Improving variant classification for genomic medicine through evidence calibration

ACMG 2026

Mar 10, 2026
Presentation

Anne O'Donnell-Luria, Daniel Zeiberg, Haneen Abderrazzaq, Vikas Pejaver, Sarah Stenton, Himanshu Sharma, Timothy Bergquist, Mugdha Singh, Emily Nadeau, Alicia Byrne, Marc Greenblatt, Steven Harrison, Sean Tavtigian, Leslie Biesecker, Steven Brenner, Predrag Radivojac

Introduction
Accurate classification of genetic variants as pathogenic or benign is crucial to the delivery of genomic medicine. The American College of Medical Genetics and Genomics and Association for Molecular Pathology, in collaboration with ClinGen, have developed and refined standards for the classification of genomic variation by evaluating population, computational, functional, and case data. These measures have substantially improved the quality and consistency of genomic results. The weighting of evidence in these assessments has been largely guided by expert opinion, which likely results in under- and over-weighting some evidence.

 

Methods
There are now sufficient data available to quantitatively assess and define the strength of certain evidence types, from population data (gnomAD), clinically classified variants (ClinVar), and multiplex assays (MaveDB). For example, we found that some computational predictors can provide up to Strong (+4 points) evidence for missense variant pathogenicity/benignity when variants reach thresholds identified by a local probability-based calibration algorithm, while initial guidelines limited computational evidence to Supporting (+1 point). This recommendation was endorsed by the ClinGen Sequence Variant Interpretation Working Group for use in clinical practice.

 

Results
We have expanded this to a number of in-frame indel predictors, which assess a category of variation that is often hard to clinically interpret. The local probability-based calibration algorithm demonstrated that some tools can reach up to Moderate evidence (+3 points). A similar approach was applied to regional missense constraint generated from gnomAD v2, which can be applied as PP2 evidence up to Moderate (+2 points), for regions where benign missense variation is depleted. The increasing availability of high-throughput functional data is aiding the resolution of variants of uncertain significance (VUS). Using a multi-sample skew-normal mixture of distributions of functional (ClinVar) and population (gnomAD) variants, we applied a constrained expectation-maximization algorithm to calculate variant-specific evidence strengths for use in the clinic. These approaches have increased the number of clinically informative classifications while also providing a rigorous basis for the quality of evidence applied, which should improve the reliability of variant classification. A validation study demonstrated that these approaches affect very few variants per individual with rare disease, consistent with the expectation that an individual’s Mendelian condition typically arises from only 1 or 2 variants.

 

Conclusion
Calibration of evidence approaches are improving the delivery of genomic medicine by resolving VUS, thereby increasing diagnostic yield and the ClinGen Computational Working Group has recently expanded its membership to continue to address computational approaches to improving variant classification.