• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of nihpaAbout Author manuscriptsSubmit a manuscriptNIH Public Access; Author Manuscript; Accepted for publication in peer reviewed journal;
Nat Methods. Author manuscript; available in PMC Aug 11, 2011.
Published in final edited form as:
PMCID: PMC3154617
NIHMSID: NIHMS308051

A Gene Expression Barcode for Microarray Data

Abstract

The ability to measure genome-wide expression holds great promise for characterizing cells and distinguishing diseased from normal tissues. Thus far, microarray technology has only been useful for measuring relative expression between two or more samples, which has handicapped its ability to classify tissue types. This paper presents the first method that can successfully predict tissue type based on data from a single hybridization. A preliminary web-tool is available at http://rafalab.jhsph.edu/barcode/

The high throughput analysis of cells and tissues is revolutionizing biological research. The ability of microarrays to measure thousands of RNA transcripts at one time allows for the characterization of cells and tissues in greater depth than was previously possible, but has not yet led to big advances in diagnosis or treatment. The main reason for this is that feature characteristics, such as probe sequence, can cloud the relationship between observed intensity and actual expression. Although this probe effect is large, it is also very consistent across different hybridizations, which implies that relative measures of expression are substantially more useful than absolute ones1, 2. To understand this, consider that when comparing intensities from different hybridizations for the same gene, the probe effect is very similar and cancels out. On the other hand, when comparing intensities for two genes from the same hybridization, the different probe effects can alter the observed differences. For this reason the overwhelming majority of results based on microarray data rely on measures of relative expression: genes are reported to be differentially expressed rather than expressed or unexpressed.

Approaches for thresholding noisy data have been successfully used in many applications including microarray studies3, 4. We used this as motivation to develop the first method that can accurately demarcate expressed from unexpressed genes and therefore defines a unique gene expression barcode for each tissue type. To do this we took advantage of the vast amount of publicly available datasets. These data were also used to assess the algorithm. With clinical data, we find near perfect predictability of normal from diseased tissue for three cancer studies and one Alzheimer’s disease study. The barcode method also discovers new tumor subsets in previously published breast cancer studies that can be used for the prognosis of tumor recurrence and survival time.

For any given gene and microarray platform, we wanted to know what intensity relates to no expression. A simple way to determine this intensity would be to hybridize tissues for which we know the gene is not expressed and look at the distribution of the observed intensities. If a new sample is provided and we want to know if this gene is expressed, we would simply compare the observed intensity to the previously formed distribution. For a single lab, creating this training dataset is logistically impossible for two reasons: 1) we do not know what genes are expressed in which tissues and 2) it would require various hybridizations for each gene. Fortunately, a preliminary version of such a dataset already exists for some platforms/organisms. Raw data were obtained for more than a hundred tissues from the public repositories and pre-processed with the same algorithm1, 57. Now, for each gene the intensity distribution is determined. Because it is expected that any given gene will only be expressed in some tissues, multiple modes should be observed. It is assumed that the lowest intensity mode is due to lack of expression. Genes that are expected to be expressed are coded with ones and the unexpressed genes are coded with zeros. This information is referred to as the gene expression barcode (see the Supplemental Methods for more details on the statistical algorithm). Barcodes were created for 118 human tissues and 44 mouse tissues (Supplemental Tables 1 and 2). A dendrogram and heatmap displaying the human and mouse barcodes and related summary statistics can be seen in Supplemental Figure 1.

We compared the barcode to the present/absent/marginal calls from the Affymetrix Microarray Suite 5.0 (MAS 5.0). With MAS 5.0, only 10% of the 22215 genes represented in the human array achieve the same call in all samples within the same tissue. This number increases to 48% using the barcode approach (Supplemental Figures 2A–B). Similar results were obtained with mouse data (Supplemental Figures 2CD). To assess sensitivity we used results from an extensive study that reported proteins present in various mouse tissues8. We mapped these proteins and found that the barcode was more sensitive at declaring genes, approximated by proteins found in the tissues, present (Supplemental Figure 2E).

