Microarray Analysis - Science topic

Simran Venkatraman Premanand A. Jagajjit Sahu thank a galaxy dear friends

Sir can you share a tutorial or workbook or something like that. I can work in R, Python and Linux.

Thanks

For PCR an expression value of > 200 is sufficient, I think. Here it is not a question of Fold Change (FC): The same FC with an expression value of 10 vs. 20 would be hardly detectable with PCR or other techniques. With expression values of 200 vs. 400, this is surely much better (identical FC of 2).

What is the software that you are planning to use for the analysis? Depends on that how you write the contrasts would vary. If you plan to use "limma" (https://www.bioconductor.org/packages/devel/bioc/vignettes/limma/inst/doc/usersguide.pdf) read the user manual and it is easy to understand. As long as you have 2 or more replicates you can do differential expression.

Regards,

Shajahan

Anyone have the actual citation for these? They are dead links all of them!!

Kind Regards

R,

As far as I know, there is no data portal for GEO allowing the simultaneous analysis of multiple microarray datasets fitting particular criteria (such as breast cancer cell line). You will probably have to identify the datasets of interest, download the microarray data, then apply an R package for microarray analysis on each dataset, then analyse the combined results. GEO does offer a web analysis tool for many microarray datasets though (GEO2R) which runs the R code for you.

You can search for your question through the following link:

Hi Tunc Morova , I am facing a similar issue right now, analysing the PAIR files. Did you have any luck?

tRNA microarrays were initially developed to assess the aminoacylation status of specific tRNAs. Although it is true that tRNA deep-seq techniques have problems (mainly due to modified bases), microarrays also present serious limitations in terms of detection sensitivity. My advice would be to use one of the several strategies for tRNA library preparation and go for small RNA sequencing.

https://www.networkanalyst.ca/. NetworkAnalyst software may help you.

You may use NetworkAnalyst software.

It is likely that the authors did not even try to check the sample and hybridization quality. To assess the sample quality you would need to have a look in the lab books or talk with the people who actually did the processing (might not be possible!). To assess the hypridization quality you need to create diagnostic plots of the raw data. I am not too deep into Illumina arrays. Boxplots of signal intensities, possibly startified by probe type (pos/neg controls, regular genes) should be a minimum. I don't know if spatial plots are possible and if unspecific/background signals are considered.

Batch correction has been described from biologist point of view in this paper: Unbiased data analytics for biomarker discovery in precision medicine.

In brief, better to have some internal controls and use comBat algorithms. If you don't have internal control, there are some algorithm too. See the paper.

The fold change or log-fold change can be used as a measure of effect size in

high throughput experiments including gene expression analysis. However, the statistical significance obtaine from repetitions is another important number. One can used multiple testing adjusted p-values or false discovery rate values and ofte it is then advisable by plotting everything as a volcano plot, where log-fold change of each gene is on the x-axis and the log10 p-value (adjusted) is on the y-axis. I recommend the limma package in R to do such an analysis.

Please see

I think if a particular study has been done for quite few times and you just want to look for the some of the facts than you can chose few replicates.In addition to this it is important to find the truth first and then consider the numbers.

Dear Humaira

Did You try to add "header=T" to Your first command (file upload)?

But in general, as Jan Lisec said, carefully inspect Your data on "NAs" instead of gene expressing values. If You have troubles to do this in R, You can do it in Excel too.

Hello Priscila,

Since you mentioned about using a GUI and not having a manual, you may try contacting the developers for the best explanation (possibly may have a look at the source code), to be exactly sure what the pearson coeffs and q values related to. In the mean time, you may have a look at the following studies where they have explained the importance of various correlation factors in RF-MDA:

https://blog.datadive.net/selecting-good-features-part-iii-random-forests/

Thanks

You can check out following two resources for gene expression analysis:

Another option is, you can browse datasets of Arabidopsis on GEO (https://www.ncbi.nlm.nih.gov/geo/). It host most publicly available datasets. It predominantly stores raw files from different transcriptome experiments. Once you have listed down GEO IDs of your interest, you can yourself analyze these datasets with click of button using portal GEO2R (https://www.ncbi.nlm.nih.gov/geo/geo2r/). It hardly takes minute and you can store result of each analysis you do. You need not have computational background to execute this program. Its as simple as plug n play.

