Inherited Disease Genetics Improves the Identification of Cancer-Associated Genes

Abstract

The identification of biologically significant variants in cancer genomes is critical to therapeutic discovery, but it is limited by the statistical power needed to discern driver from passenger. Independent biological data can be used to filter cancer exomes and increase statistical power. Large genetic databases for inherited diseases are uniquely suited to this task because they contain specific amino acid alterations with known pathogenicity and molecular mechanisms. However, no rigorous method to overlay this information onto the cancer exome exists. Here, we present a computational methodology that overlays any variant database onto the somatic mutations in all cancer exomes. We validate the computation experimentally and identify novel associations in a re-analysis of 7362 cancer exomes. This analysis identified activating SOS1 mutations associated with Noonan syndrome as significantly altered in melanoma and the first kinase-activating mutations in ACVR1 associated with adult tumors. Beyond a filter, significant variants found in both rare cancers and rare inherited diseases increase the unmet medical need for therapeutics that target these variants and may bootstrap drug discovery efforts in orphan indications.

Fig 1. Summary statistics of overlay between inherited diseases and cancer mutation datasets and computational model overview.

(A) Breakdown of the shared mutations between the two datasets, showing the percentage with an exact positional match, and of those which are mutations with the exact same or similar residue changes. (B) Refer to Methods for details of methodology. We derived a match score based on the overlay between the inherited diseases and cancer mutation datasets. A signal-to-noise ratio (SNR) was derived through bootstrapping (N = 1000)—sampling with replacement from all of the available reported cases of cancer mutations for a given gene in a given tumor type with recalculations of the match score value. A P value also was derived for the match score using the empirical null distribution, which was generated through a permutation procedure. In each iteration, the number of matches expected by chance was determined based on a binomial distribution (with probability equal to the estimated background match rate). For the number of matches determined, they are randomly assigned based on a uniform distribution to one of the available mutated positions. A match score was subsequently calculated. This process was repeated 10⁴–10⁶ times to generate the empirical null distribution for the match score.

Fig 2. Overall summary of identified mutant variants.

(A) Statistically significant hits based on a signal-to-noise ratio cutoff of 2, at least two reported cases, and a P value cutoff of 0.062 based on the first quartile of adjusted P values from the analyses of all exact matches in nonexpressed genes (see methods). The number of input exomes for given tumor type is indicated at the top of the heatmap. (B) Highlights of hits with significance in at least two tumor types, genes involved in Ras/mitogen-activated protein kinase pathways and Noonan syndrome.

Fig 3. SOS1 mutations are significantly associated with cancer and activate the RAS/MAPK pathway.

(A) Top: All four skin cutaneous melanoma datasets in cBioPortal are presented. Vertical lines indicate mutations in SOS1. Green lines identify overlap with Mendelian diseases. Bottom: A pan-cancer version of the SOS1 mutational data is presented. Data is filtered for N ≥ 4. (B) A table containing all SOS1 variants across the TCGA that match a known Noonan syndrome—associated residue. cBioPortal study ID refers to the indication name in the cBioPortal. The patient ID is the unique sample identifier for the given study. (C) HEK293T transfection experiment for phosphorylated ERK1/2. Dots indicate biological replicates (N = 8). P value was assessed by Student t test.

Fig 4. COL3A1 mutations are associated with melanoma.

(A) Reported mutations in COL3A1 for all skin cutaneous melanoma studies. Mutations overlapping with Ehlers-Danlos syndrome are shown in green. Only glycine -> charged residue mutations (E/R) in conserved regions are shown for clarity. Examination of all melanoma mutations in COL3A1 revealed numerous other glycine triple helix mutations that were distributed along the length of the protein, but are not documented to cause Ehlers-Danlos syndrome in the HUMSAVAR database (shown in black and labeled if N>2). Proline residues that also are important for the triple helix were also mutated (shown below in purple). (B) A model of collagen triple helix structure is shown. Hydrogen bonds stabilizing the triple helix conformation are shown. The tight-steric interactions of the proline residues and hydrogen bonding between the backbone nitrogen of the glycine residue and a carbonyl on an adjacent strand stabilize the triple helix conformation. Substitution of glycine with a bulkier residue directly disrupts these interactions. C) A Kaplan—Meier curve for all COL3A1 mutations in the TCGA (others non-TCGA studies did not have survival data). COL3A1 in SKCM was found to be modestly associated with decreased overall survival based upon log-rank, ExaLT and/or Cox regression. Cox regression outputs for univariate, multivariate, and reduced models with different effect sizes and P values are shown below the Kaplan-Meier curve. SKCM, skin cutaneous melanoma.

Fig 5. ACVR1 mutations in endometrial cancer.

(A) All overlapping residues between the two databases are scored for signal-to-noise ratios and nominal P values. Blue dots represent the hits that are also significant by MutSigCV. The red dots are new associations found in this study. The distribution of P values associated with the expressed genes have a long tail. (B) ACVR1 mutations in endometrial cancer. Overlap with FOP is denoted in green. GS is the glycine serine rich domain. (C) An estimate of the cumulative incidence of ACVR1-mutant diseases. Error bars are based on the difference in ACVR1 frequency estimates in pediatric high-grade glioma and represent min/max estimates of incidence. FOP, fibrodysplasia ossificans progressive; HGG, high-grade glioma; UCEC, uterine corpus endometrial cancer.