The utility of our algorithm was demonstrated by developing a classification scheme that assigns tissue types to unknown samples by comparing their barcode to predefined ones (using Euclidean distance). Various sample classification algorithms have been proposed for microarray data. Many of these were compared on the original expression estimates. Predictive Analysis of Microarrays (PAM) produced the best results (data not shown)9. We compared our approach to PAM using leave-one-out cross-validation (CV). Tissues with detailed annotation for which there were 3 or more samples were included. Table 1A shows the results, which include various clinical datasets1016. The barcode outperformed PAM in all comparisons except two, where it performed as well. Because CV has a tendency to overestimate the performance of a classification algorithm, we assessed performance on six independent datasets not included in the CV process (Table 1B). Here only the barcode performed well, with similar accuracy to that seen with CV.

Table 1
A) Percentage accuracy comparison: PAM versus the barcode approach in six data sets described in the text. In the normal tissues we assessed the ability to distinguish between tissues. For the Alzheimer’s data we assessed the ability to: 1) distinguishing ...

The fact that the barcode greatly outperformed PAM on the independent datasets is likely due to the lab/batch effect. Because studies usually target a particular tissue, a primary concern is that a strong lab effect will confound the ability to classify tissues from the ability to classify labs2. An example of the lab effect is shown in Figure 2A, where the correlations between samples from study E-AFMX-5 are high despite originating from a wide variety of tissues. The barcode approach can remove many of these effects because subtle changes in intensity values are not strong enough to make an absent gene appear present, or vice-versa. Notice that the barcode removes most of the correlations in the E-AFMX-5 study without removing the correlation between the brain tissues, both within the study and between studies (Figures 2B, 2C and 2D).

Figure 2
Demonstration of the lab effect and its removal by the barcode algorithm. The correlation between every pair of samples from studies containing 10 or more arrays is shown in different shades of blue (negative correlation in red). For each gene the across-sample ...

To assess the ability of the barcode algorithm to find undiscovered tissue subsets, we used data from three breast cancer studies that did not include normal breast tissue samples, but did include patient survival data1416. The distance to all tissue barcodes was obtained and 499 of the 500 samples were classified as breast tumor (1 as bladder cancer). When we took out the breast tumor barcode 37 of these samples were close to a variety of normal tissues and the other 463 samples to a variety of cancer tissues. We then formed good and bad prognosis barcodes using these 37 and 463 samples, respectively. This new barcode was then used to re-classify the 500 samples. We iterated this procedure until the good and bad prognosis groups did not change. The final barcodes resulted in a powerful prognosis tool that outperformed the methods described in the original papers. The details can be found in the Supplemental Results.

We expect the barcode approach to classification and discovery presented in this paper to improve in various ways. First, the classification algorithm implemented on the barcode was based on a very simple detection method and distance calculation. Many aspects can be optimized for prediction purposes. We expect the machine learning community will help improve this already powerful algorithm. Second, as microarray technology improves so will the barcode performance. In particular, the emergence of better gene annotation and arrays that probe for individual exons are the most promising developments. Finally, we have only implemented the barcode for two widely used platforms: Affymetrix HGU133A human array and MOE430 mouse array. However, as soon as enough public data is available, the barcode will be defined for other platforms.

In conclusion, we would like to acknowledge the efforts from the Microarray Gene Expression Data (MGED) Society to promote the sharing of microarray data. The work presented here would not have been possible without the existing public repositories. In particular, the availability of raw data was key as the methods used to process raw data into gene level measurements also contribute to study-to-study variability17, 18. We hope this trend continues, as we believe it to be necessary for microarray technology to fulfill its promise to help diagnose and treat disease.

Figure 1
Across sample gene expression estimates distributions. Data for two human genes are shown with the Genbank accession number on the y-axis. The left pane shows across sample distribution for all tissues. The vertical line is automatically drawn by the ...

