|Home | About | Journals | Submit | Contact Us | Français|
Rationale: Current understanding of the molecular regulation of lung development is limited and derives mostly from animal studies.
Objectives: To define global patterns of gene expression during human lung development.
Methods: Genome-wide expression profiling was used to measure the developing lung transcriptome in RNA samples derived from 38 normal human lung tissues at 53 to 154 days post conception. Principal component analysis was used to characterize global expression variation and to identify genes and bioontologic attributes contributing to these variations. Individual gene expression patterns were verified by quantitative reverse transcriptase–polymerase chain reaction analysis.
Measurements and Main Results: Gene expression analysis identified attributes not previously associated with lung development, such as chemokine-immunologic processes. Lung characteristics attributes (e.g., surfactant function) were observed at an earlier-than-anticipated age. We defined a 3,223 gene developing lung characteristic subtranscriptome capable of describing a majority of the process. In gene expression space, the samples formed a time-contiguous trajectory with transition points correlating with histological stages and suggesting the existence of novel molecular substages. Induction of surfactant gene expression characterized a pseudoglandular “molecular phase” transition. Individual gene expression patterns were independently validated. We predicted the age of independent human lung transcriptome profiles with a median absolute error of 5 days, supporting the validity of the data and modeling approach.
Conclusions: This study extends our knowledge of key gene expression patterns and bioontologic attributes underlying early human lung developmental processes. The data also suggest the existence of molecular phases of lung development.
Knowledge of lung development in humans has been largely derived from gross anatomical staging with limited molecular studies. High throughput technologies have enabled comprehensive studies of organogenesis in animal models, but corresponding studies in humans are lacking.
We provide a unique data set significantly expanding current knowledge of gene expression patterns in the developing human lung in particular and in human organogenesis in general. From these data, we have identified essential attributes characterizing the biology of the lung, which should facilitate our understanding of lung disease pathogenesis.
In utero events, such as environmental exposures during critical windows of lung development, have been demonstrated to affect lung function and susceptibility to respiratory disease in later life (1–6). One of the widely proposed mechanisms behind this “fetal origins hypothesis” is alterations in gene expression in the developing fetus (7). To best understand the critical windows of early lung development and the molecular mechanisms underlying these windows, we undertook a genome-wide study of gene expression in the normal fetal human lung.
The mechanisms regulating lung development are a temporally coordinated interaction of distinct functional modules (chemical and mechanical) that elicit the necessary morphological and physiological transformations from primordium to a differentiated, functional organ. Current knowledge of lung development has largely been derived from gross anatomical staging and limited molecular studies in humans, supplemented by careful histomorphologic, molecular, and genetic studies of animal models (8–11). The traditional foundation for describing mammalian lung development relies upon the delineation of this continuous process into distinct, morphologic stages based upon the presence of select histological features. Biological processes that transcend stages have long been appreciated. Several novel and canonical biochemical signaling pathways and morphogenic gradients previously associated with the developmental processes of other tissue–organ systems have been implicated in the lung (12–15).
Recently, high-throughput technologies (e.g., microarrays) have enabled comprehensive studies of genome-wide expression patterns and global biological processes underlying organogenesis of the mouse lung (16–18). Related studies of abnormal lung development in the mouse suggest the existence of essential regulatory modules that recur throughout development and transcend histomorphologic compartmentalization (19–22). However, there have been few systematic investigations of the global relationship between the gross morphology (macroscopic-scale features) and transcriptomic processes (microscopic-scale features) of the developing mammalian lung. Corresponding developmental studies in human are lacking.
In previous work (18), we used unsupervised principal component analysis (PCA) of genome-wide expression data to identify global transcriptomic features of mouse lung development. Gene expression variation was found to be associated with macroscopic biological features such as age and the time-to-birth. Our analyses suggested the possibility of characterizing lung development in molecular, rather than histological, terms.
