How replicable are mRNA expression QTL? Jeremy L. Peirce,1,2 Hongqiang Li,1,2 Jintao Wang,3 Kenneth F. Manly,2,3,4 Robert J. Hitzemann,5 John K. Belknap,5 Glenn D. Rosen,6 Shirlean Goodwin,7 Thomas R. Sutter,7 Robert W. Williams,1,2 Lu Lu1,2,8 1 Center for Neuroscience, Department of Anatomy and Neurobiology, University of Tennessee Health Science Center, Memphis, Tennessee 38163, USA 2 Center for Genomics and Bioinformatics, Department of Anatomy and Neurobiology, University of Tennessee Health Science Center, Memphis, Tennessee 38163, USA 3 Department of Anatomy and Neurobiology, University of Tennessee Health Science Center, Memphis, Tennessee 38163, USA 4 Department of Pathology and Laboratory Medicine, University of Tennessee Health Science Center, Memphis, Tennessee 38163, USA 5 Portland Alcohol Research Center, Department of Behavioral Neuroscience, Oregon Health & Science University, Portland, Oregon 97239, USA 6 Department of Neurology, Beth Israel Deaconess Medical Center, Boston, Massachusetts 02215, USA 7 W. Harry Feinstone Center for Genomic Research, University of Memphis, Memphis, Tennessee 38152, USA 8 Key Laboratory of Nerve Regeneration in Jiangsu Province, Nantong University, Nantong, China 226001
Received: 22 December 2005 / Accepted: 20 March 2006
Abstract
Applying quantitative trait analysis methods to genome-wide microarray-derived mRNA expression phenotypes in segregating populations is a valuable tool in the attempt to link high-level traits to their molecular causes. The massive multiple-testing issues involved in analyzing these data make the correct level of confidence to place in mRNA abundance quantitative trait loci (QTL) a difficult problem. We use a unique resource to directly test mRNA abundance QTL replicability in mice: paired recombinant inbred (RI) and F2 data sets derived from C57BL/6J (B6) and DBA/2J (D2) inbred strains and phenotyped using the same Affymetrix arrays. We have one forebrain and one striatum data set pair. We describe QTL replication at varying stringencies in these data. For instance, 78% of mRNA expression QTL (eQTL) with genome-wide adjusted p £ 0.0001 in RI data replicate at a genome-wide adjusted p < 0.05 or better. Replicated QTL are disproportionately putatively cis-acting, and approximately 75% have higher apparent expression levels associated with B6 genotypes, which may be partly due to probe set generation using B6 sequence. Finally, we note that while trans-acting QTL do not replicate well between data sets in general, at least one cluster
Correspondence to: Lu Lu, Key Laboratory of Nerve Regeneration in Jiangsu Province, Nantong University, Nantong, China 226001; E-mail:
[email protected]
of trans-acting QTL on distal Chr 1 is notably preserved between data sets.
Introduction The genetic analysis of the patterns of gene expression in segregating populations was introduced just over a decade ago in prescient and remarkably sophisticated work by Damerval and colleagues (Damerval et al. 1994; de Vienne et al. 1994). Their approach and vision did not spread quickly, and not until largely parallel methods for quantifying mRNAs and proteins (Klose et al. 2002) became far more pervasive (an unfortunate hiatus of nearly 8 years) did experimental geneticists begin to realize the potential of the methods that they so clearly demonstrated. In the few years since the clarion call by Jansen and Nap (2001) we are now being inundated with new and huge gene expression data sets (Brem et al. 2002; Chesler et al. 2005; Schadt et al. 2003) that are beginning to exploit microarray platforms capable of routinely measuring steady-state levels of every known transcript and even every known exon. These data sets are so massive that standard publishing procedures cannot accommodate them. A number of these data sets, particularly those for mouse and rat, have been integrated into GeneNetwork (www.genenetwork.org), which is essentially a web knowledgebase in which the entire data set and relevant
DOI: 10.1007/s00335-005-0187-8 Volume 17, 643 656 (2006) Springer Science+Business Media, Inc. 2006
643
644
