New network topology approaches reveal differential correlation patterns in breast cancer

Michael Bockmayr, Frederick Klauschen, B. Györffy, Carsten Denkert, Jan Budczies

Research output: Contribution to journalArticle

20 Citations (Scopus)

Abstract

Background: Analysis of genome-wide data is often carried out using standard methods such as differential expression analysis, clustering analysis and heatmaps. Beyond that, differential correlation analysis was suggested to identify changes in the correlation patterns between disease states. The detection of differential correlation is a demanding task, as the number of entries in the gene-by-gene correlation matrix is large. Currently, there is no gold standard for the detection of differential correlation and statistical validation. Results: We developed two untargeted algorithms (DCloc and DCglob) that identify differential correlation patterns by comparing the local or global topology of correlation networks. Construction of networks from correlation structures requires fixing of a correlation threshold. Instead of a single cutoff, the algorithms systematically investigate a series of correlation thresholds and permit to detect different kinds of correlation changes at the same level of significance: strong changes of a few genes and moderate changes of many genes. Comparing the correlation structure of 208 ER- breast carcinomas and 208 ER+ breast carcinomas, DCloc detected 770 differentially correlated genes with a FDR of 12.8%, while DCglob detected 630 differentially correlated genes with a FDR of 12.1%. In two-fold cross-validation, the reproducibility of the list of the top 5% differentially correlated genes in 140 ER- tumors and in 140 ER+ tumors was 49% for DCloc and 33% for DCglob. Conclusions: We developed two correlation network topology based algorithms for the detection of differential correlations in different disease states. Clusters of differentially correlated genes could be interpreted biologically and included the marker genes hydroxyprostaglandin dehydrogenase (PGDH) and acyl-CoA synthetase medium chain 1 (ACSM1) of invasive apocrine carcinomas that were differentially correlated, but not differentially expressed. Using random subsampling and cross-validation, DCloc and DCglob were shown to identify specific and reproducible lists of differentially correlated genes.

Original languageEnglish
Article number78
JournalBMC Systems Biology
Volume7
DOIs
Publication statusPublished - Aug 15 2013

Fingerprint

Breast Cancer
Network Topology
Genes
Topology
Breast Neoplasms
Gene
Correlation Structure
Cross-validation
Tumor
Hydroxyprostaglandin Dehydrogenases
Tumors
Subsampling
Clustering Analysis
Differential Expression
Correlation Analysis
Correlation Matrix
Reproducibility
Gold
Cluster Analysis
Neoplasms

Keywords

  • Breast cancer
  • Differential co-expression
  • Differential correlation
  • Microarray data
  • Molecular subtypes

ASJC Scopus subject areas

  • Molecular Biology
  • Structural Biology
  • Applied Mathematics
  • Modelling and Simulation
  • Computer Science Applications
  • Medicine(all)

Cite this

New network topology approaches reveal differential correlation patterns in breast cancer. / Bockmayr, Michael; Klauschen, Frederick; Györffy, B.; Denkert, Carsten; Budczies, Jan.

In: BMC Systems Biology, Vol. 7, 78, 15.08.2013.

Research output: Contribution to journalArticle

Bockmayr, Michael ; Klauschen, Frederick ; Györffy, B. ; Denkert, Carsten ; Budczies, Jan. / New network topology approaches reveal differential correlation patterns in breast cancer. In: BMC Systems Biology. 2013 ; Vol. 7.
@article{8cb90d2e5e08445c812e83509489a6c0,
title = "New network topology approaches reveal differential correlation patterns in breast cancer",
abstract = "Background: Analysis of genome-wide data is often carried out using standard methods such as differential expression analysis, clustering analysis and heatmaps. Beyond that, differential correlation analysis was suggested to identify changes in the correlation patterns between disease states. The detection of differential correlation is a demanding task, as the number of entries in the gene-by-gene correlation matrix is large. Currently, there is no gold standard for the detection of differential correlation and statistical validation. Results: We developed two untargeted algorithms (DCloc and DCglob) that identify differential correlation patterns by comparing the local or global topology of correlation networks. Construction of networks from correlation structures requires fixing of a correlation threshold. Instead of a single cutoff, the algorithms systematically investigate a series of correlation thresholds and permit to detect different kinds of correlation changes at the same level of significance: strong changes of a few genes and moderate changes of many genes. Comparing the correlation structure of 208 ER- breast carcinomas and 208 ER+ breast carcinomas, DCloc detected 770 differentially correlated genes with a FDR of 12.8{\%}, while DCglob detected 630 differentially correlated genes with a FDR of 12.1{\%}. In two-fold cross-validation, the reproducibility of the list of the top 5{\%} differentially correlated genes in 140 ER- tumors and in 140 ER+ tumors was 49{\%} for DCloc and 33{\%} for DCglob. Conclusions: We developed two correlation network topology based algorithms for the detection of differential correlations in different disease states. Clusters of differentially correlated genes could be interpreted biologically and included the marker genes hydroxyprostaglandin dehydrogenase (PGDH) and acyl-CoA synthetase medium chain 1 (ACSM1) of invasive apocrine carcinomas that were differentially correlated, but not differentially expressed. Using random subsampling and cross-validation, DCloc and DCglob were shown to identify specific and reproducible lists of differentially correlated genes.",
keywords = "Breast cancer, Differential co-expression, Differential correlation, Microarray data, Molecular subtypes",
author = "Michael Bockmayr and Frederick Klauschen and B. Gy{\"o}rffy and Carsten Denkert and Jan Budczies",
year = "2013",
month = "8",
day = "15",
doi = "10.1186/1752-0509-7-78",
language = "English",
volume = "7",
journal = "BMC Systems Biology",
issn = "1752-0509",
publisher = "BioMed Central",