In this study we applied a similar approach to characterize genome-wide expression patterns in the developing human lung transcriptome using a set of fetal lung tissue specimens from National Institute of Child Health and Human Development–supported repositories. We assayed expression patterns from 38 unique samples, representing 29 distinct time-points, spanning the pseudoglandular and canalicular stages (estimated 53–154 days post conception [dpc]). Analysis of this developmental time interval was sufficient to objectively identify a minimal set of 3,223 characteristic genes, including anticipated and novel molecules, capable of defining essential features of human lung development. These characteristic genes were enriched for bioontologic attributes that are organ specific and traditionally associated with later lung development states. Moreover, we find evidence that molecular phases of human lung development parallel histologically defined stages but identify distinct developmental substages that were not previously appreciated.
Human fetal lung tissues (Table 1) were obtained from two National Institute of Child Health and Human Development–supported tissue retrieval programs: the University of Washington Center for Birth Defects Research (Seattle, WA) and the University of Maryland Brain and Tissue Bank for Developmental Disorders (Baltimore, MD). Exclusion criteria include reported hepatic, renal, or pulmonary disease. The postmortem interval was less than 2 hours for samples from the University of Washington and less than 6 hours for samples from the University of Maryland. Creation of the tissue repository was approved by the University of Missouri–Kansas City Pediatric Institutional Review Board.
We studied 38 RNA samples generated from lung tissue obtained from 38 subjects of estimated gestational age between 53 and 154 days postconception (dpc). Estimated gestational age is based on the date of the last menstrual period as declared by the mother's self report at the time of consent. Between 20 and 30 mg of frozen tissue was homogenized, and total RNA was extracted with an on-column DNase I treatment according to the Qiagen RNeasy protocol (Qiagen, Valencia, CA) or the Illustra RNAspin Mini RNA Isolation Kit (GE Healthcare, Piscataway, NJ). The isolated RNA was run on an Agilent Bioanalyzer for quality and quantity assessment. RNA concentration was further assessed by measuring UV absorbance at a wavelength of 260 nm, and RNA purity was assessed using the 260/280 ratio. Only RNA samples with an RNA concentration greater than 100 ng/μl and an RNA Integrity Number (RIN) greater than 6 were used for microarray analysis.
The human lung transcriptome data are derived from whole lung tissue RNA of 38 unique human fetal samples spanning 29 distinct time points between 53 and 154 days dpc. RNA samples were profiled using the Affymetrix Human Genome GeneChip U133 Plus 2.0 microarray (Affymetrix, Santa Clara, CA) containing 54,675 probes corresponding to 19,501 unique NCBI Entrez genes. Labeled target was synthesized from purified RNA samples and hybridized according to the manufacturer's recommendations by the Harvard Medical School–Partners Healthcare Center for Genetics and Genomics using standard protocols (http://www.hpcgg.org/Microarrays/resources). Expression values were extracted from .CEL files using Robust Multi-array Average (RMA, for specific sample subsets) and Microarray Suite 5.0 (MAS 5.0) separately using BioConductor (http://www.bioconductor.org). The Affymetrix .CEL files and RMA normalized data matrix for these samples are accessible at the NCBI Gene Expression Omnibus repository (http://www.ncbi.nlm.nih.gov/geo/) (series accession number GSE14334).
We used PCA as a data reduction strategy for analysis of the data sets (see online supplement). The first principal component (PC1) defines the direction of greatest variation in the probe/gene transcriptomic feature space, with each subsequent PC explaining the next most variation. The contribution of a gene to a PC is proportional to its corresponding probe loading coefficient magnitude. PCA was applied on the standardized dataset without and with prior rank normalization separately, corresponding to the use of linear correlation and rank correlation as measures of similarity between samples, respectively. The resulting PCs are abbreviated reg (regular) PC and rank PC, respectively.
We assessed gene sets for overrepresentation of specific bioontologic attributes using DAVID 2008 (http://david.abcc.ncifcrf.gov), data freeze October 2008 (23, 24). These attributes encompass gene ontologies, sequence motifs, protein domains, and biomolecular pathways. We consider an attribute to be significant if its Expression Analysis Systematic Explorer (EASE) score (Fisher exact test P value) is less than 0.05 relative to an appropriate background gene set.
Differential expression was defined in the full transcriptome of 54,675 probe sets using Significance Analysis of Microarrays (25) at a median FDR = 0 in TIGR Multi-experiment Viewer (http://www.tm4.org/mev.html) as we have previously described (18). Genes were further restricted to those demonstrating a fold change of 2.0 or greater in average expression.
