Transcriptomics of Maternal and Fetal Membranes Can Discriminate between Gestational-Age Matched Preterm Neonates with and without Cognitive Impairment Diagnosed at 18–24 Months

Background Neurocognitive impairment among children born preterm may arise from complex interactions between genes and the intra-uterine environment. Objectives (1) To characterize the transcriptomic profiles of chorioamniotic membranes in preterm neonates with and without neurocognitive impairment via microarrays and (2) to determine if neonates with neurocognitive impairment can be identified at birth. Materials/Methods A retrospective case-control study was conducted to examine the chorioamniotic transcriptome of gestational-age matched very preterm neonates with and without neurocognitive impairment at 18–24 months’ corrected-age defined by a Bayley-III Cognitive Composite Score <80 (n = 14 each). Pathway analysis with down-weighting of overlapping genes (PADOG) was performed to identify KEGG pathways relevant to the phenotype. Select differentially expressed genes were profiled using qRT-PCR and a multi-gene disease prediction model was developed using linear discriminant analysis. The model’s predictive performance was tested on a new set of cases and controls (n = 19 each). Results 1) 117 genes were differentially expressed among neonates with and without subsequent neurocognitive impairment (p<0.05 and fold change >1.5); 2) Gene ontology analysis indicated enrichment of 19 biological processes and 3 molecular functions; 3)PADOG identified 4 significantly perturbed KEGG pathways: oxidative phosphorylation, Parkinson’s disease, Alzheimer’s disease and Huntington’s disease (q-value <0.1); 4) 48 of 90 selected differentially expressed genes were confirmed by qRT-PCR, including genes implicated in energy metabolism, neuronal signaling, vascular permeability and response to injury (e.g., up-regulation of SEPP1, APOE, DAB2, CD163, CXCL12, VWF; down-regulation of HAND1, OSR1)(p<0.05); and 5) a multi-gene model predicted 18–24 month neurocognitive impairment (using the ratios of OSR1/VWF and HAND1/VWF at birth) in a larger, independent set (sensitivity = 74%, at specificity = 83%). Conclusions Gene expression patterns in the chorioamniotic membranes link neurocognitive impairment in preterm infants to neurodegenerative disease pathways and might be used to predict neurocognitive impairment. Further prospective studies are needed.


Introduction
While advances in perinatal medicine have improved the survival and short-term outcomes of preterm neonates, rates of neurodevelopmental impairment at 18-24 month follow-up and beyond remain high [1][2][3][4][5][6][7]. Neurocognitive deficits are among the most prevalent and most debilitating forms of early childhood disabilities, reported in 23% of infants born 27-32 weeks' gestation and 37% of infants born at 22-26 weeks' gestation [4]. Cognitive impairment can impact adaptive functioning, conceptual, social, and practical domains, and lead to high personal, familial, societal and financial costs. The estimated US average lifetime costs to care for an individual with intellectual impairment is $1,014,000 [8].
Neurocognitive disorders may arise from complex interactions between genes and the environment, originating prior to birth. Though postnatal interventions have afforded limited success in preventing neurocognitive and developmental impairments associated with prematurity, prenatal interventions such as antenatal steroids [9][10][11][12][13] and magnesium sulfate [14][15][16][17][18] provide greater population impact. The search for intrauterine or perinatal disease pathways associated with fetal and neonatal brain injury may afford new insights into preventive measures and disease pathogenesis. Other investigators have utilized mRNA levels in blood samples collected soon after birth to identify children at risk for other neurodevelopmental disorders such as cerebral palsy [19] and autism [20].
The fetal membranes are an alternative source of fetal DNA and of human fetal stem cells [21] that may be impacted by intrauterine stimuli. Stem cells derived from the fetal membranes are available after every preterm birth and have pluripotent differentiation potential [22,23]. Embryonic [24,25] and pluripotent stem cells [26] have emerged as powerful tools in the study of normal neuronal development and of neuropsychiatric disorders such as Parkinson's disease [27][28][29][30], Rett syndrome [31][32][33], fragile X [34,35], Down's syndrome [36,37] and schizophrenia [38][39][40][41]. Recent data suggests that there are no significant differences between human embryonic and induced pluripotent stem cell gene expression levels [42][43][44], thus the study of pluripotent stem cells (including fetal amnion and chorion cells) [21] provides a pragmatic, yet noncontroversial methodology to readily access large numbers of relevant cells from multiple cases and controls. Changes in gene expression of the chorioamniotic membranes may capture in-utero insults and fetal response to injury in preterm infants. Our objectives were (1) to characterize the molecular profile of the chorioamniotic membranes of preterm neonates with and without neurocognitive impairment at 18-24 months' corrected age and (2) to determine if neonates with neurocognitive impairment have a molecular signature that can be used to predict future disease onset at the time of birth.

