Study of gene function based on spatial co-expression in a high-resolution mouse brain atlas
- Zheng Liu†1, 2,
- S Frank Yan†2,
- John R Walker2,
- Theresa A Zwingman3,
- Tao Jiang1,
- Jing Li4 and
- Yingyao Zhou2Email author
© Liu et al; licensee BioMed Central Ltd. 2007
Received: 21 November 2006
Accepted: 16 April 2007
Published: 16 April 2007
The Allen Brain Atlas (ABA) project systematically profiles three-dimensional high-resolution gene expression in postnatal mouse brains for thousands of genes. By unveiling gene behaviors at both the cellular and molecular levels, ABA is becoming a unique and comprehensive neuroscience data source for decoding enigmatic biological processes in the brain. Given the unprecedented volume and complexity of the in situ hybridization image data, data mining in this area is extremely challenging. Currently, the ABA database mainly serves as an online reference for visual inspection of individual genes; the underlying rich information of this large data set is yet to be explored by novel computational tools. In this proof-of-concept study, we studied the hypothesis that genes sharing similar three-dimensional expression profiles in the mouse brain are likely to share similar biological functions.
In order to address the pattern comparison challenge when analyzing the ABA database, we developed a robust image filtering method, dubbed histogram-row-column (HRC) algorithm. We demonstrated how the HRC algorithm offers the sensitivity of identifying a manageable number of gene pairs based on automatic pattern searching from an original large brain image collection. This tool enables us to quickly identify genes of similar in situ hybridization patterns in a semi-automatic fashion and consequently allows us to discover several gene expression patterns with expression neighborhoods containing genes of similar functional categories.
Given a query brain image, HRC is a fully automated algorithm that is able to quickly mine vast number of brain images and identify a manageable subset of genes that potentially shares similar spatial co-distribution patterns for further visual inspection. A three-dimensional in situ hybridization pattern, if statistically significant, could serve as a fingerprint of certain gene function. Databases such as ABA provide valuable data source for characterizing brain-related gene functions when armed with powerful image querying tools like HRC.
It is estimated that only ~1% of the genes expressed in human brain are studied in over 99% of the published neuroscience studies; we are far from understanding the enigmatic biological processes in the brain . Microarray technology has been successfully applied to profile the expression landscape of the entire transcriptome in parallel; however, the size of typical brain samples dissected for mRNA extraction only allows the detection of a globally averaged expression level over a relatively large anatomical region; therefore, standard array based gene expression data sets often lack the desirable fine resolution required for neuroscience studies [2, 3]. In order to preserve the relationships among brain circuitry, cell type, and gene expression, all of which are crucial for understanding the molecular machinery of the brain, in situ hybridization technology has been developed , which can be applied to measure the three dimensional high-resolution expression map of brain genes one at a time. The Allen Brain Atlas project [5, 6], arguably one of the most ambitious post-genome projects, aims to systematically create a detailed gene expression brain atlas for as many as 24,000 genes by 2006. For each particular gene, 25 μm thick brain sections are cut at every 100–200 μm throughout the entire mouse brain. Hybridization of gene specific antisense probes to the brain slide enables quantitative measurement of the mRNA transcription level at an unprecedented cellular resolution. At the time of this study, data on 6080 genes were released online in the Allen Brain Atlas database . At an estimated rate of generating 300 megabytes of map data per day , both the volume and the complexity of the image data present a difficult informatics challenge. Currently, the brain atlas database mainly serves as an online reference for visual examination of individual genes. The rich biological knowledge implied by this largest neuroscience database is yet to be explored–novel computational tools are essential for any such attempts.
Genes with similar expression profiles across a panel of different biological conditions are known to tend to share similar biological functions–a principle known as guilt by association (GBA) [7–11]. Extending the GBA concept to the brain atlas hypothesizes that genes share similar spatial brain expression landscapes could also imply similar biological functions. If validated, this idea will naturally become a powerful functional genomics tool for characterizing genes of unknown functions, as well as discovering new roles for known genes. One can envision a future version of the ABA database which provides an accurate pattern query and comparison tool to help neuroscientists discover genes of interesting spatial profiles and potential network partners in order to better understand the mechanism of a molecular target implicated in certain disease. In fact, ABA has made progress in this direction at the time of our writing.