I hope it helps !!

All the best.

Hi Phung,

I guess it's impossible to analyze these two types of affymetrix arrays by simple command lines. in fact U133A and U133B share only 168 probes among the 22k sondes in both designs. take a look at this post (https://www.biostars.org/p/283639/). in fact both designs are sold to be used as complementary and can't be compared.

fred

Its a while since I used this. If you refer to micro array data, here is my script for HuGene-1_0-st-v1 arrays (easily modifiable to other versions). Im using R. Just download all .CEL files to a folder, and set this folder as your working folder.

library("oligo")

library("pd.hugene.1.1.st.v1")

celFiles <- list.celfiles()

affyRaw <- read.celfiles(celFiles)

eset <- rma(affyRaw, background=TRUE, normalize=TRUE, subset=NULL, target="core")

library(annotate)

library(hugene11sttranscriptcluster.db)

annodb <- "hugene11sttranscriptcluster.db"

ID <- featureNames(eset)

Symbol <- as.character(lookUp(ID, annodb, "SYMBOL"))

Name <- as.character(lookUp(ID, annodb, "GENENAME"))

Entrez <- as.character(lookUp(ID, annodb, "ENTREZID"))

theProbes <- exprs(eset)

df <- cbind.data.frame(theProbes, ID, Symbol, Name, Entrez )

write.table(df, "something.csv", sep=";" ) # Nice to have a spreadsheet of expression values

Now you can use LIMMA.

pData <- Here you have to make a .CSV file with phenotypic data, which is described with the dataset.

library(limma)

affy <- read.table(file="something.csv", sep=";", header=T, row.names=1 )

library("Biobase")

eset <- ExpressionSet(assayData=affy)

Make the design as you which, E.g:

design <- model.matrix(~ timepoint + ID)

fit <- lmFit(eset, design)

fit <- eBayes(fit)

topTable(fit, coef="something", adjust="BH", n=Inf, confint=T)

Dear Sherif

There is no need to bother about A260/A230 ratio when you have a good RIN value. If your miRNA purity and concentration is good you can proceed for sequencing. I have got a good miRNA microarray result even from FFPE tissue which got poor RIN value. All the best.

due to the high cost of gene expression chips, it is often with limited number of samples to compare across groups. That's why bootstrap is usually used for gene expression comparison.

Dear Pranita Kamble Waghmare,

Thank you so much for your answer. It is really helpful.

Dear Francisco,

For quick GO enrichment in multiple gene lists at once, I use g:Cocoa from g:Profiler:

Its main problem is that usually the analysis goes too deep into the GO annotation and the results might be hard to "digest".

G:Cocoa also allows other enrichment analyses such as KEGG pathways.

For broad functions enrichment I use GOTermMapper:

Another useful online tool for GO enrichment is Metascape:

To analyse protein interactions / genes network from a list of genes I use Genemania:

Best,

Gautier

Machine learning works on a simple rule – if you put garbage in, you will only get garbage to come out. By garbage here, I mean noise in data.

This becomes even more important when the number of features are very large. You need not use every feature at your disposal for creating an algorithm. You can assist your algorithm by feeding in only those features that are really important. I have myself witnessed feature subsets giving better results than complete set of feature for the same algorithm. Or as Rohan Rao puts it – “Sometimes, less is better!”

Not only in the competitions but this can be very useful in industrial applications as well. You not only reduce the training time and the evaluation time, you also have less things to worry about!

Top reasons to use feature selection are:

@Hasan, I also recommend stackoverflow.com for all your R questions, it is by far the best and most up to date forum.