Study participants
A retrospective case-control study was conducted to examine the chorioamniotic membranes of 66 very preterm neonates with and without neurocognitive impairment. Cases and controls were singleton neonates born at Hutzel Women's Hospital (Detroit, MI) between January 1, 2006 and December 31, 2010 who were matched for gestational age (+ 2 weeks) and born between 23 and 32 weeks of gestation. Neurocognitive impairment was defined by a Bayley scales of infant development, 3 rd Edition cognitive composite score <80 with or without associated neuromotor impairment at 18-24 months' corrected age [45]. Control infants had normal neurodevelopmental assessments including cognitive testing, neurological examination and gross motor function [46]. The Bayley scales of infant development, 3 rd Edition has a mean (SD) cognitive composite score of 100 (15). A cut score of 80 was selected based on data from recent population studies [47,48]. The infants' mothers provided written informed consent for the collection of biological materials and clinical data for research purposes under protocols approved by the Institutional Review Boards of the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD/NIH/DHHS, Bethesda, Maryland) and the Human Investigation Committee of Wayne State University (Detroit, MI, USA). Neonatal and neurodevelopmental outcomes were abstracted from the clinical records.

Sample collection
Chorioamniotic membrane samples were retrieved from the bank of biologic samples of Wayne State University, the Detroit Medical Center, and the Perinatology Research Branch of the Eunice Kennedy Shriver National Institutes of Child Health and Human Development (NICHD) (Detroit, MI). At the time of specimen collection, the fetal membranes were dissected from the placenta, rolled, cut into small pieces and flash-frozen using liquid nitrogen [49]. In addition, a section of membranes containing maternal decidua was fixed and embedded in paraffin. 5mm paraffin sections were stained with hematoxylin and eosin and examined under bright-field light microscopy [50]. Histological examinations were reported by placental pathologists who were blinded to the group assignment and all clinical information.

RNA isolation
Total RNA was isolated from snap-frozen tissues using TRIzol reagent (Invitrogen, Carlsbad, CA, USA) and Qiagen RNeasy kit (Qiagen, Valencia, CA, USA) according to the manufacturers' recommendations. RNA concentrations and A260nm/A280nm ratios were assessed using a NanoDrop 1000 (Thermo Scientific, Wilmington, DE, USA). The 28S/18S ratios and RNA integrity numbers were assessed using a Bioanalyzer 2100 (Agilent Technologies, Wilmington, DE, USA).

