- Raychaudhuri S, Stuart JM, Altman RB,
Principal components analysis to summarize microarray experiments: application to sporulation time series,
Pac Symp Biocomput. 2000:455-66
The paper uses PCA (Principal Components Analysis) to analyze
the data from a series of microarray experiments. PCA can help
reduce the dimensionality of the data, which helps in
visualization and classification. However, there are some issues
with the methodology, including the non-uniformity of the time
intervals (which could affect the PCA analysis because PCA is a
linear transformation) and because the time points are not
independent. However, PCA results help identify qualitative
changes and patterns.
[Editor's note: see also Yeung and Ruzzo, Principal component
analysis for clustering gene expression data.
Bioinformatics,17 (9) 763-774 (2001)
http://bioinformatics.oupjournals.org/cgi/content/abstract/17/9/763
for more on PCA and microarray clustering.]
- Nagahata T, Onda M, Emi M, et al.,
Expression profiling to predict postoperative prognosis for estrogen receptor-negitive breast cancers by analysis of 25,344 genes on a cDNA microarray. ,
Cancer Sci; 95(3): 218-225
This paper uses microarrays to try to find gene expression differences between women who died of estrogen receptor-negative breast cancer and women who survived for at least five years disease free. Unlike the paper discussed in class, this study was trying to find genes that behaved in significantly different ways in two predetermined groups. Thus the computational methods were more statistically based. The two tests used were the Mann-Whitney test and the Random-permutation test. These are both well defined statistical tests that have been used and studied for a number of years. Thus the statistic significance of the results could be shown in a well-established ways. The methods used to prepare the samples and microarrays were more complex. RNA was isolated from frozen specimens and amplified twice. I lack the background to determine how freezing and amplification could have changed or biased the results, but amplifying the RNA that much almost certainly added noise.
- Heun-Sik Lee, Mie-Hee Park, Suk-Jin Yang, Hai-Young Jun et al.,
Gene Expression Analysis in Human Gastric Cancer Cell Lin Treated with Trichostatin A and S-Adenosyl-L-homocysteine Using cDNA Microarray ,
Biol. Pharm. Bull. 27(10) 1497 - 1503 (2004)
The purpose of this study is to investigate the effects of Trichostatin A (TSA) and S-Adenosyl-L-homocystine(AdoHcy) on chromatin remodeling by histone acetalyation and methylation. The gene expression profiles of TSA and AdoHcy in a gastric cancer cell line using 14K cDNA microarray are analyzed using SAM (Significant Analysis of Microarray) and hierachcial clustering analysis. The study successfully identified 98 and 43 differentially expressed genes in TSA and AdoHcy treated sets, and selected genes were functionally classified using the Database for Annotation, Visualization and Integrated Discovery (DAVID, http://david.nuaid.nih.gov/David/). What's more interesting is that genes related to cell communication, cell death, kinase activity and metal ion binding were found commonly up-regulated in both drug treatments, while various nucleic acid binding proteins were among the group of gene products commonly down-regulated by both drugs, thus suggesting a close relationship between different histone modifications such as acetylation and methylation.
The researchers were quite thorough with their study design that they conducted each experiment six times to exclude labeling bias. In addition they performed RT-PCT to verify the reliability of the expression changes detected by the profiling analysis using the cDNA microarray. However only group of 3 -5 genes were selected for the validation test. In this study they only compared the differentially and commonly expressed genes between the two drugs, but further examination of the gene expression profile when both drugs are treated at the same time might provide even more valuable information.
- Long, Mangalam, Chan, Tolleri, Hatfield, and Baldi,
Improved Statistical Inference from DNA Microarray Data Using
Analysis of Variance and a Bayesian Statistical Framework,
The Journal of Biological Chemistry; Vol. 276, No. 23, Issue of June 8, pp. 19937-19944, 2001
This paper discusses some of the issues that we talked about in class with
regards to the statistics (or lack thereof) in the 1998 Chu et al paper.
Since high density array experiments are usually replicated only a few times,
they propose using a Bayesian framework, including a prior estimate, to compute
a more reliable estimate of the variance which is then used in computing the
t statistic to identify which genes show a significant change in expression
levels. Their data show that the Bayesian statistical approach identifies a
more consistent set of up- or down-regulated genese than a simple t test.
The prior is estimated "based on a local average of standard deviations for
genes showing similar expression levels to the gene under consideration.
Local averaging is carried out by ordering all genes within a given treatment
based on their average expression level and then taking the average standard
deviation as the standard deviation observed for any given gene and the k next
higher and lower expressing genes (where k is a user defined constant)."