To carry out such a proof-of-concept study, given a query gene of interest, we first have to develop an algorithm to help filter out obviously unrelated genes and highlight a manageable subset of genes that potentially share similar spatial expression patterns. Due to the complexity of the problem, the gene candidates discovered by the algorithm are then subjected to further human visual inspection, i.e., the sensitivity of the algorithm is more important given a reasonable specificity. In this study, we developed and compared three image similarity metrics required for gene filtering with increasing sophistication: a naïve pixel-wise metric, an adjusted pixel-wise metric, and a histogram-row-column (HRC) metric based on time series summary data. These three metrics were benchmarked and the superiority of the HRC algorithm was validated by cross validation studies. The biological studies presented in the Results and Discussion section are all made possible by using the HRC algorithm as a fully automated efficient first-pass filter.
We then studied several spatial hybridization patterns and showed that, in many cases, a selective brain atlas can represent an expression neighborhood that consists of genes of statistically enriched function categories. These discoveries were then cross validated using other related databases including the GNF Tissue Atlas , GenePaint.org , and the NCBI GENSAT database . Most interestingly, our results illustrate how spatial co-expression leads to functional enrichment for the cyclic AMP (cAMP) regulatory pathway, particularly in relevance to adenylyl cyclase. We validated that substantia nigra enrichment serves as a signature pattern for the critical nigrostriatal dopaminergic pathway involved in Parkinson's disease after examining Ddc, Slc6a3, and Slc18a2 genes, which is consistent with the latest findings . We conclude that the popular guilt by association principle can be aptly applied to the brain atlas database, transforming it into a rich source for functional genomic studies in neuroscience, in addition to a reference data repository.
Results and discussion
Measure the similarity of two brain images
The goal of this study is to investigate whether genes with similar spatial mRNA expression distribution in the brain tend to be functionally related. We first need to identify all the genes with similar mRNA expression to a given query gene based on the brain image at a particular slide location, and analyze the resultant gene list for any statistically significant functional enrichment based on existing biological annotations in the literature or gene ontology databases. Measuring the similarity of two brain images is a fairly complicated computational problem for several reasons. First, the ABA database, consisting of a growing large number of brain images, makes it nearly impossible for human manual inspection. At the time when this study began, the image data for 6080 genes were posted with dozens of images per gene corresponding to different brain anatomy locations. This number is increasing quite rapidly. For a single image query, over one million image pairs would need to be compared now. Second, in addition to the requirement of sophisticated data management solutions, the complexity of a brain image poses a significant computational challenge in terms of both image processing and pattern recognition. On top of these factors, brain samples are obtained from different mice, resulting in that the overall brain size and shape, as well as the contour of each brain anatomy region, can vary significantly even if one examines the same brain region at the same section position. Third, hybridization probes of different genes have heterogeneous biochemical properties, which could lead to varying hybridization signal intensity levels and potential cross-hybridization background levels across genes. One also needs to take into account the technical factors such as different sample orientations and image scanning artifacts in brightness and contrast. It is clear any algorithm that automatically measures the similarity of two brain images should be robust against the above mentioned biological and technical variations. However, due to the complexity of the problem, we do not expect such algorithm be good enough to replace human visual inspection, but should instead act as an automatic, efficient first-pass filter to highlight a subset of candidate gene slides, which is manageable for the second-pass visual refinement.
The gene expression level of an in situ hybridization is represented as an RGB image in the ABA database. The comparison of spatial expression between different images is actually an image registration problem, whose performance depends highly on the quality of the distance metric for an image pair. Typical image registration methods can either take the pixel intensity distribution or compute the pixel-by-pixel distances using Euclidean distance, Pearson correlation coefficient, etc. The most relevant approach to our study is the use of the Gaussian mixture model for expression distribution analysis . But this method is not applicable in this case to analyze ABA images, because it lacks the capability of handling variations in anatomical regions across different brain slides. The parameters used in the global and local Gaussian mixture model matching do not reflect the gene expression property directly. Kumar et al.  uses the overlap between binarized images to measure distance. These metrics are designed to represent the global similarity or local similarity between images for different applications and are similar to our naïve pixel-wise algorithm. Here, we proposed three different alternative distance metrics for comparing a brain image pair with increasing complexity, namely naïve pixel-wise distance metric, adjusted pixel-wise distance metric, and a method based on intensity summaries by histogram, row, and column (so-called HRC method). It is noted that ABA has also released mask thumbnail images together with the original scans (mask images all have background, brightness, and contract factors corrected). Compared to the original hybridization scans, use of mask images has led to a significant performance improvement as expected. Also, at the end of our study, the ABA web site began to provide qualitative query features that enable a user to search genes based on "low/medium" or "high" expression levels in 11 selected brain regions. Carson et al. recently published a subdivision mesh technique for better pattern recognition of brain regions based on a set of reference slides; they provide web pattern query tools via GenePaint.org . Compared to these recent developments, our method still offers the advantage of quantitative description of the expression patterns in an automatic fashion. We believe that both are important aspects for the future development of a large brain image database such as ABA.
