Vision from next generation sequencing: Multi-dimensional genome-wide analysis for producing gene regulatory networks underlying retinal development, aging and disease

Abstract

Genomics and genetics have invaded all aspects of biology and medicine, opening uncharted territory for scientific exploration. The definition of “gene” itself has become ambiguous, and the central dogma is continuously being revised and expanded. Computational biology and computational medicine are no longer intellectual domains of the chosen few. Next generation sequencing (NGS) technology, together with novel methods of pattern recognition and network analyses, has revolutionized the way we think about fundamental biological mechanisms and cellular pathways. In this review, we discuss NGS-based genome-wide approaches that can provide deeper insights into retinal development, aging and disease pathogenesis. We first focus on gene regulatory networks (GRNs) that govern the differentiation of retinal photoreceptors and modulate adaptive response during aging. Then, we discuss NGS technology in the context of retinal disease and develop a vision for therapies based on network biology. We should emphasize that basic strategies for network construction and analyses can be transported to any tissue or cell type. We believe that specific and uniform guidelines are required for generation of genome, transcriptome and epigenome data to facilitate comparative analysis and integration of multi-dimensional data sets, and for constructing networks underlying complex biological processes. As cellular homeostasis and organismal survival are dependent on gene-gene and gene-environment interactions, we believe that network-based biology will provide the foundation for deciphering disease mechanisms and discovering novel drug targets for retinal neurodegenerative diseases.

1. Introduction

Ramon y Cajal was the first to recognize the retina as ‘a true nervous center, as a peripheral extension of the central nervous system’ (Ramon y Cajal, 1893 La retine des Vertebres; translation taken from Piccolino 1989 Santiago Ramon Y Cajal, the retina and the neuron theory). The retina has thus attained the distinction of being an attractive model to investigate fundamental biology and therapy of the nervous system. The vertebrate retina is composed of six major neuronal cell types that are organized in three cellular layers forming exquisite neuronal circuits for detection of visual information (Figure 1A). Light is captured by photoreceptors; visual signals then undergo enhancement, integration and processing through bipolar, horizontal and amacrine cells and by varied usage of parallel synaptic circuits, before eventually being transmitted via ganglion cells to the brain (Lamb et al., 2007; Masland, 2001). This remarkable complexity makes the retina extremely vulnerable to degeneration caused by genetic defects that eventually lead to vision loss. In fact, a vast majority of incurable blindness is caused by dysfunction or death of retinal photoreceptors (Wright et al., 2010). Thus, knowledge-based design of therapies for blinding retinal degenerative diseases (RDDs) depends on better understanding of pathways that are associated with: (i) normal development of the retinal neurons (specifically photoreceptors) from progenitors and stem cells; (ii) assembly of synaptic circuits; (iii) cellular adaptation and homeostasis; (iv) response to aging and inherited genomic changes; and (v) disease pathogenesis.

Figure 1.

The mammalian retina and systems biology approaches. A, A representative retinal anatomy is shown by hematoxylin and eosin stained cross section of an adult mouse retina (left) and by schematics (center and right). In the mammalian retina, six main neuronal classes are organized into three nuclear layers [outer nuclear layer (ONL), inner nuclear layer (INL) and ganglion cell layer (GCL)] and form synaptic connections in two plexiform layers [outer plexiform layer (OPL) and inner plexiform layer (IPL)]. Cone and rod photoreceptors comprise the outer retina with their cell bodies situated in ONL and their inner and outer segments (IS and OS, respectively) located between ONL and the retinal pigment epithelium (RPE). RPE microvilli ensheath the outer segments, supporting phototransduction and photoreceptor survival. Photoreceptors transfer visual information through bipolar and retinal interneurons to ganglion cells. B, To date, population-based genetic analyses (i) and genetic and molecular studies of known genes (ii) have been major strategies in retinal research. Such studies have long been a major driving force in identifying retinal disease genes and in revealing the function of the disease genes and their functional or structural associates. The “bottom up” approaches have been undertaken to build functional networks (iii) that are critical for retinal function (e.g., phototransduction pathway) by compiling functional/structural relationships among individual molecules. Considering widespread crosstalk between functional and/or regulatory networks (iv), system-wide measurement of various biomolecules is critical in constructing a comprehensive map of complex intermolecular regulatory interactions. Systems biology approach thus complements the traditional reductionist methodologies. The recent advent of next generation sequencing technologies has enabled system-level assessment of various biological processes. Computational analysis of next generation sequencing and other types of high throughput data, ideally by integrating multiple data sets, allows a holistic approach to elucidate cellular function (v) as well as homeostasis of tissues (vi), organs (vii), and organisms (i). RHO, rhodopsin; PDE, phosphodiesterase; CNG, cyclic nucleotide gated channel.

The human genome of almost three billion nucleotides contains the complete instruction for generating over 100 billion neurons and 150 trillion synapses (Pakkenberg et al., 2003; Williams and Herrup, 1988). Despite tremendous cellular heterogeneity and functional complexity, our genome encompasses only approximately 20,000 protein-coding genes (Ezkurdia et al., 2014)(http://www.gencodegenes.org), fewer than some of the “apparently” less-evolved organisms (Carninci et al., 2005; de Laat and Duboule, 2013; Pruitt et al., 2009). However, alternative splicing and use of alternate promoters can produce unique gene expression patterns associated with fate determination and cell-type specific functions (Wang et al., 2008). Furthermore, the non-protein-coding genomic DNA [the so-called “junk DNA” (Balakirev and Ayala, 2003; Brosius and Gould, 1992; Ohno, 1972)] is increasingly being recognized as an important “regulator” of the coding information (Palazzo and Gregory, 2014). Cis-regulatory sequences include the genomic code for binding of transcription complexes that dictate quantitatively precise as well as cell type- and stage-specific gene expression (Levine et al., 2014; Shlyueva et al., 2014). In addition, though the coding transcripts are relatively small in number, a plethora of small and long non-coding RNAs are detected in different cell types and believed to modulate gene expression and homeostasis (Morris and Mattick, 2014; Slack, 2006). Normal gene expression patterns, established by combinatorial action of multiple regulatory modules, can be disrupted in response to changes in microenvironment and/or by inherited genomic variations/mutations, resulting in altered physiology and phenotype (including disease) (Lagha et al., 2012). Thus, comprehensive understanding of integrated gene regulatory networks (GRNs) is critical for deciphering normal development and homeostatic mechanisms as well as pathways leading to disease.

Remarkable advances have been made in elucidating molecules and pathways that control retinal cell fate specification and differentiation (Agathocleous and Harris, 2009; Bassett and Wallace, 2012). Defects in over 200 genes have been associated with inherited monogenic RDDs (RetNet; https://sph.uth.edu/retnet/). We have also begun to dissect the complexities of multifactorial retinal diseases that afflict large segments of the population (Fritsche et al., 2014; Kuo et al., 2014). Impressive gene- and stem cell-based approaches are being developed for treatment of retinal disease; nonetheless, individual research projects have generally centered on single genes or molecules or on a single functional pathway, and therefore the biomedical progress has not been able to keep pace with public expectations (Bull and Martin, 2011; Cuenca et al., 2014; Lindvall and Kokaia, 2010; Rowe-Rendleman et al., 2014). We must recognize that genes/RNAs/proteins are part of complex molecular networks and biochemical pathways and that the disruption of one creates a domino effect leading to multiple changes (including in gene expression patterns) not only within the affected cell(s) but also in neighboring cells and tissues. Therefore, holistic assessment of biological components and of interaction networks constitutes essential tasks in exploring retinal development and disease in the post-genomic era (Barabasi and Oltvai, 2004; Hwang et al., 2012; Kitano, 2002; Schadt, 2009; Vidal et al., 2011; Yu et al., 2004b; Zhou et al., 2014).

The emergence of next generation sequencing (NGS) technology has dramatically broadened the scope in which diverse cellular processes can be interrogated, setting the stage for system-level approaches to comprehend retinal GRNs. In this review, we will describe the current status of genome-wide strategies as applied to the retina and discuss in which direction the field appears to be moving. As the task ahead is daunting yet feasible, we believe that a collaborative consortium-like approach, elegantly demonstrated by genetic studies of age-related macular degeneration (AMD) (Cipriani and International AMD Genomics Consortium, 2014; Fritsche et al., 2013; Fritsche and International AMD Genomics Consortium, 2014), is required to elucidate retinal GRNs that control development, maintain homeostasis and modulate responses to aging, environment and inherited disease-causing variants or mutations. We have therefore taken the liberty to put forward a framework for integrative and comparative analysis of NGS data, with a goal to build comprehensive GRNs pertaining to the retina. Here, we have primarily focused on the photoreceptors because of their association with incurable blinding retinal diseases. Rod photoreceptors, in particular, constitute nearly 80% of retinal neurons in many mammalian species, including mouse and human (Hendrickson et al., 2008; Lamb et al., 2007; Rapaport et al., 2004); thus NGS data sets from retina are especially rich in rod photoreceptor-related information. We envision the photoreceptors as an ideal paradigm for initiating multi-dimensional, system-level studies that can be widely applicable to other cell types, especially in the retina but broadly for the central nervous system.

2. Systems biology approaches

Biomedical research is mostly driven from phenotypic observations and curiosity to understand biological phenomena. For example, forward genetic screening using animal models, such as Drosophila melanogaster (fruit fly), Caenorhabditis elegans and Mus musculus (mouse), have led to identification of many essential genes and their function. Similarly, human genetic studies have focused on identifying genes linked to specific phenotypes or traits, thereby providing significant insights into biological basis of the disease. In addition, biochemical, genetic and molecular biology approaches have further elucidated mechanism(s) of action of specific molecules and cellular components. Though biological pathways can be deciphered by compiling individual molecular functions and binary relationships, current approaches are not optimal in mapping complex regulatory interactions among discrete constituents. Systems biology takes a bird’s eye view of cellular function with a goal of delineating molecular interactions and crosstalk among the pathways (Ideker et al., 2001; Ryan et al., 2013; Sorger, 2005), complementing the reductionist approach (Figure 1B). Although the concept of “systems biology” is not new (Trewavas, 2006), it has had only limited success because of unavailability of extensive data sets.

Recent advent of NGS and computational methodology has permitted biologists to generate system-wide data sets, leading to rapid advances in this field. Three major components can be assigned to systems biology, especially when we discuss GRNs (Figure 2A):

Figure 2.

Strategies and aims of system-wide, multi-dimensional data analysis. A, Networks of a tissue or a cell type of interest can be inferred from high throughput data analysis. Next generation sequencing (NGS) allows cataloguing cellular constituents at a steady state and functional interactions when combined with system perturbation and differential analysis. Molecular interactions are not confined to only one molecular type such as DNA, transcripts, chromatin marks or proteins. Thus, multi-dimensional data integration further refines the networks. In addition, comparative analysis is critical as discrete cells are subjected to temporal changes (i.e. development and aging) as well as interactions with neighboring cells and the microenvironment, which evoke physiological modulation of the tissue and eventually of the organism. These holistic approaches will lead to new discoveries of the biological systems and offer broad application. B, Cellular function is regulated at multiple levels. The DNA sequence contains the instructions of protein coding and gene regulation, and diverse gene regulatory mechanisms ensure expression of a unique set of components highly specialized for each cell identity. Intrinsic and/or exogenous damage to any level can lead to deleterious effects on function and survival of the system. TF_act, transcription activator; TF_rep, transcription repressor.

1. Prediction of regulatory networks via generation of various high throughput data sets and by computational analysis

2. Extension and/or refinement of the networks by superimposing measurements made at multiple levels of cellular constituents such as DNA, transcribed sequences, chromatin state, proteins, and metabolites.

3. Assessment of system’s response to time, risk factors or interaction with its microenvironment (e.g., neighboring cells/tissues)

A comprehensive system-level understanding of a cell/tissue/organism requires integrated analysis of all intracellular molecular interactions and pathways, including data sets from proteomic and metabolomic studies. However, high throughput data from such investigations is not readily obtainable for most tissues including the retina (or other ocular cell types). We therefore limit our discussion to genetic and epigenetic control networks that can be measured by NGS.