Quantitative reverse transcriptase–polymerase chain reaction (qPCR) was performed on a Stratagene MX3000P using Taqman chemistry as previously described (26). Inventoried (predeveloped) gene-specific assays for measuring gene expression were obtained from Applied Biosystems (Foster City, CA). Gene expression levels (dCT) were calculated relative to the measured CT value of PPIA (peptidyl prolyl isomerase A or cyclophilin A) as an internal, endogenous control.
The primary human lung development data set consists of microarray transcriptome profiles from whole lung tissue RNA of 38 unique human fetal samples spanning 29 distinct time points, (53–154 dpc; Table 1). To minimize the effects of subject or age estimation–related variation in expression measurement and to model global gene-specific developmental expression patterns, we implemented probe-wise nonparametric regression using locfit (27) in BioConductor (see online supplement for details). This approach normalizes individual gene expression patterns using a global smoothing parameter function that minimizes a weighted least squares error. The resulting data, Dataset S1, represents individual smoothened trajectories for every transcript queried on the platform modeled at 28 distinct time points (identical to the time points of the primary dataset except 154 dpc) using all 38 arrays (see Figure E1 in the online supplement). The modeled expression profiles are available as supplemental data, and the preprocessed data set has been made publicly available (NCBI Gene Expression Omnibus repository series accession number GSE14334). These data represent a fetal human lung gene expression compendium.
PCA was used to characterize the independent, dominant directions of maximal sample variation called principal components (PCs) in the developing lung transcriptome. PCA was applied to the dataset without and with prior rank normalization (18). The resulting PCs are abbreviated reg (regular) PC and rank PC. In each case, PC1–3 accounted for a substantial amount of total sample variation (75.39% for reg PC and 72.29% for rank PC). The PC1 sample coordinates were significantly correlated (P < 0.001) with fetal age (correlations of 0.9874 for reg PC1 and 0.9633 for rank PC1). This is anticipated for a data set with the major underlying feature of experimental design being subject age and supports the validity of the smoothing and PCA modeling approach.
We sought to identify the key genes contributing to expression variation over development. The contribution of a gene to a PC is proportional to the gene's corresponding microarray probe loading coefficient (Dataset S3). For each PC, we defined the top 5% (ranks 1–975) of genes with the highest loading coefficient magnitudes as its characteristic genes. Genes contributing substantially to PC1 included surfactant proteins SFTPB and SFTPC, a member of the Wnt-binding Frizzled-related proteins SFRP2, the catenin/plakophilin CTNND2, and the scavenger receptor CD36. Genes contributing substantially to PC2 included the surfactant-associated claudin CLDN18, the chemokine CXCL5, and the MHC class II molecules HLA-DRA and HLA-DRB1. Genes contributing substantially to PC3 included the general transcription factor FOSB. We observed a small set of common genes (22.5–30.5%) among the characteristic genes of PC1–3 (Table E1). These may represent a set of molecules contributing to multiple, overlapping processes important to early lung development. Forty-six genes were common to all PCs, including prominently ranked surfactant-related proteins (SFTPB, SFTPC, and CLDN18), immune molecules, and a number of molecular transporters (Table 2).
It is difficult, and potentially misleading, to resolve complex biological processes described by these PCs using individual genes. Therefore, we identified biological correlates of PC1–3 using gene attribute enrichment analysis. These biological correlates can be macroscopic (e.g., development) or microscopic (e.g., nucleic acid binding) in scale. For instance, we noted above that the PC1 coordinates of the samples were correlated with fetal age, a macroscopic variable.