Training of the HRC weighting factors
A set of image pairs with "true" distances is required to train the weighting factors in the HRC method (see Methods) and to objectively benchmark the performance of various distance metrics. To construct such an unbiased data set, we resorted to the fact that slides of a given gene have similar texture patterns if they are obtained from close vicinities, while slides are most likely different if they are taken from brain regions far apart. Therefore, the physical distance between two slides of the same gene to some extent represents their "true" similarity.
A total of 1091 thumbnail sagittal slide images for 60 genes were downloaded from the ABA web site; the number of slides per gene ranges from 15 to 20. As described in Methods, the HRC weighting factor set that has the best average performance across all the 60 genes is chosen as the final, optimal weighting factors. We subsequently carried out a final run by combining all 60 genes as the training set and our final optimal weighting factors are [1.98, 107.39, 11.91] with an average Pearson correlation coefficient of 0.58.
Comparison of the three distance metrics
Comparisons of the three similarity metrics
P-value (Student t-test)
P-value (Wilcoxon test)
Naïve vs. Adjusted
1.5 × 10-5
1.8 × 10-6
Adjusted vs. HRC (3-fold)
9.3 × 10-13
< 2.2 × 10-16
Adjusted vs. HRC (10-fold)
8.3 × 10-13
< 2.2 × 10-16
Application of the HRC method
Based on the above comparisons, the HRC algorithm is our final method of choice. Given the fact that the HRC algorithm is insensitive to the settings in the above cross validation tests, we are confident that given a particular gene slide of interest the HRC metric is able to help us filter out a large number of unrelated gene expression images without human intervention. We applied the HRC algorithm to study several genes of biological interest in order to assess the feasibility of carrying out a functional genomics study based on spatial gene expression in the mouse brain.
Since the 6080 ABA gene images were only available for online browsing, we manually downloaded 2759 sagittal brain slides of 145 genes for this proof-of-concept study. Given a gene of biological interest, we first identified a brain slide that shows interesting uncommon textual features and used it as our query image. We then applied the HRC algorithm to all the brain slides that are within 200 slide distances from the query slide position to rank these slides and locate genes with similar profiles. The HRC algorithm was applied recursively to the new set of genes that pass our visual inspection until a group of core genes with similar brain atlas expression patterns was obtained (based on visual judgment). Finally, we carried out biological functional analysis of the gene list based on literature search as well as other similar, smaller scale brain in situ hybridization databases. Several interesting examples have been found where the guilt by association principle can be applied to successfully establish the link between a characteristic gene spatial distribution pattern and a specific gene functional category. We summarized these findings in the following sections.
Type 5 adenylyl cyclase is the primary isoform accountable for striatal adenylate cyclase activity
Using a key component Adcy5 in the neuronal cyclic AMP signaling as a query pattern, the HRC algorithm is able to identify other proteins that are involved in this pathway
Top-ranked genes identified using Adcy5 as the query pattern
Relevant GO term
cAMP biosynthesis; locomotory behavior
G-protein coupled receptor protein signaling pathway
locomotory behavior; receptor guanylyl cyclase signaling pathway
dopamine receptor, adenylate cyclase inhibiting pathway
adenylate cyclase activation; locomotory behavior
regulation of cell migration
regulation of cell proliferation
G-protein coupled receptor protein signaling pathway
neurotransmitter secretion; regulation of calcium ion-dependent exocytosis
embryonic eye morphogenesis
On one hand, we notice HRC algorithm indeed effectively identifies genes of relevant expression patterns. Among the top 20 genes in Table 2, five genes are known to be involved in adenylate cyclase activity and/or locomotory behavior based on an existing gene annotation database . We further used Ingenuity Pathway Analysis (IPA) software  to study the related functions of these five genes, and it was found that they are all involved in the behavior function, mostly mouse locomotor activity, with a significance value of 10-21. Besides validating the HRC algorithm itself, the result also indicates that the Adcy5 expression pattern may be a signature pattern of the neuronal cAMP signaling pathway (Fig. 2). On the other hand, our visual inspection found that Ppp1r1b, which encodes protein phosphatase 1 regulatory subunit 1B, also shares similar expression pattern in the striatum region compared to Adcy5. The HRC algorithm was not able to identify it, despite the known fact that it is involved in the neuronal cyclic AMP signaling . A closer examination of the ABA image revealed that Ppp1r1b is indeed highly expressed in the striatum region as Adcy5, while in the current ABA image it is also widely expressed in the cerebral cortex. This might prevent the HRC algorithm from high-ranking this gene. Nonetheless, based on the GENSAT image of Ppp1r1b, it is highly expressed mainly in the striatum region, bearing significant distribution similarity to Adcy5 (Fig. 3). Gpr88 is known to be a striatum specific G-protein coupled receptor , which also shares great sequence similarity with 5-HT1D receptor. Its strikingly similar spatial distribution with Adcy5 suggests that it might also be an uncharacterized gene involved in neuronal cAMP pathway. Knockout validations are being carried out.
