Polymorphisms in Maternal and Fetal Genes Encoding for Proteins Involved in Extracellular Matrix Metabolism Alter the Risk for Small-for-Gestational-Age

Abstract

OBJECTIVE

To examine the association between maternal and fetal genetic variants and small-for-gestational-age (SGA).

METHODS

A case-control study was conducted in patients with SGA neonates (530 maternal and 436 fetal) and controls (599 maternal and 628 fetal); 190 candidate genes and 775 SNPs were studied. Single locus, multilocus and haplotype association analyses were performed on maternal and fetal data with logistic regression, multifactor dimensionality reduction (MDR) analysis, and haplotype-based association with 2 and 3 marker sliding windows, respectively. Ingenuity Pathway Analysis (IPA) software was used to assess pathways that associate with SGA.

RESULTS

The most significant single locus association in maternal data was with a SNP in tissue inhibitor of metalloproteinase 2 (TIMP2) (rs2277698 OR = 1.71 95% CI [1.26-2.32], p = 0.0006) while in the fetus it was with a SNP in fibronectin 1 isoform 3 preproprotein (FN1) (rs3796123, OR = 1.46 95% CI [1.20-1.78], p = 0.0001). Both SNPs were adjusted for potential confounders (maternal body mass index and fetal sex). Haplotype analyses resulted in associations in alpha 1 type I collagen preproprotein (COL1A1, rs1007086-rs2141279-rs17639446, global p = 0.006) in mothers and FN1 (rs2304573-rs1250204-rs1250215, global p = 0.045) in fetuses. Multilocus analyses with MDR identified a two SNP model with maternal variants collagen type V alpha 2 (COL5A2) and plasminogen activator urokinase (PLAU) predicting SGA outcome correctly 59% of the time (p = 0.035).

CONCLUSIONS

Genetic variants in extracellular matrix related genes showed significant single locus association with SGA. These data are consistent with other studies that have observed elevated circulating fibronectin concentrations in association with increased risk of SGA. The present study supports the hypothesis that DNA variants can partially explain risk of SGA in a cohort of Hispanic women.

INTRODUCTION

Neonates with a birth weight below the 10^th percentile for their gestational age at delivery are considered to be small-for-gestational-age (SGA).[1] SGA is one of the “Great Obstetrical Syndromes”[2,3] and it is often associated with other pregnancy complications such as preeclampsia (PE), preterm prelabor rupture of the membranes (pPROM), and preterm labor and delivery as well as increased risk of future health and developmental problems.[4-12] The known maternal risk factors for SGA include race, maternal age (≥35 years), low maternal body mass index (BMI), chronic hypertension, smoking, maternal occupation, substance abuse and maternal birth weight.[13-32] In addition, mechanisms of disease involved in SGA include: 1) abnormal physiologic transformation of the spiral arteries;[33-36] 2) chronic uteroplacental ischemia;[37-47] 3) endothelial cell dysfunction;[48-51] 4) increased trophoblast apoptosis/necrosis;[52] 5) intravascular inflammation [T helper 1 (Th1)-biased immune response];[53-56] and 6) an anti-angiogenic state.[57-74]

Multiple lines of evidence suggest that SGA may have a genetic component: 1) mothers born SGA have as much as a 50% increased risk of delivering an SGA infant;[15,75-78] 2) familial aggregation of SGA among siblings, especially documented sib-sib odds ratios (≥ 2.5)[79,80] 3) racial disparity exists in the incidence of SGA, which is higher among African-Americans, Indians, and Asians compared to other racial groups;[14,29,81,82] and 4) several polymorphisms have been studied in the context of intrauterine growth restriction and SGA.[83-107] Taken together, these data support a role for a genetic predisposition to SGA. Of note, in contrast to data on maternal genes, fetal genetic predisposition to SGA has been even less studied.[97,108]

In this study, a large-scale candidate gene screen was conducted to comprehensively assess the role of genetic variations in maternal and fetal DNA on SGA risk in a Hispanic population. We examined 775 single nucleotide polymorphisms (SNPs) selected from 190 candidate genes involved in pathophysiologic pathways of SGA and other adverse pregnancy outcomes including PE, pPROM, and preterm labor and delivery with intact membranes. Single locus, haplotype, and multilocus analyses tests of association were performed in both maternal and fetal DNA samples. In addition, to better understand genetic contributions to the pathophysiology of SGA, exploratory pathway analyses were performed.

MATERIALS AND METHODS

Study Design

This was a case-control study that included patients with SGA neonates (530 maternal and 436 fetal) and controls (599 maternal and 628 fetal). Gestational age was determined by last menstrual period or by ultrasound. In cases of discrepancies between these two measures, ultrasound dating was used. SGA was defined as a birth weight <10^th percentile for the gestational age at delivery according to the Chilean birth weight distribution of a Hispanic population.[109] The majority of both maternal and fetal samples were paired (406 case pairs and 578 control pairs). Controls were defined as women having a full term pregnancy (37–42 weeks of gestation) who delivered a neonate with a birth weight between 10^th and 90^th percentile for the gestational age at delivery[109] without complications including PE, eclampsia, HELLP syndrome, pPROM, preterm labor and delivery, large for gestational age, fetal demise, placental abruption, placenta previa, maternal sepsis, or chorioamnionitis.

Patients of Hispanic origin were recruited at the Sotero del Rio Hospital, in Puente Alto, Chile. All eligible mothers were enrolled in a research protocol, which requested permission to collect DNA from the mother and child for research purposes. The exclusion criteria included: 1) known major fetal anomalies or fetal demise; 2) multiple pregnancy; 3) serious medical illness (renal insufficiency, congestive heart disease, etc.); 4) refusal to provide written informed consent; and 5) a clinical emergency, which prevented counseling of the patient about participating in the study, such as fetal distress or maternal hemorrhage. A blood sample was obtained from the mother at the time of enrollment in the protocol, and from the umbilical cord from the fetus after delivery. Demographic and clinical characteristics of the mothers were obtained from a data collection form administered by trained medical and paramedical personnel. The collection of samples and their utilization for research purposes was approved by the Institutional Review Boards of the Sotero del Rio Hospital, Santiago, Chile (an affiliated of the Pontificia Catholic University of Santiago, Chile), and the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), National Institute of Health (NIH), Department of Health and Human Services (DHHS).

Genotyping

Candidate genes were selected for analysis based on biological plausibility for a role in SGA and other pregnancy complications such as PTB, pPROM and PE. Genes involved in processes such as the control of the immune response (pattern recognition receptors, cytokines, chemokines and their respective receptors), uteroplacental ischemia, or angiogenesis were considered appropriate candidates for this study. A complete list of the 190 genes evaluated is included in the supplemental materials (Supplemental Table 1).

