Systematic Functional Annotation of Somatic Mutations in Cancer

Abstract

The functional impact of the vast majority of cancer somatic mutations remains unknown, representing a critical knowledge gap for implementing precision oncology. Here, we report the development of a moderate-throughput functional genomic platform consisting of efficient mutant generation, sensitive viability assays using two growth factor-dependent cell models, and functional proteomic profiling of signaling effects for select aberrations. We apply the platform to annotate >1,000 genomic aberrations, including gene amplifications, point mutations, indels, and gene fusions, potentially doubling the number of driver mutations characterized in clinically actionable genes. Further, the platform is sufficiently sensitive to identify weak drivers. Our data are accessible through a user-friendly, public data portal. Our study will facilitate biomarker discovery, prediction algorithm improvement, and drug development.

Keywords: TCGA; cellular assay; clinical marker; driver mutation; drug sensitivity; functional genomics; functional proteomics; therapeutic target.

Figure 1. Overview of the functional genomic platform and cancer mutations tested

(A) Mutations (muts), corresponding wild-type (WT) and fusion genes were selected from TCGA projects and MD Anderson Cancer Center patient databases. Clones were generated by the HiTMMoB approach, and tested in in vitro growth-factor dependent cell viability assays with Ba/F3 and MCF10A cell models. Mutations and wild-type variants were classified into functional categories based on these results. MCF10A cell lines stably expressing selected mutations were generated for reverse-phase protein array (RPPA) analysis. The numbers of mutant, wild-type and fusion constructs are annotated at each step. (B) Pie charts showing the proportions of the mutations annotated in OncoKB or Personalized Cancer Therapy (PCT) or PubMed literature among all the 1049 mutations tested. (C) Bar plots showing the literature coverage of mutations for the top 10 genes with the greatest number of mutations tested, as shown by the percentages of tested mutations per gene annotated in OncoKB or PCT or PubMed. See also Figure S1 and Table S1.

Figure 2. Functional annotation summary of wild-type genes and mutations

(A) The numbers of missense (purple), indel (orange), nonsense (red) and silent (white) mutations tested are shown in parentheses. (B) The distribution of mutation types tested per gene for the 21 genes with >10 mutations tested is shown. (C, D) The functional annotations for wild-type genes (C) and mutations (D) in Ba/F3 (blue) and MCF10A (red) cell line models are presented based on the growth-factor independent cell viability assay results. (E) The number of mutations in each functional annotation is shown in parentheses. Eleven mutations with inconclusive functional annotations in Ba/F3 and MCF10A models were excluded. See also Figure S2 and Table S2.

Figure 3. Comparison of our functional annotation with literature data and computational predictions

(A) Activating and neutral mutations from our (non-pooled) in vitro platform results were compared to oncogenic, likely oncogenic, and likely neutral mutations annotated from OncoKB. The percentage of mutations in each category is shown. Activating mutations were further classified into strong, moderate and weak based on the degree of activating comparing with the corresponding wild-type genes. Numbers on the bars indicate the mutation numbers in each group. (B) ROC curves of 21 commonly used computational algorithms based on the functional calls in this study, with AUC scores for the top 5 algorithms. (C) Enrichment of activating mutations in three 3D computational algorithms. Numbers on the bars indicate the mutation numbers in each group. ****, p < 10^-5. See Figure also S3 and Table S3.

Figure 4. Functional proteomic profiling of select mutations in MCF10A

(A, B) A rank order plot showing the overall reverse-phase protein assay (RPPA) protein expression pattern of each BRAF mutation relative to BRAF^L584F (A) or each EGFR mutation relative to EGFR^G719D (B). Spearman rank correlations were calculated based on all the proteins profiled, and the mutants are color coded by their functional annotations. (C) RPPA unsupervised clustering analysis of 268 MCF10A cell lines stably overexpressing selected mutations was performed. Cluster names are annotated in the top row of the feature track. Gene names and functional calls are also presented in the feature track. Key differentially expressed proteins across clusters are highlighted on the right. (D) Differential cell cycle pathway activities among different clusters. (E) Differential PI3K/Akt pathway activities between activating and non-activating mutations in the PI3K cluster. (F) Differential EMT pathway activities between activating and non-activating mutations in the BRAF cluster. (D-F) The middle lines indicate median values, the top and bottom of the box indicate 25th and 75th percentiles, and whiskers indicate 10th and 90th percentiles. See also Figure S4 and Table S4.

Figure 5. Analysis of EGFR and BRAF mutation allelic series

(A) Functional annotations of EGFR (top) and BRAF (bottom) allelic series. Only recurrent mutations of the series are shown. The frequency (based on TCGA and GENIE databases) and location of mutations tested are shown in lollipop plots. In the heatmap (from top to bottom), the consensus functional annotation, OncoKB annotation, computational prediction by 3D structural cluster (HotSpot3D, HotMAPs), population-based (VEST) cancer-focus (CanDrA), Mutation Assessor and hotspot predictions (based on Chang et al., 2016) of mutations tested in this study are shown. (B) Structural clusters of activating mutations in EGFR (left) and BRAF (right). Filled color and border color of the mutation label indicate the OncoKB annotation and our consensus functional annotation, respectively. See also Figure S5.

Figure 6. Overview of FASMIC portal

(A) Data portal summary. (B) Mutation table of EGFR. (C) The 3D protein structure of p110α (encoded by PIK3CA) with residue K111 highlighted in red. (D) Bar plot of mutational frequency in different cancer types. (E) Functional predictions of various computational algorithms shown in a table with damaging mutations highlighted in dark red. (F) Differential protein expression profile of an EGFR mutant related to the wild-type gene is displayed in a sorted scatter plot.