Cyclic AMP-regulated phosphoprotein 21 isoform 1 is the only gene product of Arpp21 involved in the striatal cAMP and Ca2+/calmodulin signaling pathway
Key genes involved in the nigrostriatal dopaminergic pathway and Parkinson's disease are enriched in substantia nigra
It should be pointed out that Slc6a3 and related genes are expressed in a very small, localized region of the mouse brain (Fig. 5). This may create difficulty for the HRC method to carry out effective pattern matching for a small region, as information of all the rows and columns of the entire brain image is used to construct the row and column vectors, which may introduce noise into the H, R, C vectors. A potential improvement of this algorithm is to restrict the rows and columns used in creating the vectors based on specific region of interest of a query image. This may increase the sensitivity of the HRC method to discovery relevant matching brain images.
Guilt by association on a three-dimensional level provides more information on gene function
At the current stage, we only tested hundreds of genes in this pilot study. There are certainly more research topics in exploring this unique ABA spatial gene expression data set. For example, after we filter out those dissemble images, it is very important to develop a more sophisticated method to rank the similar images in order to identify coregulated genes. Since similar images have high global similarity scores with the query image, we could focus on investigating the local similarity and spatial information to discover the most related images with confidence. In addition, we believe that quality control and sample standardization of mouse brain slides may greatly affect our ability in applying this image processing algorithm to the data, and hence special attention needs to be considered.
We studied gene expression across the GNF and ABA atlases. With the help of our HRC filtering algorithm, we used the guilt by association approach to both confirm previous gene functional interactions and suggest new ones. Given query expression patterns of interest, we have shown that the HRC algorithm is able to produce a ranked gene list that is significantly enriched in visually confirmed positive hits and facilitates the discovery of signature patterns of important neurobiological pathways. We also highlighted the advantages of using this approach in databases of in situ hybridization images over microarray databases from tissue dissections. We believe a complete set of both coronal and sagittal mouse brain images will significantly facilitate confident (i.e. with statistical confidence) characterization of gene functions based on the unique information provided by ABA.
Brain images may have differential background intensities due to both biological and technical variations; therefore, the background effect should be removed before carrying out any meaningful comparison. As the background-subtracted mask images were made available by ABA during this study, they were used for all the calculations presented here and the background correction techniques will not be discussed. The mask brain slides outline the brain boundary, and pixels outside the brain region can be easily identified and their intensities were set to zero. The rest of the image has high contrast levels and can be closely approximated as binary black-and-white bitmap images. We first convert the mask slides into grayscale images and then into bitmap images based on the 128 intensity threshold. The resultant bitmap images led to better query results and were used for this study, although all the methods presented here are applicable to grayscale images as well.
Naïve pixel-wise distance metric
where a pixel (i, j) is identified by its location at the i th row and the j th column of the image matrix.
Adjusted pixel-wise distance metric
In addition to the basic background subtraction and contrast scaling that have been carried out for the mask slides, image pairs may require certain transformation operations such as translation and scaling in order to become more comparable. In the adjusted pixel-wise distance metric calculation, we address some of these factors, which may lead to an improvement in sensitivity. This method first linearly scales the height of foreground image F b to match the height of F a , then translates image F b horizontally with respect to F a in order to minimize d ab . We observed that for slides around the similar positions of the brain, their orientations are reasonably consistent. On the other hand, both sample size and shape differ significantly for slides with larger distance. Therefore, it is undesirable to perform the rotation optimization for such image pairs.