metadata (data about the data) are combined with sophisticated statistical and computation tools for the genetic dissection and synthesis of single traits or entire systems of traits. One challenge facing investigators in the interpretation of the massive data sets on GeneNetwork and elsewhere is deciding how much confidence to place in QTL extracted from still noisy array and proteomic platforms after having conducted many thousands of statistical tests with poorly understood degrees of dependency. Statistical approaches to address these types of challenges have been developed using either theoretically derived (Lander and Kruglyak 1995) or empirical (Churchill and Doerge 1994) p values as well as false discovery rates (Benjamini and Hochberg 1995; Benjamini and Yekutieli 2005; Storey and Tibshirani 2003). However, we still have a long way to go before we can couple statistical approaches with the huge amount of biological and biochemical data that forms a prior probability universe. Doss et al. (2005) have exploited a clever method to evaluate error rates that depends on a cis-trans test for gene eQTL data. Using the cis-trans test on B6 · D2 F1 animals, they estimated the ratio of transcript from the B6- and D2-derived chromosomes. A ratio significantly different from 1:1 indicates cis-control and is taken as a confirmation of the cis-acting eQTL. The results of this test allowed them to confirm 64% of a set of apparently cis-acting eQTL with LOD ‡ 4.3 (we use a genome-wide p value of 0.05 in our data). This is an excellent method for confirming cis-acting QTL and actually establishing their mode of action (here we assume that a QTL is cis-acting when its transcript and QTL map close together) but gives relatively low throughput and by definition is not applicable to trans-acting QTL. Traditionally, however, the best standard for deciding whether an association is real is replication, preferably using a different technology, population, or approach. For a single phenotype this is a manageable but nontrivial approach. Because of the expense of microarray data, this approach until recently has been limited to real-time polymerase chain reaction (rtPCR) or custom array confirmation of expression levels. Two recently collected pairs of data sets, one from brain and the other from striatum, offer the unique opportunity to directly address the question of replicability for mRNA abundance QTL. Both pairs consist of two mapping populations derived from B6 and D2 crosses, which means that extensive information, including sequence data, is available for both strains. Because the genomes of the
J.L. PEIRCE
ET AL.:
HOW REPLICABLE
ARE MRNA
EXPRESSION QTL?
parental strains are highly characterized, we were also able to take advantage of high-resolution genotyping data, especially for the RI strains. For each tissue (whole brain/forebrain, striatum) we have a recombinant inbred (BXD RI) and an intercross (B6D2 F2) mapping population phenotyped using the same Affymetrix array, which allows direct confirmation of results by examining the data sets for common QTL. There is increasing interest in gene expression genetics, particularly in genetic reference populations such as RIs and in segregating populations such as F2s. Currently, GeneNetworkÕs interactive knowledgebase, which includes deep phenotypic and genetic data and sophisticated analysis tools (Chesler et al. 2004; Wang et al. 2003), can be applied to mRNA microarray data for whole brain, hippocampus, striatum, cerebellum, eye, hematopoietic stem cells, and liver in the BXD RI mouse strains (and B6D2 F2 mRNA expression data on brain and striatum), and kidney and peritoneal fat in the HXB/BXH RI rat strains. In addition, hippocampal data for the Inbred Long Sleep (ILS) · Inbred Short Sleep (ISS) RI strain set (LXS), Arabidopsis RI strain data, and more BXD RI tissues are in the process of being added. While high-throughput proteome data is harder to obtain (although arguably more desirable) on a genomic scale, proteomic data sets soon will also be integrated into GeneNetwork. Expression genetics approaches, both protein and mRNA, continue to grow rapidly in power and popularity. Given this trend, estimates of the reliability of mRNA abundance QTL will assist researchers in deciding how much confidence to place in particular results when additional data sets are unavailable. Methods Array data phenotyping and normalization. All data sets were phenotyped using Affymetrix mouse M430 arrays. The brain data sets were phenotyped using the M430A/M430B two-chip version of this array, while the striatum data sets were phenotyped using the M430 v2.0 one-chip version. (While these arrays are often considered to be similar, we noted that there are actually sequence differences between data sets in 7.5% of probes. These differences are ignored in cases where comparison between data taken using these different array versions is attempted.) In all cases the data were normalized using the position-dependent nearest neighbor (PDNN) method (Zhang et al. 2003). Other normalizations are available on GeneNetwork. Detailed information about each of the BXD data sets and the whole-brain B6D2 F2 data set can be found via the ‘‘Info’’
J.L. PEIRCE
ET AL.:
HOW REPLICABLE
ARE MRNA
EXPRESSION QTL?
