|Home | About | Journals | Submit | Contact Us | Français|
The retina is a highly ordered tissue whose outermost layers are formed by subcellular compartments of photoreceptors generating light-evoked electrical responses. We studied protein distributions among individual photoreceptor compartments by separating the entire photoreceptor layer of a flat-mounted frozen retina into a series of thin tangential cryosections and analyzing protein compositions of each section by label-free quantitative mass spectrometry. Based on 5038 confidently identified peptides assigned to 896 protein database entries, we generated a quantitative proteomic database (a “map”) correlating the distribution profiles of identified proteins with the profiles of marker proteins representing individual compartments of photoreceptors and adjacent cells. We evaluated the applicability of several common peptide-to-protein quantification algorithms in the context of our database and found that the highest reliability was obtained by summing the intensities of all peptides representing a given protein, using at least the 5–6 most intense peptides when applicable. We used this proteome map to investigate the distribution of glycolytic enzymes, critical in fulfilling the extremely high metabolic demands of photoreceptor cells, and obtained two major findings. First, unlike the majority of neurons rich in hexokinase I, but similar to other highly metabolically active cells, photoreceptors express hexokinase II. Hexokinase II has a very high catalytic activity when associated with mitochondria, and indeed we found it colocalized with mitochondria in photoreceptors. Second, photoreceptors contain very little triosephosphate isomerase, an enzyme converting dihydroxyacetone phosphate into glyceraldehyde-3-phosphate. This may serve as a functional adaptation because dihydroxyacetone phosphate is a major precursor in phospholipid biosynthesis, a process particularly active in photoreceptors because of the constant renewal of their light-sensitive membrane disc stacks. Overall, our approach for proteomic profiling of very small tissue amounts at a resolution of a few microns, combining cryosectioning and liquid chromatography-tandem MS, can be applied for quantitative investigation of proteomes where spatial resolution is paramount.
The vertebrate retina is a layered tissue containing several types of neurons and supporting cells (1). The outermost half of the retina is formed by rod and cone photoreceptors, which produce electrical responses upon capturing photons. Photoreceptors are polarized neurons in which specific functions are carried out in individual highly specialized subcellular compartments. For example, the processes of light capturing and visual signaling are confined to the distal outer segment, whereas information transfer to the secondary retina neurons takes place at the synaptic terminal located at the opposite side of the cell. In this study, we analyzed the patterns of subcellular protein distribution in photoreceptors by label-free quantitative mass spectrometry.
Proteomic analysis of layered or intricately structured tissues requires a difficult analytical combination of reproducible sample preparation with high spatial resolution, significant depth of proteome coverage, and accurate quantification. These challenges have been addressed to varying degrees by laser-capture microdissection along with MS analysis and matrix assisted laser desorption ionization (MALDI)1 imaging. Laser capture microdissection is typically performed following immunostaining for specific markers to define the regions of interest (2–7); although this approach has very high resolution (microns), the amount of material available for MS analysis is exceedingly small because of the limited thickness of tissue sections. MALDI imaging has the advantage of high speed and ease of sample preparation, but suffers from lower spatial resolution (typically hundreds of microns), inherent difficulty in protein identification, and limited depth of proteome coverage (8–13).
Unbiased quantitative label-free proteomics, utilizing liquid chromatography-tandem MS (LC-MS/MS), is a rapidly emerging methodology allowing comparison of protein contents across multiple samples, when utilizing the accurate-mass and time-tag approach for alignment of peptides across the samples (14–20). This technique requires accurate-mass, high-resolution mass spectrometers for data collection and sophisticated software for data analysis. Many open-source and commercial options are now available to perform the latter task (21–24). The label-free approach also has the advantage of unlimited scalability with respect to number of samples or number of experimental variables, provided that methods for sample preparation are sufficiently quantitative and control for analysis-to-analysis variability is carefully considered.