}

TY - JOUR

T1 - New network topology approaches reveal differential correlation patterns in breast cancer

AU - Bockmayr, Michael

AU - Klauschen, Frederick

AU - Györffy, B.

AU - Denkert, Carsten

AU - Budczies, Jan

PY - 2013/8/15

Y1 - 2013/8/15

N2 - Background: Analysis of genome-wide data is often carried out using standard methods such as differential expression analysis, clustering analysis and heatmaps. Beyond that, differential correlation analysis was suggested to identify changes in the correlation patterns between disease states. The detection of differential correlation is a demanding task, as the number of entries in the gene-by-gene correlation matrix is large. Currently, there is no gold standard for the detection of differential correlation and statistical validation. Results: We developed two untargeted algorithms (DCloc and DCglob) that identify differential correlation patterns by comparing the local or global topology of correlation networks. Construction of networks from correlation structures requires fixing of a correlation threshold. Instead of a single cutoff, the algorithms systematically investigate a series of correlation thresholds and permit to detect different kinds of correlation changes at the same level of significance: strong changes of a few genes and moderate changes of many genes. Comparing the correlation structure of 208 ER- breast carcinomas and 208 ER+ breast carcinomas, DCloc detected 770 differentially correlated genes with a FDR of 12.8%, while DCglob detected 630 differentially correlated genes with a FDR of 12.1%. In two-fold cross-validation, the reproducibility of the list of the top 5% differentially correlated genes in 140 ER- tumors and in 140 ER+ tumors was 49% for DCloc and 33% for DCglob. Conclusions: We developed two correlation network topology based algorithms for the detection of differential correlations in different disease states. Clusters of differentially correlated genes could be interpreted biologically and included the marker genes hydroxyprostaglandin dehydrogenase (PGDH) and acyl-CoA synthetase medium chain 1 (ACSM1) of invasive apocrine carcinomas that were differentially correlated, but not differentially expressed. Using random subsampling and cross-validation, DCloc and DCglob were shown to identify specific and reproducible lists of differentially correlated genes.

AB - Background: Analysis of genome-wide data is often carried out using standard methods such as differential expression analysis, clustering analysis and heatmaps. Beyond that, differential correlation analysis was suggested to identify changes in the correlation patterns between disease states. The detection of differential correlation is a demanding task, as the number of entries in the gene-by-gene correlation matrix is large. Currently, there is no gold standard for the detection of differential correlation and statistical validation. Results: We developed two untargeted algorithms (DCloc and DCglob) that identify differential correlation patterns by comparing the local or global topology of correlation networks. Construction of networks from correlation structures requires fixing of a correlation threshold. Instead of a single cutoff, the algorithms systematically investigate a series of correlation thresholds and permit to detect different kinds of correlation changes at the same level of significance: strong changes of a few genes and moderate changes of many genes. Comparing the correlation structure of 208 ER- breast carcinomas and 208 ER+ breast carcinomas, DCloc detected 770 differentially correlated genes with a FDR of 12.8%, while DCglob detected 630 differentially correlated genes with a FDR of 12.1%. In two-fold cross-validation, the reproducibility of the list of the top 5% differentially correlated genes in 140 ER- tumors and in 140 ER+ tumors was 49% for DCloc and 33% for DCglob. Conclusions: We developed two correlation network topology based algorithms for the detection of differential correlations in different disease states. Clusters of differentially correlated genes could be interpreted biologically and included the marker genes hydroxyprostaglandin dehydrogenase (PGDH) and acyl-CoA synthetase medium chain 1 (ACSM1) of invasive apocrine carcinomas that were differentially correlated, but not differentially expressed. Using random subsampling and cross-validation, DCloc and DCglob were shown to identify specific and reproducible lists of differentially correlated genes.

KW - Breast cancer

KW - Differential co-expression

KW - Differential correlation

KW - Microarray data

KW - Molecular subtypes

UR - http://www.scopus.com/inward/record.url?scp=84884541854&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84884541854&partnerID=8YFLogxK

U2 - 10.1186/1752-0509-7-78

DO - 10.1186/1752-0509-7-78

M3 - Article

VL - 7

JO - BMC Systems Biology

JF - BMC Systems Biology

SN - 1752-0509

M1 - 78

ER -