Genetic Effects on Dispersion in Urinary Albumin and Creatinine in Three House Mouse (Mus musculus) Cohorts

Conventionally, quantitative genetics concerns the heredity of trait means, but there is growing evidence for the existence of architectures in which certain alleles cause random variance in phenotype, termed ‘phenotypic dispersion’ (PD) or ‘variance QTL’ (vQTL), including in physiological traits like disease signs. However, the structure of this phenomenon is still poorly known. PD for urinary albumin (PDUAlb) and creatinine (PDUCrea) was mapped using curated data from two nearly genetically identical F2 mouse (Mus musculus) cohorts (383 male F2 C57BL/6J×A/J (97 SNP) and 207 male F2 C57BL/6J×A/J ApoE knockout mice (144 SNP)) and a related mapping cohort (340 male F2 DBA/2J×C57BL/6J (83 SNP, 8 microsatellites)). PDUAlb was associated with markers in regions of Chr 1 (5-64 megabases (MB); 141-158 MB), 3 (∼113 MB), 8 (37-68 MB), 14 (92-117 MB) and 17 (14-24 MB) with several positions and quantitative architectures in common between the two C57BL/6J×A/J cohorts, most of which had a negative dominant construction. One locus for PDUCrea was detected on Chr 19 (57 MB) in the C57BL/6J×A/J ApoE−/− cohort. The large number of negative dominant loci for albuminuria dispersion relative to conventional quantitative trait loci suggests that the development of albuminuria may be largely genetically dynamic and that randomization in this development is detrimental.

phenotypic dispersion albuminuria creatinine mouse Mus musculus genetic homeostasis negative dominance Conventional quantitative genetics concerns heritable differences in mean phenotype (Roff 1997). However, there is increasing evidence that some genotypes confer significant differences in random or residual variability rather than stable mean phenotype, so that intraindividual or inter-individual randomization among genotypes or genetic groups may constitute a properly heritable genetic effect instead of sheer error (Reeve and Robertson 1953;Perry et al. 2003;Sorensen and Waagepetersen 2003;Ordas et al. 2008;Rönnegård and Valdar 2012). This effect has been described as 'phenotypic dispersion' (PD) (Perry et al. 2012a) and may reflect the effects of 'variance QTL' (vQTL) on trait variance (Rönnegård and Valdar 2012). As early as the 1950s, divergent selection experiments in Drosophila found simultaneous changes in means and variances for wing length and body size (Reeve and Robertson 1953;Clayton and Robertson 1957), suggesting the accumulation of both alternate variants and randomizing alleles via incidental inclusion of extreme individuals during selection (Hill and Zhang 2004). Since that point, genetic variation for heterogeneity has been found in plants (Hall et al. 2007;Ordas et al. 2008), fish (Perry et al. 2003), birds (Rowe et al. 2006;Wolc et al. 2009) and mammals (SanCristobal-Gaudy et al. 1998;Sorensen and Waagepetersen 2003;Rönnegård et al. 2010;Perry et al. 2012a), including rodent disease models (Ibáñez-Escriche et al. 2008) and human phenotypes and gene expression (Perry et al. 2012c;Hulse and Cai 2013;Perry et al. 2013). Theoretical investigations of residual variance suggest a genetic architecture resembling classical trait means (m i , s i ) (Hill and Zhang 2004;Hill and Mulder 2010) or the general inability of inbred individuals to buffer minor environmental perturbation (Lerner 1977). Most examples of dispersion come from common environments (i.e., Perry et al. 2003;Sorensen and Waagepetersen 2003;Wolc et al. 2009;Rönnegård et al. 2010;Perry et al. 2012a;Sell-Kubiak et al. 2015;Conley et al. 2018) so that an explanation of heredity for environmental buffering (de Visser et al. 2003) seems improbable, although an assay of dispersion in airway hyperresponsiveness (AHR) found increasing genotypic differences in PD at a Chr 10 locus with increasing methacholine dosage, suggesting environmental gradients in the expression of dispersion loci (G. M. L. Perry, unpublished data). Little, however, is known of wider trends in the quantitative construction of dispersive systems, so that the relative contributions of additivity and dominance to this phenomenon and their meaning (c.f. Roff 1997) are not understood. A strong additive basis, for example, would indicate the functional independence of individual alleles within genotypes. Disease physiology appears susceptible to the dispersive/vQTL effect (Perry et al. 2012a;Perry et al. 2013). In this work, genome-wide associations of single nucleotide polymorphism (SNP) genotype with dispersion in urinary albumin, urinary creatinine and blood urea nitrogen were tested using curated data from three mouse (Mus musculus) groups, consisting of i) a cohort of F 2 intercrosses of albuminuric A/J mice with non-albuminuric C57BL/6Js, ii) a cohort of F 2 A/J · C57BL/6J ApoE 2/2 knockout mice (Doorenbos et al. 2008) and iii) an F 2 intercross of C57BL/6J·DBA/2J mice (Sheehan et al. 2007). This data were originally used to scan for standard loci affecting albumin excretion, an early indicator of chronic kidney disease (CKD) and diabetic nephropathy resulting from podocyte damage and immune cell recruitment, and to determine the genetic role of Apoe in albuminuria (Joss et al. 2005;Doorenbos et al. 2008;Coto et al. 2013), and downregulates mesangial cell proliferation associated with renal disease (Chen et al. 2001).
