Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2018 Oct 19;9(1):4353.
doi: 10.1038/s41467-018-06867-x.

Mutational interactions define novel cancer subgroups

Jack Kuipers  1   2 , Thomas Thurnherr  3 , Giusi Moffa  4   5 , Polina Suter  3   6 , Jonas Behr  3 , Ryan Goosen  7 , Gerhard Christofori  7 , Niko Beerenwinkel  8   9
Affiliations

Mutational interactions define novel cancer subgroups

Jack Kuipers et al. Nat Commun. .

Abstract

Large-scale genomic data highlight the complexity and diversity of the molecular changes that drive cancer progression. Statistical analysis of cancer data from different tissues can guide drug repositioning as well as the design of targeted treatments. Here, we develop an improved Bayesian network model for tumour mutational profiles and apply it to 8198 patient samples across 22 cancer types from TCGA. For each cancer type, we identify the interactions between mutated genes, capturing signatures beyond mere mutational frequencies. When comparing mutation networks, we find genes which interact both within and across cancer types. To detach cancer classification from the tissue type we perform de novo clustering of the pancancer mutational profiles based on the Bayesian network models. We find 22 novel clusters which significantly improve survival prediction beyond clinical information. The models highlight key gene interactions for each cluster potentially allowing genomic stratification for clinical trials and identifying drug targets.

PubMed Disclaimer

Conflict of interest statement

The authors declare no competing interests.

Figures

Fig. 1
Fig. 1
Overview of the analyses. Starting from the mutation data, we perform two types of analysis. Supervised learning of the Bayesian network structure for each known cancer type, allowing us to uncover mutational interactions and visualise pancancer heterogeneity. Unsupervised clustering of the mutation data into components with common interactions to uncover a novel stratification of the patient samples
Fig. 2
Fig. 2
The cancer-specific connections between the genes. For clarity we limit the display to only include connections between the 20 most frequent and connected mutations per cancer type. The Bayesian networks for each cancer type are overlaid on the same union of gene nodes. Edges highlight interactions between the selected mutations and are coloured by the cancer type for which they occur (leukaemia: yellow; glioblastoma: green, etc.). Directed edges point in the inferred direction of dependency, whereas undirected edges indicate cases where the direction cannot be inferred. Solid edges indicate positive correlation (a degree of co-occurrence) while dashed edges indicate negative correlation (a degree of exclusivity) between the genes connected. Genes are coloured according to the colours of their edges. Those with edges in only one specific cancer type are grouped and labelled with the same colour. Genes with edges in different cancer types are arranged in the two circles in the centre. The size of each gene correlates with the total number of edges across all cancer types, including edges that are not shown. Black diamonds mark putatively actionable targets
Fig. 3
Fig. 3
Visualisation of pancancer heterogeneity. 2D visualisation of the similarity between patient samples based on their fit to each cancer-specific Bayesian network, highlighting the heterogeneity within and across cancer types. For example, stomach, breast and liver cancer show high inter-tumour heterogeneity with a high spread across the plot, whereas pancreatic cancer shows low inter-tumour heterogeneity and is much more localised. Ovarian and breast cancers as well as bladder cancer and lung adenocarcinomas show similar mutational profiles, while lower grade glioma is rather distinct from other cancer types, as is thyroid cancer. The solid shapes are based on contours that together contain a total of 50% of the respective cancer types. The group of samples on the lower right possess no mutations among the 201 genes while those exhibiting a mutation only in TP53 are also indicated. Versions highlighting certain cancer types are displayed in Supplementary Fig. 5
Fig. 4
Fig. 4
De novo clustering. a Assignment of the 8198 patient samples to the 22 clusters, labelled A−V, based on Bayesian network clustering of their mutation profiles. The left-hand side of the bar for each cluster indicates the number of patient samples with a given cancer type, while the right-hand side indicates the breakdown into the known subtypes. b Survival probabilities of the 22 clusters
Fig. 5
Fig. 5
The cluster-specific connections between the genes. The connections between the 20 most frequent and connected genes per cluster (rather than per cancer type, as in Fig. 2). Black diamonds near nodes indicate putatively actionable genes. Edges highlight interdependencies between the selected mutations: solid for positive and dashed for negative correlation. Directed edges point in the inferred direction of dependency, whereas undirected edges are where the direction cannot be inferred. Node size reflects the total number of edges, including edges not shown. Nodes are coloured by combining the colours of their edges from the different clusters
Fig. 6
Fig. 6
Focus on TP53. a Cancer types or b clusters are characterised by interactions of TP53 with different genes. For cancer types or clusters where TP53 shows a similarly high mutation frequency (around 0.5), we display all its interactions (positive correlation is represented by solid arrows, negative correlation by dashed). Groups are annotated by the cancer type or cluster name and the respective mutation frequency of TP53 across samples (in brackets). Several of these interactions have been studied in cancer. For example, an inverse relationship between ARID1A and TP53 was previously observed in gastro-intestinal cancers while TP53 mutations co-occur with RB1 in bladder cancer
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

See all "Cited by" articles

References

    1. Sun S, Schiller JH, Spinola M, Minna JD. New molecularly targeted therapies for lung cancer. J. Clin. Investig. 2007;117:2740–2750. doi: 10.1172/JCI31809. - DOI - PMC - PubMed
    1. Higgins MJ, Baselga J. Targeted therapies for breast cancer. J. Clin. Investig. 2011;121:3797–3803. doi: 10.1172/JCI57152. - DOI - PMC - PubMed
    1. Roock WD, Vriendt VD, Normanno N, Ciardiello F, Tejpar S. KRAS, BRAF, PIK3CA, and PTEN mutations: implications for targeted therapies in metastatic colorectal cancer. Lancet Oncol. 2011;12:594–603. doi: 10.1016/S1470-2045(10)70209-6. - DOI - PubMed
    1. Groenendijk FH, Bernards R. Drug resistance to targeted therapies: déjà vu all over again. Mol. Oncol. 2014;8:1067–1083. doi: 10.1016/j.molonc.2014.05.004. - DOI - PMC - PubMed
    1. McLendon R, et al. Comprehensive genomic characterization defines human glioblastoma genes and core pathways. Nature. 2008;455:1061–1068. doi: 10.1038/nature07385. - DOI - PMC - PubMed

Publication types