They note that genes showing very high fold changes in expression often
showed little statistical support associated with their expression change
when using the Bayesian approach. They go on to say that "large nonsignificant
fold changesin expression often result from a single extreme observation likely
to be an experimental artifact." In addition they state that "it is difficult
to accurately measure and quantify genes showing very low expression levels
using high density arrays."
Further, they find that "the variance in expression level is a strong function
of the mean". In other words, genes with very low mean expression levels tend
to have a high variance. The authors state that, in their experience, this
is not peculiar to a particular technology but seems to be a general feature
of DNA microarray experiments. They argue that "fold change in expression
may not be a good proxy for statistical significance."
They conclude that "the incorporation of a Bayesian prior into the commonly
accepted t test allows statistical inferences to be drawn from high density
array data that is not highly replicated".
- Perou CM et al.,
Molecular Portraits of Human Breast Tumors,
Nature. 406: 747-752 (2000).
In this article, the authors analyzed 65 human breast tumor samples on a cDNA microarray containing 8,102 genes. They compared each tumor sample and their supplementary samples to a reference mRNA pool. On the resulting microarray data, they used a hierarchical clustering technique to cluster the samples as well as the genes. Interestingly, several samples were from patients before and after therapy. The tumors from an individual patient usually clustered together, showing that the expression pattern for a tumor was provided a molecular portrait. Clustering with a subset of genes resulted in the separation of tumor samples into four categories: ER+/luminal-like, basal-like, Erb-B2+ and normal breast. The authors also identifed gene clusters that relate to tumor features, including proliferation genes, as well as genes that were specific to cell types within the tumor. In addition, they provided some corollary experiments, including immunohistochemistry, to test the information provided by the array. Overall, this is an excellent array paper. The authors took an excellent and complete look at the microarray data available to them, using reasonable clustering methods including gene subsets. A logical extension of this work is to examine the clinical classification of tumors in order to determine the most beneficial therapy for a tumor type. The authors did correlate tumor expression profile with survival in a subsequent paper, but a more interesting analysis may be response to treatment.
- Nadon, R.; Shi, P.; Skandalis, A.; Woody, E.; Hubschle, H.; Susko, H.; Rghei, N.; Ramm, P.,
Statistical inference methods for gene expression arrays,
Proceedings of SPIE - The International Society for Optical Engineering, v 4266, 2001, p 46-55
As we all know, there are lot of uncertainties in micro-array experiments: replicate variations, image processing bias, and fold-changing thresholds. This paper, focused on interpretate the expression as a random variable which considers true intensity value, observed random error, and unknown systamatic shift. This paper proposed a relationship among these different factors and performed statistical measure of the mean, variance of the expression random variable. Statistical tests (t-test and z-test) were also carried out so that people can have some idea how confident a "true" difference in fold-change can be.
The authors averages all the spots on the chip due to lack of data points. This average may not be propriate all the time. We can use three level of variations instead of what presented in the paper: the one accociated with gene, the systematic shift across replicates and the long-range variation across the chip.
- Inada M, Guthrie C.,
Identification of Lhp1p-associated RNAs by
microarray analysis in Saccharomyces cerevisiae reveals association
with coding and noncoding RNAs.,
Proc Natl Acad Sci U S A., Jan 2004; 101: 434 - 439.
The authors used an apparently novel combination of
immunoprecipitation (IP) and whole genome microarray analysis to
identify all of the RNA targets of Lhp1p (the yeast version of La,
which is used in eukaryotes for various stages of RNA metabolism). The
idea was to pull out of the cell all the RNA that was bound by Lhp1p,
then run the resulting cDNA on a whole genome microarray as a means of
identifying the RNA Lhp1p interacts with. The results were
interesting. La is known to stabilize a number of Pol III transcripts
(which are typically non-coding RNAs, such as tRNA, rRNA, and many
small nucleolar RNAs (snoRNA)), and a large fraction of the known
substrates were detected, along with a number of new substrates. As
further validation, three uncharacterized regions were identified in
the analysis, all of which were later identified as snoRNA genes by
independent computational and Northern analysis. More surprising were
the 26 coding mRNAs that were found to be bound by Lhp1p. The
interaction of one of these mRNAs (HAC1 RNA) with Lhp1p was determined
by knockout experiments that used Northern and Western analysis to
analyze the stability of the transcript and protein under the adverse
conditions that typically activate Lhp1p. This analysis revealed a new
type of function for Lhp1p.
