently during positive and negative selection in the B10^k mouse, only 7% had significantly different expression during negative selection in NOD^k compared to B10^k mice, and the majority showed similar expression (Figure 4b).

The presence of reduced upregulation and downregulation across the entire spectrum of negative selection-specific genes in NOD^k thymocytes indicates that upstream effects are at least partially responsible. This observation, recognizable only at a genomic level, was not predicted in previous analyses that focused solely on changes in downstream effectors, such as Bim and Nur77 1011. Such a defect may be occurring at the early signaling synapse, in line with a recognized alteration of TCR signaling components, such as enhanced Fyn kinase activity, differential activation of the Cbl pathway, impairment of membrane-translocation of Son of sevenless (mSOS) Ras GDP releasing factor, and the exclusion of mSOS and Phospholipase C (PLC)-γ1 from the TCR-Grb2-Zap70 complex, resulting in hypoactivation 54. Altered signaling in the basal TCR apparatus may be responsible for the reduced surface CD3 levels present on TCR transgenic thymocytes (without the presence of insHEL) in the NOD^k strain compared to their B10^k counterparts, with a 50% reduction at the DP stage, and a 20% reduction at the SP stage (Figure 5).

The NOD background also caused a large decrease in probesets assigned to categories B, C, H, and I, comprising genes that are preferentially or selectively altered during positive selection (Figure 2). This result has two non-exclusive explanations. First, there may be a less efficient positive selection response in NOD. Alternatively, many of the genes in this category may normally be developmentally regulated to appear at later stages of SP cell maturation, before CD69 is lost but at a stage when negative selection would have removed most such cells.

In addition to the altered negative and positive selection response above, the NOD background also had altered pre-selection gene expression in TCR^lowCD69^- DP cells, which may set the stage for altered responses when the cells encounter negative selecting antigens. Six independent pools of pre-selection DP cells were analyzed on both B10 and NOD backgrounds: three from TCR animals and three from TCR insHEL animals. There were few differences between TCR and TCR insHEL pre-selection pools within a strain background, consistent with sorting for antigen-nasïve thymocytes that had yet to display TCRs for HEL and induce CD69. Comparing pre-selection cells between the strains at a global level first (Euclidian distance, Figure 1), the difference between these states (14.2) was approximately half that of the difference between pre-selection DP and early positive selection SP cells (26.7), with a total of 1,484 probesets significantly different between NOD PreS and B10 PreS (1,484 probesets). It is unknown if this degree of pre-selection divergence is specific to the NOD^k strain, or if it is observed across multiple strains based on comparative divergence.

In terms of genomic location, these changes are particularly concentrated in 20 cytogenetic regions (Table 1). Twelve regions show increased expression in B10^k pre-selection thymocytes, eleven of which meet stringent false discovery rates: 8E, 7E, 18E, 12F, 6C, 1B, 15D, 5F, 10C, 19A and 4E. Likewise eight regions show increased activity in NOD^k pre-selection thymocytes, all of which meet stringent false discovery rates: XF, 3E, XD, 3H, 5C, 12C, 1A and XA. With regard to the phenomenon of defective negative selection in the NOD^k strain, it may be of relevance that two of these regions co-localize with genomic loci that contribute to defective negative selection 10, 7E and 15D.

Gene expression differences between NOD^k and B10^k strains induced upon negative selection were also analyzed for cytogenetic clustering. Only four cytogenetic bands show enrichment, after eliminating regions changed in the basal (that is, pre-selection thymocytes) state. Two regions, 2F and 3A, were broadly suppressed in the NOD^k strain, and two regions were broadly activated in the NOD^k strain, neither of which meet stringent false discovery rates (Table 1). The region 2F is of particular interest for several reasons. Firstly, this is the only region that was broadly activated upon negative selection in the B10^k strain (Figure 3b). Secondly, it is one of only two regions that show broad strain differences in regulation upon induction of negative selection, with lower activity in the NOD^k strain (Figure 3c). Thirdly, this region overlaps one of the six identified loci with a causative effect in defective negative selection in the NOD^k strain 10. A comparative analysis of the gene expression changes in B10^k and B10^k-NOD^k negative selection demonstrated that this locus comprises the same cluster of genes that are upregulated upon negative selection in the B10^k strain and show poor induction of gene expression in the NOD^k strain (Figure 3e). These data indicate that the NOD^k strain has a genetic defect preventing the efficient induction of the negative selection 2F cluster, including Bim, preventing initiation of apoptosis.

