Not Logged In

Combining Gene Expression and Interaction Network Data to Improve Kidney Lesion Score Prediction

Full Text: Paper-kidney-BIOT 2010Submit.pdf PDF

Abstract
Doctors often use histopathology of needle biopsies (in the form of lesion score) to help diagnose kidney rejection, which is useful for identifying the appropriate treatment. As these lesion score are subjective and error prone, some researchers have tried to train a classifier to predict the rejection or non- rejection of a renal transplant, based on the gene expression microarrays of the patient's renal biopsies. However the high dimensionality and intrinsic noisy nature of this data makes this task very challenging. The most common techniques for predicting lesion scores from microarrays just use a single regressor on a subset of genes selected by statistical feature selection methods. Due to the high dimensionality of microarray data, these models usually overfit. This paper presents a novel method for predicting lesion scores based on the majority vote of bagging regressors built on feature subsets selected by either statistical or biological feature selection approaches, including a model that uses interaction networks to select genes. Our experimental results show that focusing on genes that interact with many other genes ("genes") and also interact with statistically selected genes in interaction networks, provide significantly better results than other biological feature selection methods. These experimental results show that none of the statistical feature selection methods are significantly better than our Hub genes approach and that a simple fusion of Hub genes and the best statistical feature selected method can further increase the generalization power of the prediction model.

Citation

D. Moulavi, M. Hajiloo, J. Sander, P. Halloran, R. Greiner. "Combining Gene Expression and Interaction Network Data to Improve Kidney Lesion Score Prediction". Biotechnology and Bioinformatics Symposium, October 2010.

Keywords: Kidney transplant, Lesion score, Microarray, Biological feature selection, Hub genes, medical informatics, bioinformatics
Category: In Conference

BibTeX

@incollection{Moulavi+al:BIOT10,
  author = {Davoud Moulavi and Mohsen Hajiloo and Joerg Sander and Philip
    Halloran and Russ Greiner},
  title = {Combining Gene Expression and Interaction Network Data to Improve
    Kidney Lesion Score Prediction},
  booktitle = {Biotechnology and Bioinformatics Symposium},
  year = 2010,
}

Last Updated: February 15, 2013
Submitted by Mohsen Hajiloo

University of Alberta Logo AICML Logo