Limits...
Patient-specific driver gene prediction and risk assessment through integrated network analysis of cancer omics profiles.

Bertrand D, Chng KR, Sherbaf FG, Kiesel A, Chia BK, Sia YY, Huang SK, Hoon DS, Liu ET, Hillmer A, Nagarajan N - Nucleic Acids Res. (2015)

Bottom Line: Extensive in silico and in vitro validation helped establish OncoIMPACT's robustness, improved precision over competing approaches and verifiable patient and cell line specific predictions (2/2 and 6/7 true positives and negatives, respectively).In particular, we computationally predicted and experimentally validated the gene TRIM24 as a putative novel amplified driver in a melanoma patient.We also provide the first demonstration that computationally derived driver mutation signatures can be overall superior to single gene and gene expression based signatures in enabling patient stratification and prognostication.

View Article: PubMed Central - PubMed

Affiliation: Computational and Systems Biology, Genome Institute of Singapore, Singapore 138672, Singapore.

Show MeSH

Related in: MedlinePlus

Clustering of driver genes and GO enrichment of associated modules. (a) Violin plots showing the distribution of average distance in the gene interaction network (computed at a sample-specific level) between all pairs of genes in each class (mutated genes, predicted driver genes and random hub genes (degree ≥ 20)). The blue line represents the average distance between genes on the interaction network. The P-values are computed using Wilcoxon rank sum test. (b) Box plots depicting the distribution of the number of genes in the largest module and all other modules. The P-values are computed using Wilcoxon rank sum test. (c) Bar chart showing the frequency at which GO terms are enriched in the largest module for each patient. Enrichment analysis was done using DAVID (53) (http://david.abcc.ncifcrf.gov/) and a q-value threshold of 0.05 was used to identify enriched terms.
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4402507&req=5

Figure 4: Clustering of driver genes and GO enrichment of associated modules. (a) Violin plots showing the distribution of average distance in the gene interaction network (computed at a sample-specific level) between all pairs of genes in each class (mutated genes, predicted driver genes and random hub genes (degree ≥ 20)). The blue line represents the average distance between genes on the interaction network. The P-values are computed using Wilcoxon rank sum test. (b) Box plots depicting the distribution of the number of genes in the largest module and all other modules. The P-values are computed using Wilcoxon rank sum test. (c) Bar chart showing the frequency at which GO terms are enriched in the largest module for each patient. Enrichment analysis was done using DAVID (53) (http://david.abcc.ncifcrf.gov/) and a q-value threshold of 0.05 was used to identify enriched terms.

Mentions: The ability to generate patient-specific lists of driver genes allowed us to analyze the distributional properties of driver genes without having to resort to an aggregate analysis that may obscure its interpretation. For example, using all predicted drivers, we readily observed that driver genes tend to cluster on the gene interaction network, similar to what was observed by others (50), and distances between them were significantly lower than between all mutated genes and between random genes (Supplementary Figure S11). However, this observation has several potential explanations including, but not limited to: (i) tumors share driver mutations that affect the same functional network (29) and (ii) biases in the data (50). Analysis using patient-specific driver gene lists avoids some of these issues and our analysis using OncoIMPACT revealed a similar pattern of clustering at a sample-specific level (Figure 4a and Supplementary Figure S12), that is not explained by mutation frequency or network structure (i.e. hub genes), suggesting that an alternative explanation—that the occurrence of multiple mutations in a network module is necessary for pathway deregulation in a tumor—may be valid here. We further investigated such synergistic interactions between drivers by using an unbiased search and statistical testing to identify potential co-drivers and compared our patient-specific results with a non-patient-specific approach (Supplementary Figure S13, Supplementary File S3). Our results show that the patient-specific analysis likely identifies more meaningful co-driver gene pairs (i) identifying a smaller subset of potential gene-pairs as co-drivers (Supplementary Figure S13a), (ii) that are less likely to be enriched in false-positives due to genomic proximity (Supplementary Figure S13b) and (iii) are more enriched in genes that are likely to have similar functional roles (Supplementary Figure S13c). In addition, we identified several gene-pairs as co-drivers that were not necessarily correlated in their mutation occurrences and were not therefore detectable without a patient-specific analysis as provided by OncoIMPACT (Supplementary File S3). Interestingly, in comparison to glioblastoma, ovarian and prostate cancer, we noted only a handful of co-drivers in melanoma and bladder cancer (Supplementary Figure S13a) and we discuss this observation further in a following section (see Discussion).