link on the GeneNetwork search page (http:// www.genenetwork.org/search.html) and at http:// www.ohsu.edu/parc. Genotype data. In the BXD RI strains, a recently completed single nucleotide polymorphism (SNP) genotyping effort has yielded a very-high-density marker set of 3795 combined SNP and traditional simple sequence length polymorphism (SSLP) markers, which were used for mapping with both brain and striatum data sets. SSLP markers were used to genotype both B6D2 F2 data sets, with 306 markers in the brain F2 and 75 markers in the striatum F2 data. Whole-brain and forebrain data. Our first pair of data sets consists of two somewhat different tissues generated in different labs. The first BXD RI data set in this pair is a set of forebrain arrays generated by LL and RW for the INIA-Stress (Integrated Neuroscience Initiative on Alcoholism) consortium and is publicly available on GeneNetwork. This data set consists of 42 BXD strains, both parental strains, and the B6D2 F1, each strain being represented by two to four arrays that are balanced as well as possible across sex, and each array including the same brain regions from approximately three animals closely matched by age and sex. Nine of the 121 M430A/B array pairs were taken from the same samples used by Chesler and colleagues (2005), which were previously typed using U74Av2. In only one strain, however, (BXD5) were all samples from this older data set. In brief, cutting from the caudal border of the inferior colliculus on the dorsal side and extending the cut ventrally to the basis pedunculi and the pons on the ventral side excised the left or right hemisphere of the forebrain and midbrain. Total RNA was then extracted with RNA STAT-60 (Tel-Test) according to the manufacturerÕs instructions, and each array pair (M430A, then M430B) was hybridized in that order. We used the April 2005 data freeze available on GeneNetwork for all analyses. The brain B6D2 F2 data set was generated by RH and consisted of whole-brain, single-hemisphere samples from 56 F2 animals (25 males, 31 females), with one array used per sample. The tissue samples were dissected at the saggital midline and included forebrain, midbrain, one olfactory bulb, the cerebellum, and the rostral part of the medulla. The medulla was trimmed transversely at the caudal aspect of the cerebellum. Samples were well balanced with respect to left and right hemisphere. RNA was then extracted with Trizol reagent (Life Technologies Inc., Rockville, MD) according to the manufacturerÕs
645
protocol and purified using RNeasy (Qiagen, Inc., Valencia, CA). We used the August 2005 data freeze available on GeneNetwork for all analyses. Clearly, these are somewhat different samples, having variation not only in dissection protocols and resulting tissue but also between laboratories, animal care, and similar parameters. Some QTL that fail to replicate will then do so because of these variations, so results in these data sets and the striatal data sets described below will produce more conservative estimates of replication frequency than similar data sets gathered using the same protocols and personnel. Striatal data. Striatal data were gathered in the BXD RI strains by GR under the aegis of the Human Brain Project using a slightly different Affymetrix array (M430 v2.0) on 29 BXD strains and both parental strains. Balance by sex was attempted. Brains were hemisected midsagittally, the hippocampus removed, and both striata dissected using a medial approach developed by GR. Striata from three or four subjects matched for sex and age (P60 ± 4) were pooled for each sample. We used the April 2005 data freeze available on GeneNetwork for all analyses. The B6D2 F2 striatum data set was generated by RH and JB and consists of dorsal striatum dissected from a total of 60 B6D2 F2 animals as well as parental strains. This B6D2 F2 data set is not yet publicly released but additional information can be obtained by contacting the authors. We used the September 2005 data freeze for all analyses. Generating p values for all markers and interpretation of genome-wide p values. QTL Reaper is a program written by KM and JW that performs the QTL mapping tasks built into WebQTL, but in a rapid batch mode for all gene expression data in a data set. The output of QTL Reaper is usually the marker with the best p value for a given probe set, but was modified for our purposes to output p values for all markers for a given data set. QTL Reaper calculates p values for each probe set by permuting the gene expression values and calculating a distribution of best genome-wide likelihood ratio scores (LRS) for that probe set. (The commonly used LOD score is equal to the LRS divided by 4.6; twice the natural log of 2.) The p value is calculated by comparing the actual LRS score with the distribution of permuted best LRS scores for that probe set (Churchill and Doerge 1994) and is thus an empirical genome-wide p value (pgw) when applied to the best QTL for a given probe set. Because we are allowing multiple QTL per probe set, it should be noted that
646
