Open AccessMultiple fuzzy neural network system for outcome prediction and classification of 220 lymphoma patients on the basis of molecular profiling
Author(s) -
Ando Tatsuya,
Suguro Miyuki,
Kobayashi Takeshi,
Seto Masao,
Honda Hiroyuki
Publication year - 2003
Publication title -
cancer science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 2.035
H-Index - 141
eISSN - 1349-7006
pISSN - 1347-9032
DOI - 10.1111/j.1349-7006.2003.tb01374.x
Subject(s) - gene expression profiling , dna microarray , gene , lymphoma , artificial neural network , diffuse large b cell lymphoma , microarray , oncology , computational biology , proportional hazards model , artificial intelligence , bioinformatics , computer science , machine learning , gene expression , biology , medicine , genetics
A fuzzy neural network (FNN) using gene expression profile data can select combinations of genes from thousands of genes, and is applicable to predict outcome for cancer patients after chemotherapy. However, wide clinical heterogeneity reduces the accuracy of prediction. To overcome this problem, we have proposed an FNN system based on majoritarian decision using multiple noninferior models. We used transcriptional profiling data, which were obtained from "Lymphochip" DNA microarrays (http://llmpp.nih.gov/DLBCL), reported by Rosenwald (N Engl J Med 2002; 346: 1937-47). When the data were analyzed by our FNN system, accuracy (73.4%) of outcome prediction using only 1 FNN model with 4 genes was higher than that (68.5%) of the Cox model using 17 genes. Higher accuracy (91%) was obtained when an FNN system with 9 noninferior models, consisting of 35 independent genes, was used. The genes selected by the system included genes that are informative in the prognosis of Diffuse large B-cell lymphoma (DLBCL), such as genes showing an expression pattern similar to that of CD10 and BCL-6 or similar to that of IRF-4 and BCL-4. We classified 220 DLBCL patients into 5 groups using the prediction results of 9 FNN models. These groups may correspond to DLBCL subtypes. In group A containing half of the 220 patients, patients with poor outcome were found to satisfy 2 rules, i.e., high expression of MAX dimerization with high expression of unknown A (LC_26146), or high expression of MAX dimerization with low expression of unknown B (LC_33144). The present paper is the first to describe the multiple noninferior FNN modeling system. This system is a powerful tool for predicting outcome and classifying patients, and is applicable to other heterogeneous diseases.