Histogram-row-column (HRC) distance metric
We then apply the weighting factors obtained from a particular gene g to the slides from all the other genes to assess their extrapolating performance. The factor set with the best average performance across the whole training set is then used.
- Gewin V: A golden age of brain exploration. PLoS Biol. 2005, 3: e24-PubMed CentralPubMedView ArticleGoogle Scholar
- Su AI, Wiltshire T, Batalov S, Lapp H, Ching KA, Block D, Zhang J, Soden R, Hayakawa M, Kreiman G, Cooke MP, Walker JR, Hogenesch JB: A gene atlas of the mouse and human protein-encoding transcriptomes. Proc Natl Acad Sci USA. 2004, 101: 6062-6067.PubMed CentralPubMedView ArticleGoogle Scholar
- Walker JR, Su AI, Self DW, Hogenesch JB, Lapp H, Maier R, Hoyer D, Bilbe G: Applications of a rat multiple tissue gene expression data set. Genome Res. 2004, 14: 742-749.PubMed CentralPubMedView ArticleGoogle Scholar
- Neidhardt L, Gasca S, Wertz K, Obermayr F, Worpenberg S, Lehrach H, Herrmann BG: Large-scale screen for genes controlling mammalian embryogenesis, using high-throughput gene expression analysis in mouse embryos. Mech Dev. 2000, 98: 77-94.PubMedView ArticleGoogle Scholar
- The Allen Brain Atlas Project. http://www.brain-map.org
- Boguski MS, Jones AR: Neurogenomics: at the intersection of neurobiology and genome sciences. Nat Neurosci. 2004, 7: 429-433.PubMedView ArticleGoogle Scholar
- Walker MG, Volkmuth W, Sprinzak E, Hodgson D, Klingler T: Prediction of gene function by genome-scale expression analysis: prostate cancer-associated genes. Genome Res. 1999, 9: 1198-1203.PubMed CentralPubMedView ArticleGoogle Scholar
- Quackenbush J: Genomics. Microarrays-guilt by association. Science. 2003, 302: 240-241.PubMedView ArticleGoogle Scholar
- Joshi T, Chen Y, Becker JM, Alexandrov N, Xu D: Genome-scale gene function prediction using multiple sources of high-throughput data in yeast Saccharomyces cerevisiae. OMICS. 2004, 8: 322-333.PubMedView ArticleGoogle Scholar
- Zhou Y, Young JA, Santrosyan A, Chen K, Yan SF, Winzeler EA: In silico gene function prediction using ontology-based pattern identification. Bioinformatics. 2005, 21: 1237-1245.PubMedView ArticleGoogle Scholar
- Yanai I, Korbel JO, Boue S, McWeeney SK, Bork P, Lercher MJ: Similar gene expression profiles do not imply similar tissue functions. Trends Genet. 2006, 22: 132-138.PubMedView ArticleGoogle Scholar
- Visel A, Thaller C, Eichele G: GenePaint.org: an atlas of gene expression patterns in the mouse embryo. Nucleic Acids Res. 2004, 32: D552-D556.PubMed CentralPubMedView ArticleGoogle Scholar
- Gong S, Zheng C, Doughty ML, Losos K, Didkovsky N, B. SU, Nowak NJ, Joyner A, Leblanc G, Hattern ME, Heintz N: A gene expression atlas of the central nervous system based on bacterial artificial chromosomes. Nature. 2003, 425: 917-925.PubMedView ArticleGoogle Scholar
- Carson JP, Ju T, Lu HC, Thaller C, Xu M, Pallas SL, Crair MC, Warren J, Chiu W, Eichele G: A digital atlas to characterize the mouse brain transcriptome. PLoS Comput Biol. 2005, 1: e41-PubMed CentralPubMedView ArticleGoogle Scholar
- Peng H, Myers EW: Comparing in situ mRNA expression patterns of Drosophila embryos. Proceedings of the eighth annual international conference on research in computational molecular biology. 2004, 157-166. San Diego, CA, ACM PressGoogle Scholar