Hello

Kindly look at links here , may help you

Good Luck

If you need it in Excel, why don't you do it in Excel?

If R, you get the row means with rowMeans(). To get the variances you will have to apply() the function var() to the rows. Here is an example code, assuming that the data is in a 54675x17 data.frame or matrix "df":

rm <- rowMeans(df, na.rm=TRUE)

rv <- apply(df, MARGIN=1, FUN=var, na.rm=TRUE)

Both, rm and rv, will be numerical vectors of length 54675. You can save them as a csv and import it in Excel. You can also add them as new columns to df (df <- cbind(df, "Mean"=rm, "Variance"=rv)) and save the entire df object. If you use the package openxlsx, you can use write.xlsx(df, file="new filename.xlsx") to save it as an xlsx file.

Not directly. The might be comparable to a certain degree after a common normalization, but even then it is important to consider the experimental differences (cell types/extraction, passage numbers, RNA isolation, amplification and labelling procedures). It is very dangerous to interpret differences between groups that are taken from different experiments/series, as there might be a serieous confounding of experimental conditions and the groups.

Thank you for your answer and time Dr. Joachim. The link that you send helped me a lot and as you said the number 54675 probe set is true as well. Thanks again.

Thank you Marcus for your suggestion.

Hi Lin,

First of all please keep in mind that the MA (ratio intensity) plot is meant to compare two or two group of samples. It concludes how different your samples are in terms of signal intensities (in microarray) or read counts (in RNAseq experiments). If you try to plot sample A versus sample A (same sample) you will notice the data points converge to zero at Y-axis because log (A/A) is zero. Having said that I would recommend you to https://en.wikipedia.org/wiki/MA_plot for basics.

Coming back to your plots it look like that the data is not correct or not normalized correctly (as other people pointed out above). If it is the RNAseq experiment how come the average readcount (x-axes in all your plots) be negative??

I would suggest first normalize or pre-process your data correctly and use any of the bioconductor tools (EdgeR or DESeq2) .

Best

It seems you did not scale your data when you plotted the heatmap, and you did not use hierarchical clustering to show the relationship between your samples. So your heatmap is full of low absolute expression probes, but not the relative expression after scaling.

The low expression of microarray data is not common, but for RNA-seq is very often. Your data should be microarray data, so the best way is to obtain CEL file or other raw data from the author to repeat the normalization process by yourself.

COMBAT is a useful tool to correct batch.Or simple you can just adjust it in your model without COMBAT.

Try xena.ucsc. You will fall in love with it.

mean and sd are calcuated using all 61 samples.

Make sure you use log(signals) to calculate the z-scores.

Note that z-scores are usually calculated to show co-regulation patterns or to identify samples with differing (outlying) profiles in heatmaps. If you have just want to identify the genes with the best statistical evidence for differential expression between the groups, then you may simply run t-tests (again, using the log(signal)-values).

Depends on what your aim is, If you really want to validate the overall screening result, you should select genes at random.

If you just want to confirm that those genes you aim to continue to work on are regulated as estimated from the screening, you should better use these genes. But this is still only a technical validation. If there are not too many interesting candidate genes and if possible, you should go for a biological validation using appropriate experimental assays to determine the biological role of the regulation (knock-in, knock-out, inhibitors, enhancers).

Thanks mam. I have received data from DMET analysis. How will I present the data in various representations? What all software and applications will help in the same?

The current release of Bioconductor is version 3.5; it works with R version 3.4.0. Users of older R andBioconductor users must update their installation to take advantage of new features and to access packages that have been added to Bioconductor since the last release..

source: https://www.bioconductor.org/install/

regards,

Milan

RNA-Seq is a powerful tool if you're trying to detect novel transcripts/splice forms, go on an unbiased "fishing trip" for genes/biomarkers, detect extremely rare transcripts, or look at changes in transcript abundance that occur over a very wide dynamic range. 