2.1. High throughput data generation

NGS is a versatile technology that can be coupled with an endless list of classical assay strategies (Table 1), enabling genome-wide measurements of DNA sequence variations as well as components of GRNs at both transcriptional and epigenetic levels. High throughput genome-wide data generation requires a careful study design, which can be summarized in two complementary strategies (Figure 2A). In the first approach, the output is a nearly complete catalog of components of a given class of biomolecules with relevant quantitative information. For example, GRNs associated with rod photoreceptors can be constructed by integrating a comprehensive and quantitative catalog of mRNA transcripts (using RNA-seq), target genes for key transcription factors (using chromatin immunoprecipitation-sequencing, termed ChIP-seq), global profiles of epigenome [histone modifications using ChIP-seq, and DNA methylation using different NGS-based methods such as reduced representation bisulfite sequencing (RRBS)], and other regulatory molecules (such as miRNA, using small RNA-seq). In this review, we will not discuss NGS-based methods (summarized in Table 1), which have been subjects of excellent reviews recently (Furey, 2012; Metzker, 2010; Park, 2009; Telese et al., 2013; Wang et al., 2009). With these data, regulatory networks can be inferred based on co-expression and prior knowledge of interactions though additional experiments are needed for validation. In the second strategy, a ‘seed network’ is employed for the experimental design. A known molecular interaction or functional hub is altered experimentally, and the consequences of system perturbation are interrogated genome-wide. For example, gene knock-out studies followed by mRNA or epigenomic profiles would greatly assist in deciphering the associated GRN. While the former approach charts all the possible physical or regulatory interactions, the latter refines the networks by directly examining functional relationships.

Table 1.

High throughput techniques used in system-level data generation at multiple levels of biological pathways

	Method	Information	Strategy	Reference
Genome	WGS (whole genome sequencing)	Genetic variation in the entire genome	High throughput sequencing of the complete genome
	WES (whole exome sequencing)	Genetic variations in the exome	Exome capture followed by high throughput sequencing.	(Ng et al., 2009)
	Repli-seq (replication sequencing)	Temporal mapping of DNA replication	BrdU labeling of newly replicated DNA and FACS-mediated fractionation of cell populations according to cell cycle phases followed by high throughput sequencing.	(Hansen et al., 2010)
	ChIP-seq (chromatin immunoprecipitation sequencing), ChIP-exo, or DamID	Genome-wide protein binding sites	Genome-wide profiling of occupancy by a protein of interest. A specific antibody is used to isolate protein-bound chromatin areas (ChIP-seq), and the resolution of the assay can be further enhanced by employing exonuclease digestion of protein-bound DNA ends (ChIP-exo). In cases that the specific ChIP quality antibody is not available, the protein of interest that is fused to E. Coli DNA adenine methyltransferase (Dam) and therefore methylates DNA adjacent to the binding sites can be used instead.	(Barski et al., 2007), (Rhee and Pugh, 2011)
Transcriptome	RNA-seq	Gene expression profile	Genome-wide profiling of transcripts. Typically, polyadenylated RNAs are enriched for subsequent cDNA synthesis and high throughput sequencing.	(Mortazavi et al., 2008)
	small RNA-seq	Regulatory RNA profile	Genome-wide profiling of small RNA species by high throughput sequencing with modified cDNA synthesis protocols.	(Morin et al., 2008)
	Deep CAGE (cap analysis of gene expression)	Transcription start sites	Capturing cDNA tags corresponding 5′ ends of transcripts by cap trapping system followed by high throughput sequencing.	(de Hoon and Hayashizaki, 2008)
	RNA-PET (RNA paired end tag) sequencing	5′ and 3′ ends of transcripts	Profiling of 5′ and 3′ end tags of transcripts. This method is especially powerful in detection of fusion transcripts.	(Ruan and Ruan, 2012)
	GRO-seq (global run-on sequencing)	Nascent transcripts	Global mapping of transcriptionally active polymerase density by genome-wide profiling of nascent RNA.	(Core et al., 2008)
	Ribo-seq (ribosome profiling)	Transcripts engaged with translation	Profiling of ribosome-protected mRNA fragments by deep sequencing.	(Ingolia et al., 2009)
	CLIP-seq (Cross-linking immunoprecipitation sequencing) or RIP-seq (RNA immunoprecipitation sequencing)	Transcripts bound by a protein of interest	High throughput sequencing of cDNA made from RNA pulled down with antibody against a protein of interest. RNA-protein complexes are cross-liked with UV in CLIP-seq. In RIP-seq, no cross-linking is necessary although formaldehyde-mediated cross-linking can be included.	(Yeo et al., 2009) (Zhao et al., 2010)
Epigenome & chromatin conformation	Histone modification ChIP-seq	Genome-wide map of histone modification	ChIP-seq can be also applied for genome-wide profiling of histone modifications. For histones, ChIP can be also performed without cross-linking of proteins to DNA (native ChIP).	(Barski et al., 2007)
	MeDIP-seq (methylated DNA immunoprecipitation sequencing), WGBS (whole genome bisulfite sequencing) or RRBS (reduced representation bisulfite sequencing)	Genome-wide map of DNA methylation	Genome-wide mapping of methylated cytosine (5mC) of DNA. Methylated DNA can be detected by various methods such as immunoprecipitation using 5mC-specific antibody (MeDIP-seq), MBD2b/MBD3L1 protein complexes with high affinity to 5mC (MIRA-seq) and sodium bisulfite conversion of unmethylated C to U (WGBS or RRBS).	(Berman et al., 2012; Gu et al., 2011; Lister et al., 2009; Weber et al., 2005)
	hMeDIP-seq (hydroxymethylated DNA immunoprecipitation sequencing), TAB-seq (Tet-assisted bisulfite sequencing)	Genome-wide map of DNA hydroxymethylation	Genome-wide mapping of hydroxymethylated cytosine (5hmC) of DNA. 5hmC is either isolated by immunoprecipitation (hMeDIP-seq) or by glucosylation-mediated protection of 5hmC sites from the subsequent Tet enzyme assisted conversion of 5hmC or MspI digestion (TAB-seq and RRHP-seq, respectively).	(Jin et al., 2011; Song et al., 2011a; Wu et al., 2011)
	MNase-seq (micrococcal nuclease sequencing)	Nucleosome positioning	MNase digestion of chromatin followed by high throughput sequencing. MNase preferentially cuts at the linker DNA between nucleosomes.	(Henikoff et al., 2011)
	4C-seq (circular chromosome conformation capture sequencing) or HiC-seq	Chromatin-chromatin interaction	Variations of chromatin conformation capture (3C) assays to detect intra- or interchromosomal interaction genome-wide. 4C-seq detects chromatin areas interacting with one genomic locus of interest in genome-wide manner, whereas HiC-seq captures all detectable chromatin-to-chromatin interactions.	(Lieberma n-Aiden et al., 2009; Rao et al., 2014; Splinter et al., 2012)
	ChIA-PET (chromatin interaction analysis by paired-end tag sequencing)	Chromatin-chromatin interaction mediated by a protein of interest	Chromatin immunoprecipitation and ligation of the adjacent DNA ends followed by high throughput sequencing.	(Zhang et al., 2012)
	DNase-seq (DNase I hypersensitivity sequencing), FAIRE-seq (formaldehyde assisted isolation of regulatory sequences), or ATAC-seq (assay for transposase-accessible chromatin sequencing)	Open chromatin	Genome-wide mapping of open chromatin. Open chromatin can be selectively isolated by the following characteristics: high sensitivity to DNase I digestion (DNase-seq), segregation in aqueous phase upon phenol/chloroform extraction (FAIRE-seq), or preferential transposon integration.	(Crawford et al., 2006) (Giresi et al., 2007)

2.1.1. Gene expression profiling

Transcriptome analysis is central to comprehensive understanding of complex biological systems such as photoreceptors (Swaroop and Zack, 2002). In RNA-seq assays, total RNA isolated from a cell type or a tissue of interest is converted to cDNA and deep-sequenced to obtain a comprehensive catalogue of transcripts (Mortazavi et al., 2008). Although mRNA-seq is most commonly conducted, genome-wide profiling of regulatory RNAs is also possible and provides an additional layer of information regarding gene regulation (Morin et al., 2008). RNA-seq offers many advantages over hybridization-based expression profiling (such as microarrays), permitting greater sensitivity and dynamic range in transcript detection (Zhao et al., 2014) and more accurate identification of alternatively spliced transcripts (Shen et al., 2014). As detection and quantification of transcripts is not limited to the current annotation, RNA-seq analysis enables discovery of novel genes and of new transcripts from annotated genes (Wang et al., 2009). Finally, RNA-seq can be performed even at the level of a single cell (Shapiro et al., 2013). With increasing accessibility and affordability, RNA-seq analysis is fast becoming a routine procedure replacing microarrays.

2.1.2. Transcriptional and epigenetic regulation

Gene expression is controlled by combinatorial action of diverse regulatory programs (Figure 2B). Interaction of transcription factors with specific cis-regulatory motifs is one of the primary mechanisms underlying temporal and cell type-specific gene regulation. ChIP, a widely used procedure to study transcription factor occupancy in vivo (Gilmour and Lis, 1985; Weinmann and Farnham, 2002), can now be applied in conjunction with tiling array (ChIP-on-chip) (Ren et al., 2000) or NGS (ChIP-seq) (Barski et al., 2007) to examine genome-wide binding of a specific protein. NGS-based methods (‘-seq’ assays) are now being used to delineate global profiles of chromatin states (such as histone modification, DNA methylation, nucleosome positioning and chromatin accessibility) and to identify novel fundamental mechanisms underlying gene transcription (see Table 1).

2.1.3. DNA sequence variation

The human genome includes millions of sequence variations, primarily in the non-coding regions; of these, tens of thousands may be unique to an individual (Schaibley et al., 2013), and many can even be linked to disease phenotype. Linkage analysis has traditionally been successful in identifying the disease variant/gene in monogenic RDDs, and genome-wide association studies (GWAS) have identified variants associated with complex traits. More recently, profiling of genetic variants using NGS methods, such as whole exome sequencing (WES) and targeted or whole genome sequencing (WGS), have transformed human genetic research. The identified variants/mutations can provide valuable insights into normal biology as well as elucidate mechanism(s) of disease pathogenesis. For instance, gene ontology analysis of 142 genes implicated in photoreceptor degeneration has identified a handful of biological pathways including those associated with cilia biogenesis, lipid metabolism, and phototransduction (Wright et al., 2010). Similarly, genetic variants associated with susceptibility to AMD have revealed the involvement of complement regulation, cholesterol transport and extracellular matrix remodeling in disease pathogenesis (Fritsche et al., 2014).

2.2. Data integration

Once high throughput data is collected, computational tools are employed to predict and construct GRNs. Although system-level measurements are made at discrete levels such as transcriptome, DNA-protein interaction, epigenome and genome, intermolecular interactions that form a network are generally not confined to a single category. Thus, the experimental data pertaining to distinct aspects of the network must be analyzed simultaneously and assimilated to identify new patterns of a biological system, which single data sets fail to detect.

Generation of networks by integrating multiple data sets necessitates that data are acquired under conditions of minimal variability; e.g., from a single cell type since cellular heterogeneity in the retina would create high transcriptional noise affecting the generation of a viable and verifiable GRN across development and disease states. Ideally, transcriptional and regulatory data for network formation should be obtained from a specific cell type at distinct stages of differentiation, aging or disease. Efforts have been made to isolate individual neuronal cells of the retina for genome-wide analysis. Such attempts, however, have been few and primarily limited to gene expression profiling (Akimoto et al., 2006; Siegert et al., 2012; Trimarchi et al., 2007). Table 2 lists public databases containing high throughput data relevant to studies of the retina.

Table 2.

