Abstract:
Quantum dot encoded microspheres (QDEMs) offer much potential for bead based identification
of a variety of biomolecules via flow cytometry (FCM). To date, QDEM subpopulation
classification from FCM has required significant instrument modification or multiparameter
gating. It is unclear whether or not current data analysis approaches can handle the increased
multiplexed capacity offered by these novel encoding schemes. In this thesis the drawbacks of
currently available data analysis techniques are demonstrated and novel classification methods
proposed to overcome these limitations. A commercially available 20 code QDEM library with
fluorescent emissions at 4 distinct wavelengths and 4 different intensity levels was analysed using
flow cytometry. Multiparameter gating (MPG) a readily available classification method for
subpopulations in FCM was evaluated. A support vector machine (SVM) and two types of
artificial neural networks (ANNs), a multilayer perceptron (MLP) and probabilistic radial basis
function (PRBF) were also considered. For the supervised models rigorous parameter selection
using cross validation (CV) was used to construct the optimum models. Independent test set
validation was also carried out. As a further test, external validation of the classifiers was
performed using multiplexed QDEMs solutions.
The performance of MPG was poor (average misclassification (MC) rate = 9.7%) was a time
consuming process requiring fine adjustment of the gates, classifications made on the dataset
were poor with multiple classifications on single events and as the multiplex capacity increases
the performance is likely to decrease. The SVM had the best performance in independent test
validation with 96.33% accuracy on the independent testing (MLP = 96.12%, PRBF = 94.38%).
Furthermore the performance of the SVM was superior to both MPG and both ANNs for the
external validation set with an average MC rate for MLP = 6.1% and PRBF = 7.5% whereas the
SVM MC rate was 2.9%. Assuming that the external test solutions were homogenous the variance
between classified results should be minimal hence, the variance of correct classifications (CCs)
was used as an additional indicator of classifier performance. The SVM demonstrates the lowest
variance for each of the external validation solutions (average σ
2
= 31479) some 50% lower than
that of MPG. As a conclusion to the development of the classifier, a user friendly software system
has been developed to allow construction and evaluation of multiclass SVMs for use by FCM
practitioners in the laboratory. SVMs are a promising classifier for QDEMs that can be rapidly
trained and classifications made in real time using standard FCM instrumentation. It is hoped that
this work will advance SAT for bioanalytical applications.