However, if you're interested in studying the expression of a known panel of transcripts and none of them are expressed at an extremely high or low level, a well-designed microarray will work just as well and cost less. Microarrays can be used for most of the same experiments as targeted PCR primer sets for RNA-Seq.

Hi,

I never use before, but I see good heatmaps from this website:

Furthermore, you can use excel to do heatmaps, I think taht you need to asing a color grade to a value range:

Good Luck!!

I agree with Thomas, lear R is a great tool to do bioinformatics analysis and figures.

Dear Abhijeet KISHANPAL Mavi,

I will recommend that you use Perseus for your statistics. Perseus is a fairly userfriendly "click" software to do the kind of statistics I understand you are planning to do.

The program can be downloaded from the link below:

The developing group each year hosts an excellent summer school, which I really can recommend. However, if attending the summer school is not an option they also publish all the lectures on YouTube. Try and search for e.g. "Perseus summer school iTRAQ" for some instructions.

I hope this helps and good luck with your research.

Limma

Hello Raghuram Sir,

As I understand your question, you have read counts of the control and treatment samples but not the list of differentially expressed genes. Therefore, first, you have to perform the differential expression which can be done using Cuffdiff, DESeq or any other program. However, these programs are Linux based. I have no idea about any windows based program for differential expression studied. After performing the differential expression analysis you will get the list of genes upregulated/downregulated in the treatment vs control with the level of significance in term of P and Q values. Now with this list you can make venn diagram using the program 'Venny' . This Venny program is online tool and very simple.

All the Best

I think you can trust the "Total conc" readout.

R is a good option for open source analysis tools of a wide variety of data. If you were using GeneSpring I'm assuming you may  want to analyze Agilent array data. There are specific packages for different Agilent arrays, such as https://www.bioconductor.org/packages/release/bioc/html/agilp.html, https://www.bioconductor.org/packages/3.5/bioc/html/AgiMicroRna.html and https://www.bioconductor.org/packages/3.5/bioc/html/LVSmiRNA.html to name a few.

I have some sample code here for a paper, where the data is downloaded from the GEO database and analysed using Limma. You can modify some sections, and use the Bioconductor Array Express package instead of Geoquery. 

Hope it helps. Good luck.

we may perform Gene Ontology enrichment analysis for specific outcome in that sub-network  and correlate to the computational prediction. but before doing that we could have some clue of evidence about that specific PPI in databases like String, then we could do comparative analysis.

these steps are helpful for exporting the excel file

you can just input that excel file to this jar tool and it will do that query

This paper should answer most of your questions:

Anders, S., McCarthy, D., Chen, Y., Okoniewski, M., Smyth, G., Huber, W., & Robinson, M. (2013). Count-based differential expression analysis of RNA sequencing data using R and Bioconductor. Nature Protocols, 8(9), 1765–1786. http://doi.org/10.1038/nprot.2013.099

If you want to look at some example scripts, you can see them in my github repository for a recent data set we analysed. https://github.com/uhkniazi/BRC_Organoid_Joana

Once you get the count matrix, you have various options for performing any sort of differential expression of genes - EdgeR, DESeq2 etc. 

Hi Yee, 

Basically, using iTRAQ you can measure the absolute and relative quantitative ratios of peptides and proteins. 

Maybe you should take a look this paper. You can find more detailed answers of your questions. 

Unfortunately, there is no good general answer to this.

There are genes for which a slight regulation is biologically relevant, and others that can be considerable regulated without much biological impact.

The p-values and whatever you get from an FDR-based selection depends not only on your selected cut-off but also on the sample size.

A strategy to select "candidates" could involve two steps:

1) select the top 50 genes with largest abs. LFC and also the top 50 genes with the lowest p-values. Some may overlap, so you get a list of at most 100 genes.

2) go through this list and make a subselection based on your biological understanding of the genes in the experimental context.