Controls and validation: The primary negative control used was to make
the microarray analysis two color, where Cy5 was used to mark the cDNA
coming from the IP procedure, and Cy3 was used to mark the cDNA coming
from a negative control IP procedure in which the Lhp1p protein was
not tagged with the antigen. This allowed them to establish background
RNA levels that were not the result of Lhp1p binding. Positive
controls included the high coverage of known Lhp1p-associated RNAs,
the independent computational and biological verification of three
proposed RNAs, and the in-depth biological verification and
description of the novel Lhp1p-HAC1 RNA interaction. Finally, 18 RNA
fold enrichment values were independently confirmed by calculating the
fold-enrichment of mRNA from tagged- to untagged-Lhp1p by quantitative
PCR (QPCR). QPCR is perhaps the most accurate way of quantifying the
amount of a particular transcript. The log2 ratio from QPCR was
reasonably correlated (r-0.8) to the log2 ratio derived from the
microarray analysis. The experiment was repeated twice and the log
ratios were found to have a correlation of r=0.85. As always, more
replications would have greatly increased statistical confidence.
From the biochemical side of things, I am surprised they did not
cross-link the Lhp1p protein to the RNA. They commented that the only
interactions they could identify were those with especially strong
binding affinities; the common way to increase sensitivity is to
cross-link, which is done in similar IP experiments with protein-DNA
interactions. It may be that cross-linking makes reverse transcription
into cDNA difficult or impossible.
- K.E.Lee, N.Sha E.R.Dougherty, M.Vannucci, B.K.Mallick,
Gene selection: a Bayesian variable selection approach,
Bioinformatics, Vol 19 no 1 2003, pp. 90-97
The authors use a Bayesian statistical model to determine the
most significant genes from expression patters. Part of their motivation
is that they say previous methods used too many variables (genes) for the
number of data points. Their method's results seem quite good and do rely
on fewer variables than older methods.
Extensions: The statistical model used here was quite simple, treating
variables as independent. They authors acknowledge this and say that they
could adjust the priors used to account for dependencies. Geoing beyond
trying to design a prior manually, this seems like a perfect application
for structure learning methods using graphical models. A graphical model
approach would seems like an excellent way to model the relationships
between different genes and between genes and expressions in this case in
general.
- De Smet F, Moreau Y, Engelen K, Timmerman D, Vergote I, De Moor B,
Balancing false positives and false negatives for the detection of differential expression in malignancies,
British Journal of Cancer, 2004, 1160-1165
The potential for microarray data to differentiate between subtly different classes of malignancies would be a powerful diagnostic tool, allowing physicians to customize treatment schemes for individual patients. However, the task of identifying genes that show true differential expression is difficult, especially considering the vast number of genes typically screened in a single microarray experiment. An important consideration in microarray analysis is the rejection level, alpha, which determines the level at which a difference in expression of a given gene is significant. Depending on whether the rejection level is set too high or too low, the likelihood of false positives (Type I error) or false negatives (Type II error) increases.
De Smet et al. propose a computational method by which to balance the probability of Type I and II errors. Using their method, microarray data are adapted to a receiver-operating characteristic (ROC) curve, which allows for the estimation of an optimal rejection level, as well as an assessment of the overall quality of the data. Using their methodology, De Smet et al. compare microarray data from different classes of acute leukemia. By calculating the area under the ROC curve for each of two previously reported data sets, they find that one set is of higher quality than the other, and is therefore more likely to yield reliable predictions of differentially expressed genes between two types of acute leukemia.
This method improves upon previous methods that attempted to reduce either the probability of a Type I error, or that of a Type II error, but not both. Further, it allows for the identification of high quality data sets from which differential expression in a number of genes is likely to be accurately identified. However, as the authors warn, this method typically results in somewhat large estimates for alpha, which means that the subset of genes that may be differentially expressed remains large. From a biological perspective, a logical extension would be to screen this subset for genes known to be differentially expressed, and then to determine whether there is a biological basis for the differential expression of other genes in this subset. Such an analysis might help to elucidate the roles of previously unstudied genes in a specific disease phenotype.
- John Shaughnessy, Jr, PhD,
Primer on Medical Genomics. Part IX: Scientific and Clinical Applications of DNA Microarrays -- Multiple Myeloma as a Disease Model,
Mayo Clinic Proceedings 2003;78:1098-1109
This article discusses recent progress in diagnostics and prognostics for
Multiple Myeloma (MM) a type of cancer in antibody-screening plasma cells (PCs),
by using DNA microarrays and data analysis. The article draws several
conclusions:
1) Unsupervised hierarchical clustering of DNA microarray-derived gene
expressions shows that
a) normal PCs can be distinguished from MM-PCs;
b) there exist 4 subgroups of MM (MM1, MM2, MM3, MM4)
2) The formation of MM is related to late-stage B-call differentiation. This
conclusion (by the author's admission controversial and in need of further
clinical studies for confirmation) was reached by performing hierarchical
clustering on B cells at different stages of differentiation, and noting that
the same genes that were expressed during the late stages are also common to
the gene expressed in MM.