Public databases containing clinical or high throughput retina data sets

Database	URL	Description	Reference
Allen Brain Atlas	http://www.brain-map.org/	Public resources integrating comprehensive gene expression and anatomical data from developing or adult mouse and human brain	(Hawrylycz et al., 2012; Lein et al., 2007)
BrainCloud	http://braincloud.jhmi.edu/plots/	Database of temporal gene expression dynamics in the human prefrontal cortex across the lifespan	(Colantuoni et al., 2011)
EyeSAGE	http://neibank.nei.nih.gov/EyeSAGE/index.shtml	Resource for retina, RPE and trabecular meshwork transcriptomes	(Bowes Rickman et al., 2006)
Gene Expression Omnibus (GEO)	http://ncbi.nlm.nih.gov/geo	Public data repository of functional genomics data supporting MIAME-compliant data submissions
Gene Expression Profile Database	http://www.fmi.ch/roska.data/index.php	Microarray data across diverse adult mouse retinal cell types	(Siegert et al., 2012)
High resolution fundus image database	https://www5.cs.fau.de/research/data/fundus-images/	Database of fundus images of healthy eyes and eyes with diabetic retinopathy or glaucoma
Human Retinal Transcriptome	http://oculargenomics.meei.harvard.edu/index.php/ret-trans	The human retina transcriptome data generated from three normal human retinas using RNA-seq	(Farkas et al., 2013)
mirNEYE	http://mirneye.tigem.it	Expression atlas of 221 miRNAs in the developing and adult wild type mouse eye
Mouse Retina SAGE library	http://cepko.med.harvard.edu	Database of gene expression profiling of developing and adult mouse retina and a few other non-ocular tissues. In situ hybridization data are also available for select genes.	(Blackshaw et al., 2001; Blackshaw et al., 2004)
NEIbank	http://neibank.nei.nih.gov/index.shtml	Database of assembled EST data from ocular tissues of various organisms
Retina Central (WEBER)	http://www.retinacentral.org/	Database of genes experimentally shown to be expressed in the retina/the retinal pigment epithelium.	(Schulz et al., 2004)
Retina International	http://www.retina-international.org/sci-news/databases/mutation-database/	Database of retinal diseases, mutations, animal models and disease-associated proteins
Retinal Express	http://odin.mdacc.tmc.edu/RetinalExpress/	E14.5 mouse retina cDNA/EST database
RetinoBase	http://alnitak.u-strasbg.fr/RetinoBase/	Microarray database generated from all publicly available retina-related gene expression profiles	(Kalathur et al., 2008)
RetNet (Retinal information network)	https://sph.uth.edu/retnet/	Database of genes and loci causing inherited retinal diseases
RP Gene Expression Atlas	http://rpexp.tigem.it/	Collection of in situ hybridization data for retinitis pigmentosa genes in mouse and human retina
STARE (Structured analysis of retina)	http://www.ces.clemson.edu/~ahoover/stare/	Database of retinal images of various clinical manifestations	(Hoover and Goldbaum, 2003; Hoover et al., 1998)

2.3. System comparison

Biological systems are highly dynamic and undergo progressive transition over time (Hwang et al., 2012). During development, intrinsic and extrinsic factors sequentially restrict cell fates and specify morphology and function most suited for a given cellular identity. Cumulative adaptive responses to internal and/or environmental challenges over a lifespan also impact spatio-temporal architecture of a system. Therefore, comparisons along time series data sets represent valuable strategies to assess system dynamics. Additionally, continuous fluctuations imposed by microenvironment and/or intrinsic genetic variations may induce changes beyond the cell tolerance level and manifest as a disease. The use of animal models for retinal traits or diseases provides a powerful tool to evaluate direct influence of genetic mutation(s) or experimental stress on networks as relatively little variation is expected in congenic strains raised under uniform conditions. Another important factor in system dynamics is variability among cell types exhibiting functional interaction such as between photoreceptors and retinal pigment epithelium (RPE); for example, defects in RPE are associated with photoreceptor dysfunction or death (Strauss, 2005). Even cells of the same neuronal type may display high variability based on temporal context and spatial organization (Trimarchi et al., 2007).

Change in system architecture can be ascertained from human population genetics using control individuals and patients with a specific disease/trait (e.g., AMD). Such genetic association studies exploit naturally occurring differences that cannot necessarily be generated by experimental means. However, it should be noted that the strength of data sets depends on the size of the cohort and penetrance of a specific variant/mutation. Elucidation of multifactorial complex biological phenomena, such as aging and pathogenesis of complex disease, can especially benefit from comparative system-level analysis using large NGS-based genomic data.

In the following sections, we will elaborate on system-level analysis of rod photoreceptors in three disciplines: development (section 3), aging (section 4) and degenerative disease (section 5).

3. System-level analysis of retinal photoreceptor development

Early attempts for system-wide assessment of photoreceptor development included gene expression profiling of the retina using expressed sequence tags (ESTs), serial analysis of gene expression (SAGE) and microarrays (Figure 3A) (Akimoto et al., 2006; Blackshaw et al., 2001; Gieser and Swaroop, 1992; Livesey et al., 2000; Sharon et al., 2002; Yoshida et al., 2004; Yu et al., 2003b). NGS-based methods have dramatically expedited the pace of expression profiling and permitted the application of gene regulation assays to genome-wide scale (Figure 3B). In addition to examining the whole retina, we can now develop quantitatively precise expression profiles of individual cell types and obtain global data on transcription factor binding and epigenomic marks. Multi-dimensional data integration should therefore enable us to identify novel patterns in GRNs that control functional architecture of individual retinal cells, such as photoreceptors.

Figure 3.

Timeline of genome-wide studies of the retina biology and disease pathogenesis. The advent of genome-scale profiling technologies has been a critical step for systems biology approaches. A, From the pioneering high throughput transcript analysis, such as the initial application of microarray and serial analysis of gene expression (SAGE) to whole genome sequencing, an ever-growing number of genome-wide studies have advanced our knowledge about healthy retina and disease pathogenesis. Highlights of such innovative genome-scale studies were selected and presented chronologically. B, For more than a decade, microarray has been a widely used methodology of choice for gene expression profiling, yielding a substantial number of publications each year. RNA-seq, deep sequencing of cDNA using NGS technology, is becoming more accessible and affordable and thus expected to be applied more widely. In addition to transcriptome analysis, NGS is applicable to a variety of other conventional research techniques and has already generated numerous data sets surveying whole exomes for genetic variation (whole exome sequencing), transcription factor targetome (ChIP-seq) and epigenome (ChIP-seq for histone modifications and various DNA methylome sequencing methodologies). ¹(Livesey et al., 2000), ²(Blackshaw et al., 2001), ³(Mu et al., 2001), ⁴(Yoshida et al., 2002), ⁵(Sharon et al., 2002), ⁶(Farjo et al., 2002), ⁷(Chowers et al., 2003b), ⁸(Gustincich et al., 2004), ⁹(Klein et al., 2005), ¹⁰(Akimoto et al., 2006), ¹¹(Trimarchi et al., 2007), ¹²(Arora et al., 2010; Hackler et al., 2010; Karali et al., 2010, Wang, 2010 #304), ¹³(Chen et al., 2010b), ¹⁴(Neale et al., 2010), ¹⁵(Otto et al., 2010), ¹⁶(Tummala et al., 2010), ¹⁷(Corbo et al., 2010), ¹⁸(Grant et al., 2011), ¹⁹(Brooks et al., 2011; Mustafi et al., 2011), ²⁰(Popova et al., 2012), ²¹(Hao et al., 2012), ²²(Farkas et al., 2013), ²³(Oliver et al., 2013b), ²⁴(Fritsche et al., 2013), ²⁵(Nishiguchi et al., 2013), ^a number of publications until September 2014.

In the developing vertebrate retina, rod and cone photoreceptors differentiate from common pools of retinal progenitor cells with distinct temporal profiles. While cone generation ceases prenatally in rodents, rod birth spans a long time window from embryonic day (E)12 to postnatal day (P)10 with peak at P0-P2 (Carter-Dawson and LaVail, 1979; Rapaport et al., 2004). Photoreceptor identity is largely dictated by an intrinsic transcriptional program (Swaroop et al., 2010) that involves a number of transcription factors, including cone-rod homeobox protein CRX (Chen et al., 1997; Furukawa et al., 1997), neural retina leucine zipper protein NRL (Mears et al., 2001; Swaroop et al., 1992) and thyroid hormone receptor THRB (also called TRβ2) (Ng et al., 2001). CRX plays an essential role in photoreceptor development (Furukawa et al., 1999) by controlling the expression of both rod and cone genes (Corbo et al., 2010; Hennig et al., 2008). NRL and THRB, however, determine the genesis of rods (Mears et al., 2001) and medium wavelength sensitive (M) cones (Ng et al., 2011), respectively, from photoreceptor precursors that seem to be otherwise specified by “default” as short wavelength sensitive (S) cones (Swaroop et al., 2010). Other essential regulators of photoreceptor development include OTX2 (Nishida et al., 2003; Roger et al., 2014), RORB (Jia et al., 2009; Srinivas et al., 2006) and NR2E3 (Chen et al., 2005; Cheng et al., 2006; Cheng et al., 2011; Haider et al., 2000; Oh et al., 2008; Peng et al., 2005). These key transcriptional regulators constitute central nodes (“hubs”) in the GRN that controls photoreceptor development.

3.1. Cell type-specific approaches for generating GRN

Although few studies have been performed using purified photoreceptor cells, rod photoreceptor-specific information can be predicted by genome-wide data generated from the predominantly rod-containing mouse or human retina. Considering that NRL and CRX expression are largely photoreceptor-specific, ChIP-seq data for these two key transcription factors would broadly reflect photoreceptor biology. The same is not true for transcriptome and epigenome data obtained from the whole retina. Isolation of photoreceptors is necessary to acquire true photoreceptor-specific information. Cell type-specific approaches are even more critical for developmental studies since the proportion of photoreceptors in mouse retina continuously increases until P10, making it difficult to discern true cellular changes in gene expression or epigenetic marks. The first example demonstrating the importance of cell type-specific studies comes from microarray-based expression profiling of flow-sorted rod photoreceptors from Nrlp-GFP mice (Akimoto et al., 2006). The resulting rod-specific transcriptome data effectively detected many low level transcripts that had not been observed in the whole retina transcriptome data. Although many transgenic mouse lines expressing fluorescence reporters in specific retinal cell types have been described and are being used for transcriptome analysis (Siegert et al., 2012), identification of new cell type-specific markers (Koso et al., 2009) would greatly facilitate NGS studies of retinal neuronal subtypes.

An array of subtypes, each with unique morphology and function, can be recognized within major classes of retinal neurons (Masland, 2012). Single cell gene expression profiling has demonstrated a high degree of heterogeneity among individual retinal cells (Trimarchi et al., 2008). A better understanding of biological events thus requires delineation of GRNs specific not only at the level of individual subtypes but perhaps even at a single cell level. Due to limitations in obtaining sufficient material, conventional methods need to be miniaturized to gain enough sensitivity for single cells or cell types. However, technical hurdles including sample loss, degradation and contamination can have more pronounced impact on quality and robustness of data generated from single cells or small quantities versus large pools of cells. In addition, many rounds of PCR amplification may generate increased noise and inaccurate quantification. Nevertheless, cell type-specific or even single cell level analyses are necessary to extract gene regulatory information relevant to the cell of interest. Single cell collection methods based on micro-fluidic systems (Shalek et al., 2014), high-resolution imaging (Lubeck and Cai, 2012), and better protocols for constructing NGS libraries from small amount of starting RNA (Brooks et al., 2012; Tang et al., 2009; Tariq et al., 2011) have made it possible to attempt more robust expression profiling of individual retinal neurons. However, a major challenge remains for the generation of epigenomic and other NGS data sets with small number of cells. Our laboratory and others are currently developing such protocols (Adli et al., 2010; Guo et al., 2013; Smallwood et al., 2014).

The importance of using a purified single cell type for system-level analysis cannot be over-emphasized since biological processes such as development and aging have broad and concurrent impact on multiple cell types in a tissue and on the organism as a whole. Deciphering photoreceptor GRN from the whole retina studies would be complicated by transcriptional noise and differences in epigenome and regulatory molecules among different cell types. This is especially true for dissecting changes in GRNs during the pathogenesis of photoreceptor degeneration. As photoreceptor-specific data are not available for all NGS-based assays, we also cover studies from the whole retina in the discussion below.

3.2. Construction of photoreceptor GRN

Gene expression requires context-specific interaction among cis-regulatory DNA elements, basal transcriptional machinery and transcription regulatory proteins. In addition, transcriptional activity is under the influence of chromatin architecture as well as regulatory RNAs. Thus, basic elements needed to generate a rod photoreceptor GRN include:

1. Global mRNA profiling by RNA-seq (section 3.2.1)

2. Identification of target genes for cell specific transcription factors by ChIP-seq (sections 3.2.2 and 3.2.3)

3. Chromatin state such as histone modifications by ChIP-seq, DNA methylation profiling, and chromatin accessibility by DNase I hypersensitivity sequencing (section 3.2.4)

4. Regulatory RNA profiling (section 3.2.5)

Transcriptional profiles are also needed at multiple stages of rod development in order to build a developmental GRN. Gene expression profiling can be used in conjunction with gain- or loss-of function mutations of relevant “hub” transcription factors to further validate GRNs. For example, known hubs in the photoreceptor network such as NRL and CRX can be perturbed (e.g., in mouse mutants) before genome-wide gene expression analysis. The genes showing altered expression in the absence of a regulatory factor would represent direct or indirect (i.e., through secondary regulatory nodes) transcriptional targets. Additionally, GRN can also be inferred from co-expression data. Groups of genes that share similar expression signatures in response to experimental manipulation and/or during development would likely represent common gene regulatory pathways. Although such information is only suggestive, co-expression data allow for the identification of novel testable regulatory patterns.