Taking advantage of the highly ordered retina structure, we obtained thin serial tangential sections through the rat outer retina, which contained proteins derived from different parts of the photoreceptor cells. Because the rat retina is dominated by rod photoreceptors (outnumbering cones by >100-fold (25)), the vast majority of photoreceptor-derived proteins in these sections originated from rods. In the past, we combined this technique with Western blot detection of individual proteins in sections, which provided an alternative to immunohistochemistry to study longitudinal protein distributions in these cells (26). This method resolved several long-standing controversies regarding the subcellular localization of photoreceptor-specific proteins, which arose from conflicting results of immunohistochemical studies (26–28). We now extended this methodology to analyze distribution profiles of hundreds of proteins by combining serial sectioning of the retina with quantitative mass spectrometry. The results are summarized in a database (which we call a proteome “map”), correlating the relative abundances of all identified proteins in individual serial sections with those of marker proteins representing specific compartments of photoreceptors and neighboring cells. We used this map to assess the subcellular distribution of all glycolytic enzymes and uncovered novel patterns for the two hexokinase isoforms and triosephosphate isomerase, each likely reflecting the unusually high metabolic activity of photoreceptor cells.
Sixty-day-old pigmented Long-Evans rats (Rattus norvegicus) were purchased from Charles River Laboratories (Wilmington, MA). Animals were dark-adapted for at least 12 h and sacrificed under dim red light either immediately or following one hour exposure to bright light producing 15,000 lux on the cornea surface as described in (29).
Serial sectioning was performed as described in (26) with modifications described in (29) and (30). Briefly, eyes were enucleated from an anesthetized rat and dissected in ice-cold Ringer's solution under dim red light. A retina fragment (3 mm in diameter) was cut from the eyecup with a surgical trephine positioned right next to the optic disc (so that the optic nerve remained just outside the analyzed specimen), transferred onto PVDF membrane with the photoreceptor layer facing up, flat-mounted between two glass slides separated by plastic spacers (ca. 240 μm) and frozen on dry ice. Progressive tangential sections were then collected using a cryo-microtome following aligning the retina surface with the cutting plane of the microtome knife. The uneven edges of the retina specimen were trimmed and twelve 10-μm sections were collected.
A complete analysis was performed with two sets of retina sections, one obtained from a dark-adapted and another from a light-adapted rat referred as retinas #1 and #2, respectively, throughout the text. Each retina section was dissolved in 50 μl 0.5% v/v anionic acid liable surfactant II (Protea Biosciences, Morgantown, WV), followed by sonication and boiling for 5 min. Cysteine residues were reduced with 10 mm dithiothreitol and alkylated with 20 mm iodoacetamide. Proteins were digested with trypsin (15 ng/μl) overnight at 37 °C and anionic acid liable surfactant II was inactivated by adding trifluoroacetic acid to the final concentration of 1%. Samples were vacuum-dried and dissolved in 25 μl 0.1% trifluoroacetic acid and 2% acetonitrile. The typical amount of total protein in a section prepared as above was ~1.5 μg as determined by BCA assay in retina sections prepared identically to those analyzed by MS; no protein measurements were performed directly in samples used for MS in order to maximize the amount of material for analysis. We chose this gel-free protein digestion protocol in order to minimize any variability in peptide extraction and thus to improve quantitative aspects of the subsequent analysis.
Peptide digests obtained from each of the 24 sections (12 from each retina) where analyzed using a nanoAcquity UPLC system coupled to a Synapt HDMS mass spectrometer (Waters Corp, Milford, MA). Approximately 500 ng of peptide material in 8 μl was first trapped at 20 μl/min for 2 min in 99.9% water with 0.1% v/v formic acid on a 20 μm × 180 mm Symmetry C18 column. Separations were then performed on a 75 μm × 250 mm column with 1.7 μm C18 BEH particles (Waters) using a 120-min gradient of 5 to 40% acetonitrile with 0.1% formic acid at a flow rate of 0.3 μl/min and 45 °C column temperature. We conducted two data-independent (MSE) analyses of each retina section for simultaneous peptide quantification and identification. MSE runs of samples obtained from different sections were performed in random order, and used 0.9 s cycle time alternating between low collision energy (6 V) and high collision energy ramp (15 to 40 V). The quantitative analyses were followed by an additional, supplementary LC-MS/MS experiment in the data-dependent analysis (DDA) mode for each section, using a 0.9 s MS scan followed by MS/MS acquisition on the top three ions with charge greater than one. MS/MS scans for each ion used an isolation window of ~3 Da, a maximum of 4 s per precursor, and dynamic exclusion for 120 s within 1.2 Da.
