文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > IT/计算机 > The Analysis of Microarray Datasets Using a Genetic Programming
侵权投诉

The Analysis of Microarray Datasets Using a Genetic Programming

The Analysis of Microarray Datasets Using a Genetic Programming

representative for the whole class, this method usually leads to poor learning ability. In addition, for microarray datasets, which suffer greatly from the small sample size, it is necessary to fully utilize each sample for classification. A weight-based scheme is used to overcome the skewness among data here. Assuming there are S1 and S2 samples in two classes, respectively. Then the weights for the correct classification of the samples in two classes are:

feature set. Different from the above experiment, we choose training and test set from both datasets and the others as independent test set. Training and test set are normalized in terms of features, just the same to the independent test set. In Karnoub’s breast cancer dataset [12], there are 15 normal breast samples and 7 breast cancer samples. In Radvanyi’s breast cancer dataset [13], there are 9 normal breast samples and 47 breast cancer samples. 6695 common genes are sifted out from these two original datasets. We use W1=S2/(S1+S2)

W2=S1/(S1+S2) (1) 15 normal breast samples from the former dataset and 47

cancer samples from the last dataset as the training set. The

If an individual correctly classifies l1 samples in the first others are used as test set. Training and test set are normalized class, and l2 samples in the second, its fitness value is in terms of features, just the same to the independent test set. calculated by: In all experiments, we run GP 30 times with random Fi=(W1×l1+W2×l2)/(W1×S1+W2×S2)

(2) initialization. In each run, the corresponding parameters of

With this design, the individual with higher fitness is the GP are listed in Table 1. preferred. Due to the huge search space, only the individuals

with top fitness values are kept so as to accelerate the

TABLE I evaluation process. When two or more individuals are

THE SETTINGS OF PRIMARY GP PARAMETERS

assigned to the same fitness score, the one containing fewer

Parameter SETTING features will be selected firstly.

As there are many feature subsets generated in parallel Terminal set (T)All gene expression values and constant values

Function set (F)Boolean and floating point operators: during evolution, a set of globally optimal or at least near

gt(>),le(<=),times (*),minus(-),plus(+), max, min.

optimal trees can be obtained. In short, the GP is promising in Generation100

Crossover probability0.7obtaining predictive classifiers.

Mutation probability0.5

Termination criteriaFitness score reaches 1 or running 100 generations

III. RESULTS

Six datasets are downloaded from the website of ONCOMINE. Among the six datasets, two of them are prostate cancer datasets [8], [9], two are leukemia datasets [10], [11] and the others are breast cancer datasets.

One of the prostate datasets is from University of Michigan, and the other is from Stanford University. The former originally has 55 samples and 9,984 gene tags. After excluding samples of benign prostatic hyperplasia, prostatitis and metastatic prostate cancer, the left contains 4 normal adjacent prostate samples and 14 localized prostate cancer samples. The latter contains 112 samples, where 71 of them belong to prostate cancer and the left belong to normal prostate. We sift out the same 850 tags existing in both datasets and their corresponding expression values through the selected samples. Then samples are randomly selected from the second dataset to construct the training and test set to train and test the classifiers generated with GP, and the samples selected from the first dataset are used as independent test set. In detail, 47 tumor samples and 27 normal ones are selected randomly to build the training set and the remaining is kept as the test set.

For the leukemia datasets, one of them contains 47 ALL samples and 25 AML samples [10], and the other contains 24 ALL samples and 28 AML samples after eliminating 20 MLL samples [11]. 4880 probes are sifted out as the common

Dynamic maximum tree depth limit

Strict depth limit

56

Z-score is a widely deployed method to evaluate the frequency of different genes. Let Z = [Xi-E(Xi)]/σ

The Analysis of Microarray Datasets Using a Genetic Programming

, where Xi is the frequency that gene i is selected, E(Xi) is the expected number of times that gene i is selected, and σ is the standard deviation of variance. Let pi denote the probability of gene i being selected randomly, which is approximately equal to the total counts of frequency in a population divided by the number of individuals in the population, then divided by the σ. individuals, then E(Xi) = A×pi, and

Fig 2(A), (B) show the normalized z-score of each gene in prostate datasets and leukemia datasets.

For the reason that the more z-score some gene gets, the more it might be distinctive in classifying normal tissue from tumors or different diseases, the top genes with the highest z-scores are selected and checked in NCBI. In the prostate datasets, it’s found that 3 genes of the top 30s have been identified participating in prostate cancer genesis experimentally, such as FGFR2, RND3 and DKK3. The first gene, FGFR2 (fibroblast growth factor receptor 2), has the highest z-score. Changes in the FGFR axis are often associated with prostate cancer progression. In Freeman’s experiment, the experimenters confirm the hypothesis that the

上一页第3页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email