the ‘‘p values’’ calculated for all but the best QTL have not been shown to have all of the properties expected of a p value and should thus be interpreted carefully. We ran 1,000,000 permutations giving us at best empirical pgw < 0.000001. QTL Reaper is available for download at http://sourceforge.net/ projects/qtlreaper. It is important to note that the p values we are reporting throughout this article are genome-wide adjusted p values. It is more common to provide LRS or LOD scores or nominal p values (often calculated from the LRS score as p = 10)(LRS/4.6)) but genomewide adjusted p values can be compared more meaningfully between data sets. To clarify this distinction, we refer to genome-wide adjusted p values as pgw and to nominal p values, when applied to transcripts, as pnom. The correspondence between LRS score and pgw value depends on the distribution of the phenotype. For the BXD RI brain data, the average correspondence at levels of significance commonly assessed in this article are pgw = 0.10 fi LRS = 14.3 ± 2.0 fi pnom = 0.0008, pgw = 0.05 fi LRS = 15.6 ± 2.4 fi pnom = 0.0004, pgw = 0.01 fi LRS = 18.7 ± 3.2 fi pnom = 0.00009, pgw = 0.001 fi LRS = 22.1 ± 4.5 fi pnom = 0.00002. The regression model used in both the BXD RI and the B6D2 F2 populations for QTL mapping allows only additive components to facilitate comparison between these data sets, except where specifically noted. Generally, F2 mapping uses a free regression model that allows identification of QTL with a large dominance component, which cannot be detected with RI data since all genotypes are homozygous. Identifying QTL in each data set. The question of what constitutes one QTL is in many cases, nontrivial. We chose to implement a simple conservative approach to defining a single QTL that will tend to identify regions that include multiplelinked QTL as a single QTL. This method, implemented using the short Java program QTLSearch.java, searches first for all markers with smaller p values than both of their neighbors. From this large list, the best p value in any 50-Mb region is chosen as the best QTL candidate for that region. Parsing QTL parameters. This list of QTL candidates from each data set can then be winnowed and filtered using a number of criteria, including p-value range, expression range, and measures of probe set quality. In addition, factors such as the distance between gene and QTL that qualify as a cis-QTL can be adjusted for each data
J.L. PEIRCE
ET AL.:
HOW REPLICABLE
ARE MRNA
EXPRESSION QTL?
set. These parameters are filtered and adjusted using the short Python script QTLParse.py. The script also generates descriptive statistics and histograms of the number of QTL per probe set. It then creates output for both parsed QTL and related statistics. Both of these programs can be found at http://www.nervenet.org/papers/reliability along with documentation on parameters and required file formats. Identifying common QTL. QTLParse.py also identifies common QTL between data sets. To be considered common, QTL are required (1) to be on the same chromosome and (2) to have the same additive effect sign and (3) same putative mode of operation (cis- or trans-acting) and (4) to have peaks within a specified distance of each other in both data sets. Because QTL must have the same mode of operation, for common cis-acting QTL the sum of the maximum distances that a peak may be from the transcript in the two data sets is also the maximum distance between peaks that the two QTL could have and still be considered common. (Any value of the parameter for maximum distance between peaks larger than this will not change whether the cisacting QTL is considered common or not. In practice, we have generally set the maximum distance between peaks to a number smaller than this sum). Phenotypic variance at different levels of significance. The proportion of the phenotypic variance accounted for by a mRNA abundance QTL (v2) at these significance thresholds can be estimated as approximately LRS/N, where N is the number of genotypes and LRS is df = 1, which is true for the RI data (Rosenthal 1994) and will also be true for brain and striatum B6D2 F2 data constrained to an additive effects-only model. [The LRS for B6D2 F2 using a free regression model is df =2; but we can still obtain estimates if we assume that the observed LRS results from only the additive effects of a QTL (thus LRS has df = 1). In this case we would convert the observed LRS (df = 2) to the equivalent LRS (df = 1) in terms of the associated p values.] To calculate these thresholds, we used average LRS at different genome-wide p values. Bias toward B6 QTL. Since the sequence against which probe sets were chosen is based on the B6 inbred strain, one possible source of spurious (in the sense that it does not represent a real difference in expression rather than array signal) cis-acting QTL would be SNPs or other variations between B6 and D2 sequence in the probes themselves. This kind of bias would result in replicable QTL that are
J.L. PEIRCE
ET AL.:
HOW REPLICABLE
ARE MRNA
EXPRESSION QTL?
