Combining multiplexed functional data to improve variant classification

Abstract

With the surge in the number of variants of uncertain significance (VUS) reported in ClinVar in recent years, there is an imperative to resolve VUS at scale. Multiplexed assays of variant effect (MAVEs), which allow the functional consequence of 100s to 1000s of genetic variants to be measured in a single experiment, are emerging as a source of evidence which can be used for clinical gene variant classification. Increasingly, there are multiple published MAVEs for the same gene, sometimes measuring different aspects of variant impact. Where multiple functional consequences may need to be considered to get a more complete understanding of variant effects for a given gene, combining data from multiple MAVEs may lead to the assignment of increased evidence strength which could impact variant classifications. Here, we provide guidance for combining such multiplexed functional data, incorporating a stepwise process from data curation and collection to model generation and validation. We illustrate the potential of this approach by showing the integration of multiplexed functional data from four MAVEs for the gene TP53. By following these steps, researchers can maximize the value of MAVEs, strengthen the functional evidence for clinical variant classification, reclassify more VUS, and potentially uncover novel mechanisms of pathogenicity for clinically relevant genes.

Figure 1:. Assessment of assay performance prior to combining multiplexed functional data.

(A) Functional score distributions for a hypothetical perfect and real-world MAVEs for assumed functionally normal (Synonymous & BLB) variants and assumed functionally abnormal PTC & PLP variants. (B) An example of real-world multiplexed functional data illustrating overlap between the synonymous and PTC variants scored in the assay. (C) Critical metrics for MAVE performance.

Figure 2:. Evaluating the utility of the integrated functional score.

A method for combining data will involve merging the datasets from different MAVEs. A final score for each variant will be generated using an appropriate statistical, or machine learning method. The integrated score must be assessed for improved variant resolution compared to the multiplexed functional data from individual MAVEs based on how well the different scores distinguish BLB and PLP variants in the truth set.

Figure 3:. Example of combining multiplexed functional data for four TP53 MAVEs.

(A) Principal component analysis of the 4 datasets. The first two principal components are shown. We then used an unsupervised machine learning method, K-means clustering, to visualize candidate BLB and PLP clusters within the dataset. For each method, we calculated the OddsPath (B) or the inverse of OddsPath (C) to determine what strength of evidence could be applied towards pathogenicity (B) or benignity (C) compared to two of the best performing single MAVEs alone.