For robust peak detection and label-free alignment of individual peptides across all 72 sample injections, we utilized the Rosetta Elucidator® v3.3 software (Rosetta Biosoftware, Inc., Seattle, WA) with PeakTeller algorithm, in a similar manner to several recent publications (24, 31–37). Following alignment and annotation, chromatographic peak intensities belonging to the same precursor mass in the MSE aligned chromatograms were used to calculate the relative peptide and protein abundance on a section-by-section basis. Two MSE analytical replicate intensities for each peptide were combined by averaging following robust median scaling (top and bottom 10% excluded), to generate one intensity value per peptide for each of the 24 photoreceptor layer sections.
We utilized both MS/MS DDA and MSE to generate peptide identifications. For DDA acquisition files, .mgf searchable files were produced in Rosetta Elucidator and searches were then submitted to and retrieved from the Mascot v2.2 (Matrix Sciences, Inc) search engine in an automated fashion. For MSE data, ProteinLynx Global Server 2.4 (Waters Corporation) was used to generate searchable files that were then submitted to the IdentityE search engine (Waters Corporation, Milford, MA) (38, 39); results files were then imported back into Elucidator. To enable global spectra scoring across results from both search engines, all search results were concurrently validated using the PeptideProphet and ProteinProphet algorithms in Elucidator using independent reverse decoy database validation (40, 41). Peptides with PeptideProphet scores greater than 0.8 were then annotated; this score corresponded to a 2% peptide false discovery rate. Individual peptide scores are reported in supplemental Table 1. Each peptide identified was allowed to be assigned to a single protein entry, and these assignments were made by ProteinProphet according to the rules of parsimony, and ProteinProphet scores are provided in supplemental Table 1.
Both DDA and MSE data were searched against the NCBInr database with Rattus norvegicus taxonomy (downloaded March 4, 2009 from http://www.ncbi.nlm.nih.gov/pubmed/), with full 1× reverse database appended for peptide false discovery rate determination, and duplicates removed using Protein Digest Simulator Basic (http://omics.pnl.gov/software/ProteinDigestionSimulatorBasic.php). The final database contained 134914 sequences including reverse entries. Precursor ion mass tolerance was 20 ppm for both PLGS and Mascot searches, and product ion tolerance was 0.1 Da for Mascot and 40 ppm for PLGS. Enzyme specificity was set to semitryptic for Mascot searches and tryptic for PLGS 2.4 searches, and a maximum of two missed cleavages was allowed. Carbamidomethyl cysteine was included as a fixed modification, and variable modifications included oxidized methionine and deamidated asparagine and glutamine. Additionally, phosphorylation of serine, threonine, and tyrosine was allowed in Mascot searches (only nine phosphorylated peptides were observed).
One quantitative data file from section #11 of the light-adapted retina, upon visual inspection, appeared to have inconsistent signal where significant portions of the data were missing, and therefore this file was removed from downstream analysis. Data quality control for the remaining 71 analyses was performed within the Elucidator software package by using the retention time viewer to assess the shift required for data alignment, and principal components analysis to assess consistency of data quality. The retention time shift required to properly align data files was evaluated for every LC-MS analysis. The maximum retention time shift required at any point across all 72 analyses was 2.1 min, with 62 analyses displaying shifts not exceeding 1 min. Peptide intensities for each sample were z-score transformed and submitted to a three-dimensional principal components analysis to look for analytical outliers. This analysis showed that the technical replicate injections were always separated by the smallest difference and in many cases the coordinated sections from each retina were also closely grouped (supplemental Fig. 1). No additional outliers were observed.
To assess the reliability of several peptide-to-protein quantification algorithms we employed the intraclass correlation analysis, which is a statistical method used to describe correlation among results of similar experiments obtained by different experimental approaches or different observers (42). The degree of similarity among the results of individual experiments is reflected by the intraclass correlation coefficient (ICC). The ICC values and their 95% confidence intervals were calculated with the PASW Statistics software, version 18 (SPSS Inc., Chicago, IL), using the ICC(2,1) two-way random single measures (Consistency/Absolute agreement) option.