SNP discovery within the candidate genes was performed by DNA sequencing at Genaissance Pharmaceuticals, Inc. (New Haven, Conn., USA) using its Index Repository, which includes a total of 93 subjects with Native American, Hispanic/Latino, European, Asian, and African-American ancestry.[110] To determine which individuals in the Genaissance Index Repository were most representative of the genetic variation observed in the Chilean population, 96 unrelated Chilean individuals who are representative of the patient cohort were sequenced for 16 DNA fragments. A subset of 42 subjects from the Index Repository that is heavily weighted with the Native American and Hispanic/Latino subjects (although European, Asian, and African-American subjects contributed to the subset as well) was determined to be most representative of the variation for the Chilean population, based on the correlation in the minor allele frequencies for the SNPs in 16 DNA fragments that were sequenced in both the Index Repository and the sample of patients from Chile (mothers delivered at the same hospital and with the same ethnicity). This subset of 42 individuals was used to select polymorphisms for the candidate genes. The SNPs selected for genotyping were intended to capture at least 90% of the haplotypic diversity of each gene covering variation in the coding regions,[111] 100 bases at each end of the introns, 1000 bases upstream of the start codon, and 100 bases downstream of the stop codon.

Template DNA for genotyping was obtained by whole-genome amplification[112] of genomic DNA isolated from blood using an automated DNA isolation protocol (BioRobot 9604, Qiagen, Valencia, Calif., USA). Genotyping was carried out using the MassARRAY® TM System (Sequenom, Inc., San Diego, Calif., USA) at the high-throughput genotyping facility at Genaissance. Each genotyping assay involved PCR amplification from template DNA in a target region defined by specific primers for the respective polymorphic sites, purification of the amplification product, annealing of the indicated extension primer to one strand of the amplification product adjacent to the polymorphic site, extending the primer by one nucleotide using the MassEXTEND® TM reaction (Sequenom, Inc.) and detection of the allele-specific extension product by mass spectrometry.[113]

Quality Control

Univariate and multivariate distributions were evaluated for each variable to identify significant outliers. Outliers were removed only if found to be incorrect on reexamination. Each SNP was verified to ensure consistency between the genotypes of mother and offspring. Numerous programs are available for detecting relationship errors;[114-118] however, to produce accurate results, these programs require genotyping for a larger proportion of the genome than was available in this study. Therefore, we considered the number of Mendelian inconsistencies between mother and fetus to identify potential relationship errors (e.g. sample mix-ups or mislabeling). When an inconsistency for an individual marker was observed, those genotypes were removed at that marker. In the case of multiple inconsistencies in a given pair, the pair was excluded for further analysis (10 pairs in controls and 5 pairs in cases). Finally, we assessed the presence of genotyping errors. In some instances, genotyping errors will lead to Mendelian inconsistencies, which can be easily identified and removed from the analysis. However, most genotyping errors for SNPs will be Mendelian consistent.[119] For example, with mother-fetus pedigree structures, genotyping errors where a homozygous individual has been mistyped as a heterozygous individual will never lead to a Mendelian inconsistency for SNPs. Deviations from Hardy-Weinberg equilibrium (HWE) may indicate the presence of a genotyping error or hidden population stratification,[120] or reflect a biological effect such as natural selection (or other evolutionary force) and/or the association of disease and genotype.[121-125]

Tests for deviations from HWE were performed for mothers and fetuses separately and again separately for diagnostic subgroups. Because it is currently unclear how to unequivocally distinguish between deviations from HWE due to genotyping error, and deviations from HWE due to biological causes, such as location at or near a disease susceptibility locus, we noted SNPs that deviate from HWE, but we did not remove them from the analysis. If necessary, we could follow-up these observations with additional testing. We tested for population stratification in cases and controls using STRUCTURE,[126] which indicated that case and control Chilean samples both cluster with HapMap European samples (data not shown).

Statistical analysis

Continuous clinical and demographic characteristics of cases and controls were tested for normality using Shapiro-Wilks tests. These included gestational age, birth weight, maternal age, and BMI. All measurements deviated significantly from normality; therefore, Mann-Whitney two-sample rank sum tests were used for case-control comparisons for these traits. χ² tests were used to test for differences in smoking, and fetal gender between cases and controls. Rank sum tests were used to test for differences in the frequency of parity and Apgar scores 1 and 5. STATA 10.0 statistical software (StataCorp, College Station, TX, USA) was used for all analyses.

Single locus tests of association

Statistical tests for single locus association and for deviations from HWE were calculated using PLINK statistical software.[127] Statistical significance for deviations from HWE in cases and controls were determined using Fisher’s Exact tests. Single locus tests of association were performed with logistic regression using an additive genotypic model where the minor allele was coded as the risk allele, standard summary statistics, odds ratios (OR) and confidence intervals (CI), were reported for these tests of association. Prior to performing single locus and haplotype analyses linkage disequilibrium (LD) based SNP pruning was performed, using PLINK software, with a cutoff of r² = 0.8. Of the 775 SNPs that passed quality control (QC), we analyzed 696 maternal and 664 fetal SNPs. We also excluded a small number of X chromosome SNPs (seven total) for fetal data, as neonates included in the study were both male and female, and power is greatly reduced in the male and female samples analyzed separately. These criteria accounted for the difference in the number of SNPs tested in mothers and fetuses.

Multiple testing corrections

A false discovery rate (FDR) correction was performed to adjust for multiple comparisons using a q* of 0.15 in single locus tests of association in maternal and fetal analyses separately.[128,129] The q* indicates the expected proportion of results that are identified as interesting that are actually false. This is in contrast to α (typically set to 0.05), which indicates the probability of obtaining even one false positive result among all tests for which the null hypothesis is rejected. FDR is used to measure global error, that is, the expected number of false rejections of the null hypothesis among the total number of rejections. The critical significance level was calculated by ranking the results by p-values and then multiplying this rank by q* divided by the total number of tests using the step-up approach of Benjamini and Hochberg.[128] The threshold q* = 0.15 is deliberately generous, for the purposes of discovery, in which false acceptance of the null is more problematic than false rejection.