Several genomic regions were significantly associated with phenotypic dispersion in urinary albumin (PD UAlb ) in these cohorts after correction at the 5% False Discovery Rate (FDR), including Chr 1 (5-64 and 141-158 MB (termed PD UAlb 1 and PD UAlb 2, respectively)), 3 ($125 MB (PD UAlb 3)), 8 (37-68 MB (PD UAlb 4)), 14 (92-117 MB (PD UAlb 5)) and 17 (14-24 MB (PD UAlb 6)). A single marker on chromosome 19 (19-060823449-M, 56.5 MB (PD UCrea 1)) was associated with dispersion in urinary creatinine. Notably, a clear majority of albuminuria dispersion loci were at least partially negative-dominant in this assay, suggesting that random variance in disease physiology may be largely detrimental, or integral to the process of disease itself. These findings do not agree with the expectation that ubiquitous physiological systems underlie dispersion, but do support the case that phenotypic dispersion is physiologically relatively common and indicate a major jump in the understanding of the overall meaning of the effect to physiology and survivorship.
Spot urine samples from each mouse were quantified for urinary creatinine (UCrea; mg/dl) as an estimate of baseline kidney function/ glomerular throughput and albumin (UAlb; mg/dl). Weight (g) and blood urea nitrogen (BUN; mg/dl) were available in Doorenbos et al. Cohort 3 (Sheehan et al. 2007) The third cohort (Sheehan et al. 2007;MPD:205), consisting of male F 2 C57BL/6J (B6) · DBA/2J (D2) mice F 1 phenotyped for urinary creatinine and albumin and reported in Sheehan et al. (2007). This cohort shared only one of the source strains with the above two cohorts (DBA/ 2J) and was included for comparison to those more closely related groups. F 1 reciprocal families ([B6·D2] and [D2·B6]) were bred from B6 and D2 mice from the Jackson Laboratories. An additional F 1 [D2·B6] cohort was produced and used to breed a total of 340 F 2 C57BL/6J·DBA/2J ['BxD'] male mice from an initial cross of B6 females bred with DBA2 males. F 2 mice were phenotyped with spot urine collections and genotyped over all 19 autosomes and the X chromosome using 83 SNP, and at eight microsatellites on chromosome 2, for a mean intermarker spacing of 17 cM.
Marker location was assigned throughout based on the Cox et al. (2009) reference marker map build using base-pair (BP) distances to avoid possible mis-position from sex differences in recombination by region (Broman et al. 1998;Sakamoto et al. 2000;Popa et al. 2012).
Animal usage: Ethical animal use in the original studies was monitored and approved by the Institutional Animal Care and Use Committee (IACUC) of The Jackson Laboratory.