qRT-PCR assays with biomark system
Total RNA (500 ng) was reverse transcribed using the SuperScript III First-Strand Synthesis System and oligo(dT) 20 primers (Invitrogen, Carlsbad, CA, USA). TaqMan Assays (Applied Biosystems, Foster City, CA, USA) were used for gene expression profiling on the Biomark high-throughput qRT-PCR system (Fluidigm, San Francisco, CA, USA) according to the manufacturers' instructions. Briefly, a 0.2X pool of specific gene expression assays (S1 Table) (Applied Biosystems) was used as the source of primers. Preamplification reactions contained 1.25μl cDNA, 2.5 μl TaqMan PreAmp master mix (Applied Biosystems) and 1.25μl pooled assay mix. The reaction was performed with a thermal cycler for 14 cycles at 95°C for 15 seconds and 60°C for 4 minutes. After cycling, the reaction was diluted 1:5 with ddH 2 O to a final volume of 25μl. Next, a Fluidigm 96.96 Dynamic Array chip was primed in an Integrated Fluidic Circuit controller. Then, 2.5μl 20X TaqMan gene expression assays (Applied Biosystems) were mixed with 2.5μl 2X assay loading reagent (Fluidigm) and loaded into the assay inlet on the 96.96 array chip. 2.25μl preamplified cDNA was mixed with 2.5μl TaqMan Universal PCR master mix (Applied Biosystems) and 0.25μl 20X sample loading reagent (Fluidigm), and loaded into the sample inlet on the chip. The chip was returned to the Integrated Fluidic Circuit controller for loading. After loading the samples and assays, the chip was placed into the Biomark System to run the reactions.

Statistical analysis
Clinical Data: The maternal and neonatal demographic and clinical characteristics of the two study groups were compared using the Wilcoxon rank-sum test or t-test for between-group comparisons of continuous data as appropriate. The Chi-square or Fisher's exact tests were used for comparisons of categorical data. Statistical analyses of demographic data were performed using SPSS version 19 (SPSS Inc, Chicago, IL). A p-value <0.05 was used to designate statistical significance.
Microarray Data: Gene expression data that was measured on the first set of cases and controls (n = 14 each) was offset by adding a constant value of 1 to enable subsequent log (base 2) transformation. A quantile-normalization procedure [52] implemented in the preprocessCore package of Bioconductor (http://www.bioconductor.org) was then applied to remove non-biological systematic biases in the intensity data and hence make it comparable between arrays. A linear model was fit to the data of each probe to test the association between the gene expression and the phenotype (neurocognitive impairment vs. control). The significance of the group coefficient in the linear model was inferred using p<0.05 from a paired moderated t-test together with a minimum of a 1.5-fold-change between groups [53]. Although it is customary in microarray analyses to choose significant genes based on adjusted p-values [54], as shown in the MicroArray Quality Control (MAQC) study [55], reproducible differential expression results also can be obtained using a nominal p<0.05 cut-off provided the magnitude of changes is considered in gene selection. Since the microarray study was followed by a targeted qRT-PCR experiment involving a new set of samples that could rule out some of the eventual false positives, we used a less stringent significance cut-off in the microarray experiment to minimize false negatives. Gene Ontology analysis of significant genes was conducted using GOstats [56]. Pathway analysis with down-weighting of overlapping genes (PADOG) was conducted to identify Kyoto Encyclopedia of Genes and Genomes (KEGG) gene sets and biological pathways relevant to the group phenotype [57,58]. PADOG leverages differential evidence from all genes in a pathway while giving more weight to genes that are specific to a given pathway than to those that appear in multiple pathways.
qRT-PCR Data: Ninety differentially expressed genes based on the microarray data were selected for qRT-PCR profiling in an extended set of cases and controls (n = 33 each). There were two goals with the qRT-PCR data analysis. First, to verify the 90 genes found significant based on microarray analysis in the first set of 14 case-control pairs, and second, to determine accuracy of a multi-gene predictor based on these data. For the first goal, the extended set of 33 case-control pairs were used to verify the microarray results. Differential expression based on Cycle threshold (Ct) data was inferred using p<0.05 from a one-tailed t-test using the direction of change for each of the 90 genes in the microarray experiment as the hypothesis test alternative.
For the second goal, qRT-PCR measured expression data were split into a "training" set and a "test" set ("hold-out validation" procedure). The training/learning set included 14 cases and 14 controls that were used in the microarray experiment. The test set included 19 cases and 19 controls with blinded class membership, not used in any stage of the prediction model development. To obtain a classifier from the training data without using any information from the test data, all 90 candidate genes found significant by microarrays were considered as inputs in a classifier development pipeline that we have previously described in detail [59] and adapted for the qRT-PCR data in the current study. Briefly, with this procedure, each of the 90 genes was considered in turn as a potential normalizer and log 2 gene expression ratios between each remaining gene and the normalizer gene were computed. The log 2 ratios were simply the differences in -Ct values of each gene and the reference gene. The gene expression ratios were then ranked using p-values from a moderated t-test and those that did not change at least 1.2 fold were discarded. The top ranked gene ratios were used in a linear discriminant analysis (LDA) model implemented in the MASS package in R (http://www.r-project.org/). The number of top ratios to use in the classifier was optimized by maximizing the average model sensitivity over three cut-offs of specificity (80,85, and 90%). The sensitivity calculations were performed using a five time repeated three-fold cross validation procedure on the training data that included both the ratios ranking and LDA model fitting steps, functionality that we have made available in the maPredictDSC package [59] of Bioconductor (http://www.bioconductor.org). The sensitivity estimate of the resulting model for the optimal number of top ratios was determined for each possible normalizer gene, and the one that provided the highest sensitivity was retained. The ratios ranking and LDA model training were then performed on all training data to produce a final model. The cross-validated performance on the training set for the optimum number of predictors was determined for a quadratic discriminant analysis (QDA) and a support vector machines (SVM) classifier as well [60]. Implementations of these algorithms in R were available in the MASS and e1071 packages, respectively.
The trained model was applied on the test set to calculate an unbiased estimate of the predictive performance (sensitivity at fixed specificity). The predictive performance of the gene based classifier was compared with that of a model that used clinical information available at the time of birth: gestational age, gender, small for gestational age status, 5-minute Apgar score, labor and chorioamnionitis. Figs. 1 and 2 provide a schematic representation of the microarray and qRT-PCR study designs.

