Statistical Engineering Division
Seminar Series

Walter Liggett
Statistical Engineering Division, NIST
Admin. Building: Lecture Room A
Tuesday, January 11, 2005, 10:30-11:30 AM

Background:

Statistical analysis of replicate mass spectra can reveal sources of measurement variation in protein mass spectrometry and thereby important measurement issues.

Methods:

The measurement procedure is surface-enhanced laser desorption/ionization (SELDI) time-of-flight (TOF) mass spectrometry. Sources of variation are identified through statistical analysis of repeated measurements of a human serum standard, specifically, 88 spectra determined for mass-to-charge (m/z) values between 3300 and 30700. The statistical approach involves functional canonical correlation analysis (CCA) applied to disjoint intervals of the mass spectra for the purpose of finding long-distance correlation structure. Before CCA is applied, the spectra are normalized to remove spectrum-to-spectrum variation common to the entire 3300-30700 domain. Examination of the relation between the CCA scores and the spectra at each m/z shows the spectral peaks responsible for high canonical correlation.

Results:

We show that after normalization, the heights of some pairs of spectral peaks are correlated but others are not. Of the 17 spectral intervals considered, we choose the seven pairs of intervals with highest canonical correlation for interpretation. For some pairs, interpretation entails the singly- and doubly-charged ionization of the same protein, and for others, interpretation entails different proteins.

Conclusions:

It seems likely that sources of variation in the sample preparation step are responsible for high correlations between proteins separated widely in m/z. Non-uniformity in the crystallization on the protein chip surface is a well-known source of long-distance correlation, but normalization should remove its effect. Thus, the remaining high correlations suggest other sources of variation in sample preparation.