Association analysis: All analysis was performed in SAS (2011). In order to protect against distributional errors, the deviation of individuals from predicted multivariate values were estimated from externally Studentized residuals (Steel and Torrie 1980) in which ordinary residuals e are divided by their standard errors (ê s ¼ê=ðŝ ffiffiffiffiffiffiffiffiffiffiffiffi ffi 1 2 h ii p ) (where h ii is the observation leverage andŝ 2 ðiÞ ¼ ðn 2 m 2 1Þ), in order to satisfy the experiment-wise relations P n i¼1ê i ¼ 0 and P n i¼1ê i x i ¼ 0. Individual Studentized residuals were estimated in a general linear model of the form where y ij is albuminuria or creatinine for individual j,m is the mean phenotype for the cohort, a i is the effect of marker locus i, b MLH X MLH is the partial regression effect of multilocus heterozygosity (MLH), b Ucrea X Ucrea is the partial regression effect for glomerular filtration rate and e ij is individual residual error. MLH was included at this level to account for possible inbreeding effects and calculated as MLH = n het /n total within individuals in each group across all available genotypes. Regression effects for creatinine were only included for albuminuria. Each model was initially run without locus terms at each analytical stage order to determine covariates for the genomic models including locus terms, which were used in order to account for the effects of known and undetected conventional loci on albuminuria (see Doorenbos et al. 2008).
Individual residual error estimates (ê ij ) were absolute-transformed (|ê ij |); as absolute divergence of any particular individual from that predicted by genotype, these were then considered to be phenotypic dispersion (PD) for that trait (PD UAlb , PD UCrea ). Since absolutetransformed distributions are left-skewed with strong lower bounds at the abscissa, marker-dispersion associations were fit using Tobit quantitative and limited models (Tobin 1958) with a lower bound of zero with PD as the dependent variable and locus as an independent variable along with significant covariates. In Tobit censored distributions, the actual y of the true variable y Ã is only observed where y . t, the lower truncation value, and as y Ã otherwise (i.e., y = y Ã where y . t). The truncated PDF of such a system then is expressed as f ðy y . tÞ ¼ j f ðyÞððPðy . tÞÞ and transforming by f a ¼ 2m=s; is the cumulative distribution function (CDF; Steel and Torrie 1980;Greene 2002) of the original data so that the likelihood becomes Model terms were optimized by the default quasi-Newtonian Broyden-Fletcher-Goldfarb-Shanno algorithm (Press et al. 2007). The significance of genotypic effects in the analysis of each locus was determined via a joint nonequivalence Wald contrast against mean PD in the referential A/J homozygote ðH' ¼ hm CC Parsad 2008;SAS 2014), the last genotype being fit via default as the referential genotype against the rest of the population. MLH was included as a covariate where it was significantly associated with PD (P , 0.1) to account for the possible production of phenoaberrancy by the failure of increasingly inbred individuals to buffer phenotype against exogenous and endogenous stresses, termed genetic homeostasis (Lerner 1977). Additivity and dominance were estimated in SAS using contrast statements equivalent to Griffing's potence ratio (Griffing 1990) where Q is the quadratic dominant effect and L is the classical linear differentiation between alternate homozygotes. Additivity was tested by contrast against the midparent phenotype ððPD C þ PD A Þ=2Þ (contrast statement +1 0 -1). Dominance was tested sequentially using the vectors Q ¼ ½þ0:5 CC þ 0:5 CA 2 1:0 AA and Q ¼ ½ 2 1:0 CC þ 0:5 CA þ 0:5 AA to test positive dominance and Q ¼ ½þ1:0 CC 2 0:5 CA 2 0:5 AA and Q ¼ ½2 0:5 CC 2 0:5 CA þ 1:0 AA to test negative dominance (Aurelio et al. 2000;Lee and Sabapathy 2008).
Significance: Significance thresholds were adjusted via Benjamini-Hochberg (Verhoeven et al. 2005) by trait calculated across all markers independently without reference to linkage among markers at the classic False Discovery Rate (P # k i Ã a/m) with a 'hard floor' for rejection of H' at a nominal P i = 0.01.
SNP sites for QTL for albumin excretion: The Mouse Genome Informatics (MGI) resource curated by the Jackson Laboratories (www.informatics.jax.org) was used to identify SNP between the source strains (C57BL/6J vs. A/J; DBA/2J vs. C57BL/6J) at nonsynonymous coding sites (CNS), utranslated mRNA sequence (mRNA-UTR), splice sites (SS) and non-coding transcript variants (NTV) (see Ward and Kellis 2012) at genes closely linked (,10 MB) to consensus markers for PD Ualb as possible candidates for genetic effects on dispersion. Sequence information was based on the dbSNP (Mouse) Build 142 by MGI and the GRCm38 mouse genomic build. Annotation functions were obtained through databases from The Jackson Laboratories (www.informatics.jax.org), the European Bioinformatics Institute (www.ebi.ac. uk), UniProt (www.uniprot.org), GeneCards (www.genecards.org), WikiGenes (www.wikigenes.org) and homologs listed with the Rat Genome Database (www.rgd.mcw.edu).

