There are probably several ways to do what you want, depending on what that is.
To remove a wandering or non-zero baseline, onoe option is to use a highpass filter with relatively low frequency stopband. This will remove any D-C offsets as well as low-frequency baseline wander. The signal will be biphasic (likely with a zerp mean) and if low frequencies are part of the signal, this may not be the best option. (It works well with EKG traces.) Another option is to use either findpeaks (with a negated signal argument so that valleys become peaks) or islocalmin both with appropriate prominence settings to return the deepest valleys. After that, use polyfit on the identified valleys with the appropriate polynomial degree (usually 3) and then polyval on the entire independent variable vector and subtract that result from the signal vector to remove the baseline variation. (I have done this with spectrum signals with good results.) In this instance, the signal will not go below zero if it is performed correctly. That may require some experimentation.
Smoothing the data depends on the result you want. If the data have broadband noise, either wavelet denoising (requires the Wavelet Toolbox) or the sgolayfilt function can help with that. For band-limited noise, use a lowpass or bandpass filter to isolate the signal from the noise. It is always appropriate to use a reference electrode that simply picks up environmental noise. That signal can be subtracted from the other signals using differential amplifiers prior to the ADC stage of your instrumentation. Of course, always use shielded coaxial leads with the shield attached to the instrument ground, and the instrument connected to a true earth ground.
To plot more than one series on the same axes, use the hold function. (The vectors plotted on the same axes do not all have to have the same lengths, although (x,y) pairs in any specific plot call do need to have the same lengths.)
For the peak identification and related statistics, use findpeaks, islocalmax is also an option.
The statistical analyses can be done using the Statistics and Machine Learning Toolbox functions. Note that paired (comparing before-and-after effects on the same subjects after administering a drug, for example) and unpaired (comparing different, although matched, populations with one population undergoing the intervention such as the administration of a drug) comparisons require different functions. A set of tests that only require that the distributions betwen the groups are the same without specifying the distribution are the Wilcoxon ranksum (unpaired) and signrank (paired) tests. There are others. It may be worthwhile to consult with a statistician prior to designing your study to be certain that you have considered the statistical tests you want to do on it.
This is by no means an exhaustive discussion of all the possibilities, however it should get you started.
