When you say you have no replicates, do you mean you have a couple of patients (3+) with exon data on tumour and control samples from each patient?

If that is the case then you can do a paired t-test or SAM where the tumours are 1 class and the controls the other.

If you however have only 1 patient, then any statistics is going to be very difficult. You need to have at least 3 (the bare minimum) for some 'reasonable' statistics.

Can you give us more details on your experimental setup?

So you have 2 samples coming from the same patient (tumor vs. normal)? In this case I doubt that you can get any statistically interesting results out of the analysis, since you have no way of discriminating technical variation from biological variation.

You could try to find related data sets with the same comparison on GEO (for example) and then include your arrays in these data sets via a fRMA (frozen RMA), to increase your sample size.

But you have many patients? Then you can try to treat patients as "batches" and correct for it, see:

http://www.biomedcentral.com/1755-8794/1/42

http://biostatistics.oxfordjournals.org/content/8/1/118.abstract

(better with ComBat than pamr.batchadjust).

If just 2 samples - calculate the fold change, set high threshold and keep your fingers crossed that there are not too many false positives in the top-list.

Btw - you have exon array - you can check it separately on gene and exon levels (use eg. Brainarray CDFs) to see splicing...

if you have multiple patients then this is your replication, which is biological: you only ever need biological replication with Affy arrays.

As stated before, technical replicates are often a waste of time.

For patient-wise paired comparisons, calculate, for each gene, the patient-wise logFC (log fold change). Remember that a paired t-test is, in fact, a regular t test using the difference within the pair as the variable.

Having multiple patients and one profile (the logFC) per patient makes the whole analysis straightforward.

However, if you really have a severely limited number of replicates, you can use the limma package which implements a moderated t-statistic borrowing variance estimates across genes, thus allowing you to use effectively fewer replicates. But beware of the limits of your conclusions if you have only few biological replicates!

How many patients? This will be key. If you many say more than 10 then this might be possible.

You can use the normal samples to generate a distribution of each gene for non cancer state. Next you calculate the ratio of a cancer sample to the mean of the non cancer samples. Since you have a distribute you can calculate the t statistic of the observed value, like a z test.

You could also use expectation maximization clustering or other principal component analysis to group the patinet samples in to subgroups. Once in sub groups they can be treated as replicates and you can use Lima to calculate the differences between the subgroups and the no cancer samples. You may need to check the non cancer samples to ensure that they fall in a single group and have a normal distribution.

Researchgate needs to create a policy for a minimum experimental information. If you post here please give some experimental details. Number of samples, replicates, array platform (you gave this), type of samples, you gave these also. This is especially important when asking questions about statistical analysis, it will greatly affect the answer I would give.

Checkout this publication lapointe 2004, PNAS. Vol 101 issue 3 pgs 811-6. he does a similar analysis although the selection and validation of biomarkers are statistically flawed as he uses all samples to build the model then uses some to test. This is discussed in Lee 2008 statistical methods I medical research. Vol 17 issue 6 pages 635-42

My suggestion is for the case that neither the technical nor biological replication is available, just as your case.

Certainly, it is impossible to assess the significance for the fold change value computed from expression in the exon level or gene level, because technical variation and biological variation cannot be discriminated.

However, the analysis can be carried out in gene set level. Actually, a gene set level analysis was commonly following a gene level analysis so as to interpret the gene level results. Therefore, for you case, the best way is to conduct an over-difference test on gene set level based on the fold change values of genes. The gene set can be defined using GO annotation or KEGG pathways. It is also fine if you have a set of marker genes, you can define one or a number of gene set yourself.

The over-difference test of a gene set can conducted using non parametric test so as to avoid the issue of the null distribution for the fold change values.

Firstly, the fold change values should be presented on log2 scale and normalised by variation stabilised normalisation (VSN). This is important because it removes the dependence between variation of log2 fold change and expression level, which is commonly seen in microarray data.

Secondly, the absolute values of the normalised log2 fold change are ranked in descent order. Then, for a gene set, by taking the rank sum across the genes in this gene set as test statistic, a lower tail p-value can be evaluated by Wilcoxon rank sum test.

The significant gene sets hence can be extracted based on the p-values or adjusted p-values. Once these is done, for any a significant gene set you can say that an over-difference in gene expression is presented between tumor and normal samples.

However, due to lack of biological replication, the analytical results may be not generally correct in the population level. If there is import findings from the data, further confirmation study needs be conducted.

You need to subgroup your tumor samples, cancer is a heterogeneous disease, there is no biological rationale to group these samples unless you know that they are caused by similar underlying molecular changes.

dear dr ABD Allah you are in a university with high quality devices i think may be have a array spotter to make a slides with a low price than Affymetrix

no enough tissue is a big problem .

we can cooperate to get sample from Egypt.

Thank you all for great inputs. I have 20 patient samples, so enough biological replicates. One suggestion that is resonating here is to group the patients. I will try doing that. The objective of doing patient wise differential expression analysis was to link it with patient wise analysis for cytogenetics,epigenetics etc., As we all might have experienced, each patient has a unique genetic landscape and that is mounting evidence in favor of personalized treatment. Grouping patients does serve the statistical standards but compromise on the biological uniqueness of each patient. Please point if I am wrong ! More suggestion are welcome.

Just seen this thread. You've got tumour vs. normal, so this kind of low sample model would lend itself to using limma/eBayes in R. As has been said the expression profile of tumours is heterogenous (especially in my tumour of interest, colorectal). 20 patient:normal should be enough to give you at least an idea of some trends in the dataset.

There's a great vignette on doing this in expression data in Bioconductor here:

http://www.bioconductor.org/help/course-materials/2010/BioC2010/limma2.pdf

Another thing I would suggest is Gene Set Enrichment Analysis - this will help you pick out the "themes" in your dataset.

Andrew' suggestion is good. I would use principal component analysis or EM clustering, like in the mclust package in R, to define patient groups. While each patient is unique some will have stronger similarities, this is the rationale for clustering them, and is what is ment by personalized medicine. We do not literally make a different drug for each patient, but find those with enough similarities that they will respond to the same drug. If you have cytogenetics then this could be a justification for a priori grouping or a strong component I the PCA.

Best

Brian

PCA did separate the data into obvious groups of tumor and normal but there is no apparent way to group patients by using both tumor and normal data.

I am gonna do the GSEA analysis. I tried doing LIMMA using integromics but integromics requires at least 3 samples to begin the analysis, which does not satisfies my quest for patient wise comparison.

Dear Mohammad; given that samples are paired by patient, I would not treat them as independent samples, but rather calculate the log fold change tumor vs. normal for each patient separately (in limma, just subtract the log-intensities for tumor from the log intensities for normal on a patient to patient basis). Then you have 20 biological replicates of the logFC values and can test significance against 0 (no change).

With these values, you can also attempt to find patient subgroups using PCA or a clustering approach.

Hi All,

Many good suggestions, and especially the link posted by Dr. Andrew D Beggs:

http://www.bioconductor.org/help/course-materials/2010/BioC2010/limma2.pdf

The lecture by Dr. MacDonald pointed out:

1. Whay limma:

- Fit nearly any model

- Technical replicates

- One/Two color arrays

- Increased power

2. Why not limma

- Complexity

- Reliance on normal theory

- Can't fit linear mide model

- Can't handle multilple levels of technical replication

So, it's better to use PROC MIXED model ( a SAS procedure, a great tool) or 'lme4' a R package by Dr. Douglous Bates and others, and both these tools handle easily correlated data.

I will prefer to use PROC MIXED model.