Methods blog · Cancer Systems Biology & Drug Resistance
VToDVToD: voting-based cancer module identification from copy number, expression and PPI networks
A module-detection algorithm that lets data-driven gene–gene relationships and protein-interaction topology vote together, recovering cancer-related pathways in glioblastoma and ovarian carcinoma.
#cancer modules#data integration#copy number aberration#gene expression#PPI network#TCGA
Cancer genomes are disrupted at several levels at once: genes are amplified or deleted (copy number aberrations, CNAs) and their expression shifts. Either data type alone gives a partial view. VToD (Voting-based cancer module identification by combining Topological and Data-driven properties) integrates both with the protein–protein interaction (PPI) network to find functional modules: groups of genes that act together in cancer.
Step 1: a gene–gene relationship network from the data
Starting from a set of seed genes, we enumerate every type of pairwise correlation across patients:
- GE–GE: two genes whose expression levels co-vary;
- CNA–GE: a gene’s copy number tracks another gene’s expression, a possible cis or trans dosage effect;
- CNA–CNA: two genes that are co-amplified or co-deleted.
Each significant correlation becomes an edge, giving a data-driven network that reflects what is happening in the tumours.
Step 2: let topology and data vote
PPI networks capture what proteins can do together; the data-driven network captures what they appear to do in these patients. VToD grows modules by a voting scheme: a candidate gene joins a module when it is supported both by its topological connections in the PPI network and by its data-driven relationships to the current members. Requiring agreement filters out spurious correlations, which have no physical interaction to support them, and inactive interactions, which have no data to support them.
Results on TCGA cohorts
We applied VToD to 266 glioblastoma multiforme (GBM) and 96 ovarian carcinoma (OVC) samples that had both expression and copy number measurements.
| GBM | OVC | |
|---|---|---|
| Modules identified | 22 | 23 |
| Enriched in cancer-related KEGG pathways | 15 | 19 |
| Enriched in BioCarta pathways | 12 | 18 |
| Enriched in GO terms | 20 | 23 |
| Enriched in Cancer Gene Census genes | 9 | 15 |
VToD significantly outperformed existing module-detection methods on both functional and cancer-gene-set enrichment. Modules containing more types of gene–gene relationships were more strongly enriched for cancer genes (correlation 0.64 in GBM, 0.49 in OVC). In other words, the modules where expression changes and CNAs converge are the ones that best explain cancer biology.
Where it led
VToD grew out of my MS thesis at GIST, presented first at GIW 2011. It was later reviewed alongside other integrative module-finding methods in a Methods in Molecular Biology chapter (2017). The theme of making several kinds of evidence agree before trusting a network edge runs through our later work on DrugSimDB and pathway cross-talk.
Read the papers
-
Voting-based cancer module identification by combining topological and data-driven properties
PLOS ONE 8 (8), e70498 · 2013
-
Integrating heterogeneous datasets for cancer module identification
In: Bioinformatics: Volume II: Structure, Function, and Applications (Methods in Molecular Biology), Springer, 119-137 · 2017
-
Network Reconstruction Approach for Integrating Gene Expression & Copy Number Dataset
Poster, 22nd International Conference on Genome Informatics (GIW 2011), Busan, South Korea · 2011