Multilocus heterozygosity
Several genomic regions were significantly associated with PD UAlb in both Doorenbos et al. cohorts, with highly similar genetic architecture in these mapping groups (Table 2; Figure 1). PD UAlb was associated with SNP genotype over the approximate region of 5-64 MB on Chr 1 in Doorenbos et al. A and B, here considered to represent a locus for albuminuria dispersion ('PD UAlb 1') (Table 2; Figure 1). Contrast tests for dominance and additivity indicated that PD UAlb 1 was partially negative dominant (Lee and Sabapathy 2008) for the C57BL/6J allele, so that C57BL/6J·A/J heterozygotes and C57BL/6J homozygotes had lower PD UAlb than A/J homozygotes (P FDR , 0.01) (Table 2; Figure 2). PD UAlb was significantly associated with Chr 1 markers in the 141-158 MB range in Doorenbos et al. (Table 2; Figure 1). A post-Benjamini correction (P FDR , 0.05). Contrast tests indicated that this position was overdominant with C57BL/6J·A/J heterozygotes having marginally higher PD UAlb (P , 0.1) than C57BL/6J homozygotes and significantly higher PD UAlb (P , 0.05) than A/Js (Table 2 Figure 2). This was a wider genomic range than other consensus loci with different ranges of overlap, but based on the significance of marker-PD UAlb associations in these closest markers and the similarity of PD means by genotype, it was considered that these results represented a third locus for albuminuria dispersion ('PD UAlb 3').  Figure 1). As PD UAlb 1 and PD UAlb 3 2 5, this locus also appeared to be partially negative dominant with A/J homozygotes having significantly lower dispersion than C57BL/6J·A/J heterozygotes or A/J homozygotes (P FDR , 0.01) (Table 2; Figure 2).
There were a number of markers associated with PD UAlb in only one of the two cohorts (Chr 6, 10,   Figures 1, 3).
No genomic region was associated with PD BUN at the FDR (P . 0.1). A single marker in the anterior end of Chr 19 (SNP 19-060823449-N; 56.5 MB) was significantly associated with PD UCrea (r 2 = 0.183) (Figure 2

Distribution of genetic architectures
Of all loci significantly associated with PD traits in all three cohorts (including dual additive and dominance components for partially dominant loci), 19 were additive, one was high-dominant (heterozygote equal to the high-PD homozygote), 20 were negative dominant, five overdominant and four underdominant (Table 2). For those loci with statistical analogs in both Doorenbos et al. cohorts, there were ten loci with additive effects, 11 with negative dominance, two overdominants and one underdominant (Table 2).
C57BL/6J-vs-A/J candidate SNP SNP between the C57BL/6J and A/J strains linked to dispersion loci occurred in genes affecting cell growth/mitosis/platelet action (Arid5a, Egf, Fgl1, Fgf20, Igfals, Itpr3, Ogfrl1, Plg, Rblcc1, Rab23, Qsox1), immunology (Arid5a, Lonrf1, Msr1, Mtus1, Phf3), serine/threonine physiology (Camk2d, Dlc1, Dusp4, Pkmyt1, Prss29, Prss30, Prss33, Prss34, Prss40, Prss41, Smok2b, Srrm2), DNA repair and mitotic checkpoint maintenance (Eme2, Ercc5. Mcmdc2, Tdrd5, Telo2, Tex15, Tti2), cellular construction/morphology (Actr1b, Ank2, Cep450, Col11a1, Col5a2, Dnah7b, Dst, Ogfr1, Mdga, Mtus1), G-protein coupled receptors (Fdnc1, Fpr3, Fpr-rs3, Fpr-rs4, Fpr-rs6), calcium physiology (Bank, Dnase112, Pcdh9, Pkd1, Saraf), gene expression (transcription, splicing, translation) (Eri1, Purg, Rbm20, Rrp1b, Trmt9b, Trmt11, Trmt13). Some SNP variants occurred at genes linked to other renal diseases including autosomal dominant polycystic kidney disease (ADPKD) (Pkd1) and autosomal recessive polycystic kidney disease (ARPKD) (Pkhd1) and cystic fibrosis (Slc9A3R2). Two genes (Ccnf and Tbl3) contained WD-40 domains. There were a variety of SNP in coding sites for vomeronasal genes and in type C2H2 zinc fingers (Flywch1, Wiz1, Zgrf1, Zfp proteins) (Table S1).  (Moranne et al. 2009) and the number of independent dispersion loci in this work suggests that albuminuria distributions may be largely determined by a number of independent arrays of genes with randomizing effects on disease onset and progression. The mechanics of dispersion loci could range from ephemeral physiological 'twitches' to randomization in the progress of long-term biological insult ranging from the unaffected state to the disease state vs. retention of the unaffected status; dispersion in albuminuria, with attendant morphological changes (glomerular damage and inflammation with subsequent podocyte damage from infection, self-response or complement thrombosis) (Doorenbos et al. 2008;Coto et al. 2013;Regal et al. 2018). Only a single locus was detected for dispersion in creatinine; this sole finding against the larger number of loci for albuminuria may reflect more constant creatinine expression as a baseline estimator of kidney throughput (Stevens et al. 2013). C57BL/6J n Table 1 Marker proportions, mapping completeness and record completeness for 1) 383 male F 2 C57BL/6J 3 A/J mice (Doorenbos et al. A group), 2) 207 male F 2 C57BL/6J-Apoe 2/2 mice (Doorenbos et al. B group) (Doorenbos et al. 2008) and 3) 340 male F 2 DBA/ 2J3C57BL/6J mice (Sheehan et al. 2007). N refers to the number of records available for each phenotype (urinary albumin, blood urea nitrogen (BUN) and creatinine), m is the cohort mean for that phenotype and Range the min-max range for all observations  vs. A/J SNP variants linked to these loci included polymorphisms at various transposable element regulators, respiratory electron chain genes, G-protein coupled N-formyl peptide receptors, vomeronasal genes, serum calcium regulators, complement receptors, signal transducers, and candidates of autosomal dominant (Pkd1) and recessive (Pkhd1) polycystic kidney disease (Bergmann 2015;Ghata and Cowley 2017) and cystic fibrosis (Slc9A3R2) simultaneously mitigates the effects of the cystic fibrosis transmembrane conductance (CFTR) reducing renal cyst growth via proteostasis and reduces resting intracellular Ca 2+ (Yanda et al. 2018). Various SNP occurred in serine/threonine-enriched proteins, which have been associated with loci linked to the coefficient of variation (CV) in total RNA production (Perry, unpublished results), diabetes severity/ onset (G. M. L. Perry, unpublished results) and diabetic plasma traits (Brown 2018;G. M. L. Perry, unpublished results). Two genes (Ccnf, Tbl3) had WD-40 domains (Schapira et al. 2017); SNP in WD-40 domains were also associated with random variation in urinary calcium in a human cohort (n = 1210) (Perry et al. 2013). PD loci accounted for smaller proportions of randomized variance in Doorenbos et al. A (3-8%) than B (5-24%); this may have been due to sample size, the Beavis effect (Beavis 1998) and/or liberating effects of the Apoe KO on residual variance in the latter. The removal of mediating factors like Apoe might result in increasingly unstable physiological architecture so that downstream systems might also be subject to increasing dispersion, although the mechanics of such an effect would depend on the nature of the physiological pathway. Apoe 2/2 mice have a wide range in nephropathic outcome (Wen et al. 2002;Buzello et al. 2004). Loci detected in only a single cohort might be related to this dispersive mediation. The genetic architecture in PD UAlb appeared to be inverted between Doorenbos et al. A and B for loci on Chr 15 and 16, so that Apoe might alter the tendency to dispersion within genotypes.
A strong majority of dispersive loci were partially negative dominant, with contrasts including additive and negative dominant components. This is similar to a recent survey of PD for diabetes-related serum traits (high-and low-density lipoproteins, general cholesterol, triglycerides) in eight intercross and backcross mouse cohorts in which most PD loci were also negative dominant (Brown 2018). Not all dispersive genetic variance has this expression (Perry et al. 2013;G. M. L. Perry, unpublished results) but negative dominance-essentially recessivity for high dispersion where genetic physiological randomization is suppressed by single normalizing alleles which promote constant or stable gene activity-might be a frequent feature of this phenomenon. This propensity to suppression of randomizing variance might thus mean that 'recessive' high-PD genotypes are essentially detrimental as in other recessive systems (Charlesworth 2009), although the ecological implications of the phenomenon have not been extensively explored. A primarily recessive architecture for randomizing phenotype could also create additional complications in genetic analysis (Hildebrandt et al. 2009) similar to limited recessive penetrance (see Boone et al. 2013;Gao et al. 2015). This might include dispersive loci that influence the detection of normal genes (i.e., Perry et al. 2011). New model builds might need to be created in order to specifically address such systems.