Haplotype tests of association

Haplotype analyses were run on genes with at least 1 significantly associated SNP (p < 0.01) and at least two SNPs in the same gene. Haplotype frequencies were calculated using PLINK[127] statistical software. PLINK software was used to calculate global p-values for haplotype-based association analyses for the dichotomous PTB outcome with 2 and 3 marker sliding windows. Only haplotypes that had a frequency of 0.05 were analyzed, and only SNPs that had less than 5% missing data were used. The strongest associated haplotype windows are reported (global p < 0.05) and only these were analyzed for haplotype specific effects. We present the calculation of OR for each haplotype (using the most common haplotype as referent), as well as determination of case and control haplotype frequencies. Standard summary statistics for pairwise linkage disequilibrium (LD), r², were calculated using Haploview.[130,131] Haplotype blocks were assigned using the confidence interval algorithm created by Gabriel et al.[132]

Multi-locus analysis

Exploratory multi-locus analyses were performed, using Multifactor Dimensionality Reduction (MDR) to identify interactions among maternal, fetal, and maternal/fetal SNPs. MDR has been previously described in Ritchie et al.[133] and is available as open source software at www.epistasis.org. Briefly, MDR is a non-parametric (does not assume any statistical model) and model free (no assumption mode of genetic inheritance) tool for identifying gene-gene interactions. MDR collapses all of the genetic data into two categories (high and low risk) by comparing all single locus and all multi-locus combinations, and then categorizing each genotype into either high-risk or low-risk on the basis of the ratio of cases to controls that have that genotype. MDR ultimately selects one genetic model, either single or multi-locus, that most successfully predicts phenotype or disease status.

Analyses were performed: 1) separately for maternal and fetal data (tag SNPs only); and 2) combined for available maternal and fetal paired DNA samples. In our case, we analyzed 664 fetal and 696 maternal SNPs for a total of 1360 in the combined analysis. The different number of SNPs in maternal and fetal samples was due to different QC results for the two, and the fact that we did not want to remove possible interactions among genes in mothers and fetuses. Data were analyzed for two- and three-way interactions with 10-fold cross-validation and average balanced accuracy as the metrics for evaluating a model.[134] Several filtering steps and parameters were explored and are described in Table 1. The MDR algorithm was implemented with the full array of tag SNPs as well as after filtering, using the Tuned ReliefF (TuRF) approach as described in detail by Moore and White.[135] TuRF is a modification of ReliefF.[136] Briefly, ReliefF and similar computational algorithms have been developed, primarily in the computer science and data mining fields, to provide means to screen attributes/variables on rational bases that relate to quality in terms of ability to classify an outcome. They allow the reduction in importance of those attributes that are irrelevant to an outcome and will provide noise to any subsequent analyses. Many of the developed algorithms assume independent effects, and are therefore not appropriate when there are interactions between variables that affect outcome/disease. ReliefF does not make this assumption. It is an extension of Relief and we will briefly describe Relief in order to provide a better understanding of the attribute election process. In Relief, an instance (in this case an individual, either case or control) is randomly selected and then the two nearest neighbors, the nearest case and the nearest control, based on the multi-locus genotype data are also selected. Attributes (alleles) that are shared with those of the same status are increased in value and those that are shared among individual of opposite classes are devalued. Unshared values of an attribute in two cases are similarly decreased in value. This is repeated a preset number of times. ReliefF is done in a similar way except instead of choosing just a single individual of the same and single individual of the opposite status an arbitrary number of individuals is selected to value the attributes. In our case we use a sample of 10 nearest cases and 10 nearest controls on the genotype space. TuRF goes a step further and instead of just changing the attributes (e.g., SNP) values it systematically removes attributes (e.g. SNPs) that poorly differentiate cases and controls.[135] The motivation behind this algorithm is that the ReliefF estimates of the true associating SNPs will improve as the non-associating SNPs are removed from the dataset. We also used SuRF (Spatially Uniform ReliefF) and TuRF to filter SNPs.[137] This is an extension of TuRF that instead of using a set number of case and control nearest neighbors sets a radius around the selected individual and tests all individuals within this radius.[138] In addition, SNPs were filtered based on results of the single SNP analyses and only SNPs that had a marginal p-value of ≤ 0.1 were included, or only those with a p-value < 0.05 were analyzed separately. Permutation testing with 1,000 permutations was used to determine statistical significance of all MDR models, addressing potential multiple testing issues.

Table 1.

Parameters explored in MDR analyses

Algorithm	SNPs included	Population
	all tag SNPs	Maternal
	tag SNPs with p < 0.05
	tag SNPs with p < 0.10
Balanced Accuracy	all tag SNPs	Fetal
	tag SNPs with p < 0.05
	tag SNPs with p < 0.10
	all tag SNPs	Maternal-Fetal Combined
	tag SNPs with p < 0.05
	tag SNPs with p < 0.10
		Maternal
Balanced Accuracy with TuRF with 10 SNP filter	all tag SNPs	Fetal
		Maternal-Fetal Combined
		Maternal
Balanced Accuracy with Surf_n_Turf-25 SNPs	all tag SNPs	Fetal
		Maternal-Fetal Combined

MDR as described above is ideal for a balanced data set where the number of cases and controls are the same or close to the same. However, computational methods have been developed since the initial development of MDR to test for prediction accuracies in an imbalanced data set.[134] The method, termed balanced accuracy, corrects for imbalanced data by taking an average of the sensitivity and specificity and is defined as the arithmetic mean of sensitivity and specificity, but approaches actual accuracy as the data becomes more balanced. We tested for balanced accuracy in this manuscript.

Bioinformatics Tools

The SNPper (http://snpper.chip.org) database using dbSNP Build 125 was used to determine marker positions (bp), marker function, and identify amino acid changes.

Pathway analysis

To examine whether the SNPs found to be putatively-associated with SGA mapped to different biological networks and disease functions, an exploratory analysis was performed using Ingenuity Pathway Analyses (IPA).[139-143] The gene variants that were statistically significantly-associated with SGA in our single locus analysis (p ≤ 0.05) were entered into the IPA analysis tool. These genes were termed “focus genes.” The IPA software was used to measure associations of these molecules with other molecules and disease functions stored in its knowledge base. The knowledge base includes relationships between proteins, genes, cells, tissues, xenobiotics, and diseases. The information is scientist-curated, updated, and integrated from the published literature and other databases such as OMIM, Gene Ontology, and KEGG.

The IPA functional annotations integrate GO data (http://www.geneontology.org), but the IPA knowledge base is more extensive and complete, and therefore uses different but overlapping terminology.

We used the IPA software to identify new molecular network(s) that involve our focus genes. We mapped focus genes to networks based on their mechanistic associations with each other and partner molecules. The IPA statistical algorithm identifies molecular pathways, involving these focus genes based on their selective interconnectivity with each other and additional molecules stored in the genome-scale knowledge base. These networks are illustrated and ranked by the software for significance of focus gene enrichment.

RESULTS

Examination of demographic and clinical data (Table 2) revealed significant differences between cases and controls for gestational age at delivery, birth weight, BMI, and fetal sex. We adjusted for potential confounders BMI and fetal sex in all single locus tests of association.

Table 2.