Results
Tables 1 and 2 display the maternal and neonatal demographic and clinical characteristics. There were no significant differences between cases and controls for the maternal variables assessed. Distributions of neonatal variables assessed at delivery also did not differ between groups. However, more neonates in the neurocognitive impairment group had evidence of brain injury on post-natal cranial ultrasonography (including severe periventricular-intraventricular hemorrhage, periventricular leukomalacia and ventriculomegaly), as expected. Median cognitive scores (±SD) were 70 ± 8.8 (range  in the neurocognitive impairment group and 90 ± 8.0 (range  in the control group. Eight children in the neurocognitive impairment group had cerebral palsy.

Microarray results
Differential expression analysis revealed moderate changes in the chorioamniotic membrane transcriptome of preterm neonates with and without neurocognitive impairment at 18-24 months: 133 probes corresponding to 117 unique genes were differentially expressed (p<0.05 and fold change >1.5) (Tables 3 and 4). Gene ontology analysis indicated enrichment of 19 biological processes (e.g., positive regulation of cell proliferation by VEGF-activated platelet derived growth factor receptor signaling pathway, etc.) and 3 molecular functions (cytokine binding, vascular-endothelial growth factor receptor activity and vascular-endothelial growth factor receptor binding) as shown in Tables 5 and 6. Pathway analysis with down-weighting of overlapping genes that uses information from all genes in a pathway to compute a pathway enrichment score, indicated four significant KEGG pathways: oxidative phosphorylation, Parkinson's disease, Alzheimer's disease and Huntington's disease (q-value <0.1). Given the significant enrichment of biological processes, molecular functions and pathways identified, we selected genes involved in oxidative phosphorylation, mitochondrial function (central components of the aforementioned pathways) as well as other genes associated with neuronal development, signaling and response to injury for qRT-PCR validation. qRT-PCR study flow diagram. Differentially expressed genes identified from the microarray study were profiled using qRT-PCR for validation and construction of a multi-gene disease classifier using an extended set of 33 cases and 33 controls. Set A was used to build the prediction model and then this model was tested on set B, consisting of 19 cases and 19 controls.

qRT-PCR results
48 out of 90 selected differentially expressed genes were confirmed by qRT-PCR, including genes implicated in neuroinflammation, neurodegeneration and cognitive disorders (e.g., up regulation of SEPP1, APOE, DAB2, CD163, CXCL12, VWF and down-regulation of HAND1, OSR1) (p<0.05). (Table 7) In addition to genes previously described as playing a critical role in cognition, genes involved in neuronal differentiation, signaling, vascular permeability and cellular metabolism also were identified.