3.2.1. Expression profiling

Global profiling of gene expression in the developing and mature retina was first accomplished by Affymetrix™ microarrays (Dorrell et al., 2004; Yoshida et al., 2004). Expression profiles have also been generated from flow-sorted rod photoreceptors (Akimoto et al., 2006; Parapuram et al., 2010) and from single retinal cells (Ma et al., 2013; Roesch et al., 2008; Trimarchi et al., 2007; Xue et al., 2011). However, NGS-based RNA-seq profiling has yielded more quantitatively precise information on mRNA transcripts in the mammalian retina (Brooks et al., 2011; Farkas et al., 2013; Gamsiz et al., 2012; Kandpal et al., 2012; Kozhevnikova et al., 2013; Roger et al., 2014). More recently, we have produced extensive RNA-seq-based temporal expression profiles of purified rod photoreceptors that reveal novel insights in cellular pathways (Figure 4A) (J-W Kim, H-J Yang and A Swaroop, manuscript in preparation) as well as additional modes of gene regulation, such as alternative splicing (Figure 4B).

Figure 4.

Cell-type specific system-level analysis. A, Heat map of time course RNA-seq data generated from isolated, developing (P2 to P14) and mature (P28) mouse rod photoreceptors. Dynamic gene regulation during rod photoreceptor maturation and clusters of genes with similar temporal expression patterns are apparent. Heat map was generated based on log₂FPKM, and individual co-expression clusters were highlighted with different colors in the dendogram. FPKM, fragments per kilobase of exon per million reads. B, Transcript-level analysis of RNA-seq data enables detection of complex gene regulatory programs such as distinct splicing events during development. Alternative splicing of Bsg (basigin) gene during rod photoreceptor development is shown as an example. Coverage plots (dark grey histogram) and read alignments (grey blocks indicating individual sequence reads with thin blue horizontal lines connecting portions of sequence reads that are split between exons) show differential inclusion of the second exon (brown shade) in P2 and P28 rod photoreceptors. Expression level of each splice variant during rod development is plotted and shown on the right. C, System-level profiling of diverse chromatin signatures, including chromatin accessibility [DNase-seq (DNase I hypersensitivity sequencing)], active and repressive promoter histone modifications (H3K4me3 and H3K27me3, respectively; mini-ChIP-seq) and DNA methylation (bisulfite sequencing). Application of genome-wide analysis of chromatin states used to be limited to in vitro samples and pooled tissues of heterogeneous cell types. Bisulfite sequencing and ChIP-seq have now been miniaturized for flow-sorted single type of neurons, which are available in small amounts. Shown are active gene expression and hallmark of active chromatin state of two select photoreceptor-specific genes, Crx and Slc24a1, in P28 flow-sorted rod photoreceptors. Scale on y-axis of RNA, DNase, K4me3 and K27me3 tracks indicates RPM (reads per million reads). Average percent 5-methylcytosine (meDNA) within promoter (±1 kb from the transcription start site, highlighted with a grey shade) was plotted as a bar graph.

3.2.1.1. Perturbation studies

For elucidating GRNs, it is valuable to generate gene profiles after disrupting one of the key regulatory nodes. This strategy was first applied to Crx^−/− mice (Blackshaw et al., 2001; Hennig et al., 2008; Livesey et al., 2000). CRX is essential for both cone and rod development (Furukawa et al., 1999); however, its role in cell fate determination is debatable because of overlapping functions with another homeodomain protein OTX2, which regulates the expression of CRX as well as NRL (Nishida et al., 2003; Omori et al., 2011; Roger et al., 2014; Terrell et al., 2012). A majority of genes with altered expression in Crx^−/− mice showed photoreceptor enriched expression, and functional categorization of differentially expressed genes has revealed an array of CRX-regulated cellular functions, including those involved in phototransduction, metabolism, signal transduction and cytoskeletal components.

Another extensively studied gene regulatory program is the NRL-centered transcriptional network. The first observation that NRL is the master switch for generating rod cell fate came from studies of Nrl^−/− mice as their retina completely lacks rod photoreceptors, with concurrent enhancement of S cone function (Mears et al., 2001). In concordance, ectopic expression of NRL in photoreceptor precursors or in early-born S-cones resulted in formation of functional rods (Oh et al., 2007). Gene profiling of Nrl^−/− retina revealed little or no expression of rod-specific genes, including those associated with rod phototransduction, and a concomitant increase in the expression of S cone photopigment (encoded by Opn1sw) and other cone genes (Brooks et al., 2011; Yoshida et al., 2004; Yu et al., 2004a). Global expression analysis of purified photoreceptors from Nrl^−/− retina has validated the essential role of NRL in activating the expression of rod genes and suppressing cone genes (Akimoto et al., 2006).

NR2E3, an orphan nuclear receptor, constitutes another important hub in photoreceptor GRN and is a direct transcriptional target of NRL (Bumsted O’Brien et al., 2004; Oh et al., 2008). Mutations in NR2E3 cause enhanced S cone syndrome in humans (Haider et al., 2000; Sharon et al., 2003; Wright et al., 2004), similar to the retinal phenotype exhibited by Nrl^−/− mice (Mears et al., 2001). Retinal and rod gene expression profiling of rd7 mutant mice, in which Nr2e3 function is ablated, further validates NR2E3 function primarily as a suppressor of cone genes and as a co-activator (with NRL and CRX) of rod genes (Chen et al., 2005; Cheng et al., 2006; Cheng et al., 2011; Cheng et al., 2004; Corbo and Cepko, 2005; Haider et al., 2009; Oh et al., 2008; Peng et al., 2005).

3.2.1.2. Co-expression network analysis

Target genes of the a transcriptional regulator likely display similar expression patterns and may participate in common or related cellular functions (Yu et al., 2003a). Clusters of co-expressed genes can thus be used to model gene regulatory events. The inferred GRNs enable the detection of putative protein-DNA and protein-protein interactions without requirement of prior knowledge. Mathematical tools have been developed to predict GRNs from gene expression profile data, and putative regulatory relationships can then be validated experimentally (Hecker et al., 2009). Co-expression network inference thus represents a powerful analytical tool for expression profiling data, complementing the perturbation studies discussed above. As activity of a given transcriptional regulator is confined only to the same cell as its target, only the information obtained from a single cell or cell type provides accurate prediction without analytical noise caused by sample heterogeneity. Transcriptome analysis of flow-sorted rod photoreceptors from wild type and Nrl^−/− retina has revealed multiple clusters of co-expressed genes, which likely represent diverse regulatory events (Akimoto et al., 2006). Co-expression analysis of newer and deeper RNA-seq data sets should provide valuable insights, including those relevant for identification of disease genes and pathways (see also section 5.4).

3.2.2. Transcription targets (“targetome”) of NRL and CRX

Differential gene expression analysis does not distinguish direct transcriptional regulation from secondary regulatory events. ChIP-seq analysis, on the other hand, assays direct physical interactions between a transcription factor and the cognate DNA elements. Recent studies have superimposed transcriptome and transcription factor targetome data to identify biologically relevant target genes of CRX (Corbo et al., 2010) and NRL (Hao et al., 2012). ChIP-seq analysis of mouse retina demonstrated CRX binding to 67% of mis-regulated genes in Crx^−/− retina, likely representing direct CRX targets (Corbo et al., 2010). Similarly, only a subgroup (15%) of genes differentially expressed in Nrl^−/− retina exhibited NRL ChIP-seq peaks and thus are direct transcriptional targets of NRL (Hao et al., 2012).

One of the challenges in deciphering targetome data is assigning transcription factor binding to relevant genes. A majority of NRL- and CRX-bound genomic DNA regions are in close proximity of or within the genes. However, binding of regulatory factors far away from genes is not uncommon and may exert transcriptional control by chromatin looping. Recently, high throughput reporter gene assays have successfully validated the enhancer activity of over 1,000 CRX interacting genomic loci (White et al., 2013). Similar large-scale enhancer assays and chromatin state mapping are expected to confirm the functional significance of distant ChIP-seq peaks. Integrated analysis with gene expression data and long distance chromatin association profiling [by chromatin conformation capture (3C) assay (Peng and Chen, 2011) and its NGS applications such as 4C-seq and HiC-seq] (see Table 1) will further facilitate the identification of novel transcription targets of key photoreceptor transcription factors.

ChIP-seq analyses have identified thousands of genomic loci (sequence regions) that are occupied by NRL or CRX in vivo, and both transcription factors bind to common enhancer regions that specify photoreceptor genes (Hao et al., 2012). In embryonic stem cells (ESCs), master transcriptional regulators are reported to form super-enhancers, atypical long enhancer clusters that are associated with unusually high level of mediators as well as enhancer-associated epigenetic marks (Whyte et al., 2013). Subsequent studies have demonstrated the importance of super-enhancers in many differentiated cell types (Hnisz et al., 2013) and their implications in developmental and neurological disease (Lee and Young, 2013). Further investigations are necessary to evaluate whether NRL- and CRX-bound enhancer elements exhibit characteristics of super-enhancers and whether variants in these elements are associated with retinal disease.

3.2.3. Dynamicity and combinatorial action of transcription factors

Most of the transcription factor targetome analyses have so far been conducted at a single time point, revealing only a snapshot of gene regulation. The function of transcriptional regulators, however, is highly dynamic (Davidson and Levine, 2008; Hwang et al., 2012). Transcriptional complexes are under continuous modulation by availability of co-factors and by chemical modifications at the level of protein, DNA and/or chromatin (Berger, 2007; Voss and Hager, 2014). Thus, the regulatory outcome can vary depending on cellular context such as developmental stage, cell identity and fitness. Comparative transcriptome analysis of the retina and of flow-sorted rod photoreceptors across developmental stages has revealed a striking variation in expression kinetics among photoreceptor genes controlled by NRL (Akimoto et al., 2006; Yoshida et al., 2004; Yu et al., 2004a). Nrl expression is initiated immediately after the final mitosis of photoreceptor precursors, leading to expression of a large number of target genes including Nr2e3. Notably, a number of NRL target genes, including visual pigment rhodopsin (encoded by Rho), exhibit substantial delay in expression during rod photoreceptor development.

Although underlying factors for differential onset of expression of NRL target genes are not completely understood, analyses of multiple data sets suggest additional control mechanisms. One of the possible scenarios can be inferred by linking NRL targetome data with gene expression profiles and with previously identified retinal disease loci (Hao et al., 2012). NRL target genes with putative functional relevance can be selected based on the assumption that such genes are likely associated with rod function and cause retinal degeneration when mutated. Such analyses have identified histone demethylase KDM5B as a secondary regulatory node, and Kdm5b knock-down partially phenocopies the rod-to-cone transition observed in Nrl^−/− retina (Hao et al., 2012). KDM5B is one of the histone demethylases implicated in gene repression through demethylation of H3 lysine 4 trimethylation (H3K4me3), a histone mark associated with transcriptional activation. Thus, it can be hypothesized that NRL function is mediated through multiple as yet unidentified secondary nodes (including epigenetic factors) and that other regulatory mechanisms may define precise temporal control of gene expression. Functional interaction between NRL-centered transcriptional network(s) and the epigenetic program will be further elaborated in the following section 3.2.4.

Another important feature of GRNs is combinatorial actions of transcriptional regulators (Davidson and Levine, 2008). Numerous pre-genomic approaches have reported cooperative functions among photoreceptor transcription factors. Interaction between NRL and CRX exerts synergistic effects on transcriptional activation of rod-specific genes (Mitton et al., 2000; Pittler et al., 2004; Yoshida et al., 2004). As part of the same complex, NR2E3 exhibits dual function by inducing rod and suppressing cone gene expression (Oh et al., 2008; Webber et al., 2008), thereby constituting a critical secondary transcriptional regulatory node (Chen et al., 2005; Cheng et al., 2006; Cheng et al., 2004; Haider et al., 2000; Peng et al., 2005). Compiling system-wide surveys of putative target genes for these three regulators has revealed significant overlaps in their targets (Hao et al., 2012; Qian et al., 2005; Yu et al., 2006), suggesting that combinatorial functions among transcription factors exert tight regulation of photoreceptor development and homeostasis. Salient features of NRL transcriptional regulatory network have been assembled (Hao et al., 2012), and further computational analyses of existing and new data sets are expected to provide a comprehensive map of rod photoreceptor GRNs (Hwang et al., 2012; Wan et al., 2013).

