It depends. In situations where the species being measured constitutes a significant fraction of the total (as in e.g. recombinant expression at high levels in E. coli) it is better to normalize based on a different protein whose average concentration/cell does not change significantly between the different samples. Otherwise, it is better to use total protein contents for normalization.

As for ImageJ, you should give more detail about what you are doing before we can provide meaningful help. Are you using color images or 8-bit grayscale? How are you ensuring that the picture is not saturated? Are you normalizing contrast? Are you calibrating signal intensity, and how if so? Are you doing any kind of background subtraction or image processing? How are you measuring densities, by measuring the area under the curve of pixel intensities from a line along the measured lane, or by integrating the density of an r.o.i. encompassing the lane?

Dear Martin, thanks for your suggestions.

I am using gray scale 8-bit images taken by Gel Logic 1500 imaging system (Kodak). I don't know about 'picture is saturated or not'. I am doing background subtraction and following a protocol from page (http://howtowesternblot.net/data-analysis-3/quantification/).

The problem that I realized is if you select whole cell proteins with image J and analyze peaks, many peaks are not clear. For example, one clear peak would have many smaller shoulder peaks. So don't know how to deal with this type of situation. I think this can cause errors in normalization.

I think, as you suggested, selecting few proteins would be better rather than whole cell proteins as one can exactly calculate the area under peak for few well separated, prominent bands/peaks without any bias.

By 'saturation' I meant ensuring that all signals (pixel intensities) remain within the dynamic range, i.e. not overexposed. This is usually taken care of automatically by the imaging system, but may become a problem if using common digital cameras.

When measuring whole cell proteins, you do not have to measure individual peaks (excepting of course the 'target' peak, which has to be defined with the line tool). What you do is use the wand to measure the area of the target peak, and of the two areas comprising the remaining proteins to the left and the right of the target peak. Then you measure "expression level" as the quotient of the area of the target peak and the sum of the areas of the remaining peaks, comparing 'expression level' across lanes.

You are right that manually selecting the peaks may lead to systematic errors (Although I assure you these pale in comparison with the errors introduced by the very narrow dynamic range of Coomassie blue staining). I do not know of any plugins that can deal with this within ImageJ, although there are algorithms around to interpolate Gaussian curves (or other shapes) to do this in an objective manner. There is a great application for this purpose called PeakFit at sigmaplot.com, but it is not free.