Biomedical Engineering Reference
In-Depth Information
HPLC
NMR
MS
Excel
ICP-MS
AAA
Use tests
Figure
11.1. Information flow used to correlate analytical test results from different analytical
methods with cell-culture performance. Data mining provides a retrospective example that will
be used to gauge the probability that the performance from future raw material lots will
adequately predict performance without the need for a use test.
example still informs what future actions might be needed to create a proactive QbD
strategy for raw materials.
Figure 11.1 illustrates the strategy used during the investigation. Results from a use
test were compared with scores fromprincipal component analysis of HPLC-diode-array
detection (LC-DAD),
1
H
NMR, and LC-MS data. Results from ICP-MS and amino acid
analysis were included without recourse to multivariate data processing.
LC-DAD,
1
H
NMR, and LC-MS spectra each have noise as well as spectral regions
with no relevant signals. To reduce the probability that these low-signal regions influence
data interpretation, spectra were preprocessed to exclude uninformative regions and
noise. First, for each method, raw signals were aligned to each other using correlation
optimized warping as implemented in LineUp (Infometrix, Bothell, WA). Next, ANOVA
was performed on all the channels from the dataset to highlight channels that vary by lot
instead of replicating. Last, a threshold F-test value at 95% confidencewas applied based
on the degrees of freedom in the measurements. For LC-MS results, peak picking was
performed prior to alignment using xcms operating in R [18]. To explore the results, the
selected channels were compared using principal component analysis (PCA) or partial
least squares (PLS) as implemented inMatLab (Mathworks, Natick, MA) with PLSTool-
box (Eigenvector Research, WA) [19].
Methods were screened to seewhich could predict cell-culture results. For a result to
be “significant” and retained, it had to have statistical significance on its own,
independent of other measurement results. Factors that are significant only in combina-
tion with other factors were not included in subsequent data processing. Intuitively, this
criterion helps ensure that only informative results are included in models that predict
performance, but runs into trouble for methods that inherently combine information from
several compounds at each measurement point—methods like
1
H
NMR. For these
methods, mixed factors in the form of scores for latent variables are used in place of
individual component values.
To describe the process in detail, consider an example involving multiple hydroly-
sate lots tested using gradient reverse-phase chromatography with diode-array detection.
At each retention time, a spectrum is recorded creating a three-dimensional data cube
with retention time, wavelength, and intensity axes, for each injection (see Fig. 11.2). The
method to compare results from different lots for all wavelengths and retention times is