647
indistinguishable individually from QTL representing actual differences in gene expression. If we start with the assumption that gene expression variation between B6 and D2 is equally likely to cause higher B6 or D2 expression levels, however, our expectation will be a 1:1 ratio of QTL with higher B6 expression to QTL with higher D2 expression. Deviation from this ratio in favor of excess QTL with higher B6 expression might represent the effects of sequence variation in the array probes. Results Optimization of maximum distance between peaks and definition of cis-QTL. We first chose the maximum transcript-to-QTL peak position distance at which we would consider a QTL cis-acting. Most cis-acting regulation occurs within a few thousand base pairs of a geneÕs coding sequence (Wray et al. 2003), so the mapping resolution we could expect for a cis-QTL, rather than the position of the actual regulatory difference, which we assumed was very near the transcript start site, was our primary concern. We wanted to choose values that would reflect the marker densities of our mapping populations while minimizing the likelihood that we would misidentify a trans-acting QTL as cis-acting. Since a replicated cis-acting QTL must be identified as cis-acting (see the subsection Identifying common QTL above) in both data sets, we apply stringent criteria to the densely genotyped RI data set (3795 markers) and require that the putative cisacting QTL peak be within 5 Mb of the transcript, a distance similar to that used in other experiments that investigated cis-acting QTL (Schadt et al. 2003). This means that the QTL peak can occur only in the 0.3% of the genome within that radius of a given transcript. Because at the minimum pgw < 0.05 criterion of significance there are only 1.06 QTL per transcript, our misidentification of trans-acting QTL as cis-acting at this level of resolution should also be near 0.3%. For the brain and striatum F2 data sets, we chose somewhat larger intervals based on marker density since the brain F2 data has one marker every 10 Mb on average and the striatum F2 data has one marker every 40 Mb on average. To simplify interpretation of the maximum window between the transcript position and the QTL peak for cis-acting QTL, we chose to make this value equal to the maximum distance between QTL peaks. To decide where to set the peak-to-peak and transcript-to-peak maximum distance, we tested the reduction in cis-acting QTL in our RI data set. Reducing allowable distances from a very liberal
Fig. 1. Mean QTL peak-to-transcript distance as a function of QTL significance. This figure compares the average values of the distance between the QTL peak (best marker) and the location of the transcript. We used striatum data to emphasize the effect of marker density. BXD RI values (3795 markers) are shown in dark gray and B6D2 F2 values (75 markers) in light gray with standard deviation error bars.
baseline for both (40 Mb) to our 5-Mb criterion reduced cis-acting QTL yield by 7%. To ensure a conservative criterion, we reduced this interval in each F2 data set until 10% of cis-acting QTL had been eliminated (required pgw < 0.05, cis-QTL definition less than 5 Mb between transcript and marker in RI data). For the brain F2 data set, this occurs at a distance of 10 Mb, and for the striatum F2 data set it occurs at 25 Mb. These distances are standard for all analyses except the analysis of the effect of p values on QTL peak-to-transcript distance below. Effect of p value on QTL peak-to-transcript distance in cis-acting QTL. While we did not adjust the common QTL or transcript-to-QTL peak distance criteria for different p-value ranges, both the mean QTL-to-peak distance and standard deviation decreased as p value improved when enough markers were present, as shown in Fig. 1. (Much looser definitions of cis-QTL were used in Fig. 1 to avoid biasing results by truncating relatively inaccurately mapped cis-QTL: 25 Mb allowable transcript-to-QTL peak distance and 25 Mb allowed peak-to-peak distance. Given that there are only 1.06 QTL per transcript at pgw < 0.05, our minimum stringency, this should not result in a high frequency of misidentification of a trans-acting QTL as cis-acting.) In the BXD RI striatum and brain (not shown) data sets, for instance, the mean transcript-to-QTL peak distance is less than 1.4 Mb when pgw < 0.00001, less than a third the mean distance when 0.01 £ pgw £ 0.05, but
648
Fig. 2. Whole-brain mRNA abundance QTL number by significance level. This figure plots the number of QTL present in the whole-brain data sets for BXD RI and B6D2 F2 populations versus genome-wide significance level, given as )log pgw. (Note that these are genome-wide adjusted p values, not nominal p values.) The numbers of QTL (left axis) are represented by the solid lines, while the percentage of BXD RI QTL confirmed in the B6D2 F2 data set are indicated by the dashed line (right axis).
