A scalable approach to resolving variants of uncertain significance

This paper presents a scalable workflow that integrates large-scale experimental data and predictive models to reclassify the majority of existing variants of uncertain significance (VUS) and preclassify future unobserved variants for 40 disease-associated genes, thereby significantly advancing the clinical application of genomic medicine.

The Gist

Imagine your DNA is a massive, ancient instruction manual for building and running a human body. This manual is so long that it contains billions of letters. Sometimes, when people are born, there are tiny typos in these instructions. Most of the time, these typos don't matter. But some typos can cause serious problems, like a recipe that says "add salt" when it should say "add sugar."

For decades, when doctors found these typos (called variants) in a patient's genetic test, they often hit a wall. They couldn't tell if the typo was a harmless mistake or a dangerous one. So, they had to label it a VUS (Variant of Uncertain Significance).

Think of a VUS like a mystery box in a warehouse. You know it's there, but you don't know if it contains a bomb or a teddy bear. Because you don't know, you can't use it to treat the patient. This causes anxiety for families and leaves doctors helpless.

This paper is about a massive, international team of scientists who decided to stop guessing and start testing these mystery boxes on a huge scale.

The Problem: Too Many Mystery Boxes

The authors point out that over 90% of the typos found in disease-related genes are currently mystery boxes. This is a huge problem because:

1. It causes stress: Patients don't know if they are at risk.
2. It's unfair: People from certain backgrounds often have more mystery boxes because their DNA hasn't been studied as much.
3. It stops progress: You can't fix a problem if you don't know what it is.

The Solution: A Giant "Taste Test" Factory

Instead of studying one typo at a time (which is slow and expensive), the team built a "factory" to test thousands of typos all at once. They used two main methods:

1. The "Mass Production" Lab (MAVEs): Imagine a factory where they take a gene, introduce every possible single-letter change (like changing every letter in a word to every other letter), and see what happens.

   • VAMP-seq: They attach a glowing tag (like a glow-in-the-dark sticker) to the protein. If the protein is broken, the sticker doesn't glow or the protein disappears. This tells them if the typo broke the protein's structure.
   • SGE (Saturation Genome Editing): They use a molecular "scissors" to cut and paste these typos directly into human cells. Then, they see if the cells survive or die. If the cells die, the typo is dangerous.

2. The "Community Collection": They didn't just do their own tests; they gathered data from hundreds of other labs around the world, cleaning it up and putting it all in one giant database.

The "Translator" (Calibration)

Here is the tricky part: The lab tests give you a number (like a score of 7.5), but doctors need a clear answer: "Pathogenic" (Bad) or "Benign" (Good).

The team invented a new translator (called ExCALIBR).

• Old way: Doctors used to draw a hard line. "If the score is above 5, it's bad." But this was too blunt.
• New way: Their translator looks at the whole picture. It asks, "How does this specific typo compare to thousands of known bad ones and thousands of known good ones?" It creates a personalized map for each gene, turning a fuzzy number into a clear, confident verdict.

The Results: Opening the Mystery Boxes

The team applied this system to 40 important genes (which cover about 1% of the entire human genome, but a huge chunk of genetic diseases).

• Solving the Past: They looked at 16,000 existing mystery boxes (VUS). Using their new system, they were able to open 75% of them! They could confidently say, "This one is safe," or "This one is dangerous."
• Predicting the Future: They even looked at 90,000 typos that haven't been found in people yet. They "pre-classified" them. This means that if a doctor finds one of these in a patient tomorrow, they won't have to wait years for an answer. They will already know if it's a bomb or a teddy bear.

Why This Matters

Think of this like upgrading from a hand-drawn map to Google Maps.

• Before: Doctors were guessing, leaving patients in the dark.
• Now: They have a high-definition, automated system that can instantly tell them the status of a genetic typo with over 99% accuracy.

The Takeaway

This paper isn't just about one gene or one disease. It's a blueprint. It shows that if we build these "factories" to test our genes and use smart "translators" to understand the results, we can eventually make the term "Variant of Uncertain Significance" (VUS) disappear from medical records.

This means fewer anxious families, fairer healthcare for everyone, and doctors who can finally give clear answers to their patients. They turned a mountain of "we don't know" into a mountain of "we know."

Technical Summary

1. Problem Statement

The clinical interpretation of genetic variants is currently hindered by the prevalence of Variants of Uncertain Significance (VUS). Over 90% of missense variants in disease-associated genes are classified as VUS, creating a major barrier to genomic medicine.

• Clinical Impact: VUS cannot be used for diagnosis or treatment, leading to patient distress and inequitable care, particularly for individuals from ancestries underrepresented in genomic databases.
• Data Gap: While experimental data (e.g., MAVEs) and computational predictors exist, they are often gene-specific, lack standardized calibration, and are difficult to translate into clinical evidence according to ACMG/AMP guidelines.
• Goal: To develop a scalable, systematic workflow that resolves existing VUS and "preclassifies" future unobserved variants using only experimental and predictive evidence, aiming to make the term "VUS" obsolete for a significant portion of the clinical genome.

2. Methodology

The authors, part of the IGVF (Impact of Genomic Variation on Function) Consortium, developed an end-to-end pipeline involving data generation, curation, calibration, and classification.