3) Virtually all genes from chromosome 13 associated with MM are
down-regulated. This can have profound implications in diagnosis and in seeking
links to other types of cancer which exhibit the same pattern of
gene down-regulation.
4) Gene expression profiles can act as surrogates (predictors) for the results
of FISH (Fluorescent in situ hybridization) and point to possible
tumor-suppressor genes. This result was obtained via multivariate step-wise
linear discriminant analysis (MSDA) and confirmed via leave-one-out
cross-validation(85% correct prediction)
5) Additional testing using MSDA and leave-one-out cross-validation shows that
the effectiveness of certain experimental treatments can be predicted by
the expression of certain genes.
I had trouble reading this paper due to its heavy use of medical
terminology: I had to read it twice, and I still do not understand the biology.
This article is aimed at other biologists rather than
computational scientists, so the numerical methods used in supporting the
conclusions are mentioned only in passing, within the broader discussion. Since
I do not understand the biological relevance of, for example, two particular
genes being over-expressed together, I do not have a criterion to evaluate the
conclusions obtained via clustering of discriminant analysis. Something I
noticed, however, is the relatively low number of "samples" (meaning test
subjects). I might be biased by my "maximum-likelihood" way of thinking here,
but it is important to analyze the statistical significance of the results
when the number of samples is low. This also suggests that more modern data
analysis methods that take into account a small number of samples may be
better suited than more traditional techniques.
[Additional note: This article is part of an excellent primer on medical
genomics published by the Mayo Clinic. I skimmed the earlier chapters and found
that it provided an in-depth introduction the the biology and the technology
of medial genomics and I would recommend it to those seeking a more in-depth
introduction than some other introductory materials.]
- Troyanskaya,O. and et al.,
Missing value estimation methods for DNA microarrays,
http://www.cs.princeton.edu/courses/archive/fall04/cos557/Articles/missing-value-estimation.pdf
This paper investigated three different algorithms, SVD, KNN and Row Average, on dealing with the missing values in DNA microarrays. According to their results, Row Average has the worst performance and KNN is more robust than SVD to noisy data (more missing values). Moreover, the performance of KNN was shown to be less sensitive to the number of neighbors (the value of K). Generally speaking, KNN outperforms SDV on the task though SDV performs better than KNN on time-series data with low noise level. Sometimes, simpler algorithms are better.
I think, instead of using row average values as the initial values for the missing data in the EM iterative process for SVD estimations, they could use the values estimated by KNN as the initial guess. A better initial guess may lead to a better finalized estimation in the EM algorithm. On the other hand, since only the normalized root mean squared difference between the imputed values and the true values was used as the objective function, the different biological importance of the data was not reflected. It is hard to say an algorithms is biologically better the others.
- Furey TS, Cristianini N, Duffy N, Bednarski DW, Schummer M, Haussler D.,
Support vector machine classification and validation of cancer tissue samples using microarray expression data.,
Bioinformatics.2000 Oct;16(10):906-14.
As the title suggests, this paper by Furey et al. presents a support vector machine (SVM) classification
and validation of cancer tissue samples using microarray expression data.
The paper stresses the importance of a new analysing method for DNA microarray experiments in diagnosing
diseases while this analysis consists of the classification of the tissue samples as well as an
exploration of the data for mislabelled or questionable tissue results. The investigators used 3 samples:
ovarian cancer, normal ovarian and other normal tissues. In addition, they analysed two other previously
published datasets to reveal the robustness of their SVM method. Worth mentioning is the fact that they
compared their algorithm’s results to the ones of other learning methods (SOM, perceptron algorithm
and p-norm algorithm.
- Dharmadi, Y., Gonzalez, R.,
DNA Microarrays: Experimental Issues, Data Analysis, and Application to Bacterial Systems,
Entrez PubMed
Overall, I found this to be a very good survey of the uses of microarrays in the field of biological study. Particularly strong was its level of breadth - the paper not only covered the basic science behind how microarrays work, it also went relatively deeply into how microarray data analysis is conducted, and finished with real-life examples of microarray analyses unearthing revealing data on bacterial organisms.
Much of the beginning of the paper focused on the design aspect of microarray experiments and statistical analysis. Although no great detail was given on each step of experiment design, the authors made reference to many techniques that I thought were interesting and worth reading later on (particularly some of the statistical approachs, such as using a Bayesian approach).