Combinatorial action of regulatory factors is highly context-dependent and cell type-specific, and the composition of transcriptional complexes is expected to influence the selectivity of downstream target genes. For instance, although CRX expression is common to both rods and cones, transcriptional regulatory outcomes show a striking difference. CRX cooperates with RORB to induce S cone photopigment (Opn1sw) expression in cone precursors (Srinivas et al., 2006), whereas, in the presence of NRL and NR2E3, CRX enhances rod-specific gene expression in rod photoreceptors (Hao et al., 2012). ChIP-seq analysis has shown that OTX2-bound cis-regulatory elements are quite different in RPE and neural retina (Samuel et al., 2014). Only retina-specific OTX2 target sites, but not RPE-specific ones, show redundancy with CRX occupancy, further demonstrating the importance of combinatorial transcriptional programs in cell fate specification.

3.2.4. Chromatin state and gene regulation

DNA within a cell is packaged into DNA-protein complexes (termed chromatin), which constitute a dynamic structure. DNA itself and the histones, which are primary protein components in the chromatin, are subjected to chemical alterations that can influence transcriptional activity. Regulation of chromatin state is an important mechanism for modulating biological states, including lineage restriction and cell fate specification during development (Hirabayashi and Gotoh, 2010; Meshorer and Misteli, 2006; Sasaki and Matsui, 2008; Wan et al., 2013). Studies of epigenetic marks such as DNA methylation and posttranslational modifications of histones in the retina have revealed cell type- and stage-specific chromatin states, recapitulating context-dependent differential gene regulation.

3.2.4.1. DNA methylation

DNA methylation at gene promoters has been negatively associated with gene expression and is considered a key mechanism for maintenance of chromatin state (Suzuki and Bird, 2008). Integration of global DNA methylation profiles with gene expression is critical for deciphering the overall structure of a GRN. System-level analysis of DNA methylome has implicated differential 5-methylcytosine in altering gene expression during photoreceptor degeneration in rd1 retina (Farinelli et al., 2014) as well as in tissue-specific splicing in the retina and the brain (Wan et al., 2013). More restricted approaches examining selected genes further support the association of DNA methylation with photoreceptor gene regulation. Bisulfite sequencing of targeted genomic areas from WERI and Y9 human retinoblastoma cell lines and laser-capture micro-dissected mouse photoreceptors and non-photoreceptor cell types has revealed photoreceptor-specific cytosine hypomethylation within opsin genes such as Rho, Opn1sw and Opn1mw (Merbs et al., 2012), demonstrating cell type-specific patterns of DNA methylation. Conditional knock-out of DNA methyltransferase 1 (Dnmt1) shows lack of outer segments despite nearly normal expression of phototransduction and cilia genes in photoreceptors (Nasonkin et al., 2013), suggesting a broad impact of DNA methylation on morphological specification.

Contrary to rapidly growing number of gene expression profiling studies, only few genome-wide DNA methylome profiles have been generated for the retina (Figure 3B). New highly sensitive detection assays for 5-methylcytosine and 5-hydroxymethylcytosine (see Table 1)(Oliver et al., 2013b; Smallwood et al., 2014) and with single cell-type resolution (Powell et al., 2013) are expected to facilitate such investigations at the level of a single retinal neuron. We have recently completed the generation of DNA methylome for flow-sorted rod photoreceptors and are examining its relationship to RNA-seq and histone modification data sets (Figure 4C) (H-J Yang, J-W Kim and A Swaroop, manuscript in preparation). Further technical improvements should therefore unravel system-wide interactions between DNA methylation and cell type-specific dynamics of gene expression.

3.2.4.2. Histone modifications

More than 60 amino acid residues on different histones undergo diverse posttranslational modifications including acetylation and methylation (Kouzarides, 2007; Zentner and Henikoff, 2013). While histone acetylation is almost strictly associated with active gene expression, methylation represents more complex modes in gene regulation. Transcriptional activity is shown to depend on the degree of methylation – mono-, di-, and trimethylation, the specific modified amino acid residue and associated DNA elements (e.g., promoter, gene body, enhancers). Among multitudes of reported histone methylations, H3K4 and H3K27 methylation are closely correlated with transcription status and thus constitute the most commonly examined histone modifications. Promoters of active and silent genes are enriched with H3K4me2/me3 and H3K27me3, respectively, while active and repressive enhancers are demarcated with H3K4me1/H3K27ac and H3K4me1/H3K27me3, respectively. Dynamic regulation of histone methylation patterns in rod or cone photoreceptors is expected to have significant impact on photoreceptor GRNs. ChIP-seq is fast becoming a routine procedure for genome-wide profiling of histone modifications. Histone signature, however, is highly heterogeneous across distinct cell types (ENCODE Project Consortium, 2012); thus, cell type-specific histone modification profiling is necessary to obtain useful information on the chromatin state. Purified cells do not easily meet the requirement of sample size for reported epigenome profiling assays as yet.

The whole retina histone ChIP-seq performed with appropriate control tissue such as rd1 mutant retina, a degenerative mouse model that lacks photoreceptors, has offered significant photoreceptor-specific epigenomic information as rod photoreceptors account for a vast majority of retinal neurons in mature retina (Popova et al., 2012). In this study, genome-wide signatures of H3K4me2 and H3K27me3 in the mouse retina were mapped and compared with transcriptome at various developmental stages. Distinct clusters of chromatin modification dynamics were identified among rod-specific genes (Popova et al., 2012). Surprisingly, rod genes showed varied H3K4me2 signature. While some rod genes have a detectable level of H3K4me2 as early as E17.5 and at all subsequent developmental stages, another group of rod genes including phototransduction genes shows delayed de novo accumulation of H3K4me2 from P7 to P15. Concordantly, phototransduction genes gain full chromatin accessibility only after P10 as revealed by DNase I hypersensitivity sequencing (DNase-seq) of the mouse retina (ENCODE, www.mouseencode.org). ChIP experiments have demonstrated a limited interaction of NRL with cognate DNA binding site in Rho at P2 (Hao et al., 2011), suggesting that inactive chromatin state at Rho and other phototransduction genes may interfere with NRL binding to its target genes in newborn rods. These results, however, should be interpreted with caution as the detected change in histone marks, chromatin accessibility and NRL recruitment may simply reflect a dramatic increase in the proportion of rod photoreceptors in the retina from P2 to P28. Assays using purified rod photoreceptors offer an ideal alternative. Our initial analysis of histone ChIP-seq data from purified rod photoreceptors has revealed promising insights into rod-specific histone signature (Figure 4C) (H-J Yang et al., manuscript in preparation).

3.2.5. miRNA and other transcribed sequences

Non-coding RNAs (ncRNAs) constitute as much as 80% of the human transcriptome (Rosenbloom et al., 2013) and contribute additional complexity to GRNs (reviewed in (Mortimer et al., 2014). Small ncRNAs (<200 bp) include microRNAs (miRNAs) of 18-24 nucleotides. Long ncRNAs (lncRNAs) are over 200 bp in length but can be up to hundreds of kb. Each of the reported 2,000-3,000 miRNAs (miRBASE Release 21, June 2014) can regulate hundreds of target genes, and each mRNA can be targeted by several miRNAs, fine tuning gene expression of an estimated 30–80% of the human genome (John et al., 2004; Lim et al., 2005; Lu and Clark, 2012). Tens of thousands of lncRNAs are expressed at low levels in a tissue-restricted manner and believed to regulate the expression of related genes associated with most biological processes in a given cell (Li and Chang, 2014).

Although photoreceptor-specific profiling of ncRNA species is yet to be conducted, dynamic regulation of specific ncRNAs has been reported during retinal development and disease. In silico, RT-PCR and in situ hybridization studies have revealed spatio-temporal expression of miRNAs in the developing and mature retina (Arora et al., 2010; Deo et al., 2006; Hackler et al., 2010; Karali et al., 2010; Ryan et al., 2006; Xu, 2009). In addition, miRNAs exhibit altered expression during circadian cycle (Krol et al., 2010; Xu et al., 2007), cell survival (Damiani et al., 2008; Lumayag et al., 2013; Sundermeier et al., 2014), and in disease conditions (Loscher et al., 2008; Shen et al., 2008). The effect of miRNAs at the intersection of several GRN is exemplified by the “miR-96, miR-182 and miR-183” cluster, which contributes to circadian regulation of gene expression in the retina and has a more general role in neurosensory cell-specific GRNs (Xu et al., 2007).

Eighteen lncRNAs are highly conserved in mammalian eyes; of these, fourteen are expressed in the macular region and others in rest of the retina (Mustafi et al., 2013). Furthermore, lncRNAs are shown to play a role in retinal development (Rapicavoli and Blackshaw, 2009); e.g., in regulation of cell cycle (Vax2os1), modulation of transcription factor activity (Six3OS), and specification of cell fate (RNCR2) (Meola et al., 2012; Rapicavoli et al., 2011), highlighting their potential involvement in GRNs. As transcriptome databases expand, functional studies of existing and newly-identified small and lncRNAs will provide novel insights into their role in fine-tuning GRNs.

3.2.6. Proteome analysis

Transcriptome analysis does not accurately reflect protein expression in a cell or tissue since protein content is also regulated by translational and post-translational events and reflects the equilibrium between synthesis, stability and degradation. Complex protein-protein interactions add another level of complexity to how cellular physiology and homeostasis are controlled. Protein profiles of the retina (Barnhill et al., 2010; Finnegan et al., 2008) have been generated using mass spectrometry, validating the presence of a multitude of retinal proteins as previously predicted by microarray results. Studies have also reported photoreceptor-enriched proteome acquired from planar cryosections of the photoreceptor layer (McKay et al., 2004) or rod outer segment disk purification (Elliott et al., 2008; Kwok et al., 2008; Panfoli et al., 2008). However, global proteomic methodologies are more qualitative than quantitative and not suitable for the analyses that require high detection sensitivity and comprehensiveness as in genomics approaches. Various micro-scale protein detection techniques are being developed for protein analysis. Quantification of protein expression using nanopore technology (Wei et al., 2012b), single protein molecule counting on a two-dimensional surface (Tessler et al., 2009), and single cell western blotting (Hughes et al., 2014) are among promising new technologies, which can potentially be adapted for increased throughput.

A targeted proteomic approach has now begun to provide insights into photoreceptor gene regulation. A recent study focused on NRL-containing transcriptional complexes and identified the transcription-splicing protein, NonO, and other binding partners of NRL as important regulators of rod gene expression (Yadav et al., 2014). We should note that translation of mRNA into protein has been successfully monitored indirectly through ribosomal profiling (Ribo-seq) (Ingolia et al., 2009). Application of newly evolving proteomic approaches would also be valuable to complement the gene expression profiling and epigenome data. More targeted or global proteomic analysis is expected to completely define photoreceptor GRNs.

4. System-level analysis of retinal aging

Aging in higher organisms reflects the interplay of genetic factors with the environment and influence of stochastic events (Figure 5), requiring investigation(s) of interaction networks underlying biological processes rather than focus on a single gene or protein (Kriete et al., 2006). Although aging does not cause a clinical disease per se, phenotypes of aging can resemble disease phenotypes. Conversely, aging-associated changes increase vulnerability and instability of the system, making it more susceptible to further damage and to the effects of genetic mutations. Thus, aging is a leading risk factor for several common late-onset diseases. In this section, we explore genome-wide studies of aging as a time-dependent event relevant for eye function and pathology, with focus on the retina.

Figure 5.

Interface between aging and disease. The environment influences photoreceptor homeostasis throughout life. The cell adaptive response helps maintain a balanced homeostasis. As the adaptive response becomes insufficient to overcome “insults” to the system, damage accumulates in aging photoreceptors. Major metabolic failure is observed for the ubiquitin-proteasome system and mitochondria. Epigenetic changes and stochastic changes in gene expression combined with the presence of susceptibility variants, eventually tilt the equilibrium towards the disease state. At which point, inflammation and angiogenesis become pathologic, further aggravating disease manifestations.

Each component of the eye contributes differently to overall decline in visual function with aging. Progressive deterioration in retinal anatomy (Curcio, 2001; Curcio et al., 1993; Gao and Hollyfield, 1992) and psychophysical parameters (Birch and Anderson, 1992; Bonnel et al., 2003; Freund et al., 2011; Shinomori and Werner, 2012) is observed with advanced age. Structural and functional changes have also been reported in photoreceptors (Curcio, 2001; Gresh et al., 2003; Kolesnikov et al., 2010; Parapuram et al., 2010; Samuel et al., 2011).