Combining the global transcription profiles for NOD^k and B10^k pre-, positive and negative selection with linkage data for efficiency of negative selection in the same in vivo conditions 10 provides a way to identify candidate genes responsible for the NOD trait of defective negative selection. While expression differences are promising candidates for quantitative traits, we recognize that this approach is unable to detect allelic variants arising from differential mRNA splicing, such as the Idd5 allele of Ctla4 55, or from amino acid substitutions, such as the Idd13 polymorphism in β2 m 56.

Expression pattern was first used to identify six categories of high priority candidate genes (Figure 6, Additional data file 2). Group 1 consists of probesets that were preferentially increased in B10^k negatively selecting thymocytes (patterns D, E and L), but were significantly less increased in NOD^k counterparts. There are 77 probesets in this group, of which 14 show poor induction and 63 show no induction. Defective apoptotic initiators could fall in this group. Group 2 consists of probesets that were increased in B10^k thymocytes upon selection (pattern A), but were significantly less increased in NOD^k mice (p < 0.05). Only 8 probesets are in this group, of which seven show no increase upon maturation in the NOD^k strain. Defective functional prerequisites for negative selection switched on during positive selection could fall in this category. Group 3 consists of probesets that were significantly lower in NOD^k thymocytes compared to B10^k thymocytes for each biological condition. There are 151 probesets in this group, 80 of which show no development or selection-induced differences within the NOD^k or B10^k groups. Defective constitutively expressed prerequisites for negative selection could fall in this group. Group 4 is the reverse of group 1. It consists of probesets that were increased in the NOD^k negatively selecting thymocytes, but were significantly less increased in B10^k thymocytes. This group is designed to catch candidate genes that were more strongly induced in NOD^k negatively selecting thymocytes and provided protection from negative selection. Only two probesets fall in this category. Group 5 is the reverse of group 2. It consists of probesets that were increased in NOD^k thymocytes upon selection (pattern A), but were significantly less increased in B10^k thymocytes. This group is designed to catch NOD^k over-expressed maturation-induced protective genes. Group 6 is the reverse of group 3, comprising probesets that were significantly higher in NOD^k thymocytes compared to B10^k thymocytes for each biological condition. There are 98 probesets in this group, 48 of which show no developmental or selection-induced differences within NOD^k or B10^k thymocytes. Constitutively over-expressed protective factors could fall in this group.

A matrix comprising the probesets in these six categories and the genomic location to a 30 cM bracket surrounding peak linkage to the four regions of NOD^k susceptibility to defective negative selection markers (D7mit101, D15mit229, D2mit490/Idd13 and D1mit181/Idd5) 10 identified 44 candidate probesets. Each region has 6 to 10 candidates using this method, except for the region centered on D2mit490, which includes the 2F cluster and has 20 candidates. These are discussed in more detail below, as summarized in Table 2.

As previously described 10, cell populations were purified from healthy 6-8-week old female mice, stained in 5 μg/ml actinomycin D and 2 μg/ml α-amanitin (Sigma-Aldrich, St Louis, MO) with CD4- fluorescein isothiocyanate (FITC), CD69-phosphatidylethanolamine (PE), CD8- peridinin chlorophyll protein (PerCP) and 1G12 indirectly labeled with Allophycocyanin (APC), then sorted with a Becton Dickinson (Franklin Lakes, NJ) FACVantage. Sorted populations were 'early DP' (CD4⁺CD8⁺CD69^-1G12^-) and 'early SP' (CD4⁺CD8^lowCD69⁺1G12⁺). Purified RNA underwent two rounds of in vitro RNA amplification before fragmentation and hybridization to Affymetrix GeneChip 430A arrays (Santa Clara, CA, USA).

The Affymetrix 430A chips were scanned using standard Affymetrix protocols. All arrays passed routine quality control assessment for hybridization and data quality. Expression values, referred to as probeset 'signal', were calculated using the Affymetrix GeneChip analysis software MAS 5.0, with a scaling chosen so that each array has a trimmed mean of 150. Following examination, the endogenous control probesets were removed. The MAS 5.0 signal values were then transformed to logarithms (base 2). Signal values of 0 were assigned a value of 0.1 before taking logs. Finally, the arrays were standardized to mean zero and variance one. Initial gene filtering kept only those probesets that were called 'present' by the Affymetrix signal detection algorithm in all replicates for at least one biological group. All statistical analyses were carried out using these transformed signal values. Statistical analysis was carried out in three data sets, one consisting only of the B10^k biological groups, one consisting only of the NOD^k biological groups, and one consisting of both B10^k and NOD^k biological groups.