The bulk of the paper, however, focused on showing the reader what information can be gleaned from microarray experiments, by providing numerous examples of microarray data in the discovery of baterial stress-testing (seeing what genes are expressed in B. subtilis and E. coli under harsh conditions, for instance) and host-pathogen interactions (using cholera and meningitis). On genome-wide level, the authors mentioned that microarray data could also be used to make revelations on gene organization, by detecting transcription factors within intergenic regions.
The one drawback I did see to this paper was that in terms of the application to biology, the authors did not go into too much detail in terms of how the microarray data was actually used to infer the information about, say, bacterial stress-testing or gene organization. That left me wondering exactly what techniques and methods they might have used to reveal this information (design-wise, statistically, etc). Overall, though, I found it to be very informative, and a great primer on not only the use and analysis of microarray data, but their application as well.
- I. Hedenfalk et al.,
Gene-Expression Profiles in Hereditary Breast Cancer,
2001 New England Journal of Medicine 344(8):539-548
I chose this paper primarily because it was cited as one of the top 7 core papers in the area of DNA microarray analysis. I also thought the application to breast cancer would be interesting. Overall, this was a good paper. It had a nice mix of biology and statistical/computational analysis. The microarray analysis was used to differentiate gene expression levels between two types of specific hereditary breast cancers and general, sporadic breast cancers. The methodology was very similar to what we saw in class for the Chu et al. paper. However, this study was very thorough in its statistical analysis. It compared three different statistical methods for generating lists of genes with different levels of expression between the cancer types to better confirm the results. It also used clustering to visualize the correlation of expression of gene groups between the different cancers. While the paper left out much of the detail of these analyses, it provided nice links to further reading and web page tutorials which I found to be helpful. I particularly liked how they used tissue microarrays to verify some of their findings. I had trouble following some of the biology jargon, but that was to be expected. My only complaint was that the number of patient's sampled for the test was rather low. Their results, as with the Chu et al. paper, could be strengthened greatly by having more samples and by repeating the experiment multiple times.
- Iyer VR, Horak CE, Scafe CS, Botstein D, Snyder M, Brown PO,
Genomic binding sites of the yeast cell-cycle transcription factors SBF and MBF,
Nature. 2001 Jan 25;409(6819):533-8.
The yeast transcription factors SBF and MBF activate certain genes during the G1/S phase of the cell cycle by performing regulatory functions once they have bound to specific DNA sequences in the promoters of those genes. In this paper, the authors seek to identify those specific sequences to which SBF and MBF bind by using an approach that draws from both
biochemistry and from bioinformatics. Their first step was to perform a biochemical technique known as immunoprecipitation, in which an antibody against a transcription factor was used to purify the transcription factor and the DNA to which it was bound from DNA to which the transcription factor was not bound. The second step was to apply this sample of DNA to a microarray containing the all of the intergenic (non-coding) sequences from yeast, where transcription factor binding sites are located. The third step was to analyze this data, and they did so in three ways. First, they examined how their set of sequences varied over several experiments to identify true positives (they hope). Second, given this set of sequences which hypothetically contain a specific subsequence to which one of SBF or MBF binds, the authors asked what this sequence might be, and used the programs MEME and Consensus (described elsewhere) to do so. (This methodology is not the only way to answer this question, but the authors' results were in line with the results determined by other approaches.) Third, they attempted to specify the particular function of SBF and MBF by classifing the genes that were regulated by each, and they propose a hypothesis for these functions.
- Xiangqin Cui, M. Kathleen Kerr, and Gary A. Churchill,
Transformations for cDNA Microarray Data,
Statistical Applications in Genetics and Molecular Biology, 2003: Vol. 2: No. 1, Article 4.
This paper discusses several mathematical methods to correct microarray data inaccuracies due to known variations in the medium. The transforms are applied in post-processing to normalize, smooth, and reduce variance among the data; after removing local bias inherent in the reading. A normal Ratio by Intensity (RI) plot should display data along a horizontal line although most exhibit some curvature or skew. Raw signal intensity data from an RI plot exhibit a higher degree of log ratios as the intensity increases although low intensity areas can display wide variability. Signal intensity models such as linear (array fluorescence corresponds to mRNA concentration) and RI plots are discussed along with sample graphs showing effects due to scanner saturation, additives and multiplicative error, background variance, slop and heterogeneity.