Gene-based classifier
A gene-based classifier was developed using qRT-PCR measured expression data from 14 controls and 14 cases using a vetted pipeline that automatically determines the appropriate number of markers using an internal cross-validation procedure. The candidate predictors considered as inputs in this pipeline were the 90 genes selected based on the microarray data. Instead of using the Ct value of a reference gene (e.g., GAPDH) as an internal normalizer for each sample, we searched for the best normalizer among the 90 candidate genes. In doing so, we attempted to accomplish three goals: 1) convert the Ct values into a platform-independent measure (the ratio of the expression of two genes), 2) explore potential gene interactions in predicting neurocognitive outcome, and 3) avoid inclusion of a gene that brings no discrimination between phenotypes in the model. Our methodology identified the ratios OSR1/VWF and HAND1/VWF as providing the best cross-validated performance on the training data when used as the inputs in a linear discriminant analysis model (70% average sensitivity over three specificity values: 80, 85, and 90%). Other classifiers such as quadratic discriminant analysis  and support vector machines had similar, yet lower performance (62 and 68% respectively), and hence the linear discriminant analysis model was retained as the final model. This model had a sensitivity of 74% and a specificity of 83% when applied to a new set of patients (18 Controls and 19 Disease, as 1 Control was discarded due to PCR failures for multiple genes) (See Fig. 3). Although most of the misclassified samples are close to the decision boundary (black line in Fig. 3), two of the misclassified cases with neurocognitive impairment had very high OSR1/VWF and HAND1/VWF gene ratios; these cases had multiple post-natal complications (specifically, severe bronchopulmonary dysplasia, necrotizing enterocolitis and postnatal sepsis) often associated with neurodevelopmental impairment. When compared with clinical covariates available at the time of birth, the molecular prediction model had superior Area Under the Receiver Operating Characteristic curve (AUC 0.77 vs 0.57, p = 0.049) as determined by a bootstrap based test implemented in the pROC package [61]. (Fig. 4).
As suggested by an anonymous reviewer, we also examined the effect of gestational age (GA) at delivery on the quality of predictions of the gene based classifier. The subjects in the test dataset were divided based on gestational age at delivery: extremely preterm (GA from 23.9 to 26.9) and very preterm (GA from 27.0 to 32.1), where the cut-off point of 27.0 weeks was the median of gestational age at delivery in the test set. Although the point estimate of the area under the receiver operating characteristic curve (AUC) was higher for the extremely preterm group (AUC = 82%) compared to the one for the very preterm group (AUC = 70%) the difference was not significant, and including GA at delivery as a covariate in the LDA model (either as a main effect or interaction with the two gene based predictors) did not improve the classifier prediction.

Discussion
The principal findings of our study were: 1) the chorioamniotic membrane transcriptome of preterm neonates with cognitive impairment differed significantly from that of neonates with normal neurodevelopment; 2) Gene ontology analysis indicated enrichment of 19 biological processes (e.g., positive regulation of cell proliferation by VEGF-activated platelet derived growth factor receptor signaling pathway, etc.) and 3 molecular functions (cytokine binding, vascular-endothelial growth factor receptor activity and vascular-endothelial growth factor receptor binding); 3) PADOG identified 4 significantly enriched KEGG pathways: oxidative phosphorylation, Parkinson's disease, Alzheimer's disease and Huntington's disease (q-value <0.1); 4) 48 out of 90 selected differentially expressed genes were confirmed by qRT-PCR, including genes implicated in energy metabolism, neuronal differentiation, signaling, vascular permeability and response to injury (e.g., up-regulation of SEPP1, APOE, DAB2, CD163, CXCL12, VWF; down-regulation of HAND1, OSR1) (p<0.05 and fold change >1.5); and 5) cases with neurocognitive impairment at 18-24 months could be identified at birth apart from gestational agematched controls with a sensitivity of 74% at a specificity of 83% using a molecular signature determined in an independent sample (using ratios of OSR1/VWF and HAND1/VWF in the chorioamniotic membranes). These findings support the view that disturbances in oxidative metabolism, synaptic signaling and response to injury identified at birth, contribute elevated risk of neurocognitive impairment assessed in early childhood.

