Baseline_Subtraction.ppt -- powerpoint file showing a single spectrum
   before and after baseline subtraction, both for the whole spectrum
   and for a zoom in on the low mass region. This was produced from 
   the Ciphergen software.

base_sub_all.[png,eps] -- a single spectrum
   before and after baseline subtraction, both for the whole spectrum
   and for a zoom in on the low mass region. This was produced from 
   the Ciphergen software and saved as a PowerPoint file 
   (Baseline_Subtraction.ppt) and these images are simply file 
   conversions. 

base_sub_2.png -- the same spectrum as in base_sub_all, but without
   the zoom on the low mass region.

data_set_1.png -- heat map of the spectra in DS1, divided into 
   Cancers, Normals, and Benign Disease. There is a clear difference
   with the Benign Disease group. 

data_set_2.png -- heat map of the spectra in DS2, divided into 
   Cancers, Normals, and Benign Disease. There is no clear difference
   with the Benign Disease group. This is in contrast with what 
   was found in data_set_1.png for DS1, and is perplexing as these
   data sets involve the same biological samples run on different
   chip types. 

data_sets_1_2[,_gray].[png,eps] -- simultaneous heat maps of the
   spectra from both DS1 and DS2. The benign disease group from 
   DS1 clearly belongs with the spectra from DS2.

data_sets_2_3_offset[,_gray].[png,eps] -- various views of two
   big peaks taken from Data Set 2 (Top is all of DS2, and the 
   middle row is a set of zooms on the peaks indicated at at the
   top) and the corresponding regions of DS3. These two peaks 
   were chosen for two reasons: they were big, and there was 
   essentially nothing else near them. Thus, we are fairly certain
   we are seeing the same proteins in DS2 and DS3, but they are
   clearly offset. The magnitude of the offset is substantially
   larger than would be expected given the nominal mass accuracy
   of the machine. 

distmat_theirpeaks.png -- Euclidean distance matrix for Data Set 3
   using the intensities at their 7 supplied peaks. The bifurcation
   between cancers and normals is readily apparent.

ds1_distmat_theirpeaks[.png,_bw.png,_bw.eps] -- The pairwise Euclidean
   distance matrix for all spectra in data set 1, with a flag shown 
   for those pairs where the distance exceeds the nominal cutoff
   threshold for clustering. The small number of pairs falling into
   this category suggests that the data have been processed before
   posting, with the most obvious candidate being baseline subtraction.

ds3_bivar_split.png -- Perfect separation of the normals from the 
   cancers in DS3 using just the intensities at M/Z 435.46 and 
   465.57. These are the first two peaks in their list of 7, and 
   are also the two in that list with the largest magnitude t-values. 
   This suggests that the remainder of the list may be "noise".
   We note that both of the masses involved here are quite low. 

ds3_bivar_split_2.png -- Perfect separation of the normals from the 
   cancers in DS3 using just the intensities at M/Z 2.79 and 
   245.24. These two peaks were chosen from a list of those 
   features with t-scores in excess of 10. The masses involved
   here are exceedingly low, and the mass at 2.79 should be in
   the noise range of the spectra. We are seeing structure in the
   noise. 

ds3_bivar_split_3[,_gray].[png,eps] -- Perfect separation of the 
   normals from the cancers in DS3 using just two pairs of
   intensities: M/Z (435.46,465.57) and  (2.79, 245.24).
   Perfect separation is not difficult with this data set, 
   but it is also not difficult in the noise region of the 
   spectra, and it should be. 

ds3_distmats_bw.[png,eps] -- Side by side plot of the Euclidean distance
   matrices produced from the intensities in Data Set 3 at the peak
   locations supplied for Data Sets 3 and 2, respectively. Separation
   is clear using the DS3 peaks, and nonexistent using the DS2 peaks. 

ds3_dists_ds2_peaks.png -- the Euclidean distance matrix produced using
   the spectrum intensities from Data Set 3 at the peak locations found
   to be useful for Data Set 2. The clear separation of cancers and 
   normals has disappeared, suggesting that the peaks found are not
   stable due to one cause or another. 

ds3_peak1_various[,_gray].[png,eps] -- Several views of the single
   peak in DS3 from their list that was most effective at splitting 
   cancers from normals. Views on the left correspond to the peak
   in DS3, and views on the right correspond to the same views of
   the parallel mass range from DS2. While the peak is quite visible
   in DS3, the DS2 region shows no clear evidence of the same peak,
   and shows clear signs of local saturation of the spectra.

ds3_tstats.png -- plot of the two sample t-statistics contrasting
   normals and cancer for every index value in DS3. There are several
   peaks with very high t-scores (some in excess of 20 in absolute
   magnitude), indicating that perfect separation should be fairly
   easy to achieve. We find it slightly disturbing that many of the
   biggest t-values are in the mass range of the spectra that we tend
   to suspect of matrix contamination.

maxima_loc[,2,3]_ds1.png -- three views of the maximum intensity
   peaks in data set 1. The first view shows the index of the 
   maximum intensity peak within each of the 216 spectra (we 
   excluded the early regions), the second view shows a single
   spectrum from each of the 3 groups: normal, cancer, and benign,
   and the third shows the spectrum intensities at the rightmost 
   maximum found across all spectra. These 3 views showed us that 
   the benign disease spectra were very different from the rest
   in data set 1, but we couldn't explain this until we had access
   to data set 2 (and even then it took a while).

sample_normal.png -- A sample spectrum from data set 1, showing
   the negative intensities over at the low mass end. The negative
   intensities vary widely from spectrum to spectrum, and they make
   normalizing to the range of intensities shown a very bad idea,
   as there will be systematic biases for large parts of the spectra. 
   This type of normalization makes a bit more sense if the spectra
   have not been baseline subtracted. 

saturate[_mz,_tick].png -- Two views of a single spectrum from 
   Data Set 3 showing clear saturation of the signal in the low
   mass range. The extent of saturation is more visible when the
   values are plotted on the clock tick scale; plotting the values
   on the m/z scale compresses the observations at the low masses. 
   Saturation of the spectra makes the surrounding values also
   somewhat suspect, as saturation effects can extend beyond 
   just the highest values. This is due in part to the fact that
   SELDI spectra are actually averages of readings produced from
   several different laser blasts of the sample (typically 65) 
   and high intensities can arise when just some of the blasts 
   have yielded saturated readings. 
