0-7803-9387-2/05/$20.00 ©2005 IEEE
Utilizing Artificial Neural Networks to Elucidate
Serum Biomarker Patterns Which Discriminate
Between Clinical Stages in Melanoma
Lee Lancashire
1
Selma Ugurel
2
Colin Creaser
1
Dirk Schadendorf
2
Robert Rees
1
Graham Ball
1
1
Interdisciplinary Biomedical Research Centre,
School of Science,
Nottingham Trent University,
Clifton Lane, Clifton, Nottingham NG11 8NS, United Kingdom.
2
Department of Dermatology,
Skin Cancer unit, German Cancer Research Centre,
Univeritatsklinkum Mannheim,
Theodor Kutzer Ufer1, D-68135, Mannheim, Germany.
Email: Lee.Lancashire@ntu.ac.uk
Abstract- The identification of proteomic patterns from
biomarkers in diseases such as cancer could lead to the
determination of novel prognostic and diagnostic markers
fundamental to the treatment of patients. We apply a recently
developed approach utilizing artificial neural networks as a data
mining tool to identify and characterize the best subset of
biomarkers associated with melanoma. These were capable of
predicting whether a sample is from a patient diagnosed with
stage I or stage IV melanoma to median accuracies of 98 % on an
independent subset of data used for validation. Furthermore,
individual response curves have been generated allowing the
investigation of whether these markers are up or down regulated
with regards to tumor progression.
I. INTRODUCTION
The ability to identify biomarker patterns indicative of
onset or progression of diseases such as cancer would lead to
the ability to identify the disease more rapidly, thus increasing
the probability of successfully treating the patient whilst the
disease is still in its early stages. Furthermore, this may lead
to the development of novel assays capable of assisting in the
diagnosis and management of patients on an individual basis,
recently described as ‘personalised medicine’ [1]. Current
advances in high throughput mass spectrometry (MS)
technologies such as surface enhanced laser/desorption
ionization (SELDI) time of flight (TOF) MS has permitted for
the rapid identification of potential biomarkers associated with
disease [2-7]. SELDI MS allows for the potential analysis of
the proteome as a whole and generates patterns that these
masses of proteins produce, thus facilitating the identification
of proteins that are being expressed differently in different
conditions, such as tumor stage [8].
We have previously shown that by using pattern recognition
algorithms such as artificial neural networks (ANNs) as a data
mining tool, it is possible to accurately predict disease
outcome from complex SELDI MS datasets in both cancer
patients [9-11], and in the classification of diseases from
micro-organisms [12].
Here we apply a novel ANN methodology to re-analyze a
dataset recently analyzed by our group [13]. This dataset
consists of 205 melanoma serum samples, 106 of these being
from patients with stage I disease, and the remaining 99 from
stage IV patients, with the aim being to determine possible
biomarkers present in the patients which could accurately
predict disease stage, and would therefore have the potential
to be used to detect disease onset at an earlier time point.
A. ANNs
ANNs are a form of machine learning and are capable of
modelling for complex systems. They are based on the way a
biological neuron operates in organising and processing
information. This gives ANNs several advantages; for
example, they are able to handle data which contains high
levels of noise or redundancy, or that which contains non-
linear elements. Perhaps the most popular form of ANN used
is the multi-layer perceptron (MLP), which has been shown to
provide a robust approach to outcome approximation, and has
been used to great effect in the biological sciences [14-19].
The MLP in particular is being commonly used for the data
mining of highly complex datasets [20-22], and models