Patient-specific driver gene prediction and risk assessment through integrated network analysis of cancer omics profiles.

Bertrand D, Chng KR, Sherbaf FG, Kiesel A, Chia BK, Sia YY, Huang SK, Hoon DS, Liu ET, Hillmer A, Nagarajan N - Nucleic Acids Res. (2015)

Clustering of driver genes and GO enrichment of associated modules. (a) Violin plots showing the distribution of average distance in the gene interaction network (computed at a sample-specific level) between all pairs of genes in each class (mutated genes, predicted driver genes and random hub genes (degree ≥ 20)). The blue line represents the average distance between genes on the interaction network. The P-values are computed using Wilcoxon rank sum test. (b) Box plots depicting the distribution of the number of genes in the largest module and all other modules. The P-values are computed using Wilcoxon rank sum test. (c) Bar chart showing the frequency at which GO terms are enriched in the largest module for each patient. Enrichment analysis was done using DAVID (53) (http://david.abcc.ncifcrf.gov/) and a q-value threshold of 0.05 was used to identify enriched terms.
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4402507&req=5

Figure 4: Clustering of driver genes and GO enrichment of associated modules. (a) Violin plots showing the distribution of average distance in the gene interaction network (computed at a sample-specific level) between all pairs of genes in each class (mutated genes, predicted driver genes and random hub genes (degree ≥ 20)). The blue line represents the average distance between genes on the interaction network. The P-values are computed using Wilcoxon rank sum test. (b) Box plots depicting the distribution of the number of genes in the largest module and all other modules. The P-values are computed using Wilcoxon rank sum test. (c) Bar chart showing the frequency at which GO terms are enriched in the largest module for each patient. Enrichment analysis was done using DAVID (53) (http://david.abcc.ncifcrf.gov/) and a q-value threshold of 0.05 was used to identify enriched terms.
Mentions: The ability to generate patient-specific lists of driver genes allowed us to analyze the distributional properties of driver genes without having to resort to an aggregate analysis that may obscure its interpretation. For example, using all predicted drivers, we readily observed that driver genes tend to cluster on the gene interaction network, similar to what was observed by others (50), and distances between them were significantly lower than between all mutated genes and between random genes (Supplementary Figure S11). However, this observation has several potential explanations including, but not limited to: (i) tumors share driver mutations that affect the same functional network (29) and (ii) biases in the data (50). Analysis using patient-specific driver gene lists avoids some of these issues and our analysis using OncoIMPACT revealed a similar pattern of clustering at a sample-specific level (Figure 4a and Supplementary Figure S12), that is not explained by mutation frequency or network structure (i.e. hub genes), suggesting that an alternative explanation—that the occurrence of multiple mutations in a network module is necessary for pathway deregulation in a tumor—may be valid here. We further investigated such synergistic interactions between drivers by using an unbiased search and statistical testing to identify potential co-drivers and compared our patient-specific results with a non-patient-specific approach (Supplementary Figure S13, Supplementary File S3). Our results show that the patient-specific analysis likely identifies more meaningful co-driver gene pairs (i) identifying a smaller subset of potential gene-pairs as co-drivers (Supplementary Figure S13a), (ii) that are less likely to be enriched in false-positives due to genomic proximity (Supplementary Figure S13b) and (iii) are more enriched in genes that are likely to have similar functional roles (Supplementary Figure S13c). In addition, we identified several gene-pairs as co-drivers that were not necessarily correlated in their mutation occurrences and were not therefore detectable without a patient-specific analysis as provided by OncoIMPACT (Supplementary File S3). Interestingly, in comparison to glioblastoma, ovarian and prostate cancer, we noted only a handful of co-drivers in melanoma and bladder cancer (Supplementary Figure S13a) and we discuss this observation further in a following section (see Discussion).

Bottom Line: Extensive in silico and in vitro validation helped establish OncoIMPACT's robustness, improved precision over competing approaches and verifiable patient and cell line specific predictions (2/2 and 6/7 true positives and negatives, respectively).In particular, we computationally predicted and experimentally validated the gene TRIM24 as a putative novel amplified driver in a melanoma patient.We also provide the first demonstration that computationally derived driver mutation signatures can be overall superior to single gene and gene expression based signatures in enabling patient stratification and prognostication.

View Article: PubMed Central - PubMed

Affiliation: Computational and Systems Biology, Genome Institute of Singapore, Singapore 138672, Singapore.

Show MeSH
Related in: MedlinePlus