Then you should have a list that should allow you to get an idea what experiments to plan next.

we may use  this shiny app

as well as plenty of R packages like

Thank you very much Mr. Kamstra and Dr. Muley! 

I am looking at human samples from infant cord blood lymphocytes. 

So far, I have a smaller list of CpG sites that are relevant based on p-values and fold-change (as what Dr.Muley suggested). 

Mr. Kamstra, I have associated CpG sites with genes using GenomeStudio, but would I have to use biomart to convert these gene names to ensemble IDs or entrez IDs? 

Thank you for your help!

Check out these tools, may provide you desired data

Hello Tinku,

First, you have to divide the FPKM of the second value (of the second group) on the FPKM of the first value to get the Fold Change (FC). then, put the equation in Excel =Log(FC, 2) to get the log2 fold change value from FPKM value. 

Batch correction only works when the experimental conditions are (nearly) evely distributed over the batches. Otherwise, batch correction models will make things even worst (see link).

This is a severe problem when selecting data from different experiments, where many experimental issues are 100% confounded with the batches/experiments. If additionally biological groups are counfounded with experiments (like: I take group A from GSEx and group B from GSEy and compare A against B), the result is almost completely arbitrary (usually, results will already cluster nicely within experiments/batches, indicating considerable (but artificial!) differences in the expression profiles between the groups, but after batch-correction, these differences will be even exaggerated.

Further, if there are many different batches and relatively few samples per batch, a correction with combat will lead to an overestimate of the residual degrees of freedom for tests for DE (what might be corrected manually, but ignorance would be a bad guide). This prblem is also discussed in the attached link.

@Björn Abendroth typhoon 9410

thank you @ Audrey.  

Thx :)

BSA and SDS salted out when it storaged at 4 degree, but it's completely solved at RT.

Hi Xiao-Tao,

DESeq2 and edgeR, indeed, have a problem when global changes come into play (e.g. if cells produce different overall amounts of RNA under different conditions - a fairly normal situation among prokaryotes), because they actually assume that the majority of genes DO NOT change. As you said, there are apparently no genes that keep their expression levels under all possible conditions in all possible genetic backgrounds, even ribosomal RNAs vary hugely between different growth phases.

I can see to possible solutions here. 1) Use spike-in RNA. When you know from how many cells you isolate RNA, the spike-in will always permit you to estimate the actual abundances of transcripts independently of the depth of sequencing. 2) Use RNA-seq in conjunction with other quantitative techniques (e.g. northern blotting or RT-qPCR) applied to the same samples. Again, if you know from how many cells you isolate RNA and you deposit on the gel accordingly (e.g. 1/10 of the amount you isolated, whatever it be), by probing for several RNAs you will get a precise idea about how the abundance of these RNAs in one sample relates to that in another. Like this you will derive normalisation factors and will be able to account for the differential sequencing depth.

In next gen sequencing the first crucial step is mapping your sequence fragments, the reads, to the reference sequence (genome or transcriptome). However, reads are short (typically 75 bp or 100 bp) and it is not always possible to get a single best hit.  It is common to use only those reads with a single best match (= "unique hits") in downstream  analyses.

RNA seq is different from microarrays. Microarrays are based on hybridization and produce a continous fluorescence signal, whereas RNA-seq is based on short read sequencing and mapping and produces as a discrete signal the number of unique reads mapped per gene. The normalization and statistical analysis for RNA seq differs from microarrays. The most commonly used analysis packages for RNA seq are EdgeR and DESeq2. You can find those on Bioconductor.

Hi Mathavan,

If you are asking about this particular case, it can be explained by the fact that INO80B and WBP1 genes can sometimes occur as a read-through transcript that cannot be distinguished from either of the genes individually by the probes in question (see https://genome.ucsc.edu/cgi-bin/hgc?hgsid=579057461_Tm25CaPexEPPpKrMhRyrGSqaNmCR&c=chr2&l=74456725&r=74462493&o=74455022&t=74460891&g=refGene&i=NR_037849).

