I work with datasets in which protein abundances are reported across samples. I have some measurements of biological samples that should be more or less equal in protein abundance.

After getting the data as intensities, they are normalized by subtracting the log$_2$ median of each sample. If I plot these normalized and log transformed intensities, I get a Pearson correlation of around 0.98. But when I divide each row by the row median and log$_2$ transform of the ratio, I get a Pearson correlation of about 0.3. I need to process my data by row normalizing them.

So I am clearly doing something very wrong when performing the correlation analysis after row normalization. How do I perform a correlation analysis on these values?


It seems to me you're wandering towards correspondence analysis, which is based on modeling row normalized (and column normalized) measures in a low dimensional space. That literature should provide some principled normalization strategies. Greenacre has a very readable book on the topic:

Greenacre, Michael. Correspondence Analysis in Practice. Chapman and Hall/CRC, 2007.

There is also a related sub-literature on 'log-ratio' analysis that might be helpful.

For explicitly correlational questions, an algebraically identical (so far as I can see) strategy is called canonical correlation analysis. It's described in Section 9.6 of

Agresti, Alan. Categorical Data Analysis. 2nd ed. New York: Wiley-Interscience, 2002.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.