Molecular marker odd-skipped related 1 (OSR1) gene, downregulated in cases
The Osr1 gene encodes a zinc finger transcription factor that modulates embryonic patterning and morphogenesis as a pair-rule gene; mutations in this gene in Drosophila lead to a loss of odd segments during embryogenesis, thus the name odd-skip related [62]. Human Osr1  consists of three exons located on chromosome 2p24 [63], a region recently implicated as a candidate risk susceptibility locus for autism spectrum disorder [64,65]. The Osr1 transcription factor encoded by this gene may have functional protein-protein interactions with Oxidative stress responsive 1 kinase (OXSR1) [66,67]. This kinase phosphorylates the Na + -K + -2Cl -(NKCC1) co-transporter, responsible for the high intra-cellular chloride concentration of immature neurons contributing to their depolarization (and excitation) on GABA binding [68]. GABA, the main inhibitory neurotransmitter in adults, serves as an excitatory neurotransmitter during fetal and early neonatal life [69,70]. GABA excitation stimulates giant depolarizing potentials which mediate activity-dependent stimulation of neuronal growth, migration, synapse formation and development of functional brain networks. Dysregulation of early GABAergic signaling may lead to aberrant neuronal circuitry and impaired cognitive functioning and is implicated in autism. Interestingly, the Osr1 transcription factor plays a role in regulating renal development, a tissue rich in chloride transporters and vascular tight junctions [71,72].

Molecular marker heart and neural crest derivatives expressed 1 (HAND1), downregulated in cases
Heart and neural crest derivatives expressed 1 (Hand1, also known as eHand) encodes a Twistfamily basic helix-loop-helix transcription factor that plays a role in placentation, trophoblast differentiation [73][74][75][76], fetal cardiac [77,78] and fetal sympathetic nervous system development [79][80][81]: tissues all known for their high energy demands and sensitivity to hypoxia signaling. Until recently, understanding of Hand1 function remained poor due to the early embryonic lethality of Hand1-knockout animals [80,82]. Null and hypomorphic mouse models revealed defects in extraembryonic mesoderm and trophoblast giant cells associated with reduced expression of Placental Lactogen I (Pl1), a placental hormone important for maintenance of the corpus luteum and normal progesterone levels [82,83]. In addition, a developmental arrest in yolk sac vasculogenesis was revealed, followed by increased expression of angiogenic growth factors [75]. Despite these findings, Hand1 deficient embryos had downregulated HIF-1α mRNA expression.
Recently, Breckenridge et al reported that Hand1 was induced by hypoxia and HIF-1α binding upstream of the Hand1 transcriptional start site in mice; additionally, Hand1 expression was associated with decreased oxygen consumption via down-regulation of lipid metabolism and uptake in fetal and adult cardiomyocytes [84]. Thus, Hand1 appears to mediate mitochondrial energy generation and the fetal to neonatal switch from glycolytic to oxidative metabolism [84]. Decreased Hand1 expression is associated with up-regulation of genes encoding proteins involved in lipid uptake and mitochondrial β-oxidation. Whether Hand1 plays a similar role in the central nervous system (CNS) remains to be elucidated. In the CNS, Hand1 is expressed in sympathetic neurons [80] and regulates neuronal sympathetic survival and differentiation along with Hand2 and homeodomain transcription factor Phox2b [81].