Gene attribute enrichment analysis was performed on each characteristic gene set of reg and rank PC1-3, relative to the background gene set of all 19,501 unique Entrez genes measured on the U133 Plus 2.0 microarray. A total of 1,084 unique attributes form the union of all six characteristic gene sets (Table E2), while 35 attributes were common to all six PCs (Tables 2 and E3). Among the common characteristic attributes, we noted the abundance of surfactant–gas exchange attributes, even though the function of surfactant proteins in these early stages of lung development is unclear. Also prominent are immunologic-major histocompatibility complex (MHC), in particular MHC class II genes, whose role in early lung development is similarly unclear. Regarding attributes that are unique to individual PCs, PC1 was enriched for DNA and nuclear organization and replication and cell cycle processes. PC2 was enriched for cytoskeletal protein domains (e.g., spectrins) and transcription factor activity and binding motifs. PC3 was enriched for cell-specific differentiation processes, cellular motion, and extracellular matrix interaction (e.g., chondroitin sulfate proteoglycans) attributes.
We defined a developing lung characteristic subtranscriptome (DLCS, Table E4) as the union of characteristic genes of PC1–3. This set is comprised of 3,223 unique genes, corresponding to 11,558 probes. The DLCS constitutes 16.53% of the total 19,501-gene transcriptome measured by the microarray, and its corresponding 11,588 probes account for 72.73% (reg PC) or 69.31% (rank PC) of the variation of PC1–3 on average. Temporal expression patterns for a subset of these genes are presented in Figure 1A. Many of these DLCS genes showed marked changes in expression at the transition from the pseudoglandular to canalicular stage of lung development.
Each array can be viewed as a point in space defined by genome-wide variation in expression (e.g., as summarized by PCA), demonstrating the relationship between samples and ages. When the developing human lung samples are projected in PCA gene expression space, there is a strong and significant correlation between PC1 and age. However, the configuration of the samples plotted in principal component space is highly complex in structure (Figures 2 and E2). One prominent feature of this visualization is the identification of two periods where progression along the PC1 (e.g., age) axis “slows” dramatically, resulting in a transition point of the lung development trajectory. The first of these is between Days 91 and 96, and the second is between Days 110 and 117. The second transition point coincides with the transition from the pseudoglandular to canalicular stages of lung development (16 weeks or 112 days). The first transition point occurs within the middle of the pseudoglandular stage, suggesting a previously unappreciated molecular subphase of early human lung development.
To confirm this trajectory is not stochastic, PCA was performed as above but was restricted to the DLCS feature set. PC1–3 here accounted for 80.41% (reg PC) or 79.04% (rank PC) of total sample variation. The DLCS PC1 sample coordinates were significantly (P < 0.001) correlated with fetal age (correlation 0.9956 for reg PC1 and 0.9956 for rank PC1). The resulting PC1–3 sample configurations were highly similar to those observed using the full transcriptome feature set (Figures 2B and E3). These observations suggest that a small and specific subset of the transcriptome—the DLCS—recapitulates the essential macroscopic features of human lung development from 53 to 140 dpc. Consistent with the hypothesis that two distinct molecular phases of the pseudoglandular stage exist, recreating the time-contiguous trajectory of the samples using the DLCS alone, a set enriched in informative genes (and presumably with reduced noise), identifies the identical transition points.
We identified genes with significantly different expression between these putative molecular phases of the pseudoglandular stage of lung development. To avoid confounding this analysis by including transitional time points, we defined the early pseudoglandular molecular phase as between 63 and 85 days and the late pseudoglandular molecular phase as between 101 and 108 days. We identified 245 genes with significantly higher expression from Days 63 to 85 and 230 genes with significantly higher expression from Days 101 to 108 (Figure 1B and Table E5). Attribute analyses of these genes demonstrated that the early pseudoglandular genes were enriched with chromosomal organization processes associated with mitosis, and the late pseudoglandular genes were enriched with surfactant function–gas exchange and immunological-MHC class II attributes (Table E6). All surfactant protein genes measured (SFTPA, B, C, and, D) were significantly increased from the early to late pseudoglandular phase. Together, these data suggest the existence of two distinct molecular phases within the pseudoglandular stage of human lung development, which can be distinguished by induction of surfactant gene expression.