The retinal sectioning is not exactly the same from sample to sample reflecting imperfections in specimen preparation: freezing the retina may result in layer distortions causing three-dimensional waving and small variations in tissue compression (26). In addition, the thickness of the first section cannot be precisely controlled on the microtome. Therefore, averaging protein quantification data obtained from two retinas required an alignment of protein distribution profiles in sections obtained from each of them, so at least two of these variations (tissue compression and the first section thickness) could be accounted for. This was accomplished using phosducin as a representative protein, because it is an abundant protein distributed through the entire photoreceptor length and absent from the neighboring cell types (27, 43). The distribution profiles of phosducin were therefore modeled as piecewise cubic polynomials. To reduce boundary artifacts, the original 12 elements of the phosducin profile measurement vector obtained from each retina were padded by adding four extra elements in the beginning and at the end, creating a 20 element discrete representation of this profile (44). We set the values for the first and the last four elements of this vector equal to the fifth and 16th elements, respectively. We used “cftool” function of the MATLAB software (Mathworks, Natick, MA) to fit a piecewise cubic spline curve to each 20 elements profile. Next, we searched for the scaled affine transform that warps the phosducin profile in retina #1 to that in retina #2: , where Y is the value of this profile at a distance x from the origin. We estimated the values of coefficients a, b and c by minimizing the following constrained least-squares cost function:
where ‖…‖22 is the Euclidean norm. The constrain over b is evoked because the mismatch between the two section sets cannot be larger than a single section. The constrain over horizontal (x) search space is evoked because the first and last real sample points (elements 5 and 16 in our padded measurement vectors) and their replicas (elements 1 to 4 and 17 to 20) are most prone to the measurement noise. The calculated coefficients were: â = 0.8, = 0.95, ĉ = 1.1. This procedure allowed us to define the data acquired in 10 μm imperfect physical sections as a corrected distance from the distal end of the photoreceptor cell.
Immunofluorescence detection of selected proteins in the rat retina was performed as described previously (45); all experiments were performed in triplicate. Briefly, eye cups were fixed in paraformaldehyde and embedded in 4.5% low-melt agarose (Invitrogen). One hundred and fifty-μm-thick cross-sections of the retina were collected with a vibratome (Leica VT1200 S), incubated with blocking solution (3% goat serum and 0.1% Triton X-100 in phosphate-buffered saline), treated with primary antibodies overnight, washed and incubated with goat anti-rabbit Alexa Fluor 488 secondary antibodies (1:400). Stained sections were mounted with Fluoromount G (Electron Microscopy Sciences) under glass coverslips and visualized using a Nikon Eclipse 90i confocal microscope. Polyclonal rabbit antibodies against hexokinase I and against hexokinase II (1:200) were provided by J.E. Wilson (Department of Biochemistry, Michigan State University, East Lansing, MI), and antibodies against triosephosphate isomerase (Fl-249) and GAPDH (sc-47724) were purchased from Santa Cruz Biotechnology(Santa Cruz, CA) (1:100). No signals were observed in control sections incubated with the secondary antibody only.
To analyze the patterns of protein distribution along the longitudinal axis of the photoreceptor cell, we separated the entire photoreceptor layer from a frozen flat-mounted rat retina into 12 serial 10-μm-thick tangential sections and determined the relative abundance of individual proteins in each section by LC/MS (our workflow is illustrated in Fig. 1). The analysis of 24 sections obtained from two individual retinas yielded a total of 5038 unique peptides corresponding to 896 protein database entries (supplemental Table 1). This table contains intensity values for all peptides from each section, which were used for protein quantification. There are several approaches to translate information contained in peptide intensities determined in multiple biological samples into relative protein abundance in each sample. Many studies rely on averaging the peptide intensity ratios of multiple peptides representing each protein in different samples (e.g. (36, 46–48)). Others average signals from the three most intense peptides (49) or sum signals from all peptides identified for a given protein (50–52). We, therefore, evaluated the accuracies of these protein quantification algorithms in application to our case.