Creatinine
One locus for PD UCrea was detected on Chr 19 (57 MB) in the C57BL/ 6J·A/J ApoE 2/2 group. As a product of lean muscle mass, creatinine should be relatively stable, but intraindividual CVs for creatinine approximate 9% (Bingham and Cummings 1985) and the heritability of individual CV in urinary creatinine was significant (h 2 s = 8.7%) in a three-generation cohort of 949 kidney stone probands and first-degree relatives (Perry et al. 2012b). Dispersion effects in fitness or survivorship from creatinine might operate through physiological related to  (Doorenbos et al. 2008) and c) a cohort of 340 male F 2 DBA/2J · C57BL/6J mice (white symbols) (Sheehan et al. 2007). Significant points with maximal association with PD UAlb are indicated (i.e., 'PD UAlb 1').
disease state: Gibb et al. (1989) found higher between-individual variance in creatinine clearance in diabetic children than non-diabetics.  (Brock et al. 2015) and across ontogeny (Gillingham et al. 2013;Annavi et al. 2014), due to contextual variance in selection gradients. This implies some tractable, functional variability in HTCs, but the cohorts used here were lab-reared under no known selective pressure, suggesting that differences in MLH  correlation were endogenous in origin. Some work indicates that HTCs at specific regions are more important than total individual heterozygosity (Rodríguez-Quilón et al. 2015) so that genetic differences between strains might be expected to generate differences in both HTCs and MLH-PD correlations. Differences in heterozygosity for specific regions enriched for immunological or other functional groups (i.e., the MHC complex on human Chr 6 and mouse Chr 17 (cytoband B-C)) might, for example, be key to variation in this effect.

Conclusions
The incidence of heritable dispersion appears to be growing (SanCristobal-Gaudy et al. 1998;Perry et al. 2003;Sorensen and Waagepetersen 2003;Hill and Zhang 2004;Ordas et al. 2008;Hill and Mulder 2010;Perry et al. 2012a;Rönnegård and Valdar 2012;Wang et al. 2014). Analytically, this represents an enormous potential area of genetic interest: dispersive systems, themselves a series of random risk factors invisible to conventional analysis, could render critical elements of genetic control conventionally undetectable (Perry et al. 2011) or mask major elements of trait distributions from genetic decomposition of architecture. High similarity of position and effect for albuminuria dispersion markers across the two cohorts strongly supports the existence of dispersion loci underlying the effect, but there was little evidence that the basis of the effect was in general physiological systems like transcription regulation or splicing. A genetic architecture ranging from negative dominance to additivity (Brown 2018) indicates that high heritable randomization values tend to be recessive so that stable expression is 'rescued' by a single normalizing allele. Randomization in signs or elements of disease physiology such as albuminuria might be particularly unfit, generally.
Additionally, the evolutionary consequences of such systems could be profound: dispersion loci could create 'fuzzy' surfaces on fitness landscapes, permitting individuals or subpopulations to transit between local adaptive peaks without the risk of intervening saddles, or mitigate competition among siblings by dispersing phenotype in close relatives, or allow single parents to produce an array of heterogenous progeny to exploit new niche space or variable environments. Albuminuria (Syme et al. 2006) and creatinuria (Gibb et al. 1989) are linked to survivorship so that dispersion in proteinuria may indeed have direct fitness relevance. The elaboration of dispersion in this system and others may provide a powerful insight into the construction of phenotype, elucidating unseen spandrels of distribution in medicine, evolution and agriculture.

ACKNOWLEDGMENTS
I acknowledge the comments of Dr. Gary Churchill and Dr. Ron Korstanje. The initial work on which this manuscript was based was supported by National Institute of Diabetes and Digestive and Kidney Diseases grant DK-69381 and National Institute of General Medical Sciences grant GM-070683. There are no conflicts of interest.