4.1 Expression profiling of aging retina

Systems biology of the aging retina starts with gene expression profiling to detect chronological events that alter cellular homeostasis (Table 3). Early studies were performed using SAGE (Sharon et al., 2002) or microarray platforms (Cai et al., 2012; Chowers et al., 2003b; Yoshida et al., 2002) on post-mortem human retinas, which have inherent variability associated with individual genetic variations, pre- and post-mortem conditions, and quality of RNA. Little to no concordance among studies was observed at the gene level, due to differences in samples, platforms and analyses. However, when genes are considered in the context of ontologies and gene networks, characteristic patterns begin to emerge in the aging retina, suggesting a general metabolic slow down (Cai et al., 2012; Chowers et al., 2003b; Yoshida et al., 2002) and decrease in expression of phototransduction genes (Chowers et al., 2003b; Sharon et al., 2002). Genes associated with stress response and inflammation are generally up-regulated (Cai et al., 2012; Chowers et al., 2003b; Yoshida et al., 2002), thus highlighting manifestations of retinal aging that are shared with the brain (Lee et al., 2000). Notably, there appears to be a trend towards increased expression in older age of extracellular matrix genes, with relevance to the ontogenesis of AMD (Chowers et al., 2003b) and differential aging mechanisms in the macula compared to the peripheral retina (Cai et al., 2012).

Table 3.

Gene expression profiling studies of aging ocular tissues

Organism	Tissue/cell type	Platform	Number of genes	Age range	Reference
Human	Neural retina/macula and periphery	SAGE	320,998 tags	44–88 yr	(Sharon et al., 2002)
Human	Neural retina	Microarray, Micromax	2,400 probes	13–72 yr	(Yoshida et al., 2002)
Human	Neural retina	Microarray, Custom	10,034 genes	29–90 yr	(Chowers et al., 2003a)
Human	Neural retina/macula and periphery	Microarray, Affymetrix Human Genome U133 plus 2	54,600 gene probes	18–79 yr	(Cai et al., 2012)
Mouse	Neural retina	Microarray, Agilent Technologies, Whole Mouse Genome Oligo Microarrays 4x44K	43,379 gene features	3–20 mo	(Chen et al., 2010a)
Mouse - senescence	Neural retina	Microarray, Affymetrix MG_U74Av2 GeneChip	36,000 full-length genes and EST clusters	3–21 mo	(Carter et al., 2005)
Mouse	Rod photoreceptors (flow-sorted Nrlp-EGFP retina)	Microarray, Affymetrix Mouse Exon 1.0ST Array GeneChip	>28,000 coding, >7,000 non-coding transcripts	1.5–12 mo	(Parapuram et al., 2010)
Mouse	Microglia (flow-sorted, CD11b-immunopositive cells)	Microarray, Affymetrix Mouse Exon 1.0ST Array GeneChip	>28,000 coding, >7,000 non-coding transcripts	3–24 mo	(Ma et al., 2013)
Rat - senescence	Neural retina	RNA-Seq, Illumina GA IIx	15,300 identified transcriptsa	3–18 mo	(Kozhevniko va et al., 2013)

^aApproximate number of transcripts detected in at least three samples with FPKM>1

Some of the variability reported by the human studies can be overcome by molecular studies of aging in mouse. Comparison of two senescence-accelerated mouse prone strains (SAMP8 and SAMP10) to the senescence-accelerated mouse resistant strain (SAMR1) and the commonly inbred C57Bl/6J strain using Affymetrix™ oligonucleotide microarrays (Table 3) has highlighted the strong genetic background effect on the transcriptional response to aging (Carter et al., 2005). Expression of genes involved in inflammation and response to injury/stress is higher in aging mouse retina (Chen et al., 2010a), in concordance with the studies on human aging and with SAMR mice.

A limitation of gene expression data in the aging retina is that it represents average changes in six neuronal and two glial cell types, which contribute differently to tissue homeostasis. While these data may provide insights into the behavior of the system at the tissue level, expression profiling of the aging mouse rod photoreceptors (Parapuram et al., 2010) and of microglia (Ma et al., 2013) has expanded our understanding at the cellular level. Notably, genes associated with inflammatory signaling and stress response pathways constitute predominant hubs in network analysis of photoreceptor aging data (Parapuram et al., 2010). Furthermore, expression of genes involved in oxidative phosphorylation is reduced in aging photoreceptors (J. Barb, N. Gotoh, J-W Kim, T. Cogliati and A. Swaroop, manuscript in preparation), highlighting an age-related pathway that was not detected by whole retina studies but is a common feature of aging (de Magalhaes et al., 2009). Finally, angiogenesis and retinoid/lipid metabolism (regulated through signaling from the retinoic acid receptor) have emerged as unique pathways to rod photoreceptor aging (Parapuram et al., 2010). Focusing on photoreceptors as a single cell type, the question is whether gene expression changes observed in aging are intrinsic or an adaptive response to modification of retinal microenvironment, and whether these changes are protective, permissive or causative of the functional decline (Parapuram et al., 2010). Transcriptome analysis of isolated aging microglia shows changes in genes involved in cell metabolism, shape and motility, and in neurotrophic signaling, which could compromise the supporting role of microglia in the retina (Ma et al., 2013). Furthermore, expression changes in genes associated with microglia activation and inflammatory signaling suggest a contribution to the general retinal neuro-inflammatory process (von Bernhardi et al., 2010; Xie et al., 2003). As in case of rod photoreceptors, expression changes also affect genes involved with lipid metabolism and angiogenesis. The latter underscores the complexity of angiogenic events in AMD, in which signaling from multiple cell types, including RPE, photoreceptors and microglia, may converge to determine pathological changes, further demonstrating the importance of approaching retinal aging from a systems biology perspective.

From existing expression data, it is reasonable to assume that aging involves small changes in several genes that contribute to common pathways rather than major changes in individual genes. It becomes therefore critical to detect even small expression shifts with good statistical confidence. RNA-seq using NGS, with much higher reproducibility and greater dynamic range compared to microarray, is the ideal platform for more in depth molecular studies of retinal aging. RNA-seq has been employed in expression studies of human and mouse retina, however, not in relation to aging (Brooks et al., 2012; Farkas et al., 2013; Gamsiz et al., 2012; Li et al., 2014). One RNA-seq study so far published in the aging retina investigates the senescence-accelerated OXYS rats compared to Wistar controls (Kozhevnikova et al., 2013). Because of the profound pathological phenotype observed in the OXYS rats, it is difficult to distinguish aging from overt strain-dependent disease. However, differential expression of genes between 3 and 18 month old Wistar rats shows significant down-regulation of genes associated with extracellular matrix, response to nutrient levels and aminoacyl-tRNA biosynthesis, together with up-regulation of genes encoding negative regulators of transcription and components of synaptic transmission pathways. Immune response genes appear in both up-regulated and down-regulated gene ontologies, suggesting potentially impaired immune function in the aging rat retina, similar to the findings by microarrays in mouse and human.

4.2. Limitations and potential of aging studies in the retina

The scenario emerging from gene expression analyses of the aging retina is one of a tissue that is metabolically slowing down, activates stress and signaling pathways, and presents a complex immune and inflammatory activity. Because of the nature of aging, data collection from the same individual over time would be ideal but it is not a viable option for the retina. For human studies, a random selection of subjects from different age groups (post-mortem donors) suffers from large individual variations that can only be overcome if enough replicates can be collected and if experimental variables are minimally altered. This underscores the importance of strict collection and transportation protocols that minimize the time from death to enucleation and processing. Eliminating technical variability becomes even more critical if one wants to take an unbiased approach and use principal component analysis (PCA) to partition on the first component of the samples between young and old.

RNA-seq has revealed the expression of novel transcripts and genes in the retina (Farkas et al., 2013; Gamsiz et al., 2012; Li et al., 2014) suggesting the need to extend this analysis to aging individual cell types in mouse and, where possible, in human. In depth analysis of RNA-seq data will allow the identification and study of emerging modulators of aging such as regulatory ncRNAs described in section 3.2.5 (reviewed in (Montano and Long, 2011)). Reproducibility of RNA-seq will also facilitate compilation of different studies to increase sample number and statistical significance. Finally, comprehensive cross-comparisons of existing databases (Table 2) may provide deeper insight than that achieved by individual analyses.

5. Systems biology of retinal degeneration

Retinal degenerative diseases (RDDs) are genetically heterogeneous, ranging from monogenic forms such as retinitis pigmentosa to complex disorders such as AMD. Since the first linkage analysis that identified a genomic locus associated with X-linked retinitis pigmentosa (Bhattacharya et al., 1984), we have witnessed a remarkable progress in elucidating the genetic basis of retinal diseases ((Swaroop and Sieving, 2013); RetNet, http://sph.uth.edu/retnet/). Photoreceptor cell death remains central to many RDDs (Wright et al., 2010). Functional analyses of the identified disease genes through cell-based and animal models have enhanced our understanding of their normal function in ocular development and homeostasis. However, elucidation of underlying mechanism(s) has not kept pace with disease gene identification.

The advent of NGS is now transforming, at an unprecedented rate, the way we study genetic basis of human disease. Holistic approaches to RDDs have become possible through a combination of NGS-based strategies including disease-associated variant identification, transcriptome analysis, and high throughput search for biomarkers (Figure 6). We can learn about regulatory networks, disease-relevant pathways and their crosstalk, which can be invaluable for translational and personalized medicine. In this section, we discuss the following areas, which will provide the necessary impetus:

Figure 6.

Integration of multi-level data sets for system-level understanding. Simultaneous analysis of genomics, transcriptomics and proteomics data can provide a comprehensive multidimensional view of the system and can help identify pathways and networks that are specific to development and disease. This can lead to better understanding of the biology as well as to identification of disease associated nodes, which can be potential drug targets (e.g., hub depicted as a purple node in the example network).

1. Whole exome and whole genome sequencing (WES and WGS) of patients with monogenic RDDs for better genotype-phenotype correlation and search for genetic modifiers

2. Fine mapping and functional dissection of associated loci

3. Genetics of RDDs beyond genetic variants

5.1. NGS applications for disease variant/mutation discovery

NGS methods have expedited the process of cataloging human genetic variation. While precise correlation of variants associated with disease phenotypes is still challenging, NGS approaches, such as WES, WGS and targeted sequencing, have become preferred methods for disease gene discovery (Figure 7A). A number of novel genes have been identified through WES in small Mendelian RDD families and in sporadic cases that were not suitable for customary linkage studies (Ratnapriya and Swaroop, 2013).

Figure 7.

General recommendations for genetic analysis of monogenic and complex diseases using next-generation approaches. A. Families with a monogenic (i.e., Mendelian) disease, even those with a small number of individuals, can be analyzed by whole exome or genome sequencing. In case of identification of mutation(s) in a novel gene, additional validation is required, including genetic analysis of more families/patients as well as functional analysis using animal models to understand the role in disease causation. For complex diseases, genome wide association study is performed on a large number of samples including patients and population-matched controls, to identify common susceptibility variants associated with the disease. Targeted sequencing, whole exome or genome sequencing can be applied to identify the causal variant at the susceptibility locus. Finally, multiple molecular genetic and cellular assays as well as studies in animal models are required to identify functional variants underlying trait association. WES, whole exome sequencing; WGS, whole genome sequencing, GWAS, genome wide association study. B. Schematic representation of follow-up on a GWAS hit to identify the causal gene/variant. Causal variants could be common regulatory variants, rare coding variants or copy number variants. Targeted sequencing around the associated locus can help identify the causal variant. Generally, more than one candidate gene is found at an associated locus, and eQTL analysis in disease relevant tissues can help identify the specific genotypes affecting gene expression within the causal gene. Additional functional annotation, including histone modifications, transcription factor binding sites and DNase I hypersensitivity sites can also help in identifying the causal gene. Ultimately, a high throughput functional assay in animal and/or cell-based models is highly recommended to understand the role of the gene in disease pathophysiology. cM/Mb, centimorgan (cM) per megabase pair (Mb); eQTL, expression quantitative trait locus, TF, transcription factor.

Progress in dissecting genetic architecture of complex RDDs has been relatively slow since these multifactorial diseases are more common in the population and often do not exhibit familial segregation. Furthermore, non-genetic factors, such as environment, can contribute to disease onset and severity. We have come a long way since the initial association of CFH variant with AMD (Edwards et al., 2005; Hageman et al., 2005; Haines et al., 2005; Jakobsdottir et al., 2005; Zareparsi et al., 2005a). A total of 19 AMD susceptibility loci have been detected so far, including 7 new loci identified in the meta-analysis of 17,000 cases and 60,000 controls (Fritsche et al., 2013). These discoveries have provided clues about the underlying disease processes and have implicated complement cascade, high-density lipoprotein cholesterol and extracellular matrix pathways in AMD pathogenesis (Fritsche et al., 2014; Priya et al., 2012). However, associated locus variants are usually not causal and often reside outside the gene region. Thus, more than one gene at each associated locus can be potential disease gene. The nature of causal variants/genes at these loci remains debatable and needs further explorations as the causal variants could be an associated variant, another common variant, a rare variant or even a structural variant (Figure 7B).