Despite the unique protein composition of each section, there are several protein groups expected to preserve a constant molar ratio in all sections where they are found. These are proteins confined to specific cellular organelles, which maintain constant composition of their constituents regardless of their abundance in different sections. We applied four different quantification algorithms to generate the distribution profiles of proteins representing two organelles, the rod outer segment and the mitochondrion: (1) we expressed the peak area for the single most intense peptide from each protein as percentage distribution among all twelve sections; (2) we summed the peak areas for the three most intense peptides representing a given protein; (3) we summed the peak areas for all peptides from a given protein; (4) we calculated the fraction of each peptide present in each section and averaged the resulting values among all peptides representing a protein of interest. Note that in algorithms #2 and #3 peptides producing more intense peaks contributed to the resulting value with higher weights, whereas in algorithm #4 each peptide made an equal contribution to the resulting value. The reliability of protein quantification obtained using each algorithm was assessed by cross-correlating the resulting profiles of proteins representing each organelle. We used the intraclass correlation analysis (42), an ANOVA-based statistical methodology in which the degree of similarity among the individual protein profiles was reflected by the ICC, ranging from 0 to 1.
The first protein group included six phototransduction proteins localized exclusively to rod outer segments (Fig. 2; Table I), and indeed each protein quantification algorithm yielded distribution profiles peaking in the several most distal sections expected to represent rod outer segments. Interestingly, even algorithm #1 based on the single most intense peptide showed a fair degree of cross-correlation among individual proteins (ICC = 0.88; Fig. 2A). However, the ICC values were higher for each other algorithm we used, with the best cross-correlation obtained with algorithm #3 summing the intensities of all peptides from a given protein (Fig. 2E; ICC = 0.96), and the other two falling slightly behind (ICCs of 0.92 and 0.89 for algorithms #2 and #4, respectively; Fig. 2C, G). Repeating this analysis for sections obtained from the second retina yielded comparable results (Table I).
We next quantified a randomly chosen subset of eight mitochondrial proteins (Figs. 2B, D, F, H; Table I). Their distribution profiles in photoreceptors are rather complex: the majority resides in the inner segment and a minority in the synaptic terminal (see cartoon above Fig. 2B), reflected by lower absolute ICC values than for phototransduction proteins. However, the relative ranking among the four protein quantification algorithms was essentially the same: the best cross-correlation was obtained by summing the intensities of all peptides, algorithms #2 and #4 provided intermediate results, and the worst cross-correlation was obtained using the single-peptide approach.
We also addressed how the reliability of the “sum of intensities” protein quantification algorithm depends on the number of peptides used in calculations. The quantities of phototransduction and mitochondrial proteins in each section were recalculated taking between one and nine most intense peptides into account, and the ICC values were determined as above (Fig. 3). All plots displayed an upward trend with ICC values improving significantly by including a second and third peptide and reaching saturation at ~5–6 peptides, which appears to be an optimal number for this analysis, at least for the type of tissue samples and instrumentation used in our study. These results are generally consistent with predictions by Carrillo et al. (52), whose bioinformatics analysis and experiments using a relatively simple protein mixture indicated that the “sum of intensities” algorithm (#3) produces more accurate estimates of relative protein abundance than the “average of ratios” approach (#4). Our data extended this conclusion to an example of a more complex biological system. We should also stress that the idea to validate protein quantification algorithms by correlating profiles of proteins known to exist in fixed stoichiometric ratios in multiple samples is universally applicable to essentially any type of biological sample.
We used the sum of intensities algorithm to generate distribution profiles for all identified proteins in each retina (supplemental Table 1, bold values). To preserve any potentially useful information, we did not set any thresholds on the minimal number of peptides or minimal peptide intensity, although we realize that the accuracy of protein profiles obtained on the basis of one or few peptides might require particularly careful verification in any follow-up studies.
To improve reliability of this analysis, we averaged the protein profiles between two retinas. This necessitated accounting for any differences in tissue alignment encountered upon freezing and sectioning of each retina. As described under “Experimental Procedures,” two kinds of tissue distortion (tissue compression and unpredictability of the first section thickness; Fig. 4A), could be compensated mathematically. This was achieved by aligning the distribution profiles of phosducin, a marker protein chosen on the criteria of high abundance, exclusive expression in photoreceptors and distribution throughout the entire photoreceptor length (27, 43). Examination of each phosducin profile (Fig. 4C) suggested that the photoreceptor layer in retina #2 was more compressed than in retina #1 and slightly shifted to a lower section number. To account for both distortions, we modeled phosducin profiles as piecewise cubic polynomials and applied a scaled linear model to align the profiles, as described under “Experimental Procedures” (Fig. 4D), reducing the mean squared error between the retinas from 5×106 to 1.4×106. Applying the same correction parameters to other proteins yielded very closely matching profiles, as exemplified in Fig. 4E–4H for two abundant proteins (CNG-α1 and COX5B) residing in different parts of the photoreceptor cell. In all subsequent figures, we present averaged corrected protein profiles in the physical space of retina #1.