Assignment of probesets to different patterns of gene expression was separately carried out on the B10^k and NOD^k datasets. The patterns are defined in terms of direction of change between means, rather than the extent of change, with assignment of a change having a statistical cut-off rather than a fold-change cut-off. For the separate B10^k and NOD^k datasets, two way analyses of variance (ANOVAs) were performed on each of the selected probesets. For each probeset, if the overall F-test for a test of mean difference was significant (p < 0.005), the probeset was considered significantly changed. Having established that the means were different, t-tests for the contrasts in means were used to determine significantly different means. A significance level of less than 0.05 was used for these tests. Contrasts between 'pre-selection' thymocytes (CD4⁺CD8⁺1G12^- CD69^-) from TCR transgenic and TCR insHEL double transgenic mice showed very few differentially expressed probesets, as is expected for populations that have not been exposed to HEL antigen due to low expression of TCR and anatomical segregation of antigen in the corticomedullary junction. As a consequence, these two groups were merged and the model re-estimated for all the selected probesets. Since there are now three groups, there are twelve distinct patterns of up, down and no difference. Contrasts between the means using the re-estimated models were used to assign probesets to a gene expression pattern, based on a logical set of significant changes between the various conditions. A three-way ANOVA model was used to analyze the combined B10^k-NOD^k dataset for differential expression between strains. The p values are used in this research as indicative of 'evidence' and, except in rare instances, will not be an exact measure of probability. Model assumptions for the use of the different tests were examined for a small number of significant probesets and found to hold. Following analysis, each probeset was annotated for genomic location and gene function using the FACTS (Functional Association/Annotation of cDNA Clones from Text/Sequence Sources) program 82.

Euclidian distance is the most common measure of metric distance, which is an approximation of the 'distance' between gene expression from replicates in one condition to replicates in other conditions. Euclidean distance is calculated by treating the expression of a group of probesets as a point in n-dimensional space, where the distance from the axis in each dimension represents the expression (the technical mean of transformed signal value for the replicates) of a single probeset. This produces a single point in n dimensions (where n is the number of probesets in the group) that represents one set of conditions, and a single point in the same n-dimensional space that represents the same group of probesets under different conditions. The Euclidean distance between each point (representing each condition) was calculated by using the square root of the sums of squared differences between the points in each dimension, using the formula:

Euclidean distance = √(x₁ - y₁)² + (x₂ - y₂)²+ (x₃ - y₃)² + ...

where x is the point representing the probeset group in condition x, with each dimension x₁, x₂, x₃, and so on being the expression of individual probesets within the group, and y is the point representing the probeset group in condition y, with each dimension y₁, y₂, y₃, and so on being the expression of individual probesets within the group 21.

The straight line distance between the two points thus calculated represents the difference in expression of the included probesets between the two conditions. The probeset group used to calculate the Euclidean distance here included all probesets present in all replicates for at least one condition, thus creating an approximation of the genome-wide similarity between two conditions 218384.

Gene Set Enrichment Analysis (GSEA) was used to examine the genome for areas of differential expression by cytogenetic band. GSEA R script (script defaults, standard method) 85 was used with the 'gene label' permutation method. Genes represented by multiple probesets were reduced to a single probeset with the smallest overall p value. A lower cut-off of 20 genes per cytogenetic band was used, which resulted in a total of 113 gene sets being tested for enrichment. A false discovery rate cut-off of 0.4 was initially used, with the stricter criterion of 0.25 for regions listed. Regions with enriched gene expression during positive selection (B10^k) refer to cytogenetic bands with an enriched number of genes with expression differences between B10^k pre-selection thymocytes and B10^k positive selection thymocytes (TCR transgenic early SP cells). Regions with enriched gene expression during negative selection (B10^k) refer to cytogenetic bands with an enriched number of genes with expression differences between B10^k positive selection thymocytes (TCR transgenic early SP cells) and B10^k negative selection thymocytes (insHEL:TCR double transgenic early SP cells). Regions with enriched strain differences in pre-selection thymocytes refer to cytogenetic bands with an enriched number of genes with expression differences between B10^k pre-selection thymocytes and NOD^k pre-selection thymocytes. Regions with enriched strain differences induced during negative selection refer to cytogenetic bands with an enriched number of genes with expression differences between B10^k negative selection thymocytes (insHEL:TCR double transgenic early SP cells) and NOD^k negative selection thymocytes (insHEL:TCR double transgenic early SP cells), with any regions showing enrichment at the pre-selection thymocyte stage removed.

We analyzed 6-10-week old mice (or 6 week post-reconstitution chimeric mice) as described previously 510 using the following antibodies: 1G12 anti-clonotype 86 (gift of E Unanue and D Peterson, Washington University, St Louis, MO, USA) culture supernatant followed by rat anti-mouse IgG1 allo-phycocyanin; anti-CD8α-PerCP; anti-CD4-FITC or PE; anti-Ly5a-FITC; anti-CD3-PE; and anti-B220-PE (all from BD PharMingen, San Jose, CA, USA).