Apolipoprotein E (APOE) upregulated in cases
We identified several other genes involved in neuronal differentiation, lipid uptake and CNS signaling. The Apolipoprotein E gene (APOE) encodes a multifunctional glycoprotein that transports lipids and cholesterol in the plasma and CNS by binding to low-density lipoprotein receptors [85]. APOE/lipoprotein particles are produced by astrocytes [86] and to some extent microglia [87]. With brain injury, APOE expression is up-regulated. APOE interacts with cytokines and alters macrophage function, suppresses T cell proliferation, up-regulates platelet nitric oxide production and increases lipid antigen presentation by CD1 molecules; it also maintains the integrity of the blood brain and blood nerve barriers [88]. In the brain, APOE binds to the very low-density lipoprotein receptor (VLDLR) and APOE receptor 2 (APOER2), the two main reelin signaling receptors. APOE can significantly inhibit reelin binding and subsequent phosphorylation of the adapter molecule disabled 1 protein (dab1) which initiates the intracellular transduction of reelin signaling [89]. Mutations in the RELN and DAB1 genes that disrupt reelin signaling are associated with cerebellar hypoplasia and severe abnormalities in neuronal organization and migration [90][91][92]; mutations in either VLDLR or APOER2 in isolation result in more subtle defects in cell positioning, synapse and dendritic spine formation [93]. APOE genotype/expression has been linked to human neurocognitive and neuroinflammatory disorders (e.g., Alzheimer's disease [94], Parkinson's disease [95][96][97] multiple sclerosis [98] and HIV disease progression [99]). Preliminary evidence also suggests an association between APOE gene expression and altered brain structure at birth (alterations in regional cortical brain volumes) [100].

Selenoprotein P, plasma 1 (SEPP1) upregulated in cases
The selenoprotein P, Plasma 1 (Sepp1) gene encodes a selenium rich extracellular protein involved in selenium transport and antioxidant defense mechanisms in the brain [101]. In the setting of brain injury, Sepp1 is up-regulated and secreted by astrocytes [102]; it is then taken up by neurons via the APOER2 which also binds reelin [103][104][105]. Interruption of the reelin signaling pathway may have devastating effects on brain development as previously noted. Increased Sepp1 expression is reported in neuroinflammatory disorders associated with impaired cognition. Bellinger et al reported increased expression in post-mortem Parkinson's disease brain tissues (in Lewy bodies and the substantia nigra, relative to neuron count) [106] and in Alzheimer's disease (in amyloid beta plaques and neurofibrillary tangles) [107].
Disabled homolog 2, mitogen responsive phosphoprotein (DAB2) upregulated in cases The disabled homolog 2, mitogen responsive phosphoprotein (Drosophila) gene (Dab2) encodes disabled protein 2. Murine disabled-2 (initially termed p96) was isolated as a 96 kD phosphoprotein involved in macrophage signaling via colony stimulating factor-1 [108]. Sequence homology suggested that p96 was an ortholog of the Drosophila disabled gene [109] and this was the origin for the names of neuronally expressed Dab1 and the more broadly expressed Dab2. Recently, Dab2 expression was shown to be up-regulated by macrophages and astrocytes in various CNS injury models (e.g., cryoinjury [110] and autoimmune encephalomyelitis [111]). In humans, Dab2 was up-regulated in a microarray study of autopsy specimens in multiple sclerosis lesions. Dab2 has several functions. Similar to Dab1, it serves as an intracellular signaling protein that mediates cell organization and positioning by regulating Src activity and the mitogen-activated protein kinase signaling pathway [112]. Dab2 also is involved in lipid receptor endocytosis (including APOER2) [113] and neurotransmitter release [114]. In addition, Dab2 interacts with the N-terminal domain of Dab2 interacting protein [115]; this protein regulates dendrite development, synapse formation and neuronal migration in the cerebellum and developing cortex [116,117].
Other genes involved in vascular endothelial function and response to injury: von Willebrand factor (VWF), Cluster of Differentiation 163 (CD-163) and C-X-C motif chemokine 12 (CXCL-12) upregulated in cases compared with controls VWF gene is upregulated in cases. The von Willebrand factor gene encodes a large plasma glycoprotein that is synthesized by vascular endothelial cells and megakaryocytes in response to endothelial injury. It plays a central role in platelet adhesion, activation and thrombin generation [118]. Compared with older individuals, preterm infants and fetuses have higher concentrations and larger multimers of VWF [118][119][120]. By combining the signal from this upregulated gene with that of the OSR1 and HAND1 genes that are down-regulated in disease, we obtain two dimensionless variables (OSR1/VWF and HAND1/VWF) that 1) differ to a greater extent between cases and controls, 2) are platform-independent and 3) lead to a gene classifier that is more cost-effective as compared to one that uses a standard gene normalizer (e.g., GAPDH) with no discriminatory power.
CD 163 gene is upregulated in cases. The protein encoded by the CD163 gene is a member of the scavenger receptor cysteine-rich (SRCR) superfamily that is expressed by monocytes/ macrophages. In the CNS, CD163 is localized to perivascular macrophages and microglia. It is thought to function as an innate immune sensor for bacteria altering local immune responsiveness and an acute phase-regulated receptor involved in the clearance and endocytosis of hemoglobin/haptoglobin complexes protecting tissues from free hemoglobin-mediated oxidative damage [121]. CD163 expression is regulated by both proinflammatory and anti-inflammatory mediators (suppressed by lipopolysaccharide, interferon-gamma and tumor necrosis factor alpha and strongly up-regulated by IL-6 and IL-10) [122]. Increased CD163 expression has been linked to neuroinflammatory disease states such as multiple sclerosis [123], Alzheimer's disease [124]) HIV-associated neurocognitive disorders [125] and schizophrenia [126].
CXCL12 gene is upregulated in cases. The C-X-C motif chemokine 12 gene also known as stromal cell-derived factor 1 alpha (SDF-1α) encodes a chemokine protein that binds to chemokine receptor 4 and 7 (G-protein coupled receptors). CXCL12 is induced by proinflammatory stimuli (e.g., lipopolysaccharide, IL1β, TNFα) and has many diverse functions. In the brain, CXCL12 is produced by both neurons and glial cells and is involved in neurogenesis, axonal guidance, neurite outgrowth, modulation of neuronal excitability, neurotransmitter release (particularly GABA release), and neurotransmitter systems cross-talk (e.g., GABA, glutamate, opioids) [127,128].
CXCL12 also is involved in immune functions (immune surveillance, response to inflammation, leukocyte activation) and vasculogenesis.