qPCR was used to independently verify the expression profile of 18 DLCS genes in the developing human lung at 28 distinct time points. Genes were selected for their prominent ranking in the characteristic gene sets of PC1–3 or for their corresponding, significantly overrepresented bioontologic attributes, mainly surfactants and immunologic/stress response. To test the accuracy of the array-based global expression profiles, we computed the correlation of microarray-reported signal and negative qPCR delta CT (dCT) values across these 26 time points (Figure 3 and Table E7). Fifteen of 18 genes showed significantly correlated expression profiles: CCL20 (r = 0.572; P = 0.002), CD36 (r = 0.863; P < 0.001), CFTR (r = 0.519; P = 0.007), CLDN18 (r = 0.6985; P = 0.001), CLIC5 (r = 0.430; P = 0.029), CTNND2 (r = 0.757; P < 0.001), CXCL3 (r = 0.6527; P < 0.001), CXCL5 (r = 0.5795; P = 0.001), CXCL6 (r = 0.533; P = 0.005), HPGD (r = 0.501; P = 0.009), KITLG (r = 0.619; P < 0.001), SERPINA1 (r = 0.511; P = 0.007), SFTPB (r = 0.7880; P < 0.001), SFTPC (r = 0.8927; P < 0.001), and TUBB2B (r = 0.8051; P < 0.001).
We also used qPCR to test the validity of the novel pseudoglandular substage, or molecular phase, and the ability of individual genes to define this phase transition. We evaluated 10 genes predicted to be differentially expressed across the substages (Figure 3 and Table E8), confirming that seven have significantly different expression levels: CCL20 (P = 0.024), CD36 (P < 0.001), CLDN18 (P = 0.011), CTNND2 (P = 0.002), SFTPB (P = 0.035), SFTPC (P = 0.005), and TUBB2B (P < 0.001).
We investigated the robustness of this characterization of early lung development by assessing its capacity to (1) estimate the age of independent human fetal lung samples and (2) chronologically order developing mouse lung samples on the basis of their lung transcriptome profile alone. To do this, we performed a holdout cross validation, whereby the developing human lung data were split into training (22 arrays) and test (16 arrays) sets. Here, the training set (only) was used to regenerate smoothed time-dependent, whole-genome expression trajectories using probewise, nonparametric regression (as for Dataset S1). Dataset S2 represents individual smoothed trajectories for every transcript queried on the platform modeled at 18 distinct time points, at regular 5-day intervals from 59 to 140 dpc, using the 22 training arrays (Figure E1). Dataset S2 was used to rederive the transcriptomic PC1–3 feature space of lung development, and the human or mouse lung test sample transcriptome profiles were projected into this space as we have previously described (28, 29).
To estimate the age of human samples, PCA was performed on the training set full transcriptome (19,501 genes, 54,675 probes) or the training set limited to the DLCS (3,223 genes, 11,558 probes). Next, all 38 human fetal lung samples were projected into the transcriptomic PC1–3 space of the training set (Figures E4 and E5). The estimated age of all human samples was inferred from the age of its closest training sample in PC1–3 space (Figure 4A). The median absolute differences between estimated and true age of the 22 training samples were 5 days for the full transcriptome and the DLCS. For the 16 primary test samples, the median absolute differences were 7 days (full transcriptome) and 5 days (DLCS), respectively. These results demonstrate the ability of the PCA-based time-dependent trajectory to accurately define human fetal lung age and confirm the added utility of the DLCS genes to describe this biological process.
To assess the age of mouse lung samples, PCA was performed on the training set (Dataset S2) limited to the set of 7,683 orthologous genes (in common between the human and mouse microarray platforms) or those orthologs in the DLCS (1,568 genes). The test data for this purpose consisted of 21 developing mouse whole lung transcriptome profiles at six discrete time points: embryonic Days 14 (×4), 17 (×3), and 19 (×4) and postnatal Days 7 (×3), 14 (×3), and 28 (×4) (16, 18). When projected in human lung development transcriptomic PC space, the mouse samples, especially the embryonic time points (14, 17, and 19 dpc), were correlated with their true developmental age (Figures 4B, E6, and E7). The rank correlation of PC1 coordinates of mouse samples (embryonic and postnatal, median per time point) with their true age is 0.9429 (P < 0.01). These results further validate the biological foundation of this PCA-based, time-contiguous trajectory and suggest that it captures conserved biological features of lung development.