Transforms of acquired data from the microarray usually are run on the raw dataset to enhance statistical properties. Previously developed transformations including log, shift, and curve fitting are discussed with useful properties and limitations on each. Variance stabilizing transformations are reviewed which attempt to reduce error by keeping all data in the set (including outliers). The authors introduce the linlog transform which uses a linear transform for low intensity spots and a log transform for those of high intensity. By using linlog in conjunction with additional transforms such as shift-log, both variance and graph curvature may be optimized.
Simulated and actual microarray data were used with the various transforms to show suitability for correction of RI, IOR, and Bias plots. In actual lab usage, the balancing of channels in the scanning operation should be ensured as the transforms are best used to remove only the residual effects after acquisition. The paper concludes with recommendations on appropriate application of transforms to correct for RI plot curvature, variation, and spatial heterogeneity.
While this paper was thorough and did a good job of reviewing several well known effects, it would have been interesting to see additional information (particularly raw data) on the inputs to the transforms. It's difficult to visualize the application of the transforms without several examples with hard data points instead of just the graphs presented. In the interest of space, this could be expanded to either a follow-on paper or one well chosen example. Understanding how difficult it is to display intensities in the black on white paper medium, several illustrations would help exhibit the intensity variances as well as local clustering which would impact the results of the transforms.
- Ingrid Hedenfalk et al.,
Gene-Expression Profiles in Hereditary Breast Cancer,
N Engl J Med. 2001 Feb 22;344(8):539-48
In this study, the authors used cDNA microarray technology to identify different gene expression profiles for patients with diagnosed hereditary breast cancers caused by mutations in gene BRCA1 and BRCA2. RNA samples from 7 patents with BRCA1 mutation, 7 patents with BRCA2 mutation, and 7 patients with sporadic breast cancers were hybridized to microarrays with 5361 genes, and then statistical analyses were used to identify the genes that were expressed differentially among the three types of samples. They concluded that there are significant differences in the global patterns of gene expression for these types.
What interests me most is the various statistical methods they used to analyze the data, including
Compound covariate predictor for class-prediction
Leave-one-out cross-validation for estimating the misclassification
Random permulations of the class-membership indicators to determine the significance of the results
Modified F-tests and t-tests, weighted gene analysis, and mutual information scoring (InfoScore)
Agglomerative hierarchical clustering
Multidimensional scaling plot
I think one limitation of their study is that the number of samples is too small, and a larger number of patients for experiment and genes for anlysis might improve the validity and precision of the estimation.
- Mayer, H., Bilban, M., Kurtev, V., Gruber, F., Wagner, O., Binder, B. R., and de Martin, R.,
Deciphering regulatory patterns of inflammatory gene expression from interleukin-1-stimulated human endothelial cells.,
Arteriocler. Thromb. Vasc. Biol. 2004; 24:1192-1198.
This was a very well done study that effectively combined computational methods and traditional experimental assays for purposes of confirmation. I also think it was very interesting, particularly because it is exactly the type of research I would like to do, namely, using time series microarray data and in silico regulatory analysis to identify regulatory relationships between genes. The authors used three complimentary approaches for promoter analysis of groups of genes showing similar temporal patterns of activation or repression. One interesting result was the identification of an over-represented sequence motif in one cluster of genes that did not match any known transcription factor binding sites contained in TRANSFAC. By electrophoretic mobility shift, a protein present in IL-1-stimulated human endothelial cells was shown to bind to this sequence. The presence of this protein in IL-1-stimulated cells was 1.8 fold higher than in cells not stimulated by IL-1. The authors note that this protein could represent a novel transcription factor involved in inflammation.
- Schena M; Shalon D; Davis Rw; Brown Po,
Quantitative Monitoring Of Gene-Expression Patterns With A Complementary-Dna Microarray,
SCIENCE 270(5235), 1995 , pages 67-470
This paper is similar to, but distinct from, the paper by Chu et al that was discussed in class. While the paper from class sought to establish that microarrays can be used to scientifically measure what's happening in an experiment, they did so by measuring the level of gene expression relative to the first measurement they took. In the early (1995) paper that I read, the authors seek to establish that microarrays can be used to measure the level of gene expression in absolute, rather than relative, terms.
I find the essence of their approach both simple, and elegant. On the microarray plate, they put a spot onto it that will bind to something known to be absent from the sample being measured. They then add a known quantity of the normally absent material to the sample, and run the sample on the microarray plate like normal. Once that's done, they've got data on all the spots on the plate, including their measurements of the normally absent material. They then use their measurements from that material to calibrate their measurements of the other, unknown materials in the sample.
In this experiment, they used Arabidopsis as their sample source, and were attempting to measure the expression of 45 known Arabidopsis genes. The added some rat, yeast, and human DNA, all of which are known to be absent from the plant DNA sample. The rat and yeast DNA were used as controls ; a known concentration (1:10,000 (w/w)) of the human DNA was added to the sample of plant material and used to calibrate the readings that were obtained.
