Modeling the ACMG/AMP variant classification guidelines as a Bayesian classification framework

Abstract

Purpose: We evaluated the American College of Medical Genetics and Genomics/Association for Molecular Pathology (ACMG/AMP) variant pathogenicity guidelines for internal consistency and compatibility with Bayesian statistical reasoning.

Methods: The ACMG/AMP criteria were translated into a naive Bayesian classifier, assuming four levels of evidence and exponentially scaled odds of pathogenicity. We tested this framework with a range of prior probabilities and odds of pathogenicity.

Results: We modeled the ACMG/AMP guidelines using biologically plausible assumptions. Most ACMG/AMP combining criteria were compatible. One ACMG/AMP likely pathogenic combination was mathematically equivalent to pathogenic and one ACMG/AMP pathogenic combination was actually likely pathogenic. We modeled combinations that include evidence for and against pathogenicity, showing that our approach scored some combinations as pathogenic or likely pathogenic that ACMG/AMP would designate as variant of uncertain significance (VUS).

Conclusion: By transforming the ACMG/AMP guidelines into a Bayesian framework, we provide a mathematical foundation for what was a qualitative heuristic. Only 2 of the 18 existing ACMG/AMP evidence combinations were mathematically inconsistent with the overall framework. Mixed combinations of pathogenic and benign evidence could yield a likely pathogenic, likely benign, or VUS result. This quantitative framework validates the approach adopted by the ACMG/AMP, provides opportunities to further refine evidence categories and combining rules, and supports efforts to automate components of variant pathogenicity assessments.

Keywords: Bayesian framework; medical genetics; unclassified variants; variant classification; variants of uncertain significance.

Figure 1

Permissible solutions to Likely Pathogenic (LP) and Likely Benign (LB) equations. X–Y Combinations of Prior probability of pathogenicity (Prior_P) and Odds very strong (O_Vst) lying to the right of the blue curve satisfy combining rules Likely Pathogenic (ii–vi). X–Y combinations of Prior_P and O_Vst lying to the left of the red curve satisfy the Likely Benign combining rule (ii). Values between the two curves and above their intersection at (Prior_P=0.25, O_Vst=81) simultaneously meet LP and LB criteria. Values outside of the two curves and below their intersection at (Prior_P=0.25, O_Vst=81) meet neither LP nor LB criteria. The black triangle marks the minimum simultaneous solution of LP and LB at Prior_P=0.25, O_Vst=81. The black circle marks the solution of LP and LB at Prior_P=0.10, O_Vst=350 illustrated in Table 1.