Systematic analysis of somatic mutations in phosphorylation signaling predicts novel cancer drivers

Abstract

Large-scale cancer genome sequencing has uncovered thousands of gene mutations, but distinguishing tumor driver genes from functionally neutral passenger mutations is a major challenge. We analyzed 800 cancer genomes of eight types to find single-nucleotide variants (SNVs) that precisely target phosphorylation machinery, important in cancer development and drug targeting. Assuming that cancer-related biological systems involve unexpectedly frequent mutations, we used novel algorithms to identify genes with significant phosphorylation-associated SNVs (pSNVs), phospho-mutated pathways, kinase networks, drug targets, and clinically correlated signaling modules. We highlight increased survival of patients with TP53 pSNVs, hierarchically organized cancer kinase modules, a novel pSNV in EGFR, and an immune-related network of pSNVs that correlates with prolonged survival in ovarian cancer. Our findings include multiple actionable cancer gene candidates (FLNB, GRM1, POU2F1), protein complexes (HCF1, ASF1), and kinases (PRKCZ). This study demonstrates new ways of interpreting cancer genomes and presents new leads for cancer research.

Figure 1

Analysis overview. (A) Missense SNVs (crosses) were extracted from cancer genomes and classified as phosphorylation-associated (pSNVs; red crosses) if they affected phosphosites (red P-circles) and their flanking regions (pink rectangles) or kinase domains (blue rectangles). We designed the statistical model ActiveDriver to find cancer genes with significantly enriched or depleted pSNVs. Using pathway enrichment analysis, we identified GO terms, pathways and protein complexes with over-represented pSNVs. (B) Phosphorylation network composed of experimentally determined kinase–substrate interactions. To find kinases important in cancer, all kinase-centric signaling modules (light blue star) were tested for statistical enrichment of pSNVs. Each such module comprised a fixed central kinase (blue diamond) and its direct upstream kinases and downstream substrates (black diamonds and circles within the light blue star). (C) To find clinically relevant signaling modules, we designed a novel local network search algorithm HyperModules that combines pSNVs, kinase–substrate interactions, and patient survival. (D) Distribution of cancer samples across cancer types. Two glioblastoma data sets are shown separately in green (Parsons et al, 2008) and purple (Cancer Genome Atlas Research Network, 2008). (E) Distribution of genes with pSNVs across cancer types. (F) Phosphosites are enriched in somatic cancer mutations in comparison to genome-wide mutation rate averaged across cancer genomes (binomial test, error bars show s.d.).

Figure 2

Genes with significant phosphosite mutations (pSNVs). (A) ActiveDriver analysis revealed 58 genes with significant mutation rates in phosphosite regions. Top barplot shows gene significance (log10 P-value) by cancer type, and bottom color-strip shows cancer types with related pSNVs. Rightmost panel represents 14 additional genes found in a composite analysis of somatic mutations of all cancer types. No pSNVs are known for KRAS, it is listed due to significant depletion of phosphosite mutations. Known cancer genes (*) and drug targets (o) are enriched in the list of discovered genes. (B) N-terminal protein sequence region 16–52 of CTNNB1 includes seven phosphosites (blue letters) and is significantly enriched in pSNVs. Four out of five SNVs in CTTNB1 are found in the region, including three direct mutations (red letters) and one phosphomimetic mutation (green letter). Seven phosphosites (blue letters) are known targets of several kinases, shown in boxes below the sequence. Names of known cancer genes are underlined and printed in bold. (C) Comparison of ActiveDriver results with standard mutation frequency-based gene ranking. Top plot shows 6182 genes with at least one missense SNV, ranked in decreasing order by number of missense point mutations across all cancer samples, followed by increasing order by gene length. Bottom plot shows the position of 58 ActiveDriver-predicted genes with significant pSNVs in the global SNV-based ranking. Black lines represent known cancer genes and red lines represent candidate cancer genes.