They tested the accuracy of their measurements several ways. The first was to use the well-established Northern blot test to verify that their microarray measurements were reasonable. The second was to compare and contrast two different types of plant tissue (leaf and root), and confirmed that the microarray measured significantly higher levels of expression of a gene that was previously known to be expressed more in the leaf.
The first experiment looked to measure overexpression of a gene (HAT4) in both wild-type plant, and a transgenic plant known to be overexpressing the gene. They ran a sample of each type of plant through the microarray, and got results that matched what they expected - the wild-type plant was expressing it as normal levels, while transgenic was expressing it 50x more. They then checked this result by using a traditional Northern blot test to measure the level of HAT4 expression in each of these two samples. The Northern blot test confirmed the microarray's measurements to within a factor of 2x. The paper said that "Expression of all the other genes monitored on the array differed by less than a factor of 5" between the two types of plants, although it wasn't clear to me what (if any) of these other genes had been checking using the Northern blot test.
The second experiment tried to measure differences in gene expression when comparing leaf and root tissue (it's not clear to me what type of plant the two tissue types were drawn from). A gene known to be light-regulated (CABI) was about 500x more abundant in the leaf than in the root, according to measurements taken from the microarray. This is consistent with what was previously known from earlier, biological experiments. Interestingly, 26 other genes were also found to change in expression by more than 5x.
There are a number of good aspects to the paper, and several aspects that I have questions about. On the good side, they used simultaneous, two-color hybridizations, put each target into multiple, albeit adjacent, spots on the microarray, and they verified their results using other means. I'm not sure whether using human DNA in the experiment was a good thing (since the Arabidopsis shouldn't have it), or a bad thing (since it seems like it would be easier to contaminate the samples). It seems bad that so many of the results are accurate only to within wide margins of error (5x, etc). Also, I don't quite understand how they're able to fully calibrate their measurements with a single spot. On the one hand, they do mention that they're assuming that they're detecting things linearly, but that would still seem to indicate that one would need at least two points to get good results. This is, of course, in addition to the normal concerns that microarrays have, such as variability in binding rates between genes & the microarray, genes binding to other, similar spots on the microarray, etc.
They specifically mentioned that they used 'simultaneous, two-color hybridizations, which served to minimize experimental variation inherent in the comparison of independent hybridizations." So when they compared the wild-type and HAT4 plants, or the leaf and root sample, they put both samples (each with a different marker) on the same plate, and ran the experiment once. This seems to be an undeniably good thing, since they have run the two experiments on a single plate, and thus don't have to worry as much about differences between plates introducing further noise into their experiments.
The fact that they put each gene target into multiple spots on the same plate helped to reduce the amount of noise in the data. They said that they put them into 3 spots (per target), but that they spots were next to each other, in a row. It seems like it would have been better to put the spots further away, so that any noise from local anomalies would be further minimized.
They also made a point of doing PCR on the all cDNAs, including the known concentration of human DNA, which seems good, since it will allow one to calibrate one's measurements to the original quantities in the sample, rather than calibrating it to what's observed in the post-PCR sample. This assumes, of course, that PCR affects everything in the sample equally. While this might not always be true, it seems like for these 48 cDNAs, primers can be picked to specifically amplify them.
They also made a point of verifying their measurements from the microarray with a Northern blot test, which is extremely good. It would have been good to try to measure some other genes using the Northern blot, and if they did so, to have been more clear about it in the paper. At the same time, it's possible that the paper is clear, and I'm missing it, instead.
On a somewhat more suspect note, most of their verification is accurate to within fairly ranges. The paper mentions that the level of expression of the leaf-specific gene is more than 500x what was measured in the root, but many other genes are measured to only within 5x. On the one hand, these biological experiments are often quite noisy ; on the other hand, it seems like it should be possible to get better results.
They didn't seem to verify the absence of the human DNA in the plant samples prior to their experiment. They did verify that the microarray gave them no reading on the rat/yeast material, which is good (and it seems good that they've chosen species substantially different than the plant to use as controls, and this would seem to reduce the likelihood of accidentally having something from the plant binding to some rat DNA), but given that they're humans, and used human DNA as a control, it seems like it would make the experiment more susceptible to contamination. My guess is that for a single experiment like this, it's probably not an issue, but would need to be addressed for a production system.