We present here, for the first time, a comprehensive assessment of gene expression patterns in the developing human lung between 53 and 154 dpc, a time interval spanning the pseudoglandular and canalicular stages. These data, including thousands of individual gene expression patterns, are publicly available. We have applied novel nonparametric, regression-based smoothing to the primary dataset in an effort to minimize subject/sample and age estimation-dependent effects. This approach effectively improves dataset quality as objectively defined by increasing similarity between trajectories for multiple probe sets interrogating the same transcript (data not shown). The resulting dataset significantly expands our current understanding of changes in the expression of genes during human lung development.
We distilled the developing human lung transcriptome into independent, dominant directions of gene expression variation using unsupervised PCA. Analyzing the developmental time series in this way, we can objectively resolve the contribution of different biological/functional modules throughout the course of development. For each PC, a characteristic gene set was defined and its corresponding bioontologic attribute profile identified. The characteristic gene sets were significantly enriched for many attributes known to be involved in lung development, but whose roles in these early stages have not been well characterized (e.g., surfactant). In addition, a number of genes previously unappreciated as playing a role in lung development were discovered using this approach.
Our PCA characterization facilitated visualization of the time-series as a linear trajectory defined by gene expression variation. The most dominant direction of sample variation in this developing human lung transcriptome (PC1) was strongly correlated with estimated gestational age. This is intuitive and consistent with our earlier studies of the developing mouse lung transcriptome time series (18). However, the short gestational period of the mouse coupled with the low sampling frequency of that dataset limited the ability to detect finer transcriptomic granularity in that process.
Traditional embryology partitions growth of the developing human lung into five stages: embryonic (26 dpc to 5 weeks), pseudoglandular (5–16 weeks), canalicular (16–26 weeks), saccular (26 weeks to birth), and alveolar (birth to 6 months), based upon selective morphological features that are observed within these intervals. However, it has been long appreciated that complex biological processes occurring during lung development are not specifically associated with this classical stage definition. In the human data, we observed a gene expression transition point at approximately 117 dpc, reflecting the age associated with the morphologic transition from the pseudoglandular to canalicular stages. A second transition point was noted at approximately 94 dpc, suggesting the presence of at least two distinct molecular phases within the pseudoglandular stage. The substages identified by this PCA characterization might reflect critical molecular windows of lung development (3) and suggest that molecular staging in this manner captures information that is complementary but distinct from classic morphologic staging alone. Given sufficient sampling/measurement frequency, PCs based upon genome-wide expression data could form a rational basis for an alternative and objective molecular taxonomy of lung development.
To further explore the biology identified by our analysis of the developing human lung transcriptome, we focused on characteristic gene sets (the highest 5% loading coefficient magnitude genes) of PC1–3 and their ontological attributes. We defined the 3,223–gene union of PC1–3 characteristic genes as the developing lung characteristic subtranscriptome (DLCS), which may be regarded as the minimal set of genes that accounts for a significant proportion of the transcriptomic variation in the early fetal developing human lung. Among the 46 genes common to all PC1–3 characteristic gene sets are three surfactant-associated genes (STFPB, SFTPC, and CLDN18), three MHC class II genes, and 12 substrate-specific transporter genes. Overall, the DLCS contained 28 of 77 genes previously identified to be functionally involved in general lung development (14), which is a 2.9-fold enrichment (odds ratio P < 0. 5 × 10−4 by two-tailed Fisher exact test; Table E9). This enrichment for functionally relevant genes and the prominence of surfactants strongly suggests the PC characterization to be a meaningful abstraction of the specific molecular biology underlying lung development.
There were 35 bioontologic attributes common to all PC1–3 characteristic gene sets. Not surprising among these were attributes related to cell cycle and cellular division, universal to general developmental processes. Additionally, there was a significant presence of surfactant–gaseous exchange and immunologic–MHC class II attributes. Genes with these attributes possess expression profiles that were largely increasing from 53 to 140 dpc. Although the production of surfactant is associated with later gestational development, our observations that these surfactant proteins are expressed relatively early in fetal lung development is consistent with a prior description of SFTPB and SFTPC expression as early as Week 13 in humans (30) and before the end of the pseudoglandular stage in mice (31). The early expression of surfactant proteins may also indicate a necessity of early molecular programming of the lung for subsequent development. This is supported by recent work revealing the potential of embryonic stem cells to form glandular respiratory epithelium after coculture with dissociated fetal lung (32). Additionally, given the recent description of forkhead box M1 (FOXM1) as a key regulator of surfactant protein expression (33), we found that FOXM1 is a PC1 characteristic gene here, further supporting the role of surfactants in early lung development.
