1. Introduction
In the era of high-throughput sequencing, the bottleneck of genomic medicine has shifted decisively from data generation to data interpretation. We can now sequence a human genome for under $200, generating 3-4 million variants per individual. Yet, the clinical utility of this data hinges entirely on a single, high-stakes question: Does this specific variant cause the patient's disease?
Variant classification is the rigorous process of assigning a probability of pathogenicity to a genetic alteration. It is not merely a labeling exercise; it is a medical diagnostic procedure with profound consequences. A "Pathogenic" classification can trigger irreversible surgeries (e.g., prophylactic mastectomy in BRCA1 carriers) or lifelong surveillance. Conversely, a missed diagnosis leaves a patient in a diagnostic odyssey. This series explores the scientific, statistical, and ethical frameworks that govern this critical discipline.
2. The Pre-ACMG Chaos: A Tower of Babel
Prior to 2015, the field of clinical genetics lacked a standardized language. Laboratories used idiosyncratic terms like "mutation," "polymorphism," "variant of unclear significance," or "favor deleterious."
The Consequence of Inconsistency
A 2013 study showed that laboratories agreed on variant classification only 34% of the time. One lab might classify a variant as "Pathogenic" based on a single case report, while another called it "VUS" due to lack of functional data. This discordance eroded clinician trust and endangered patients.
The lack of population data was a major culprit. As discussed in our previous post on population databases, the "Manrai artifact" revealed that many variants classified as pathogenic in the literature were actually common benign polymorphisms in non-European populations.
3. The 2015 ACMG/AMP Framework
In 2015, the American College of Medical Genetics and Genomics (ACMG) and the Association for Molecular Pathology (AMP) published a landmark joint consensus recommendation (Richards et al., 2015). This document fundamentally transformed the field by establishing:
- Standardized Terminology: The 5-tier system (Pathogenic to Benign).
- Evidence Criteria: 28 specific criteria (e.g., PVS1, BS1) weighted by strength (Supporting, Moderate, Strong, Very Strong).
- Combination Rules: Explicit algorithms for combining criteria to reach a classification.
Figure 1: The Hierarchy of Evidence
The framework requires a "pyramid" of evidence. A single piece of strong evidence is rarely enough; it must be supported by the base.
4. The 5-Tier Classification System
The output of the ACMG framework is one of five classes. While these are categorical labels, they represent underlying statistical probabilities of pathogenicity (Plon et al., 2008).
Class 5: Pathogenic
Probability of Pathogenicity: > 99%
Clinical Action: Actionable. Can be used for clinical decision making, predictive testing in at-risk relatives, and prenatal diagnosis.
Class 4: Likely Pathogenic
Probability of Pathogenicity: > 90%
Clinical Action: Actionable. In most clinical settings, treated the same as Pathogenic. However, caution is advised for irreversible decisions (e.g., termination of pregnancy).
Class 3: Variant of Uncertain Significance (VUS)
Probability of Pathogenicity: 10% - 90%
Clinical Action: NOT Actionable. Should not be used for clinical decision making. Family segregation studies or research testing may be pursued. This is the "clinical purgatory" of genomics.
Class 2: Likely Benign
Probability of Pathogenicity: < 10%
Clinical Action: Not reported clinically (usually). Treated as negative.
Class 1: Benign
Probability of Pathogenicity: < 0.1%
Clinical Action: Not reported.
5. The Clinical Impact of Classification
The distinction between these classes is not academic. In oncology, a BRCA1 VUS cannot justify a mastectomy. In cardiology, an MYH7 VUS cannot justify implanting an ICD.
The high rate of VUS results (often 30-40% in large panels) remains the single biggest barrier to the widespread adoption of genomic medicine. Reducing this rate requires better data sharing (ClinVar), better functional assays, and better algorithms—topics we will cover in the rest of this series.
6. The Future: Bayesian Frameworks
While the 2015 guidelines were rule-based (e.g., "1 Strong + 2 Moderate"), the field is moving toward a quantitative Bayesian framework (Tavtigian et al., 2018). This approach assigns numerical points to each evidence code based on its odds of pathogenicity, summing them to reach a posterior probability.
This shift allows for more nuance—a "Strong" piece of evidence isn't just a label; it's a mathematical weight (e.g., Odds of Pathogenicity = 18.7). This evolution is critical for integrating AI scores and continuous data types.