We defined the positions of major photoreceptor compartments in the proteome map using protein markers displaying highly restrictive localization patterns. In principle, this approach is similar to that of mapping organelles separated by centrifugation on a sucrose density gradient (53). For example, we generated a template profile for rod outer segments by averaging the individual profiles of five phototransduction proteins (Fig. 5A; solid line). Dashed lines in the same figure demonstrate that this template is a strong predictor of the distribution profiles for other proteins known to be confined to outer segments. We should note that a small group of signaling proteins, most importantly arrestin and transducin, change their distributions between the outer segments and the rest of the photoreceptor cell in a light-dependent manner, with arrestin moving into and transducin out of the outer segment (54). Although a detailed analysis of this phenomenon was not a focus of the present study, the distribution patterns of both arrestin and transducin were in good agreement with results obtained by other methods (supplemental Table 1), serving as an internal control that the expected protein distribution differences between the two retinas were clearly detected.
A similar approach was used to generate templates for proteins representing photoreceptor inner segments, the nuclear layer and the synaptic region (Fig. 5B). Although generating nuclear and synaptic templates was a straightforward task accomplished on the basis of randomly chosen abundant proteins representing each compartment, generating a template for the inner segment was challenging because most proteins residing there are also present in other parts of the cell. A rare exception is rootletin, a cytoskeletal protein thought to support the structural integrity of the inner-outer segment junction (55). The distribution profile of rootletin (Fig. 5B; orange line) was similar to the inner segment-specific peak of mitochondrial protein profiles, but unlike most mitochondrial proteins did not display a notable second peak at the synaptic terminal (e.g. Fig. 2F). To determine whether other proteins in our database have profiles similar to rootletin, we calculated the level of overlap between the rootletin template and averaged profiles of all other database entries. The proteins quantified on the basis of at least two peptides and displaying at least 80% overlap with rootletin are listed in supplemental Table 2. Although most represent mitochondrial proteins, because of their dominant peak in inner segments, several other hits were revealed, most notably the voltage-gated K+ channel and retinoschisin (Fig. 5C). The former was shown to be expressed in photoreceptors (56) and the latter is known to be a part of the interphotoreceptor matrix (57). However the highly restricted localization patterns observed in Fig. 5C have not previously been established for either protein. On the other hand, the β2 subunit of Na+/K+ ATPase was previously shown to be restricted to the inner segments (58), and this pattern was indeed confirmed in Fig. 5C. Another interesting example, hexokinase II, will be analyzed in detail below.
In addition to photoreceptors, the outer retina contains parts of two non-neuronal cell types, the Müller glial cells and the retinal pigment epithelium. The processes of the pigment epithelium extend into the space between rod and cone outer segments and, indeed, several proteins from the pigment epithelium were found in the first two sections of our map representing the outer segment tips (Fig. 6A). Müller cells span their cell bodies throughout the entire retina up to the border between photoreceptor inner segments and nuclei (called the outer limiting membrane); their thin microvilli penetrate even further into the space between inner segments. Accordingly, several Müller cell-specific proteins were present in our map, displaying distribution profiles reflecting this anatomical pattern (Fig. 6B). One protein, the cellular retinal aldehyde binding protein (CRALBP), is known to be expressed in both Müller cells and the pigment epithelium (59); in fact, the profile of CRALBP, but not other proteins from Müller cells contained an additional small peak in the outmost sections.