it did not appreciably gain resolution below this level of significance. The B6D2 F2 striatum data set, which was genotyped at 75 markers, did not gain resolution below pgw = 0.001, suggesting that marker density was limiting. Expression QTL yield. About the same number of QTL, approximately 4700 4900, were identified in brain at the lowest stringency (pgw = 0.05) from the B6D2 F2 population and the BXD RI population (Fig. 2). The BXD RI population in striatum also yielded approximately the same number of QTL (5300), while the striatum B6D2 F2 yielded only approximately 600 QTL at this stringency, probably because of the considerably smaller number of markers (75). This will often cause nearby QTL to be identified as a single QTL, and the best markers available for a given QTL will be a greater average distance from their associated genes of effect (as indicated in Fig. 1). For consistency and simplicity in generating statistics and figures, we have generally considered the F2 population to be our confirmatory population. In other words, the percentage of QTL confirmed is always calculated as the ratio of common and total QTL in the RI strain set. Obviously, this is an arbitrary perspective and will shift depending on the intent of the researcher. To compare equivalent data sets, we have also chosen to restrict our F2 mapping model to additive effects only. Removing this restriction and performing a free (additive and dominance effects allowed) regression in the F2 population greatly
J.L. PEIRCE
ET AL.:
HOW REPLICABLE
ARE MRNA
EXPRESSION QTL?
Fig. 3. Overlap of brain QTL between BXD RI and B6D2 F2
data sets. This Venn diagram shows the overlap of QTL at a genome-wide significance level of Pgw £ 0.001 for the whole-brain data sets. (QTL were required to reach at least pgw £ 0.001 in both the B6D2 F2 and BXD RI data sets for inclusion as common QTL.) There are a total of 176 QTL in the B6D2 F2 and 1357 QTL in the BXD RI data set, 830 of which are common.
increases the yield of QTL. At a genome-wide adjusted threshold of pgw = 0.05 in the additive-only model, for instance, the brain B6D2 F2 data set has 4709 QTL. A free regression under the same criteria has 16,358 QTL, a gain of more than threefold. Confirmation rates are also slightly higher using the free regression model. The percentage of confirmed QTL in the RI data set climbs relatively rapidly as the genome-wide p-value criterion is made more conservative (Fig. 2). At pgw £ 0.001 in both data sets, for instance, 61% of QTL identified in the RI data set are also found in the brain F2 data set, as shown in the Venn diagram in Fig. 3. Different confirmation rates of cis- and trans-acting QTL. Trans- and putative cis-acting QTL identified in our BXD RI populations are confirmed in our B6D2 F2 populations at very different rates, resulting in a net increase in the percentage of cis-acting QTL in confirmed data relative to individual data sets. For instance, in the brain BXD RI data set at a genome-wide threshold of pgw = 0.05 there are 1824 cis-acting QTL and 2885 trans-acting QTL. After attempting to confirm these QTL at a genome-wide threshold of pgw = 0.05 in the B6D2 F2 data set, there are 1184 cis- and 85 trans-acting QTL (approximately 14:1) in the common (confirmed) set and 640 cis- and 2800 trans-acting QTL in the noncommon set. Because we expect a ratio of 1 cis:1.6
J.L. PEIRCE
ET AL.:
HOW REPLICABLE
ARE MRNA
EXPRESSION QTL?