You can also look up more information about the U133 array's probes sensitivity and specificity to certain transcripts to understand why this could be for other probes and genes. For this particular example you can look at the information for probe 65133_i_at here https://genecards.weizmann.ac.il/cgi-bin/geneannot/GA_search.pl?keyword_type=probe_set_id&array=HG-U95&target=genecards&keyword=65133_i_at and for the 223072_s_at probe  see https://genecards.weizmann.ac.il/cgi-bin/geneannot/GA_search.pl?keyword_type=probe_set_id&array=HG-U133&target=genecards&keyword=223072_s_at.

I guess separate analyses need be done for men and women as men have XY chromosomes and would differ from the normal (MFA) analysis.

I think software PLINK should help you do the needful.

There is plenty of literature on GWAS in different organisms. The field is exploding. For fruitful guidance in your case, a journal like American Journal of Human Genetics would be very useful to search problems and solutions similar to yours.

Dear Denise,

the basic problem of high-througput data (common to metabolomics, transcriptomics and protein arrays) is the huge number of variables (protein species in your case) with respect to statistical units (the 14 samples) that open the way to a plethora of chance correlations. Thus the main lane IS TO EXCHANGE THE ROLE OF VARIABLES AND STATISTICAL UNITS. Simply operate on the transpose of tour original data matrix, i,e. the matrox having as rows the protein species and as columns (variables) the samples. On this matrix operate a Principal Component Analysis (that allow a dual representation of the same data set in terms of loadings (correlation coefficients of the variables with components) and scores (values of different component for each sample). So operate this pC, if you have 14 variables (7 control + 7 treatment) you will have in principle 14 components but, due to the mutual correlation between the abundance of different protein species you will end up into very few (2 or 3) principal acomponents explaining the by far major part of total variance.

Look at the component loadings and you will surely get a PC1 with all loadings of the same sign (size component) this is the signature of a common global profile shared by all the samples. Then go to PC2, PC3..PC4. You must see if there is a component in which the loadig values are significantly different between control and treated (ideally you will get  a component in which control and treated have opposite sign loadings) allowing for a perfect separation of the two groups in the loading space (shape components). If this is the case you will go to the scores of the discriminating component and look for protein species having the higher (in absolute value) scores on the component : those protein species are the ones allowing for the separation of the two groups and you solved your problem, If you will run a PCA on the correlation matrix you do not need to normalize protein values, normalization is implicit in the correlation metrics.

See:

What "difference" do you mean?

- a difference in distribution?

- a stocastic difference in magnitude?

- a more specific difference of the distributions (e.g. difference in concentration, in variation, or in some quantile?)

Hi, sorry for kinda late reply.

I think your suggested workflow could work. Be careful when imputating data. Always check the histogram of log2 LFQ intensities that you do not introduce too many data (e.g. for normal distribution). You would be able to see that by a kinda bimodal distribution. I think in your dataset this will be difficult when comparing different isolation methods because alone by comparing 4 and 1 you would have to imputate more values than you actually measured in method 4. Again, comparing between similiar isolation methods is fine I guess. :)

To compare different methods I really would simply compare the numbers of quantified proteins for your replicates.

For ibaq- indeed it was "developed" for aboslute quantification but it is also used for relative comparision in some studies and immunoprecipitations.

Best

Hendrik

Hello

Look at this link 

Good Luck 

Have a look at

That depends on the probe sets.

If they are against different isoforms / splice variants, the difference may be due to a variant-specific expression. If the probes are designed against different builds of the genome/transcriptome there might be a more "reliable" build.

Eventually you may consider sequencing the respective genes and confirm the regulation with a different assay (ISH, northern blot, qPCR, ...).

If you are sure that these genes are in the operon and can confirm these results using qRT-PCR, then the difference can be due transcriptional attenuation, e. g.  premature termination of transcription. Transcriptional termination is one of the mechanisms to regulate gene expression.