Identification of rare variant(s) in a specific gene at the associated locus can provide functional clues. For AMD, a rare penetrant mutation, R1210C was first identified at the CFH locus (Raychaudhuri et al., 2011). Subsequently, rare variants in several complement pathway genes have been associated with AMD through targeted and whole genome sequencing (Helgason et al., 2013; Seddon et al., 2013; van de Ven et al., 2013; Yu et al., 2014; Zhan et al., 2013). However, the search for rare variants in RDDs has been limited to known loci and sequencing of the candidate association regions. Whole exome and whole genome sequencing studies in large patient cohorts may facilitate cataloging of rare and common causal variants, identification of novel disease genes and assist in biological investigations. However, rare variant association study design requires a large sample size since, as predicted, not many individuals will harbor a rare variant. Genetic refinement of associated signals with improved imputation methods can further help in identifying potential causal variants. Recently, a new SNP array (“exome-chip”) was developed to explore the role of common as well as rare variants in complex traits (Huyghe et al., 2013). Sixteen additional AMD loci have now been identified by exome-chip analysis and imputation of variants in approximately 26,000 cases and 22,000 controls (Cipriani and International AMD Genomics Consortium., 2014; Fritsche and International AMD Genomics Consortium., 2014)(International AMD Genomics Consortium, manuscript in preparation). Notably, only a small number of rare variants were identified in this study, and whole genome sequencing might be necessary for evaluating the role of rare variants in AMD.

5.2. Less-recognized complexities associated with genetic variations and human retinal diseases

Human genome harbors a large number of genetic variations, and NGS has made it possible to systematically catalog genome-wide sequence variants. However, methods for distinguishing a neutral variant from the causal one are not yet fully established. The current criteria include nature of change, allele frequency and nucleotide conservation across species. However, these criteria do not seem sufficient as we are learning more about the architecture of variants in the human genome. A single human genome can harbor thousands of variants (10,000–11,000 non-synonymous and 10,000–12,000 synonymous variations)(Genomes Project et al., 2012), with as many as 300 loss-of-function alleles having little or no clinical consequence (Kukurba et al., 2014; MacArthur et al., 2012). Incorrect assignment of variants can have immediate implication for diagnosis, therapeutic advice and research direction. The recommendations for identifying causal genes/variants associated with specific phenotypes through NGS studies have been reviewed recently (Chakravarti et al., 2013; MacArthur et al., 2014; Samocha et al., 2014). Here, we revisit some of the issues that have implications for retinal genetics.

5.2.1. Are mutations rare genetic variations?

Mutations causing monogenic diseases are believed to be rare, and often a blanket minor allele frequency cut-off is applied to filter out disease variants. Such an approach can miss cases where disease variants are common and/or likely modify the disease phenotype (Ebermann et al., 2010; Khanna et al., 2009; Louie et al., 2010; Venturini et al., 2012). Notably, carrier frequency of the null mutations responsible for RDDs is unusually high in general population (Nishiguchi and Rivolta, 2012). A vast majority of known genetic variants associated with complex diseases (such as AMD) are common and yet some may have functional relevance to pathology (such as Y402H in CFH). As investigations of genetic variation underlying complex traits continue to mature, we are seeing examples of rare and common regulatory as well as copy number variants associated with disease. Thus, no general analysis pipeline should be applied to identify the variants of interest, and one must customize based on genetic architecture of the disease under study.

5.2.2. Do disease-causing mutations have to be conserved?

Computational prediction tools incorporate evolutionary conservation across species as a criterion when assigning causality to variants. While a majority of disease-related variants follow this pattern, less conserved residues, silent variants or changes in the non-coding region have also been implicated in disease (Cartegni et al., 2002; Sauna and Kimchi-Sarfaty, 2011). Furthermore, the effect of a variant is likely to be influenced by size, polarity, codon-usage, neighboring amino-acid residues and location with respect to the overall protein conformation. Prediction programs fail to account for many of these important parameters. Thus, filtering out any “potentially interesting” variant based on prediction criteria should not be a common practice. Development of more sophisticated analytical tools is therefore necessary for better predictive power.

5.2.3. Are monogenic diseases caused by defects in only one gene?

Until recently, a majority of disease genes for monogenic traits have been identified through traditional linkage approach, where loci that segregate with the disease phenotype are screened for mutation(s) in candidate genes. Once a segregating disease mutation is identified, the researchers usually no longer search for mutations/variations. However, extensive genetic and even phenotypic heterogeneity observed in RDDs has been an obstacle to diagnosis and counseling. Different RDDs also share overlapping clinical findings; thus, mutations in many different genes may result in similar phenotypes and/or mutations in one gene can lead to distinct phenotypes. In Mendelian diseases, inheritance of a genetic defect implies that one will get the disease. However, many individuals carrying a mutation may show no clinical phenotype (incomplete penetrance) or variable expressivity. As every individual carries thousands of unique variants, one can envisage compensatory variations in interacting genes (epistasis) that can alter the clinical phenotype(s) (Lehner et al., 2006). Thus, in addition to the “primary” gene defect, many patients may harbor alleles that can modify the disease phenotype (modifier alleles). For example, a common allele in the RPGRIP1L gene, associated with Joubert and Meckel syndromes (Arts et al., 2007; Delous et al., 2007), is shown to contribute to retinal degeneration phenotype in individuals with ciliopathies caused by mutations in other genes (Khanna et al., 2009). In other instances, mutations in more than one gene might be necessary to observe a disease phenotype (e.g., “digenic” RP caused by ROM1 and RDS alleles) (Kajiwara et al., 1994). NGS offers an unbiased approach for unveiling genetic modifiers of retinal degeneration across the genome and can help in refining the genetic definition of phenotypic variations in RDD patients.

The spectrum of potentially disease-causing variants makes the correlation of the variant with the phenotype challenging. Given that causative variants could be rare, common, regulatory, intronic, silent or compensatory, establishing the role of a gene or a variant in disease pathogenesis is not straightforward. Genes that are identified through genetic studies should be referred as “candidates” until their association with the disease phenotype is validated by additional evidence including functional data. Further support can emerge from enrichment of genetic variants (gene-level evidence) in patients/families compared with controls. Disruption of the gene in a model system can, in some instances, reproduce the relevant human phenotype. In addition, one should be able to rescue the phenotype with the wild type allele but not with mutant allele(s), thereby mimicking the “Koch’s postulates” (Chakravarti et al., 2013). While mouse (or zebrafish) models do not always recapitulate the human phenotype, such studies could provide disease mechanisms and opportunities for evaluation of treatments (Veleri et al., 2015). Alternative models using patient-derived human induced pluripotent stem (iPS) cells also appear promising.

5.2.4. Missing heritability of complex traits

In addition to common and rare genetic variants, other components including copy number variation, epigenetic signature and gene-environment interaction contribute to complex disease phenotype. While many studies have looked at copy number variations in candidate genes with limited success and replication (Cantsilieris and White, 2013; Liu et al., 2012), genome-wide surveys have not been useful in identifying such variations associated with major risk of AMD (Fritsche et al., 2013; Meyer et al., 2011; Neale et al., 2010; Sobrin et al., 2012).

Epigenetics has gained immense interest because of its role in human diseases such as cancer (Feinberg et al., 2006). Disease-related changes include stochastic modifications in epigenetic marks as well as local gene-specific alterations (Pujadas and Feinberg, 2012). Beyond cancer, epigenetic changes have been widely associated with human diseases involving chromosomal abnormalities and intellectual disability (Brookes and Shi, 2014). Retinal diseases have also been explored for possible involvement of epigenetic alterations, and genes such as IL17RC (Wei et al., 2012a) and Apo J (Suuronen et al., 2007) have been suggested to contribute to the pathogenesis of AMD. However, such studies require further validation (Oliver et al., 2013a). More extensive analysis of the epigenome must be performed using normal or disease retina (or target cell types).

Complex diseases, such as AMD, exhibit tremendous phenotypic heterogeneity, and multiple classification and grading systems have been put forward (Bird et al., 1995; Ferris et al., 2005; Ferris et al., 2013). However, much of the genetic analysis of AMD has been performed on cohorts with advanced stage of disease, with phenotypes broadly classified as geographic atrophy and choroidal neovascularization. While an ARMS2 risk variant is reportedly more strongly associated to neovascularization compared to geographic atrophy (Chen et al., 2010b; Sobrin et al., 2011), it has been difficult to identify genes and pathways associated with distinct clinical subtypes (Fritsche et al., 2014). The recently completed exome-chip analysis of the large case-control cohort may provide additional insights on subtype specific genetic variants. Whole genome sequencing of 6,000 AMD cases and controls is currently in progress (Kwong, 2014) and should be valuable for deciphering susceptibility and protective alleles for AMD subtypes.

5.3. Integration of transcriptome with genetic data

Though heritable mutations are present in every cell of an individual, disease manifestation is often tissue or cell specific because of differences in spatio-temporal expression or function of the responsible gene(s). Thus, it would be important to correlate genetic variations to gene expression in the retina (or photoreceptors) to obtain insights into the regulatory network/pathways perturbed in RDDs. NGS methods allow high throughput analysis of genotypes as well as provide an opportunity to sequence all species of RNA, including the splice variants and relatively less studied miRNAs and lncRNAs.

A majority of variants associated with complex diseases often reside outside the coding region of a gene or even in non-transcribed regions. It is assumed that such variants would have subtle effects on gene expression or function. The expression quantitative trait locus (eQTL) analysis combines gene expression data with genotype information and can assist in identifying risk alleles (Cheung et al., 2003; Lappalainen et al., 2013). Expression profiling of disease-relevant tissue or cells in case-control samples may reveal functional variants with biological relevance to the phenotype. eQTLs are primarily cis-acting, suggesting a common mechanism at many associated loci (Edwards et al., 2013). Identification of target genes modulated by cis-regulatory risk variants will allow better understanding of disease network(s) and/or mechanism(s) through which they act. In fact, ENCODE project has demonstrated that a large proportion of associated GWAS SNPs appear to map to functional regulatory elements (Schaub et al., 2012). As noted earlier, gene expression patterns are cell type- and tissue-specific, and eQTL analyses are more useful in cells/tissue types relevant to the human disease of interest. The current databases do not include eQTL information on the retina or other ocular cell types owing to limitations in obtaining human tissues.

GWAS studies have yielded a large number of new genetic loci associated with AMD susceptibility. However, several genes that seem to have biological relevance to AMD pathology have not been validated even in large GWAS meta-analyses, in part because of false positive results or phenotypic overlap together with etiological heterogeneity; these include ABCA4 (Allikmets, 2000), TLR4 (Zareparsi et al., 2005b), TLR3 (Kleinman et al., 2008; Yang et al., 2008), CX3CR1 (Tuo et al., 2004), DICER (Kaneko et al., 2011), and AHR (Choudhary et al., 2014; Kim et al., 2014). Another explanation could be that the effect size of these variants is small, and much larger sample size is needed to achieve genome-wide significance. We should also note that the biological significance of many genes/variants identified by GWAS is not yet clear. The genetic variants identified through different approaches may have subtle yet meaningful influence on specific pathways and networks that when disrupted lead to AMD pathology. We recently suggested that cumulative impact of few strong or multiple weaker alleles must reach a threshold of disruption for producing a clinically identifiable phenotype (Fritsche et al., 2014). An integrated approach of spatial and temporal mapping of disease genes and variants onto the expression data generated from disease-relevant tissues would be necessary to generate disease-specific networks. Distinct candidates and pathways may converge in networks of co-expressed genes in disease-relevant cell/tissue at a specific time. Such disease networks can help in identifying novel regulatory nodes and therapeutic targets.