Figure 3

Phosphosite mutations in TP53 correlate with increased patient survival. (A) ActiveDriver analysis of TP53 pSNVs identifies a mosaic of nine phosphosite mutation hotspots (red columns) and deserts (blue columns) across multiple cancer types (top panel). Middle panel shows the protein sequence of TP53 with 29 phosphosites (black bars) and number of flanking phosphosites per residue (yellow and orange). Bottom panel shows number of SNVs per position, with the majority of mutations accumulating to the DNA-binding domain in the central region of the sequence. Protein domains of TP53 are shown below the chart (TAD, transcriptional activation; PRO, proline-rich region; NLS, nuclear localization signal; HO, homo-oligomerization; C, C-terminus). (B) Kaplan–Meier analysis of clinical data of ovarian cancer patients shows that TP53 pSNVs correlate with increased survival. Survival of patients with pSNVs (black solid line) is compared to survival of patients with other, non-phosphorylation-associated SNVs (red dashed line) as well as patients with wild-type pSNVs (blue dashed line). (C) Long-term survivors of glioblastoma with TP53 pSNVs (black solid line) show a similar correlation with increased survival; however, these observations have borderline statistical significance due to small sample size.

Figure 4

Functional enrichment analysis of pSNVs. (A) Top 50 non-redundant, phosphorylation-specific GO categories with statistically significant enrichment of pSNVs (FDR P⩽0.05, TP53 and EGFR excluded). Categories are ranked according to number of samples, shown on the X-axis. Colors denote different types of cancer. (B) Toll signaling pathway is enriched in pSNVs, with genes defined by GO and connections from the kinase–substrate network. Network shows genes with pSNVs (colored circles) and their non-mutated phosphorylation partners (small black circles) in the Toll pathway. Arrows denote kinase–substrate relationships, known cancer genes are labeled using bold and underline, and known drug targets are shown on yellow background. (C) The set of 17 non-redundant Reactome pathways with enriched pSNVs, ranked by number of affected samples (X-axis). (D) Six of 21 CORUM protein complexes with significant pSNV enrichment.

Figure 5

pSNVs in the kinase–substrate network. (A) The set of 59 kinase-centric signaling modules with significant pSNV enrichment (FDR P⩽0.05), grouped according to their positions in the defined kinase hierarchy. Bars show number of samples with pSNVs in upstream kinases (light blue), downstream substrates (dark blue), and central kinase (yellow). Feedback loop mutations (orange) occur in proteins that are both upstream kinases and downstream substrates of the central kinase. Color-strip under the bars shows statistical significance of pSNV enrichment in each module, asterisks denote known cancer genes, and circles denote known drug targets. (B) Tissue-specific phosphosite mutations in EGFR. 14 lung cancer mutations occur in the kinase domain (diamonds), while 14 glioblastoma (GBM) pSNVs associate to phosphosites. (C) Eleven glioblastoma pSNVs in EGFR (A289, C291) flank a novel extracellular phosphosite at T290 (shown in red). (D) Signaling network of the master kinase PRKCZ and its downstream target STK11 is frequently phospho-mutated and involves several tumor suppressors such as PTEN, TP53, and LATS1. The network involves pSNVs in ∼25% of cancer samples. Colored circles denote genes with pSNVs, small black circles denote genes with no pSNVs, and arrows denote kinase–substrate phosphorylation events. Names of known cancer genes are underlined and printed in bold, and drug targets are shown on yellow background.

Figure 6

Top survival-associated network module in ovarian cancer. (A) One of the top significant survival-associated kinase–substrate signaling modules involves 11 genes and eight pSNVs found in the neighborhood of HCK kinase. (B) Kaplan–Meier survival analysis shows a statistically significant difference of module-associated ovarian cancer patients (red line) and other patients (black line), as all patients with mutations in the module were alive at the end of the study. (C) Permutation test shows that the observed correlation with increased survival is unlikely to be found in equivalently structured networks with shuffled mutations. (D) Survival module involves mutually exclusive pSNVs in eight genes and patients, including the active phosphosite and kinase domain pSNV in the gene of HCK kinase (filled diamond). (E–G) Additional clinical variables show significant correlation with mutations in the module, including enrichment of alive patients (left) and tumor-free patients (middle) and depletion of subjects of additional chemotherapy (right). Expected values are shown with s.d. from binomial sampling. Names of known cancer genes are underlined and printed in bold.