One thing that did irritate me is that they seem to be calibrating their measurements off of a single point of data – the readings they get from the human gene target. It seems like for any given gene, there's both a given amount of it in the sample, and also a function that describes at what rate it binds to the spot. In the paper, they assume it's linear, but even with the assumption that the binding function is linear, it seems like they'd need another point to figure out what that function is. If it isn't linear, then they'd need a bunch more points. In order to get these points, they'd need to find a number of different targets, all of which are absent from the sample, and put known, different, concentrations into the sample. All of which assumes that there's a single function describing the rate of binding of all these different genes.
Lastly, there are the normal concerns with using a microarray - What if the DNA in the sample binds to itself (instead of to the target spots)? What if it binds to multiple spots? These seem to be more generic concerns about microarray, though, and so aren't specific to this experiment.
In summary, this paper explains how the authors tried to verify that one can obtain quantitative (absolute) measurements from use a microarray, in addition to relative measurements. They did so by adding a target spot to the microarray which binds to something not found in the sample, then adding that something to the sample in a known amount, and then calibrating their measurements off of that spot's readings. They checked their results using both a Northern blot test, and against previously obtained biological knowledge, and produced experiments that are interesting, well-thought out, and fundamentally sound.
- Mark Schena, et.al,
Quantitative monitoring of gene expression patterns with a complementary DNA microarray,
Science 1995, Vol(270) pp467-470
The author selected Arabidopsis thaliana as the model organism by analyzing the 45 gene sequences from its cDNA library. The experiment setup of microarray was introduced and fluorescent method is used to characterize the expression level. The author compared the gene expression level between HAT4-transgenic and normal wild plant, highl similarity was found between micro array results and northern plot. The gene expression level of root and leaf are also compared.
It is an old paper at the moment when microarray was just shown to be a promising experimental tool. As Chu's paper, this paper just presented qualitive analysis of the data and lacked statistics. Considering the pioneer work of those biologists in mid-90s, it is still tolerable. At least it showed that it is feasible to analyze multiple gene sequence at high throughput.
- Chu, et al.,
The transcriptional program of sporulation in budding yeast,
Science. 1998 Oct 23;282(5389):699-705
I found that the paper, while heavily inundated with genome science, was
understandable to one such as myself who has no biological training beyond the university sophomore level. It was an invigorating intellectual process to follow the development of sporulation transcription built from the ground up with minimal preconceptions. The fact that identification of actively induced genes during sporulation increased by an order of magnitude from this study is certainly a strong quantitative testament to the import of utilizing temporal expression data to track developmental processes; this can be extended perhaps even to gametogenesis in vertebrates with similar homologs.
- Kuo WP, Jenssen TK, Butte AJ, Ohno-Machado L, Kohane IS,
Analysis of
matched mRNA measurements from two different microarray technologies,
Bioinformatics. 2002 Mar;18(3):405-12
Kuo et al set out to analyze how well the results from two different
mRNA platforms can be compared and synthesized. The main motivation
was to test reproducibility of microarray experiments and the
potential for combining data from different experiments. They compared
previously published results from Standford cDNA arrays to Affymetrix
oligonucleotide arrays on 56 cell lines using ~3000 genes. They also
compared multiple measurement methods for the cDNA arrays, including
ratio and non-ratio based methods. On average, no correlation was
found between the two Arrays (2=0.33). The correlation only slightly
improved when only "well-behaved" probes/genes were analyzed. In
general, correlations converged on r=0.1 as the lengths of cDNA probes
increased, the length of the target gene increased, the sequence
similarities between probes increased, and the spot intensities
decreased. Correlation was more dependent on qualities of the cDNA
chips than the Affy chips. There was no difference between ratio and
intensity measurements of the cDNA chips. Only a couple same cell line
replicates were available, and the correlations were highly variable.
Comparisons of clustering results from hierarchical clustering
(Euclidean distance) showed no statistically significant similarities
between the two data sets.
This paper was very thorough in its analysis. Underlying statistical
assumptions (eg Normality) were tested and adjusted for, and many
conceivable sources of variation were accounted for. Unfortunately, as
the analysis was on previously published data from different labs (one
each for cDNA and Affy), it was impossible to measure the variation
due to lab-specific protocols. At the time of publishing, this was the
first large-scale microarray platform correlation analysis. It would
be interesting to extend this to other platforms and to do the
experiments in-house so that protocols could be tightly controlled. On
the other hand, it would be interesting to do a similar analysis on
the correlation between similar experiments on the same platform
between different labs to test the reproducibility of microarray
experiments in general. On the biological end, this paper suggests
that more research needs to be done on the chemistry of hybridization
etc in order to understand the dynamics of microarrays and the source
for such egregious variation.
CSE Home