We addressed the robustness of our findings in multiple ways. First, we noted that PCA using the DLCS, a feature set enriched in informative expression profiles likely to increase the signal-to-noise ratio, returned a time-contiguous trajectory very similar to that of the full transcriptome. qPCR for DLCS genes demonstrated a very high rate of global validation for individual gene expression trajectories (83%) and for differential expression of substage marker genes (70%). Next, we assessed the validity of this PCA-based characterization by testing its ability to predict the age of independent developing human or mouse lung samples. We divided the 38 fetal human lung samples into training (22 samples) and test (16 samples) sets and used the principal components of the training set to predict gestational age in the test set. We found that we could accurately estimate the gestational age of the test sample using the 3,223 gene DLCS or the full (genome-wide) transcriptome. The ability to independently predict gestational age from a lung transcriptomic profile may be viewed as proof of the concept that the PC characterization captures the essential biology of the system because one expects age to be critical parameter in the lung development process. The ability to chronologically order developing mouse lung profiles by their age in the human lung development PC space further suggests the DLCS identifies conserved molecular mechanisms and bioontologic attributes.
There are several limitations to this study. First, the dataset is limited to a relatively small number of samples from a specific window of time. We analyzed only 38 samples, including 29 distinct time points, over a 101-day sampling age range solely from early fetal stages. A larger sampling window and a higher sampling frequency would add to the details of the analysis and may lend further support for the presence of additional molecular stages that we are unable to investigate more comprehensively with the current data. However, the lack of available human samples during late fetal gestation means that extending the embryonic time points beyond the current age window is essentially impossible. We have performed preliminary microarray analysis of newborn and juvenile human lung tissue. Data from these samples is confounded by multiple variables, including varying age at birth and primary or secondary pulmonary complications. Preliminary analysis of data from these samples suggests they represent a poor comparison group for the fetal data set, and thus we have decided against including them in the current study. Second, there may be a degree of error in the estimation of fetal gestational age. This and other factors, such as maternal smoking history (34), may affect gene expression in individual samples. However, our findings were robust in the prediction of gestational age. In fact, we implemented an expression profile smoothing approach specifically to minimize variability due to sampling. Third, we did not dissect individual cellular contributors to fetal lung development. Instead, we defined the process as the dynamic sum of all cellular constituents. Whole lung tissue transcriptome profiles may have limited success in resolving cell/tissue specific molecular events, particularly during the later stages of development when the lung cell population is more heterogeneous with distinct cooccurring and specialized molecular processes. Nonetheless, our data are representative of the overall molecular events occurring in the early fetal human lung.
In conclusion, we present novel expression patterns for thousands of genes in the developing normal human lung. Our analyses of the human fetal lung transcriptome support current developmental paradigms and have revealed the existence of molecular phases of development. Our results have been validated via different modalities and appear robust. These observations, and the application of these analytic strategies to more comprehensive data sets, will further our understanding of essential molecular events in normal development and pathological derangements of the lung and will provide further insights into critical windows of development wherein environmental exposures may lead to subsequent disease.
The authors thank Alyssa McCoy for technical assistance and B. Raby and R. Lazarus for analytical advice.
Supported by National Institutes of Health grants K25 HL91124, R01 HL88028, and P50 NS40828 (A.T.K.); K23 HG3983 (K.G.T.); R01 ES10855 (J.S.L.); U54 LM8748 (I.S.K.); U01 HL65899 (S.T.W.); and R01 HL71885 (T.J.M.).
This article has an online supplement, which is accessible in this issue's table of contents at www.atsjournals.org
Originally Published in Press as DOI: 10.1164/rccm.200907-1063OC on October 8, 2009
Conflict of Interest Statement: None of the authors has a financial relationship with a commercial entity that has an interest in the subject of this manuscript.