Supplementary Material

01

REFERENCES

1. Irizarry RA, Gautier L, Cope LM. In: The Analysis of Gene Expression Data: Methods and Software. Parmigiani G, Garrett ES, Irizarry RA, Zeger SI, editors. New York: Springer-Verlag; 2003. pp. 102–119.
2. Irizarry RA, et al. Multiple-laboratory comparison of microarray platforms. Nat. Methods. 2005;2:345–350. [PubMed]
3. Kim S, et al. Multivariate measurement of gene expression relationships. Genomics. 2000;67:201–209. [PubMed]
4. Pal R, Datta A, Fornace AJ, Jr, Bittner ML, Dougherty ER. Boolean relationships among genes responsive to ionizing radiation in the NCI 60 ACDS. Bioinformatics. 2005;21:1542–1549. [PubMed]
5. Barrett T, et al. NCBI GEO: mining millions of expression profiles--database and tools. Nucleic Acids Res. 2005;33:D562–D566. [PMC free article] [PubMed]
6. Parkinson H, et al. ArrayExpress--a public repository for microarray gene expression data at the EBI. Nucleic Acids Res. 2005;33:D553–D555. [PMC free article] [PubMed]
7. Carter SL, Eklund AC, Mecham BH, Kohane IS, Szallasi Z. Redefinition of Affymetrix probe sets by sequence overlap with cDNA microarray probes reduces cross-platform inconsistencies in cancer-associated gene expression measurements. BMC Bioinformatics. 2005;6:107. [PMC free article] [PubMed]
8. Kislinger T, et al. Global survey of organ and organelle protein expression in mouse: combined proteomic and transcriptomic profiling. Cell. 2006;125:173–186. [PubMed]
9. Tibshirani R, Hastie T, Narasimhan B, Chu G. Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proc. Natl. Acad. Sci. U.S.A. 2002;99:6567–6572. [PMC free article] [PubMed]
10. Blalock EM, et al. Incipient Alzheimer's disease: microarray correlation analyses reveal major transcriptional and tumor suppressor responses. Proc. Natl. Acad. Sci. U.S.A. 2004;101:2173–2178. [PMC free article] [PubMed]
11. Kimchi ET, et al. Progression of Barrett's metaplasia to adenocarcinoma is associated with the suppression of the transcriptional programs of epidermal differentiation. Cancer Res. 2005;65:3146–3154. [PubMed]
12. Dyrskjot L, et al. Gene expression in the urinary bladder: a common carcinoma in situ gene expression signature exists disregarding histopathological classification. Cancer Res. 2004;64:4040–4048. [PubMed]
13. Lenburg ME, et al. Previously unidentified changes in renal cell carcinoma gene expression identified by parametric analysis of microarray data. BMC Cancer. 2003;3:31. [PMC free article] [PubMed]
14. Miller LD, et al. An expression signature for p53 status in human breast cancer predicts mutation status, transcriptional effects, and patient survival. Proc. Natl. Acad. Sci. U.S.A. 2005;102:13550–13555. [PMC free article] [PubMed]
15. Pawitan Y, et al. Gene expression profiling spares early breast cancer patients from adjuvant therapy: derived and validated in two population-based cohorts. Breast Cancer Res. 2005;7:R953–R964. [PMC free article] [PubMed]
16. Sotiriou C, et al. Gene expression profiling in breast cancer: understanding the molecular basis of histologic grade to improve prognosis. J. Natl. Cancer Inst. 2006;98:262–272. [PubMed]
17. Irizarry RA, Wu Z, Jaffee HA. Comparison of Affymetrix GeneChip expression measures. Bioinformatics. 2006;22:789–794. [PubMed]
18. Shi L, et al. Cross-platform comparability of microarray technology: intra-platform consistency and appropriate data analysis procedures are essential. BMC Bioinformatics. 2005;6 Suppl 2:S12. [PMC free article] [PubMed]

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...

Links

  • MedGen
    MedGen
    Related information in MedGen
  • PubMed
    PubMed
    PubMed citations for these articles

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...