Photoreceptors are among the most metabolically active cells in the body. In the dark, they need energy to power ion pumps maintaining both the circulating dark current and Ca2+ flux for synaptic transmission. In the light, they expend significant energy to support phototransduction (see (60, 61) for recent detailed analyses). It has been long known that glycolysis plays a critical role in fulfilling these energy needs (62). Glucose reaches the outer retina from the blood, passing the layer of pigment epithelium cells with the aid of facilitative glucose transporter Glut1 (63, 64). The same transporter is used for glucose entry into photoreceptor cells, where it is expressed in inner segments and synaptic terminals, but not outer segments, as demonstrated in our recent study (65). We now used the photoreceptor proteome map to investigate the intracellular sites of subsequent glucose metabolism by analyzing distribution patterns for the entire set of glycolytic enzymes (Fig. 7).
Our immediate surprising finding was that, unlike the majority of neurons, photoreceptors express hexokinase II in addition to the previously found hexokinase I (66). Furthermore, the distribution profiles of each hexokinase isoform were distinctly different from those for the majority of glycolytic enzymes (Fig. 7A). The latter displayed profiles suggesting an even distribution throughout the entire cytoplasmic volumes of photoreceptors and neighboring cells. They resembled the distribution of phosducin (Fig. 4), except for a higher abundance in sections 10–12 containing significant amounts of material from non-photoreceptor cells, consistent with these proteins being expressed in adjacent cell types as well. In contrast, the profiles for each hexokinase had a large peak in the inner segment, likely reflecting their established pattern of mitochondrial association ((67); reviewed in (68, 69)). Hexokinase I displayed a second, even larger peak in sections 10–12, whereas hexokinase II was confined almost entirely to the inner segments, resembling the profile of rootletin (Fig. 5C).
To elaborate on these observations, we immunostained retina cross-sections with antibodies specifically recognizing each hexokinase isoform and found that each staining pattern closely matched its proteomic profile. Hexokinase I was expressed most abundantly outside the photoreceptor layer, although it was also found in photoreceptor inner segments (Fig. 7D; see also (66)). In contrast, almost all hexokinase II was found in inner segments, with a very small additional signal in photoreceptor synapses and essentially no staining in other retinal cells (Fig. 7E). For comparison, we immunostained retina sections with an antibody against glyceraldehyde-3-phosphate dehydrogenase (GAPDH), the proteomic profile of which was representative of the majority of glycolytic enzymes (Fig. 7G). The staining pattern for GAPDH differed from both hexokinase isoforms and was indeed consistent with an even cytoplasmic distribution in photoreceptors and other retina cells.
The presence of hexokinase II and the mitochondria-associated localization pattern for both isoforms likely reflect the unique properties of glucose metabolism in photoreceptors. Hexokinases catalyze the first step in glycolysis—the phosphorylation of glucose to produce glucose-6-phosphate (Fig. 7B)—which is essential for intracellular glucose retention because phosphorylated glucose cannot undergo reversible transport through the plasma membrane. Hexokinase I is ubiquitously expressed in neuronal tissues, whereas hexokinase II is typically found in myocytes and cancer cells known to be particularly dependent on anaerobic energy metabolism (reviewed in (70, 71)). The latter is attributed to the higher catalytic activity of hexokinase II and its activation through mitochondrial association (70). Our finding that photoreceptors, but not other cells in the retina, express predominantly hexokinase II suggests that they also benefit from utilizing this more versatile isoform in order to keep up with the high level of energy metabolism.
We should also note that the lack of either hexokinase in the outer segment hints why outer segments are also deprived of the Glut1 glucose transporter (65). Glucose phosphorylation by hexokinase is required to trap glucose inside the cell because glucose-6-phosphate cannot escape the cell through the same transporter. Therefore, direct glucose entry into the outer segment compartment devoid of hexokinases would not result in its intracellular accumulation. Instead, glucose enters the cell in the inner segment, where it is converted into glucose-6-phosphate by hexokinases and can undergo further metabolism in both the inner and outer segments.
The third glycolytic enzyme with a profile distinctly different from the majority of this group was triosephosphate isomerase. Its proteomic profile displayed a steady upward trend from outer segments to synapses (Fig. 7A). Immunostaining revealed that the abundance of this enzyme in photoreceptors is very low compared with other cells. At a staining intensity providing bright signals in the inner retina, we observed essentially no staining in outer segments and little staining around the outer limiting membrane and around the nuclei, a pattern more consistent with staining of Müller cells than photoreceptors (Fig. 7F).