649
Fig. 4. Replication of a cluster of trans-acting factors affecting brain mRNA expression. This figure shows positions of QTL plotted against positions of transcripts. For ease of viewing, the graph is plotted as if all chromosomes are laid end to end along each axis. QTL significant at pgw £ 0.05 in the BXD RI brain expression data set are indicated as small gray dots. Larger black dots indicate QTL that were also confirmed at a threshold of pgw £ 0.05 in the B6D2 F2 brain data set. The thick diagonal line formed from black dots indicates confirmed cis-acting QTL with QTL positions very near the transcript position. There are also a number of trans-acting clusters indicated by vertical bands (trans-bands) of QTL in the BXD RI data that can be seen in the plot. Of these, only the trans-band on distal Chr 1 was substantially replicated in the B6D2 F2 data.
trans, this is a 22-fold enrichment of cis-acting QTL relative to trans-acting QTL in the confirmed set. This is partly because cis-acting QTL tend to have smaller p values and larger effect sizes than transacting QTL. In the brain RI data, for instance, at a genome-wide threshold of pgw = 0.05, the average LRS for a cis-acting QTL is 32, while the average for a trans-acting QTL is only 21. This does not entirely account for the difference in confirmation rates, however. Constraining the QTL under consideration to be 0.001 £ pgw £ 0.05 for the identifying (RI) population and confirming QTL in the B6D2 F2 population at the least stringent level (pgw £ 0.05) still results in 18-fold enrichment for confirmed cis-acting QTL. Narrower constraints such as 0.01 £ pgw £ 0.05 for discovery (and 0 £ pgw £ 0.05 for confirmation) produce similar enrichments. While 38% of cis-acting QTL in this interval are confirmed, only 0.6% of trans-acting QTL are confirmed in this pvalue range. Because the interval contains the majority of all trans-acting QTL, only a small fraction of trans-acting QTL overall are confirmed (3% with pgw £ 0.05 in both populations). The fraction of trans-acting QTL that are confirmed does improve with higher stringency, as does the fraction of cis-acting QTL, though the fraction of confirmed trans-acting QTL remains lower. Even at a threshold of pgw = 0.001 for discovery and confirmation (Fig. 3), however, 67% (792/1187) of cis-acting QTL are confirmed in brain data, but only 23% (72/391) of trans-acting QTL are.
As Fig. 4 shows, replication of trans-QTL is not entirely random. As noted in the work of Chesler and colleagues (2005) using an earlier BXD RI-based data set based on Affymetrix U74Av2 arrays (an earlier version), trans-QTL clusters by position can be seen. Such bands were also present in our brain BXD RI data set (vertical ‘‘trans-bands’’ in Fig. 4). Of the visible bands present in our data set, only one has any overlap with trans-bands reported by Chesler et al. (2005), despite the small overlap in samples between our data set and theirs, as described in the Methods section. There was a substantial number of transcripts on Chr 10 (50/4709) and smaller clusters throughout the genome in our data set, but they are not as pronounced as and do not have substantial overlap with those reported by Chesler et al., and none of the QTL are replicated in the F2 brain data. Only one of the trans-bands present in our data is also clearly present in the B6D2 F2 brain data set. This band lies between 170 and 174 Mb on Chr 1 and has a larger total number of QTL (84/4709). In addition, 25 of the QTL in this band are replicated in the F2 data, an impressive finding since at this level of stringency (pgw = 0.05 for discovery and confirmation) there are only 85 trans-acting QTL in the entire genome. Because we would expect 0.11 trans-acting QTL on average to reside in this interval, this is a dramatic overrepresentation of confirmed trans-acting QTL. In addition, the replication of the rate for tran-acting QTL in this region
650
J.L. PEIRCE
ET AL.:
HOW REPLICABLE
ARE MRNA
EXPRESSION QTL?
(25/84) is tenfold higher than for the genome in general (p 8) range. This difference is significant (p < 0.03) but does not represent a large percentage increase. When a somewhat less stringent identification criterion (pgw = 0.05) in the RI set is substituted, however, expression levels have a larger effect. In this case, there are 578 (22%) common and 2074 unshared QTL associated with genes with below-average expression and 687 (34%) common and 1362 unshared QTL associated with genes of above-average expression. This is a highly significant result (p