Demographic and clinical characteristics of the study population

Variable	Cases*		Controls*		p-value
	Median (25th-75th)	Mean (SD)	Median (25th-75th)	Mean (SD)
Parity (number of previous pregnancies)	1 (0-2)	0.95 (1.25)	1 (0-1)	0.91 (1.08)	0.500
Maternal age (years)	24 (20-31)	26 (7.18)	24 (20-30)	25 (6.03)	0.509
BMI	23 (21-26)	24 (4.19)	24 (22-26)	24 (3.76)	0.026
Smoking	18%		14%		0.143
Gestational age at delivery (weeks)	39 (37-40)	38 (3.37)	40 (39-41)	40 (1.12)	<0.0001
Birth weight (grams)	2610 (2190-2800)	2401 (592)	3440 (3230-3650)	3449 (287)	<0.0001
1st Minute Apgar score	9 (8-9)	8 (1.83)	9 (9-9)	8 (1.28)	0.0001
5th Minute Apgar score	9 (9-9)	9 (1.09)	9 (9-9)	9 (0.30)	0.0003
Fetal sex (male)	44%		51%		0.024

BMI: body mass index

^*Maternal samples: 530 cases and 599 controls; fetal samples: 436 cases and 628 controls

Single Locus Tests of Association

A summary of the gene information for the SNPs with the most significant associations with SGA (p < 0.01) in maternal and fetal DNA is provided on Table 3. There were no significant deviations from HWE in maternal controls. There was one deviation from HWE in fetal controls at rs2304573 (p = 0.007), but this marker did not have a statistically significant deviation from HWE in cases (Table 4). The most significant single locus association in maternal DNA was at a synonymous coding (S101S) SNP in tissue inhibitor of metalloproteinase 2 (TIMP2) (rs2277698 OR = 1.71 95% CI [1.26-2.32], p = 0.0006). The minor allele frequency for this SNP (A) was 0.11 in cases and 0.07 in controls. This association was not significant after FDR correction (Table 4).

Table 3.

Gene summary information strongest associations (p < 0.01)

Population	Gene Name	Gene Code	rs#	Chromosome	Position (bp)	Function
Maternal	Prostaglandin E receptor 3, subtype EP3 isoform	PTGER3	rs5673	1	71250429	Intron
	Interleukin 6 receptor isoform 1 precursor	IL6R	rs8192282	1	152668303	Coding exon (A/A 31)
	Toll-like receptor 9 isoform B	TLR9	rs352140	3	52231737	Coding exon (P/P 488)
	Corticotropin releasing hormone receptor 2	CRHR2	rs4723000	7	30668439	Intron
	Elastin	ELN	rs2071307	7	73108650	Coding exon
	Fibroblast growth factor 4 precursor	FGF4	rs3740640	11	69296764	Downstream
	Alpha 1 type I collagen preproprotein	COL1A1	rs17639446	17	45628902	Intron
	Tissue inhibitor of metalloproteinase 2	TIMP2	rs2277698	17	74378612	Coding exon (S/S 101)
	Matrix metalloproteinase 9 preproprotein	MMP9	rs2274755	20	44073099	Intron
Fetal	Tumor necrosis factor receptor 2 precursor	TNFRSF1B	rs5746016	1	12173566	Intron
	Von Willebrand factor preproprotein	VWF	GNSC_635285421	1	635285421
	Fibronectin 1 isoform 3 preproprotein	FN1	rs2304573a	2	215951895	Intron
	Fibronectin 1 isoform 3 preproprotein	FN1	rs3796123	2	215992203	Intron
	Fms-related tyrosine kinase 4 isoform 1	FLT4	rs2242213	5	179980922	Intron
	Coagulation factor XIII A1 subunit precursor	F13A1	rs5982	6	6119865	Coding exon (P/L 565)
	Toll-like receptor 4	TLR4	rs4986790	9	119515123	Coding exon (D/G 299)
	Matrix metalloproteinase 10 preproprotein	MMP10	rs486055	11	102155634	Coding exon (R/K 53)
	Heparan sulfate proteoglycan 2	HSPG2	GNSC_634098268	12	634098268
	Platelet glycoprotein Ib alpha polypeptide	GP1BA	rs6066	17	4777453	Coding exon (N/N 258)
	Tissue inhibitor of metalloproteinase 2	TIMP2	rs2277698	17	74378612	Coding exon (S/S 101)

OR is the odds ratio for the additive genotypic model. {95% CI is the 95% confidence interval of the odds ratio.

^aDeviated from HWE in fetal controls (p1/40.007), but not cases.

^*Significant after FDR correction.

Table 4.

Single locus tests of association (p < 0.01) adjusted for fetal sex and BMI

Population	Gene Code	rs#	Minor Allele	Minor Allele Frequency		OR1	95% CI2		p-value
				Cases	Controls		Lower	Upper
Maternal (Cases n = 530; Controls n = 599)	PTGER3	rs5673	A	0.07	0.04	1.76	1.19	2.61	0.005
	IL6R	rs8192282	A	0.13	0.08	1.57	1.18	2.1	0.002
	TLR9	rs352140	A	0.5	0.43	1.26	1.07	1.5	0.007
	CRHR2	rs4723000	T	0.13	0.09	1.46	1.11	1.92	0.007
	ELN	rs2071307	A	0.25	0.2	1.36	1.09	1.68	0.005
	FGF4	rs3740640	G	0.08	0.05	1.68	1.17	2.4	0.005
	COL1A1	rs17639446	G	0.09	0.05	1.62	1.17	2.27	0.004
	TIMP2	rs2277698	A	0.11	0.07	1.71	1.26	2.32	0.0006
	MMP9	rs2274755	T	0.09	0.05	1.62	1.16	2.27	0.005
Fetal (Cases n = 436; Controls n = 628)	TNFRSF1B	rs5746016	T	0.01	0.03	0.3	0.12	0.73	0.008
	VWF	GNSC_635285421	A	0.02	0.01	3.81	1.47	9.84	0.006
	FN1	rs23045733	C	0.39	0.46	0.72	0.59	0.87	0.001
	FN1	rs3796123	A	0.51	0.44	1.46	1.20	1.78	0.0001*
	FLT4	rs2242213	C	0.12	0.07	1.62	1.19	2.2	0.002
	F13A1	rs5982	T	0.12	0.08	1.52	1.12	2.05	0.007
	TLR4	rs4986790	G	0.05	0.03	1.96	1.20	3.23	0.008
	MMP10	rs486055	A	0.07	0.04	1.73	1.16	2.59	0.007
	HSPG2	GNSC_634098268	T	0.23	0.18	1.45	1.15	1.82	0.002
	GP1BA	rs6066	T	0.04	0.02	2.09	1.21	3.62	0.008
	TIMP2	rs2277698	A	0.13	0.10	1.51	1.13	2.03	0.005

¹OR is the Odds Ratio for the additive genotypic model;

²95% CI is the 95% confidence interval of the Odds Ratio;

³Deviated from HWE in fetal controls (p=0.007), but not cases.

^*Significant after FDR correction.

The most significant single locus association in fetal DNA was an intronic SNP in fibronectin 1 isoform 3 (FN1) (rs3796123, OR = 1.46 95% CI [1.20-1.78], p = 0.0001) (Table 4). The minor allele frequency (A) for rs3796123 in fetal DNA was reversed in cases and controls (allele frequency of 0.51 in cases and 0.44 in controls). Fetal SNP rs3796123 remained statistically significant after FDR correction. Additional SNPs that were associated with SGA at a p-value of < 0.05 are presented in Table 5. Although we have not emphasized these findings in the present report, in some instances they represent associations in pathways previously reported, (e.g., IGF1R and IGF1 in the present report and IGF2 in Nagaya et al.[108]) that may lend support to previous findings or may be additional SNPs in genes reported with a p-value < 0.01. Such findings strengthen the likelihood of an association because they are based on multiple SNPs for the same gene.

Table 5.

Single locus tests of association (p < 0.05) adjusted for fetal sex and BMI

Population	Gene Code	rs#	Minor Allele	Minor Allele Frequency		OR1	95% CI2		p-value
				Cases	Controls		Lower	Upper
Maternal	APOE	rs405509	G	0.45	0.48	0.84	0.70	0.99	0.043
	CCL2	rs2857657	G	0.12	0.09	1.35	1.01	1.79	0.042
	CD55	rs28371583	G	0.24	0.28	0.81	0.67	0.99	0.036
	COL1A1	rs17639446	G	0.09	0.05	1.62	1.17	2.27	0.004
	COL1A1	rs11079898	G	0.02	0.03	0.54	0.30	0.95	0.032
	COL3A1	rs7579903	A	0.11	0.08	1.35	1.01	1.81	0.044
	COL4A2	rs41315048	T	0.13	0.10	1.32	1.01	1.73	0.046
	COL4A3	rs1882435	A	0.32	0.26	1.27	1.05	1.54	0.014
	COL4A3	rs12621551	T	0.15	0.19	0.78	0.62	0.98	0.033
	CRHR2	rs4723000	T	0.13	0.09	1.46	1.11	1.92	0.007
	CSF1	rs1058885	C	0.25	0.22	1.22	1.00	1.48	0.050
	CSF2	rs1469149	C	0.21	0.24	0.79	0.65	0.97	0.026
	ELN	rs2071307	A	0.25	0.20	1.36	1.09	1.68	0.005
	F2	rs5896	T	0.30	0.26	1.23	1.01	1.50	0.038
	FGF4	rs3740640	G	0.08	0.05	1.68	1.17	2.40	0.005
	FN1	rs3796123	A	0.50	0.46	1.27	1.05	1.54	0.013
	GP1BA	rs6066	T	0.03	0.02	1.97	1.08	3.57	0.026
	IGF1R	rs3743262	T	0.17	0.14	1.28	1.01	1.62	0.045
	IL10	rs1800872	A	0.37	0.32	1.25	1.04	1.50	0.018
	IL1B	rs1143634	T	0.13	0.17	0.73	0.57	0.93	0.012
	IL4R	rs1029489	A	0.43	0.48	0.80	0.67	0.95	0.011
	IL6R	rs8192282	A	0.13	0.08	1.57	1.18	2.10	0.002
	MGC4093	rs1800471	C	0.05	0.03	1.55	1.02	2.37	0.042
	MMP1	rs470132	T	0.22	0.18	1.33	1.05	1.68	0.016
	MMP10	rs17860950	C	0.04	0.03	1.71	1.05	2.78	0.030
	MMP10	rs486055	A	0.06	0.04	1.71	1.12	2.60	0.013
	MMP11	rs28363645	A	0.02	0.03	0.50	0.26	0.94	0.033
	MMP15	rs28733439	T	0.02	0.04	0.56	0.34	0.93	0.024
	MMP16	rs3739382	T	0.02	0.01	2.16	1.10	4.21	0.025
	MMP9	rs2274755	T	0.09	0.05	1.62	1.16	2.27	0.005
	NPPA	rs17376426	A	0.01	0.02	0.38	0.15	0.98	0.045
	PLAU	rs2227562	A	0.06	0.08	0.67	0.47	0.95	0.023
	PTGER3	rs5673	A	0.07	0.04	1.76	1.19	2.61	0.005
	PTGS1	rs1236913	T	0.07	0.04	1.57	1.07	2.30	0.020
	PTGS1	rs3842798	C	0.29	0.25	1.27	1.05	1.54	0.013
	RNASE4	rs11701	G	0.14	0.18	0.75	0.59	0.96	0.023
	SERPINC1	rs677	C	0.07	0.04	1.51	1.03	2.22	0.036
	TAP1	rs2071538	T	0.20	0.16	1.28	1.02	1.60	0.035
	THBS4	rs17885983	C	0.18	0.16	1.27	1.00	1.59	0.046
	TIMP2	rs2277698	A	0.11	0.07	1.71	1.26	2.32	0.0006
	TLR3	rs5743312	T	0.15	0.12	1.28	1.01	1.63	0.042
	TLR3	rs3775290	T	0.30	0.26	1.22	1.01	1.47	0.036
	TLR9	rs352140	A	0.50	0.43	1.26	1.07	1.50	0.007
	TNR	rs1385540	T	0.21	0.17	1.28	1.02	1.60	0.032
	TBXAS1	GNSC_53711588	T	0.02	0.01	2.75	1.12	6.76	0.028
	VWF	GNSC_635156524	A	0.02	0.01	2.56	1.10	5.97	0.029
	IL12B	GNSC_632292870	G	0.02	0.01	2.28	1.01	5.14	0.047
Fetal	APOC3	rs2070666	A	0.28	0.23	1.25	1.02	1.53	0.033
	APOE	rs769449	A	0.10	0.07	1.46	1.05	2.05	0.026
	CCR2	rs3749461	G	0.05	0.03	1.61	1.00	2.59	0.049
	COL1A2	rs13240759	C	0.37	0.42	0.81	0.67	0.99	0.036
	COL4A1	rs1816884	G	0.34	0.30	1.22	1.00	1.47	0.046
	EDN1	rs1800543	C	0.23	0.27	0.79	0.64	0.98	0.032
	F13A1	rs5982	T	0.12	0.08	1.52	1.12	2.05	0.007
	FLT4	rs2242213	C	0.12	0.07	1.62	1.19	2.20	0.002
	FN1	rs2304573	C	0.39	0.46	0.72	0.59	0.87	0.001
	FN1	rs3796123	A	0.51	0.44	1.46	1.20	1.78	0.0001
	GP1BA	rs6066	T	0.04	0.02	2.09	1.21	3.62	0.008
	IGF1	rs5742620	A	0.04	0.02	1.94	1.14	3.30	0.014
	IL1B	rs1143643	A	0.28	0.24	1.24	1.00	1.53	0.049
	IL2RA	rs7899538	T	0.07	0.05	1.50	1.03	2.19	0.036
	IL2RA	rs10752175	A	0.35	0.31	1.24	1.02	1.52	0.029
	IL2RA	rs12358961	T	0.50	0.45	0.82	0.68	0.99	0.037
	IL6R	rs8192282	A	0.11	0.08	1.49	1.10	2.04	0.011
	IL9	rs2069885	T	0.08	0.06	1.51	1.05	2.18	0.026
	IMP5	rs28364026	A	0.11	0.15	0.77	0.60	0.99	0.045
	LPL	rs11570891	T	0.10	0.07	1.42	1.03	1.97	0.032
	MMP10	rs486055	A	0.07	0.04	1.73	1.16	2.59	0.007
	MMP11	rs28363665	A	0.01	0.02	0.33	0.13	0.83	0.019
	MMP17	rs11835665	A	0.16	0.21	0.74	0.59	0.94	0.013
	MMP9	rs2274755	T	0.07	0.05	1.59	1.09	2.33	0.017
	PTGER1	rs10421765	A	0.01	0.02	0.41	0.18	0.95	0.037
	PTGER4	rs2228058	T	0.05	0.03	1.63	1.04	2.54	0.033
	REN	rs3730102	T	0.21	0.17	1.31	1.04	1.64	0.020
	TIMP2	rs2277698	A	0.13	0.10	1.51	1.13	2.03	0.005
	TLR4	rs4986790	G	0.05	0.03	1.96	1.20	3.23	0.008
	TLR6	rs5743818	G	0.11	0.08	1.38	1.01	1.88	0.043
	TNFRSF1B	rs5746016	T	0.01	0.03	0.30	0.12	0.73	0.008
	HSPG2	GNSC_634098268	T	0.23	0.18	1.45	1.15	1.82	0.002
	VWF	GNSC_635285421	A	0.02	0.01	3.81	1.47	9.84	0.006
	SERPINE1	GNSC_629203538	T	0.01	0.01	3.02	1.12	8.14	0.029
	LFT	GNSC_633838704	C	0.13	0.10	1.34	1.02	1.76	0.037

Haplotype Tests of Association

Haplotype analyses identified six genes in maternal DNA [interleukin 6 receptor isoform 1 (IL6R), matrix metalloproteinase 9 (MMP9), alpha 1 type I collagen preprotein (COL1A1), fibroblast growth factor 4 precursor (FGF4), coricotropin releasing hormone receptor 2 (CRHR2), and elastin (ELN)] that associated with SGA and one in fetal DNA, fibronectin 1 (FN1) (Table 6, Figure 1 and Figure 2). For maternal DNA haplotypes the most significant association was a two SNP haplotype in COL1A1 rs2141279-rs17639446 (global p = 0.005). The G-G haplotype was the most significant haplotype with the strongest effect (OR = 1.92 95% CI [1.35-2.74], p = 0.0001). This haplotype included COL1A1 SNP rs17639446, a SNP that independently associated with SGA (Table 4) (p = 0.004) where the G allele was the risk allele. These two SNPs are in weak LD (r² = 0.02 in cases and r² = 0.01 in controls, Figure 1).

Table 6.

Haplotype sliding windows association results unadjusted for covariates (p < 0.05)

Population	Gene Code	SNP rs#	Haplotype	Frequency		OR	95% CI		p-value
				Cases	Controls		Lower	Upper
Maternal	IL6R	rs8192498-rs4723000-rs8192496	Global p						0.012
			GCT	0.71	0.75	-	-	-	-
			GCC	0.16	0.17	0.99	0.79	1.25	0.965
			GTC	0.13	0.09	1.53	1.15	2.02	0.002
	MMP9	rs3918253-rs2274755	Global p						0.009
			CG (Referent)	0.64	0.68	-	-	-	-
			TG	0.27	0.27	1.06	0.88	1.29	0.518
			CT	0.09	0.06	1.59	1.13	2.22	0.005
	COL1A1	rs2141279-rs17639446	Global p						0.005
			GT (Referent)	0.68	0.73	-	-	-	-
			AT	0.23	0.22	1.12	0.91	1.38	0.261
			GG	0.09	0.05	1.92	1.35	2.74	0.0001
	FGF4	rs3740640-rs2073464	Global p						0.033
			AT (Referent)	0.49	0.50	-	-	-	-
			AC	0.43	0.45	0.98	0.82	1.16	0.785
			GT	0.08	0.05	1.64	1.14	2.36	0.006
	CRHR2	rs8192282-rs7521458	Global p						0.042
			GC (Referent)	0.64	0.66	-	-	-	-
			GT	0.25	0.26	0.99	0.82	1.21	0.952
			AT	0.12	0.08	1.54	1.15	2.07	0.003
	ELN	rs2071307-rs2856728	Global p						0.032
			GT (Referent)	0.50	0.53	-	-	-	-
			GC	0.26	0.27	1.02	0.84	1.25	0.816
			AT	0.25	0.20	1.32	1.07	1.64	0.009
Fetal	FN1	rs2304573-rs1250204-rs1250215	Global p						0.045
			TGG (Referent)	0.67	0.43	-	-	-	-
			CGC	0.29	0.52	0.36	0.30	0.43	<0.0001
			TGC	0.04	0.05	0.51	0.32	0.80	0.002

Figure 1. Haploview plots of genes identified in analyses of haplotype tests of association in maternal samples.

LD plots were generated in Haploview and are presented for: A) MMP9 controls r²; B) MMP9 cases r²; C) COL1A1 controls r²; D) COL1A1 cases r²; E) IL6R controls r²; F) IL6R cases r²; G) FGF4 controls r²; H) FGF4 cases r²; I) CRHR2 r² controls; J) CRHR2 cases r²; K) ELN r² controls; L) ELN r² cases. Within each triangle is presented the pairwise correlation coefficient (r²) LD plots white, (r² = 0), shades of grey, (0 < r² < 1), black, (r² = 1).

Figure 2. Haploview plots of genes identified in analyses of haplotype tests of association in fetal samples.

LD plots were generated in Haploview and are presented for: A) FN1 controls r²; B) FN1 cases r². Within each triangle is presented the pairwise correlation coefficient (r²) LD plots white, (r² = 0), shades of grey, (0 < r² < 1), black, (r² = 1).

The only haplotype that showed an association in fetal DNA was in FN1, rs2304573-rs1250204-rs1250215 (global p = 0.045) with the C-G-C haplotype having the strongest association (OR = 0.36 95% CI [0.30-0.43], p < 0.0001). One of these SNPs, rs2304573, was independently associated with SGA, although this is not the most significant FN1 SNP in single locus analyses; the other two did not associate with predisposition to SGA (p > 0.05 for both).

Multilocus analysis

Exploratory MDR analyses were performed using different filtering approaches as described in statistical methods. In maternal DNA, one significant 2 locus interaction model was identified after SuRF and TuRF filtering (GNSC_635148713 and rs2227562; testing balanced accuracy 0.59, p = 0.035; CVC = 10/10) (Table 7A, Figure 3). GNSC_635148713 is in collagen type V alpha 2 (COL5A2) and rs2227562 is in plasminogen activator urokinase (PLAU), suggesting an interaction between genes in the extracellular matrix metabolism and hemostasis.

Table 7.

Summary of MDR analyses

A. Maternal results with tagged SNPs
Model	Training Balance Accuracy	Testing Balance Accuracy	Cross Validation Consistency	p-value
All SNPs
rs352140	0.5616	0.5073	4/10	0.8713
rs243832 RS1800774	0.5969	0.5613	3/10	0.1899
Surf_n_Turf- 25 SNPs
rs1143634	0.5463	0.5310	3/10	0.4934
GNSC_635148713 RS2227562	0.5899	0.5899	10/10	0.0352
GNSC_635148713 RS8178674 RS2227562	0.6177	0.5766	4/10	0.0897
Turf Option - 10 SNPs
rs352140	0.5577	0.5417	8/10	0.3498
rs352140 RS2621208	0.5920	0.5491	5/10	0.2897
rs7521458 RS352140 RS2621208	0.6238	0.5798	5/10	0.0822
Genotypic p-value cutoff - 0.05
rs352140	0.5616	0.5073	4/10	0.8713
rs352140 RS1800774	0.5937	0.5675	5/10	0.0928
rs1882435 RS352140 RS7719521	0.6356	0.5188	2/10	0.6883
Genotypic p-value cutoff - 0.10
rs352140	0.5616	0.5073	4/10	0.8713
rs352140 RS1800774	0.5937	0.5675	5/10	0.0928
rs4674259 RS352140 RS598893	0.6341	0.4886	2/10	0.8085

B. Fetal results with tagged SNPs
Model	Training Balance Accuracy	Testing Balance Accuracy	Cross Validation Consistency	p-value
All SNPs
rs3796123	0.5571	0.5355	5/10	0.4326
rs1250215 RS2621208	0.6006	0.5350	4/10	0.4557
rs1250215 RS956732 RS11603042	0.6497	0.5403	3/10	0.4123
Surf_n_Turf- 25 SNPs
GNSC_623246114	0.5524	0.5231	6/10	0.5643
rs11541998 GNSC_623246114	0.5906	0.5755	5/10	0.1110
rs11835665 RS2293117 GNSC_623246114	0.6223	0.5606	8/10	0.2071
Turf Option - 10 SNPs
rs1250215	0.5545	0.5267	5/10	0.5649
rs1250215 RS11541998	0.5935	0.5741	6/10	0.0883
rs2076059 RS1250215 RS1077835	0.6432	0.5873	9/10	0.0768
Genotypic p-value cutoff - 0.05
rs3796123	0.5571	0.5355	5/10	0.4326
rs1049033 RS2293117	0.5921	0.5511	6/10	0.2996
rs9332607 RS1049033 RS2293117	0.6319	0.5631	3/10	0.1881
Genotypic p-value cutoff - 0.10
rs3796123	0.5571	0.5355	5/10	0.4326
rs1049033 RS2293117	0.5925	0.5456	5/10	0.3541
rs2304573 RS1049033 RS405509	0.6345	0.4874	2/10	0.7984

C. Combined results with tagged SNPs
Model	Training Balance Accuracy	Testing Balance Accuracy	Cross Validation Consistency	p-value
All SNPs
rs352140	0.5654	0.5182	5/10	0.7045
rs352140 GNSC_30713312	0.6102	0.5234	3/10	0.6406
Surf_n_Turf- 25 SNPs
rs6909681	0.5560	0.5152	4/10	0.7657
GNSC_635148713 RS11541998_2	0.5943	0.5454	4/10	0.3196
GNSC_635148713 RS11541998_2 GNSC_63246114_2	0.6271	0.5791	4/10	0.1132
Turf Option - 10 SNPs
rs7719521	0.5586	0.5232	8/10	0.6509
rs7719521 RS12358961_2	0.5891	0.5513	5/10	0.3138
rs7719521 RS1049033 RS243864	0.6411	0.5448	5/10	0.3803
Genotypic p-value cutoff - 0.05
rs352140	0.5654	0.5182	5/10	0.7045
rs352140 GNSC_30713312	0.6052	0.5680	7/10	0.1785
rs1049033_2 Rs3842798 RS2293117_2	0.6467	0.5539	2/10	0.2886
Genotypic p-value cutoff - 0.10
rs352140	0.5654	0.5182	5/10	0.7045
rs352140 GNSC_30713312	0.6064	0.5576	5/10	0.2534
rs3730103_2 RS352140 RS7719521	0.6477	0.5736	4/10	0.1451

Figure 3. MDR schematic for 2-Way interaction involving maternal SNPs GNSC_635148713 in collagen type V alpha 2 (COL5A2) and rs2227562 in plasminogen activator urokinase (PLAU).

Dark gray boxes represent high risk genotypes and light gray boxes low risk genotypes. The left hand bar in each box is the number of PTB cases with that genotype and the right bar is the number of term controls with that genotypes.

There were no statistically significant multilocus interactions (p < 0.05) in fetal DNA; however, there was one interaction between rs2076059, rs1250215, and rs1077835 that had a balanced accuracy of 0.59, p = 0.077 and a CVC of 9/10 (Table 7B). rs2076059 is in selectin E (SELE), rs1250215 is in fibronectin 1 isoform 3 (FN1) and rs1077835 is in hepatic lipase (LIPC). None of these SNPs were among our strongest (p < 0.01) single locus associations, although the fibronectin SNP rs1250215 is in LD with SNPs that associated either individually (rs2304573) or in haplotypes in our samples (r² = 0.58 in cases and r² = 0.64 in controls, Figure 2).

Pathway analysis

To discover novel networks of interacting molecules that could be involved in SGA, the IPA was seeded with our focus genes meeting the p < 0.05 criterion. In the mothers, the IPA network algorithm discovered that our focus biomarkers are incorporated into four highly significant networks with scores ranging between 8 and 42 (corresponding to IPA network p-values of p = 10⁻⁸ to p = 10⁻⁴² (Supplemental Table 2A). The network diagram for the top scoring network is shown in Figure 4. The top ranked function of this network was connective tissue disorders, which involved regulatory interactions of TIMP2, MMPs, PLAU, collagen and PTGS1.

Figure 4. Connection map for first ranked maternal network generated by IPA from focus gene input.

The biomarkers passing the p < 0.05 significance threshold (focus molecules) were entered in to the IPA software for an unsupervised functional analysis to discern regulatory networks involving these molecules. Focus molecules are depicted in pink, and the most statistically-significant molecule is in red. The asterisk indicates that there was more than one SNP probe for the gene tested and the most significant value was placed into the analysis. Solid lines show direct interaction (binding/physical contact); dashed line, indirect interaction supported by the literature but possibly involving one or more intermediate molecules that have not been investigated definitively. Molecular interactions involving only binding are connected with a line without an arrowhead since directionality cannot be inferred.

Using the fetal genes as input, two highly significant networks (scoring 24 and 44), and two marginally significant networks (each scoring 3) were detected (Supplemental Table 2B). The most significant functional roles for this network were cell signaling, tissue development (including extracellular matrix remodeling), and reproductive system development. Fibronectin 1 participated in both the top scoring network (Figure 5) and the third-ranked network.

Figure 5. Connection map for first ranked fetal network generated by IPA from focus gene input.

The biomarkers passing the p < 0.05 significance threshold (focus molecules) were entered in to the IPA software for an unsupervised functional analysis to discern regulatory networks involving these molecules. Focus molecules are depicted in pink, and the most statistically-significant molecule is in red. The asterisk indicates that there was more than one SNP probe for the gene tested and the most significant value was placed into the analysis. Solid lines show direct interaction (binding/physical contact); dashed line, indirect interaction supported by the literature but possibly involving one or more intermediate molecules that have not been investigated definitively. Molecular interactions involving only binding are connected with a line without an arrowhead since directionality cannot be inferred.

DISCUSSION

Principal findings of the study

We report the results of the first, large scale genetic association study, of carefully phenotyped women with an SGA neonate from a homogeneous Hispanic population. This study identified DNA variants that predispose to SGA. The main observations were: 1) a SNP in FN1 in the fetus increased the risk for SGA; 2) haplotypes in COL1A1 and MMP9 in the mother and FN1 in the fetus were associated with SGA; 3) multilocus analysis with MDR revealed a significant interaction between PLAU and COL5A2 in mothers; 4) pathway analysis (Supplemental Table 2) suggested that maternal and fetal genes involved in extracellular matrix function predispose to SGA; 5) importantly, TIMP2 and FN1 were found to participate in both maternal and fetal networks along with other molecules associated with SGA. Collectively, these findings support the concept that there is a genetic component in the risk of SGA and that this risk is mediated in part by extracellular matrix function (e.g. cell adhesion, chemotaxis, collagen remodeling).

Single locus analysis

The findings of this study support a significant role for a fibronectin genetic variant in the fetal DNA. Fibronectin is a well documented glycoprotein present in the intrauterine tissues and cervix that interacts with extracellular matrix components including collagen and integrins.[144,145] It is localized at the maternal-fetal interface of the fetal membranes and decidua, acting as “glue” between the chorioamniotic membranes and the uterus. The presence of high concentrations of fetal fibronectin in the amniotic fluid of asymptomatic women during second trimester has been associated with the subsequent development of severe preeclampsia and intrauterine growth restriction, suggesting involvement of extracellular matrix.[146,147] In vitro studies documented that excess fibronectin stimulates cell growth and reverse the established growth arrest of tumor cells.[148] This supports the hypothesis that FN1 can have a role in fetal growth. Although we identified a putative association between TIMP2 and SGA in mothers, it did not remain significant after correcting for multiple testing. However, TIMP2 appears to play a central role as part of the extracellular matrix metabolism associated with SGA as documented by IPA (see below). The significance of the fetal FN1 association and the weaker association with maternal gene variants supports the view that isolated SGA is predominantly a disease of fetal origin. However, our multilocus analyses described below do not necessarily support this.

Haplotype analyses identified novel genes predisposing to SGA

Haplotype analysis of maternal and fetal DNA revealed association between SGA and genes involved in the extracellular metabolism. Specifically, we identified significant haplotypes in MMP9 and COL1A1 in the maternal genome and FN1 in the fetal genome. In the FN1 analysis, 3 SNP haplotypes that associated with SGA did not include the one SNP that showed single locus significance after testing for multiple testing, suggesting that multiple variants in this gene may affect the risk of SGA.

MMP9 is specifically involved in the degradation of the basement membrane collagen type IV and has been documented to be associated with pregnancy complication such as spontaneous preterm birth and pPROM.[149-161] MMP9 activity is essential for the of trophoblast cells invasion into the maternal endometrium and myometrium during embryo implantation.[149]

Exploratory Analysis

SGA is syndromic in nature, and multiple mechanisms of disease are likely to be involved.[2,3] To address the complexity of the genetic predisposition to this phenotype, we performed two exploratory analyses. MDR explicitly addressed the potential role of interactions among genes (maternal, fetal, and maternal-fetal). The MDR results in maternal DNA indicate a significant interaction between PLAU and COL5A2. This finding is of note, because in monocytes PLAU has been shown to induce the expression of MMP9 that can promote proteolysis. In addition, maternal MMP9SNP (rs2274755) was nominally associated with SGA. It is of note that the most significant model in maternal DNA is a two locus model, implying that the role of the maternal genome is more complex than that of the fetal genome in affecting risk of SGA. In fetal DNA, no MDR model was statistically significant. However, MDR data provided weak evidence (p = 0.077) of epistatic interaction between SNPs in the fibronectin, selectin E and hepatic lipase genes. Taken with our other results, this lends support to the involvement of extracellular matrix and adhesion related molecules especially through FN1.

There is an increasing realization of the importance of pathways in determining the phenotype of complex diseases. We utilized IPA to examine the contribution of genetic variants in determining networks and disease functions. In maternal IPA analysis, the top ranking network incorporated connective tissue disorders and included 18 of the genes that we found to be significantly associated with SGA (p < 0.05) in single locus analysis. This network was highly enriched with collagens, MMPs, and their regulators including PLAU and TIMP2, further emphasizing the importance of extracellular matrix metabolism in predisposing to SGA. This network also included growth hormone, indicating a functional connection between this hormone and extracellular matrix molecules associated with SGA.

Strengths and limitations of the study

This is the largest study to examine the genetic predisposition to SGA in terms of number of subjects studied and number of genetic variants analyzed in both maternal and fetal DNA. Other strengths include a well-defined phenotype (SGA) and a homogeneous population. Importantly, we identified that both maternal and fetal DNA variants contributed to modify the risk of SGA, although only fetal single locus results remained significant after multiple testing correction. As all studies of this type, our findings require confirmation in additional cohorts. In addition, we have not incorporated environmental variables that can potentially interact with genetic variations in our analysis that are known to play a role in the risk of SGA. Functional studies are also needed to assess the precise physiologic implications of the DNA variants identified in this study.

Conclusion

This large genetic association study of candidate genes suspected to be involved in adverse pregnancy outcome revealed that different maternal and fetal DNA variants are associated with SGA.