The role of triosephosphate isomerase in glycolysis is to convert dihydroxyacetone phosphate into glyceraldehyde-3-phosphate, after both are produced upon the cleavage of fructose-1,6-bisphosphate by fructose bisphosphate aldolase (Fig. 7B). Therefore, low abundance of triosephosphate isomerase in photoreceptors is predicted to result in an elevated level of dihydroxyacetone phosphate compared with other retina cells. Although we can only speculate about the functional significance of such a phenomenon, one possibility is that it relates to phospholipid biosynthesis, as dihydroxyacetone phosphate serves as a precursor of glycerol-3-phosphate, the phospholipid backbone (72). Photoreceptors synthesize phospholipids at an unusually high rate because of the constant renewal of their outer segment membranes (73, 74). The surface area of these membranes produced daily in a rat retina exceeds the surface area of the eyeball by over 50-fold (75). This makes it plausible that an elevated dihydroxyacetone phosphate concentration would facilitate this biosynthetic pathway.
As a final note in regards to glycolytic enzymes in the outer retina, we should mention that our data reject a recent claim that photoreceptors, retinal pigment epithelium and Müller cells lack glucose-6-phosphate isomerase-1 (76). We identified 36 peptides from this protein, showed that its distribution profile matches the majority of glycolytic enzymes and found no evidence that any other glucose-6-phosphate isomerase isoform is present in the outer retina within the detectability limits of our assay.
The data presented in this study illustrate the power of combining label-free quantitative proteomics with subcellular fragmentation of a layered tissue by serial sectioning. Based on this approach, we generated a proteome map describing the distribution profiles of hundreds of proteins representing the photoreceptor layer of the retina. We used this map to examine the subcellular composition of glycolytic enzymes and revealed previously unknown patterns of isoform composition and distribution, all suggesting unique mechanistic features employed by photoreceptor cells to meet their high energetic and metabolic needs. This is unlikely to be the sole discovery embedded within our protein map, which is presented in a format that can be mined by other investigators studying other aspects of photoreceptor function.
Our approach to studying the photoreceptor proteome differs from that used in several recent reports. Most previous studies focused on analyzing the protein composition of rod outer segments, the organelle readily separable from the rest of the retina by centrifugation procedures (77–79). Others attempted to obtain proteomes of the whole photoreceptor cell by mechanical separation of the retina layers (80), obtaining tangential sections containing the entire photoreceptor layer (81), or by analyzing the difference between proteomes of the normal retinas and those in which rods underwent inherited degeneration (82). Unlike any of these strategies, thin serial sectioning allowed us to conduct protein identification and quantification in all the individual subcellular compartments of these cells, each defined by profile shapes of marker proteins. Our approach is also different from any “snapshot” proteomes of fractionated organelles, which rarely contain any intrinsic hints to distinguish between the proteins actually residing in an organelle of interest and those introduced from contaminations. For example, each published rod outer segment proteome contained multiple entries of mitochondrial and nuclear proteins, despite their actual absence from outer segments (77–79). Sections representing the outer segments in our study were not free of contamination either, but unlike in snapshot proteomes, the distinction between outer segment-specific proteins and contaminants from other cellular compartments could be made on the basis of each protein's overall distribution profile.
Finally, we would like to stress that the approach developed in this study can serve as a guide for generating similar proteome databases addressing protein distribution profiles in essentially any layered tissue. Our work represents a case study exploring both general methodology of protein quantification and simple computational corrections for the unavoidable problem of tissue distortion during cryosectioning. In fact, the retina is among the smallest layered tissues in our body; applying the same approach to larger tissues is likely to yield deeper protein coverage, while being less technically challenging. Furthermore, a similar strategy could be very useful in the proteomic investigation of specimens where quantification of proteins with high spatial resolution at a margin is important, for example the margins of invasive tumors.
We are grateful to P. A. Ferreira for helpful discussions and S. M. Gospe for critically reading the manuscript.
* This work has been supported by NIH grants EY010336, EY017648 (VYA) and NIH Core Grant for Vision Research EY5722 to Duke University.
This article contains supplemental Tables S1 and S2 and supplemental Fig S1.
1 The abbreviations used are: