Discovering biomarkers from gene expression data for predicting cancer subgroups using neural networks and relational fuzzy clustering

Table 2 The list of twenty best genes selected by FSMLP in the first stage

Gene ID	Image ID	Name
FGFR4	(*)784224	fibroblast growth factor receptor 4
EST	208699	EST
FCGRT	(*)770394	Fc fragment of IgG, receptor, transporter, alpha
AF1Q	(*)812105	transmembrane protein
HCLS1	(*)767183	hematopoietic cell-specific Lyn substrate 1
NAB2	(*)770868	NGFI-A binding protein 2 (ERG1 binding protein 2)
CDH2	(*)325182	cadherin 2, N-cadherin (neuronal)
EHD1	(*)745019	EH domain containing 1
HLA-DQA1	80109	major histocompatibility complex, class II, DQ alpha 1
LGALS3BP	811000	lectin, galactoside-binding, soluble, 3 binding protein (galectin 6 binding protein)
BAT3	898237	HLA-B associated transcript-3
SGCA	796258	sarcoglycan, alpha (50 kD dystrophin-associated glycoprotein)
ESTs	244618	ESTs
NOE1	52076	olfactomedinrelated ER localized protein
LSP1	(*)143306	lymphocyte-specific protein 1
IFG2	296448	insulin-like growth factor 2 (somatomedin A)
PMS2L12	(*)878652	postmeiotic segregation increased 2-like 12
NA	450152	NA
FVT1	(*)814260	follicular lymphoma variant translocation 1
CCNE1	68950	cyclin E1

We used a network with 2308 input nodes, 150 hidden nodes and 4 output nodes. These twenty genes are selected based on the gate opening values. We made several runs of ordinary MLP using these twenty genes and in each case the network was able to correctly classify all training and blind test examples. In the second stage, the FSMLP network is used to select ten best genes from amongst the twenty genes selected in the first stage. These ten genes are marked by asterisks. This set of ten genes has adequate cancer specific signatures to categorize the four types of SRBCTs.

ISSN: 1471-2105