Massively Parallel Functional Analysis of BRCA1 RING Domain Variants

Abstract

Interpreting variants of uncertain significance (VUS) is a central challenge in medical genetics. One approach is to experimentally measure the functional consequences of VUS, but to date this approach has been post hoc and low throughput. Here we use massively parallel assays to measure the effects of nearly 2000 missense substitutions in the RING domain of BRCA1 on its E3 ubiquitin ligase activity and its binding to the BARD1 RING domain. From the resulting scores, we generate a model to predict the capacities of full-length BRCA1 variants to support homology-directed DNA repair, the essential role of BRCA1 in tumor suppression, and show that it outperforms widely used biological-effect prediction algorithms. We envision that massively parallel functional assays may facilitate the prospective interpretation of variants observed in clinical sequencing.

Keywords: BRCA1; deep mutational scanning; human genetic variation; protein function; variants of uncertain significance.

Figure 1

(A) Scheme for leveraging scores from parallelized assays for BRCA1 RING function into predictions for the function of the full-length BRCA1 protein in homology-directed DNA repair. (B-F) Scoring the E3 ligase and BARD1-binding activities of BRCA1 RING domain variants. (B) A sequence-function map of the effect of missense mutations in the BRCA1 RING domain on E3 ligase function. The functional score for each variant is the slope of the fit curve, normalized by setting stop codons to a score of 0 and the wild-type to a score of 1. Each position in BRCA1(2-103) is arranged along the x-axes, structural features of the RING domain are diagrammed above. The amino acid substitutions, grouped by side-chain properties, are on the y-axes. The E3 ligase scores range from improved activity versus wild-type (red), equivalent to wild-type (white), to less than wild-type (blue). Yellow represents the wild-type residue and gray missing or low confidence data. (C) A sequence-function map of the effect of missense mutations in the BRCA1 RING domain on BARD1-RING binding. Coloring as in panel B. (D) Comparison of the variant scores from the deep mutational scan for E3 ligase activity versus literature-reported E3 ligase activities for the same BRCA1 variants (Brzovic et al., 2003; Morris et al., 2006). The Wilcoxon rank sum test (WRST) was used to test for significant differences between the categories. The biggest outlier in the wild type-like category, D96N, not only performed poorly as an E3 ligase score but also failed to bind to BARD1 and to support homology-directed repair in cells (Table S2). (E) Comparison of BARD1-binding scores from the two-hybrid experiment versus literature-reported BARD1 binding by the same BRCA1 variants (Brzovic et al., 2003; Ransburgh et al., 2010). The WRST was used to test for significant differences between categories. (F) The relationship between the quality-filtered E3 ligase functional scores and the BARD1-binding scores. Colors indicate the clinical classification or database of origin for each variant.

Figure 2

Testing BRCA1 variants for their ability to rescue homology-directed DNA repair. (A) Integrated into the genome of the HDR reporter strain is one copy of the GFP gene containing an I-SceI homing endonuclease site that introduces an in-frame stop codon, along with another copy of the GFP gene lacking both its start and stop codons that functions as a donor for DNA repair (Pierce et al. 2001). The cell line is depleted for BRCA1 by transfection with an siRNA that targets the 3′-UTR of the endogenous gene. I-SceI and a full-length variant of BRCA1 are then transfected into the cells. After 3 days, the GFP+ population is assessed by flow cytometry. (B) The wild-type normalized percentage of cells that were GFP+ in the BRCA1 HDR rescue assay is shown. Experiments were performed in triplicate and error bars represent the standard error. siRNA, BRCA1 variant, and clinical classification or database of origin is indicated by color. Dashed horizontal line represents the midpoint between the average HDR scores for known pathogenic and benign variants.

Figure 3

Scores from massively parallel E3 ligase and BARD1-binding assays on BRCA1 RING domain variants are better predictors of the HDR activity of the full-length protein. The linear relationship of the E3 ligase scores (A), BARD1-binding scores (B), and HDR scores. (C) A support vector regression (SVR) model of HDR rescue scores from the combination of the E3 ligase and BARD1 binding functional scores. Variants are colored by database of origin. The blue line represents the least-squares fit of the displayed data. Known pathogenic splice variant R71G is marked with an asterisk. (D) Experimentally or computationally derived values for the effect of missense variants on protein function were used to predict the effect on HDR. The LOOCV R² and RMSE for each model is indicated. The RMSE of LOOCV indicates the average distance between the HDR rescue predictions and the true HDR rescue scores, and the LOOCV R² is the overall correlation between predicted and observed values; low RMSE and high R² indicate better predictive power. For A-GVGD, the Grantham deviation value was used. Source of HDR predictions is indicated by color, linear model (LM), and SVR.

Figure 4

Predicted HDR rescue scores for 1287 BRCA1 RING variants create a prospective map of the effect of missense substitutions. (A) A histogram of the predicted HDR scores for the 1287 BRCA1 RING variants with both high-confidence E3 ligase and BARD1-binding scores. (B) A histogram of the 59 of the 62 BRCA1 RING variants found in the human population, clinical classification, or database of origin is indicated by color. Known splice variants R71G and R71K are marked with an asterisk. (C) Sequence-function map of the predicted HDR rescue scores for BRCA1 RING variants. Colors are centered on 1.0 as wild type (white). Structural features of the BRCA1 RING domain are diagrammed above.