5.4. Approaches for elucidating disease networks

Identifying disease-associated genetic variants is an important piece of the puzzle in understanding pathogenic mechanism(s). Rather than examining one gene (and its function) at a given time, our goal must be to understand the “disease” itself and elucidate complex interplay among multiple factors (genes, epigenetic changes, and interacting environment). Initially, one can investigate all genes associated with a retinal disease phenotype (such as Leber Congenital Amaurosis or AMD) to deduce relevant biological pathways and then integrate genes and pathways into networks. For examples, 19 AMD loci, published so far, have implicated complement, extracellular matrix and lipid metabolism pathways (Fritsche et al., 2014; Priya et al., 2012). Other studies on GWAS loci have hinted towards localization of cis-regulatory variation at enhancer elements (Pennacchio et al., 2013). Integrating high throughput functional annotations, such as histone marks, transcription factor binding sites and DNase I hypersensitive sites (DHSs), relevant to AMD cell types can help in prioritization of non-coding variants at associated loci for subsequent functional evaluation. Combining this information with the genetic data can provide novel insights into disease mechanisms; for example, exome sequencing in small families can identify large number of rare and novel potential disease variants, and ChIP-seq data can allow filtering of genetic variants, as shown in case of MAK (Ozgul et al., 2011) where CRX Chip-seq data was used to rank the candidate genes. Combined systems level analyses of multiple datasets have already begun to reveal novel insights in disease mechanisms and targets for drug discovery (Cohen et al., 2006; Dadu and Ballantyne, 2014; Sahebkar, 2014).

Rare variants or common loci with small effect often require large sample size to reach genome-wide statistical significance. However, there could be alternate ways to extract meaningful information about the disease process. For example, exome sequencing of families where the disease is clustered can identify rare segregating variants that can impact susceptibility. Enrichment analysis of candidate variants can either link to existing disease-associated pathways or reveal novel insights into the disease process (Ratnapriya, 2014). Alternatively, pathogenic mechanisms associated with a specific phenotype within the complex disease (such as AMD) can also be derived from delineating the biology of Mendelian RDDs that include overlapping clinical features (such as Stargardt disease) (Ratnapriya et al., 2014).

6. Next generation diagnosis and drug development

The promise of next generation genomics to medicine is exciting and unprecedented; yet, we have a long way before incorporating NGS into routine clinical practice. Meanwhile, NGS has created new opportunities for faster genetic diagnosis that can help in improved disease management. Given the allelic heterogeneity of RDDs, the current sequencing methods are not very efficient. Decreasing cost of sequencing and approval from U.S. Food and Drug Administration allowing NGS methods for clinical diagnosis constitute steps in the rapid realization of personalized medicine. NGS-based screening methods are now being utilized for RDDs as well (Booij et al., 2011; Bowne et al., 2011; Neveling et al., 2012; Song et al., 2011b). The National Ophthalmic Disease Genotyping and Phenotyping Network (eyeGene), a CLIA certified community resource designed to bring together patients, providers, molecular testing laboratories and researchers, is incorporating NGS tools for investigating disease causing mutations (Blain et al., 2013).

Another major application of NGS is in the area of pharmacogenomics. Patients with the same disease often respond differently to drug efficacy and safety because of unique genetic variations they inherit. Identification of large number of susceptibility alleles has propelled the field of pharmacogenomics to a potential future in personalized medicine. In AMD, the effect of GWAS susceptibility loci, mainly CFH and ARMS2 variants on treatment outcome of Avastin (bevacizumab) and Lucentis (ranibizumab), has received a lot of attention, though such studies have not yielded consistent results (Hagstrom et al., 2013; Kanoff and Miller, 2013). Small sample size, incorporation of only few disease markers, and short term-follow up represent some of the challenges. More comprehensive, large-scale prospective studies are necessary to establish definitive associations between genotypes and drug response/efficacy.

Relevance of NGS datasets to development of translational strategies has been limited so far because of inadequate understanding of underlying biology of the disease. GWAS hits are rarely direct drug targets (Teslovich et al., 2010). Limited predictive value of the disease models and inability to accurately estimate the efficacy, toxicity and long-term side effects have made discovery of new drugs very expensive and time consuming (Plenge et al., 2013). NGS technology has remarkable promise for genomic medicine, impacting both predictive and preventive medicine. Few studies have already shown novel means for drug discovery using genetic data and bioinformatics tools, such as functional annotation, cis-acting eQTLs and pathway analyses (Okada et al., 2014). We need to better define and map interactome networks with its components (nodes) and interactions (edges) (Vidal et al., 2011). These biological networks can be integrated with known drug and disease information to build drug-target networks (Yildirim et al., 2007). As NGS technology becomes widely available, pharmacological interventions will have personalized focus on disease treatment. Identification of clinically important variants, together with transcriptome and epigenetic profiling, should permit dissection of human disease genome that is amenable to pharmacological interventions.

7. Challenges and further considerations

Here, we have examined new developments in photoreceptor biology stemming from NGS technologies. Despite much progress, challenges remain; these include development of effective bioinformatics tools for multi-layer data integration, standardization of procedures to facilitate collaborative works, data sharing within the research community and application of novel findings from system-level analysis to develop new hypothesis.

7.1. Multi-layer data integration

Most high throughput data generation and analysis have so far been conducted at one or two layers of biological information. For instance, transcriptome analysis has been the most dominant form of genome-wide information, and other types of data such as transcription factor targetome or epigenome profiling have been individually compared to transcriptome data but not to one another. With increasing amount of studies examining various layers of biological pathways, including regulatory RNAs, epigenome and proteome, the resulting higher order integrative analysis is expected to further refine the current knowledge of molecular networks. To this end, applying robust computational analysis of the data sets is critical. The large-scale ENCODE (encyclopedia of DNA elements) consortium probably provides the best model. Upon completion of the human genome project, ENCODE was launched to investigate functional elements of human and mouse genome (http://encodeproject.org). By employing more than two-dozen assays, a large number of data sets have been generated from 147 distinct cell types or tissues (ENCODE Project Consortium, 2012). Several powerful software tools were developed with a goal of effective data integration across heterogeneous data sets (http://encodeproject.org/ENCODE/encodeTools.html). Additional open source bioinformatics tools are freely available.

7.2. Standardization of procedures and quality control

Given the significant amount of expertise, workloads, and funding required for high throughput data collection, the practice of comprehensive and multi-dimensional data analysis demands collaboration among various disciplines. Independently generated data sets will often have to be compiled together for integrative analysis. The vision research community, therefore, needs standardized procedures for experimentation and data analysis. Detailed instructions for high throughput assay protocols and quality control are also provided as a part of the ENCODE project (ENCODE Project Consortium, 2011), and all the upcoming data generation should refer to those. The retina, however, consists of highly specialized sets of neurons and supporting cells; thus, additional guidelines specific to the vision field must be implemented. Directions and relevant resources for retinal sample preparation should be established. As retina is rich with tissue-specific transcript isoforms, genome annotations tailored for the retina should be established for standardized data analysis.

7.3. Data sharing

A fundamental consideration for systems approaches is to facilitate data sharing. High throughput studies are unique in that the generated data can be re-analyzed to discover new patterns and to test new hypotheses distinct from those posed in the original study. Therefore, after the initial analysis and publication, high throughput data sets should be readily accessible for any research group to view, verify and reanalyze. More than 200 retina-related data series can be retrieved from public databases, such as Gene Expression Omnibus (GEO), which archive genome-wide information generated by microarray and NGS technologies. Microarray data deposited to GEO follows, for the most part, the guidelines from Minimum Information About a Microarray Experiment (MIAME) (Brazma et al., 2001) put forward by the microarray gene expression data (MGED) society. However, standardized directions specific for NGS data have not been implemented yet, nor the strict measures for quality control are in place. We have recently undertaken the establishment of a web-based database for retinal NGS data repository with the following goals:

1. Implementation of uniform guidelines for sample collection, assay procedures and data analysis

2. Ensure the quality of the deposited data and encourage data sharing in a timely manner

3. Consolidation of high throughput data sets related to retina (or other ocular tissues) commonly affected in retinal neurodegenerative diseases

4. Provide structured and comprehensive annotation of each data set

Our archive of high throughput data sets should provide a critical resource to facilitate higher-order integrated data analysis and/or novel discovery through re-analysis.

7.4. From in silico analysis to the bench

The observations made from system-wide analyses, in most instances, are descriptive and hypothesis-generating rather than hypothesis-testing. Wet bench research is needed to verify and further develop the knowledge obtained from systems level computational analyses. Networks and/or patterns implicated in retinal development, aging or disease must be tested by traditional in vivo and in vitro model systems. The three-dimensional retinal cultures derived from human or mouse stem cells (Eiraku et al., 2011; Nakano et al., 2012) provide a useful model for human retina in the context of evaluating the role of microenvironment and genetic pathways. Patient-specific induced pluripotent stem cells provide hope for the development of personalized medicine.

8. The future of systems biology approaches: biology of the organism as a whole

In this review, we have primarily focused on GRNs that maintain photoreceptor function and on RDDs; nonetheless, we do realize that a comprehensive systems biology strategy would also incorporate population-wide genetic variations, protein interaction networks and metabolomics. In general, we believe that reproducible topologies in the networks with principal regulators would constitute the hubs of sub-networks. It is not difficult to imagine that perturbation of such primary hubs (e.g., key regulators, rate-limiting substrates and enzymes) likely imposes more pronounced impact on photoreceptor health than the defect in locally confined nodes with fewer interactions (e.g., those with limited physiological reach) (Figure 8). Targeting “key” hubs would therefore represent an attractive strategy for drug discovery.

Figure 8.

Model of multi-dimensional network of photoreceptors. Networks consist of nodes and edges, representing cellular components (genes, transcripts, proteins, etc.) and their interactions, respectively. The nodes with the greatest number of interactions constitute primary hubs. Secondary hubs themselves are the interacting partners of primary hubs and simultaneously connect with many other nodes, thereby partially mediating the function of the primary hubs. Local sub-networks are also prevalent. Perturbation of primary or secondary hubs of the network likely poses detrimental impact to the system, while perturbation of a more locally confined network or a node with few interacting partners may not exert substantial effect on the system integrity. Identification of commonly affected networks in diverse retinal diseases will allow the better drug design.

One of the major bottlenecks in therapeutic development has been the overwhelming number of disease-associated nodes (i.e., disease genes) due to heterogeneity of RDDs. As suggested earlier (Swaroop et al., 2010; Yu et al., 2004b), it is conceivable that distinct mutations resulting in a broad spectrum of retinal diseases converge onto a manageable number of “key” hubs (with overlapping disease-associated pathways/networks) prior to photoreceptor death. Thus, one may not need to target individual disease causing genes to rescue photoreceptor integrity. Identification of common retinal disease-related networks and associated hubs may now be within reach for discovering better drug targets.

We have come a long way in defining the biology of individual genes. Nonetheless, we now must provide an appropriate cellular context to individual genes and their functional products. We have already started to examine genetic, transcriptional and epigenetic architectures and await integrated analysis to incorporate proteome and metabolome data. Networks constructed from multi-dimensional data analysis will allow us to simulate functional and physiological outcome of photoreceptors in response to alteration in the microenvironment and variations in the genetic makeup. Human genomes are highly polymorphic, and even monozygotic twins may carry genetic differences (Bruder et al., 2008). Notably, individuals are estimated to carry over 2,500 functionally relevant genetic variations (Genomes Project et al., 2012). Such variability is what makes each of us unique and would be expected to influence the network structure. Genetic polymorphism, together with aging and environmental stress, may have profound impact on susceptibility to diseases and drug response. The holistic approaches thus have the key to accurate diagnosis and personalized medicine.

In this review, we have primarily discussed system-level analysis of the whole cell (rather than focusing on a gene or a pathway in isolation). We recognize that cells in higher organisms do not function independently but constitute higher order structures that are morphologically and functionally distinct (e.g., retina, lens, cerebellum), which in turn are in constant communication with structures farther away (e.g., via systemic circulation or neuronal connectivity). Thus, aberrant function of a gene in a specific cell may have wide-ranging impact on other tissues and the whole organism. A true system-level analysis will eventually incorporate both limited and long-range effects on GRNs and cell function. The retinal research has an exciting future, as we are closer to better predictive and treatment strategies for every individual.

List of abbreviations

AMD

Age-related macular degeneration

ChIP

Chromatin immunoprecipitation

ChIP-seq

Chromatin immunoprecipitation-sequencing

CLIA

Clinical laboratory improvement amendments

eQTL

Expression quantitative trait locus

GA

Geographic atrophy

GRN

Gene regulatory network

GWAS

Genome-wide association study

H3K4me2(3)

Histone H3 lysine 4 di(tri)methylation

H3K27me3

Histone H3 lysine 27 trimethylation

M cone

Medium wavelength sensitive cone

MIAME

Minimum information about a microarray experiment

NGS

Next generation sequencing

ONL

Outer nuclear layer

PCA

Principal component analysis

RDD

Retinal degenerative disease

RNA-seq

RNA-sequencing

RPE

Retinal pigment epithelium

S cone

Short wave length sensitive cone

SNP

Single nucleotide polymorphism

WES

Whole exome sequencing

WGS

Whole genome sequencing