You need to do both statistical analysis and filter based on the fold changes. Say you have a treated population and a control population (untreated) with replicates for each group. With t-test, you can see whether the mean value of the treated population significantly differs from the mean value of the control population. Usually, people use p < 0.05 as a cutoff for the p-value. You can also calculate the fold changes from the mean values of both groups (you need to determine your own fold change cutoff; For example, you choose 2). These two criteria combined will be the threshold for up-regulation and down-regulation of the proteins: If the p value < 0.05 and fold change lower or larger than 2, then you will say the protein is deferentially expressed. Usually, people have their own preferences in choosing the value for the fold changes factor, some choose 1.5 and others choose 2. This is okay as long as you report the fold change factor you pick to the audience so they will know how large or small is the up/down-regulation.

Note that this is just one of many ways to set the threshold for up/down-regulations of proteins. You can even include some other statistical tests to filter the protein.

the answer is not simple - there are several factors to consider, and many publications do not take into account p-values.

1. number of replicates - with more replicates you can observe significant changes in smaller fold changes.

2. the variability of the data - for each protein you may observe difference CVs, and so for one protein you could observe significant difference when the fold change is 1.3, while another protein does not achieve significance, even though the fold change is 4 (for example).

what you should do is run a significance test on your data and use it to determine what to report.

Cheers,

D.

Thanks a lot

Thank you dear Ajit kumar

You mean,I don't need to calculate fold changes (compared to control) and filter those more than 2. I just should do statistical analysis (T test) to see which expression differences are significant. Right?

Thanks

You need to do both statistical analysis and filter based on the fold changes. Say you have a treated population and a control population (untreated) with replicates for each group. With t-test, you can see whether the mean value of the treated population significantly differs from the mean value of the control population. Usually, people use p < 0.05 as a cutoff for the p-value. You can also calculate the fold changes from the mean values of both groups (you need to determine your own fold change cutoff; For example, you choose 2). These two criteria combined will be the threshold for up-regulation and down-regulation of the proteins: If the p value < 0.05 and fold change lower or larger than 2, then you will say the protein is deferentially expressed. Usually, people have their own preferences in choosing the value for the fold changes factor, some choose 1.5 and others choose 2. This is okay as long as you report the fold change factor you pick to the audience so they will know how large or small is the up/down-regulation.

Note that this is just one of many ways to set the threshold for up/down-regulations of proteins. You can even include some other statistical tests to filter the protein.

Very clear and complete answer.

Thanks a lot.

It depends on your experiment design and what you are trying to achieve (e.g. is it a pilot? Are you proving a hypothesis? Are you doing hypothesis free discovery?) but there are quite a lot of statistical issues with omics data that can make simple approaches misleading.

I would have a look at this http://www.nonlinear.com/progenesis/qi-for-proteomics/v2.0/faq/pq-values.aspx

...and then do a google search for "proteomics multiple testing" and have a look over some of the articles.

All informative answers, and David Bramwell raises an important point with respect to multiple testing considerations. Unfortunately, many researchers use 1.5 fold change with p < 0.05 on extremely small samples sizes (n = 3 per group), or simply just one or the other of these two, none of which are particularly stringent. You would find many proteins just reaching this threshold would be completely different if you were to repeat the experiment a second time. I've also reviewed papers where an unusually small cutoff was used, e.g 1.2 fold (no p-value), which was obviously selected to include a protein of interest that barely met this threshold, e.g. 1.22 fold change, an approach also termed 'cherry picking'.

A paper I often refer others to in regard to applying cutoffs for proteomic analysis is by Yishai Levin (2011), The role of statistical power analysis in quantitative proteomics, Proteomics 11:2565-67. This paper gives you a sense of what fold change and p-value might be necessary to confidently distinguish between groups based on sample sizes available for your study. I believe it is accessible here...Article The role of statistical power analysis in quantitative proteomics

Another statistical approach you might find useful to analyze your data is called limma. It is popular for RNA-seq and microarray studies, and I know some proteomics researchers who are adapting it for their use as well. The benefit of this approach is that it uses methods that help overcome a number of issues that can be inherent in small experiments. If you know how to use R, there is a limma package in Bioconductor. This paper from Gordon Smyth's group explains limma quite well...

https://academic.oup.com/nar/article/43/7/e47/2414268

By the way, the 1.6/0.6 and 2.0/0.5 numbers you state are ratios, not fold change, but to convert these to fold change is straightforward. For upregulated proteins the number remains the same, and for downregulated proteins (ratio < 1.0) simply convert to the negative inverse (-1/ratio).

I concur with David and Kent.

The habit of imposing fold-change thresholds originated from early studies that used a simple 2-sample t-test to determine significance. The standard t-test is ill-suited to the "omic" realm, particularly proteomics experiments, which are most often ridiculously underpowered. This is because many significant hit for small fold changes arise by happenstance. Researchers reflexively understand that 5% changes are not going to fund grants, so a large fold change threshold is imposed to increase likelihood of focusing on something more biologically relevant.

So, on one hand, we do statistics to show statistical rigor, but then we add another threshold because we don't believe the statistics are yielding anything meaningful.

As Kent correctly points out, modern emprical Bayesian approaches to statistics, as in LIMMA, go a long way to addressing this. LIMMA "borrows power" from the rest of the dataset to get a better estimate of the variance of particular protein. The net effect of implementing LIMMA is to reduces both the false positive and false negative rate. The smallest fold changes are no longer significant and some of the whoppers that weren't now are.

Kammers et al developed R code to perform eBayes-modified t-tests on labeled MS2 data :

Article Detecting Significant Changes in Protein Abundance

See also

Article Assessment and Improvement of Statistical Tools for Comparat...

This reduces the compulsion to impose an arbitrary fold change threshold on top of an already arbitrary p-value threshold. Imposing a fold change threshold also messes with rigorous determination of the false discovery rate.

Of course, in the end, p-values are not a metric of reproducibility and nature doesn't abide by fold change thresholds. This realization underlies the methods like GSEA and weighted gene/protein coexpression networks:

Subramanian et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles

Proc Natl Acad Sci U S A. 2005 Oct 25;102(43):15545-50. Epub 2005 Sep 30

Article Weighted Gene Coexpression Network Analysis: State of the Art

Good thread.

Brian