- Kumar S, Jayaraman K, Panchanathan S, Gurunathan R, Marti-Subirana A, Newfeld SJ: BEST: a novel computational approach for comparing gene expression patterns from early stages of Drosophilia melanogaster development. Genetics. 2002, 162: 2037-2047.PubMed CentralPubMedGoogle Scholar
- Iwamoto T, Okumura S, Iwatsubo K, Kawabe J, Ohtsu K, Sakai I, Hashimoto Y, Izumitani A, Sango K, Ajiki K, Toya Y, Umemura S, Goshima Y, Arai N, Vatner SF, Ishikawa Y: Motor dysfunction in type 5 adenylyl cyclase-null mice. J Biol Chem. 2003, 278: 16936-16940.PubMedView ArticleGoogle Scholar
- Lee KW, Hong JH, Choi IY, Che Y, Lee JK, Yang SD, Song CW, Kang HS, Lee JH, Noh JS, Shin HS, Han PL: Impaired D2 dopamine receptor function in mice lacking type 5 adenylyl cyclase. J Neurosci. 2002, 22: 7931-7940.PubMedGoogle Scholar
- Schaefer ML, Wong ST, Wozniak DF, Muglia LM, Liauw JA, Zhuo M, Nardi A, Hartman RE, Vogt SK, Luedke CE, Storm DR, Muglia LJ: Altered stress-induced anxiety in adenylyl cyclase type VIII-deficient mice. J Neurosci. 2000, 20: 4809-4820.PubMedGoogle Scholar
- The Jackson Laboratory. http://www.jax.org
- Gray PA, Fu H, Luo P, Zhao Q, Yu J, Ferrari A, Tenzen T, Yuk DI, Tsung EF, Cai Z, Alberta JA, Cheng LP, Liu Y, Stenman JM, Valerius MT, Billings N, Kim HA, Greenberg ME, McMahon AP, Rowitch DH, Stiles CD, Ma Q: Mouse brain organization revealed through direct genome-scale TF expression analysis. Science. 2004, 306: 2255-2257.PubMedView ArticleGoogle Scholar
- Rakhilin SV, Olson PA, Nishi A, Starkova NN, Fienberg AA, Nairn AC, Surmeier DJ, Greengard P: A network of control mediated by regulator of calcium/calmodulin-dependent signaling. Science. 2004, 306: 698-701.PubMedView ArticleGoogle Scholar
- Reed TM, Repaske D, Snyder GL, Greengard P, Vorhees CV: Phosphodiesterase 1B knock-out mice exhibit exaggerated locomotor hyperactivity and DARPP-32 phosphorylation in response to dopamine agonists and display impaired spatial learning. J Neurosci. 2002, 22: 5188-5197.PubMedGoogle Scholar
- Schwindinger WF, Betz KS, Giger KE, Sabol A, Bronson SK, Robishaw JD: Loss of G protein gamma 7 alters behavior and reduces striatal alpha(olf) level and cAMP production. J Biol Chem. 2003, 278: 6575-6579.PubMedView ArticleGoogle Scholar
- The Gene Ontology. http://www.geneontology.org
- Ingenuity Pathway Analysis. http://www.ingenuity.com
- Mizushima K, Miyamoto Y, Tsukahara F, Hirai M, Sakaki Y, Ito T: A novel G-protein coupled receptor gene expression in striatum. Genomics. 2000, 69: 314-321.PubMedView ArticleGoogle Scholar
- GNF SymAtlas. http://symatlas.gnf.org
- Brooks DJ, Frey KA, Marek KL, Oakes D, Paty D, Prentice R, Shults CW, Stoessl AJ: Assessment of neuroimaging techniques as biomarkers of the progression of Parkinson's disease. Exp Neurol. 2003, 184 Suppl: S68-S79.View ArticleGoogle Scholar
- Storch A, Ludolph AC, Schwarz J: Dopamine transporter: involvement in selective dopaminergic neurotoxicity and degeneration. J Neural Transm. 2004, 111: 1267-1286.PubMedView ArticleGoogle Scholar
- Ito Y, Fujita M, Shimada S, Watanabe Y, Okada T, Kusuoka H, Tohyama M, Nishimura T: Comparison between the decrease of dopamine transporter and that of L-DOPA uptake for detection of early to advanced stage of Parkinson's disease in animal models. Synapse. 1999, 31: 178-185.PubMedView ArticleGoogle Scholar
- Bayer L, Mairet-Coello G, Risold PY, Griffond B: Orexin/hypocretin neurons: chemical phenotype and possible interactions with melanin-concentrating hormone neurons. Regul Pept. 2002, 104: 33-39.PubMedView ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.