Strengths and limitations
The strengths of our study include (1) the novel study design using transcriptomics of the chorioamniotic membranes, an abundant source of fetal DNA and of fetal stem cells that may be impacted by the intrauterine environment and (2) the use of state-of-the-art analytics to identify perturbed disease pathways and to develop a multi-gene disease classifier. Limitations to be addressed in future studies include the assessment of neurocognitive outcomes at a later time point when follow-up outcomes are deemed more stable, testing of the molecular disease classifier using a second independent sample with the full spectrum of neurocognitive outcomes collected prospectively, and exploration of alternative or complementary biomarkers (additional biological specimens coupled with advanced neuroimaging techniques).

Conclusion
Impaired brain function in preterm neonates is thought to arise from (1) inflammation and/or hypoxic-ischemic injury to developing preoligodendrocytes and cortical neurons, (2) secondary atrophy after sublethal axonal injury and (3) an arrest or alteration of the developmental trajectory postnatally [129,130]. We propose that this alteration in the developmental trajectory also arises prenatally from stimuli that alter cellular metabolism, neuronal differentiation, signaling, vascular permeability and response to injury. Together, the genes and biological pathways that we have identified provide important preliminary data for the mechanistic processes that may mediate brain injury and aberrant neuronal development in utero. Prospective cohort studies are needed to determine whether this information can be used to identify newborns that will develop neurocognitive impairment in early childhood and might benefit from early intervention or neuro-protective strategies.
Supporting Information S1 Table. Primers used in qRT-PCR assays with the Biomark system.