Multimodal framework to resolve variants of uncertain significance in TSC2

Abstract

Efforts to resolve the functional impact of variants of uncertain significance (VUS) have lagged behind the identification of new VUS; as such, there is a critical need for scalable VUS resolution technologies. Computational variant effect predictors (VEPs), once trained, can predict pathogenicity for all missense variants in a gene, set of genes, or the exome. Existing tools have employed information on known pathogenic and benign variants throughout the genome to predict pathogenicity of VUS. We hypothesize that taking a gene-specific approach will improve pathogenicity prediction over globally-trained VEPs. We tested this hypothesis using the gene TSC2, whose loss of function results in tuberous sclerosis, a multisystem mTORopathy affecting about 1 in 6,000 individuals born in the United States. TSC2 has been identified as a high-priority target for VUS resolution, with (1) well-characterized molecular and patient phenotypes associated with loss-of-function variants, and (2) more than 2,700 VUS already documented in ClinVar. We developed Tuberous sclerosis classifier to Resolve variants of Uncertain Significance in T SC2 (TRUST), a machine learning model to predict pathogenicity of TSC2 missense VUS. To test whether these predictions are accurate, we further introduce curated loci prime editing (cliPE) as an accessible strategy for performing scalable multiplexed assays of variant effect (MAVEs). Using cliPE, we tested the effects of more than 200 TSC2 variants, including 106 VUS. It is highly likely this functional data alone would be sufficient to reclassify 92 VUS with most being reclassified as likely benign. We found that TRUST's classifications were correlated with the functional data, providing additional validation for the in silico predictions. We provide our pathogenicity predictions and MAVE data to aid with VUS resolution. In the near future, we plan to host these data on a public website and deposit into relevant databases such as MAVEdb as a community resource. Ultimately, this study provides a framework to complete variant effect maps of TSC1 and TSC2 and adapt this approach to other mTORopathy genes.

Figure 1:. Development and initial evaluation of a TSC2-specific classifier for predicting pathogenicity of missense variants.

(A) Receiver operator characteristic (ROC) curve for TRUST classifier on the test set (n=151 variants). (B) Area under the ROC curve (AUROC) for TRUST and some of the best-performing globally-trained VEPs on the test set. Globally-trained VEPs trained on the ClinVar dataset are displayed in gray bars. (C) ROC curve for TRUST classifier on the holdout dataset (n=206 variants). (D) AUROC for TRUST and some of the best-performing globally-trained VEPs on the holdout dataset. (E) Plot of the TRUST output probability of pathogenicity by TSC2 amino acid position for all variants in the study (n=11,728). Truth set variants are displayed as blue (BLB) or red (PLP) circles. All other variants are displayed as either a blue (predicted BLB) or red (predicted PLP) ‘X’ symbol. A cartoon diagram of TSC2 functional domains is provided for reference. (F) Violin plot showing the distribution of TRUST probability output by variant class for the full truth set as well as missense VUS.

Figure 2:. Overview of curated loci prime editing (cliPE) method for TSC2 MAVE.

(A) Cartoon overview showing a simplified version of the cliPE method for performing the TSC2 MAVE reported herein. (B) epegRNA screen design based on regions of high local VUS density in TSC2. Exons targeted are shown and the codons overlapped by the epegRNA are provided. Codon numbering is based on TSC2 transcript NM_000548.5.

Figure 3:. cliPE mapping of variant effect in more than 200 variants, including 106 missense VUS.

(A) CMAS scores plotted by amino acid position and grouped by epegRNA target region. Scores are derived from 3–4 biological replicates. Dashed lines indicate the median score for either synonymous (Syn) or PTC variants for that particular exon. Plots of each variant with standard error are provided in Figures S10 to S15. *Exon 42 is the final exon in TSC2 and as such PTC variants likely escape nonsense-mediated decay and encode functional proteins, as evidenced by the median score for synonymous and PTC variants being equivalent. (B & C) Histograms of CMAS, bars colored by variant type. Synonymous (n=41) and PTC (n=41) variants (excluding final exon 42 PTCs) used to determine pathogenicity thresholds are shown in the top plot (B). Missense BLB (n=13), PLP (n=12), and VUS (n=106) are shown in the bottom plot (C). Full details of the truth set can be found in Table S4.

Figure 4:. TSC2 MAVE has the potential to reclassify many missense VUS.

Sankey diagram showing that with moderate strength evidence of benignity or pathogenicity, the current TSC2 MAVE data could enable reclassification of up to 86.8% of missense VUS. In agreement with our in silico classifier, most TSC2 missense VUS appear to be rare benign variants.

Figure 5:. Validation of gene-specific VEP performance on MAVE data.

(A) Receiver operator characteristic (ROC) curve for TRUST classifier on missense variants classified unambiguously by CMAS score from the TSC2 MAVE (n=114 variants). (B) Area under the ROC curve (AUROC) for TRUST and some of the best-performing globally-trained VEPs on the MAVE dataset. (C) Correlation (Pearson) plot of TRUST, CMAS, and some of the bestperforming globally-trained VEPs. (D) Heatmap of overall performance of gene-specific in silico classifier and best-performing globally-trained VEPs averaged across all three classification datasets (ClinVar test set, holdout set, and MAVE data).

Figure 6:. Global explanations provide insight into TSC2 gene-specific in silico pathogenicity prediction.

Contributions of features to final model as determined by SHAP values. Other globally-trained VEPs including AlphaMissense, PROVEAN, SIFT, and PolyPhen-2 are among the top features. Notably, a structure-related feature (relative solvent accessibility; RSA) and an in silico predictor of protein stability (MAESTRO) are also among the top features, suggesting that missense variants in TSC2 may impact protein folding or abundance.