A. Data Generation (Experimental)

• Multiplexed Assays of Variant Effect (MAVEs):
  • VAMP-seq (Variant Abundance by Massively Parallel Sequencing): Measured steady-state protein abundance for 64,963 variants across 3 genes (G6PD, TSC2, F9). It identifies pathogenicity via protein destabilization/degradation.
  • Saturation Genome Editing (SGE): Measured cellular fitness effects of 33,280 SNVs across 7 genes (including BRCA1, BRCA2, PALB2, BARD1). This captures effects on RNA expression, splicing, and protein function in haploid human cells.
• Arrayed Phenotypic Assays: High-throughput assays measured protein abundance (DUAL-IPA), subcellular localization (Variant Painting), and protein-protein interactions (sqY2H) for 1,407 variants across 163 genes to identify mechanistic hallmarks of pathogenicity.

B. Data Curation and Integration

• Curated 68 community-generated datasets (68 published assays) covering 30 additional genes.
• Integrated these with IGVF data to create a harmonized dataset of 255,354 unique variants across 40 clinically relevant genes (including 18 ACMG Secondary Findings genes).
• Annotated variants with scores from top predictors: AlphaMissense, MutPred2, and REVEL.

C. Automated Calibration (The Core Innovation)

To translate raw scores into clinical evidence (ACMG/AMP points), the authors developed new calibration methods to replace subjective, threshold-based approaches:

• ExCALIBR (Experimental score CALIBRator): A statistical method for experimental data. It fits skew-normal mixture models to score distributions of pathogenic, benign, gnomAD, and synonymous variants. It calculates gene-specific priors and posterior probabilities to assign evidence strength (ranging from -8 to +8 points).
• Gene-Specific Predictive Calibration: Addressed the flaws of "genome-wide" calibration by developing a gene-specific framework for predictors (REVEL, MutPred2, AlphaMissense). This prevents misassignment of evidence due to gene-specific score distribution shifts.

D. Scalable Classification Workflow

• A workflow that combines calibrated experimental and predictive evidence points.
• Filters: Excluded variants with conflicting evidence, start-loss variants (non-SGE), and splice-altering variants (where predictors are untrained).
• Classification: Summed points to assign ACMG/AMP categories: Benign, Likely Benign, VUS, Likely Pathogenic, or Pathogenic.

3. Key Results

A. Resolution of Existing VUS

• Applied to 16,115 ClinVar VUS across the 40 genes.
• 75% (12,135 variants) were reclassified as either Pathogenic/Likely Pathogenic (PLP) or Benign/Likely Benign (BLB).
  • 69% reclassified as BLB.
  • 6% reclassified as PLP.
• Accuracy: The workflow showed extremely low discordance with established ClinVar classifications (<1% error rate). Specifically, only 0.66% of definitive classifications disagreed with ClinVar when using REVEL as the predictor.
• Validation: Pathogenic classifications showed strong association with disease phenotypes in the All of Us biobank (elevated odds ratios for cancer and other conditions).

B. Preclassification of Unobserved Variants

• Analyzed 90,530 unobserved missense variants (not yet seen in ClinVar or gnomAD).
• 62% received sufficient evidence to be "preclassified" as PLP or BLB immediately upon observation.
• Notably, 8% of these unobserved variants were preclassified as PLP (higher than the 6% for ClinVar VUS), consistent with the hypothesis that unobserved variants are more likely to be deleterious.

C. Mechanistic Insights

• Integration of arrayed assays revealed that pathogenic variants disrupt diverse protein properties (abundance, localization, interactions).
• Orthogonality: Different genes require different assays to detect pathogenicity (e.g., some show mislocalization without abundance loss).
• Case Study (Factor IX): Combined VAMP-seq and Variant Painting distinguished between variants causing protein retention in the Golgi (rescuable) versus rapid degradation (non-rescuable), offering therapeutic implications.

4. Key Contributions

1. Production-Scale Data: Generated high-quality experimental data for 62,215 variants across 10 genes and curated data for 193,139 variants across 30 additional genes.
2. Novel Calibration Methods: Introduced ExCALIBR and gene-specific predictive calibration, moving away from arbitrary thresholds to empirical, unbiased evidence assignment.
3. Scalable Workflow: Demonstrated that experimental and predictive evidence alone can resolve 75% of VUS with high accuracy, without needing family segregation or population frequency data for the initial classification.
4. Resource Creation: Developed MaveMD (clinical interface for experimental data) and PredictMD (for predictive data) to make this evidence accessible to clinicians.
5. Preclassification Concept: Established a framework to classify unobserved variants before they appear in clinical testing, preventing the accumulation of future VUS.

5. Significance

• Clinical Impact: This work provides a roadmap to eliminate the "VUS" bottleneck for a significant portion of the clinical genome (representing ~1% of all coding variants but covering high-priority disease genes).
• Equity: By relying on functional data rather than population frequency (which is biased toward European ancestry), this approach offers more equitable variant interpretation for diverse populations.
• Future-Proofing: The "preclassification" strategy ensures that as sequencing scales up, new variants can be interpreted immediately, accelerating the transition to precision healthcare.
• Scalability: The automated calibration and workflow are designed to scale to thousands of genes as new MAVE technologies and predictors emerge.

Conclusion: The study proves that systematic, scalable generation of functional data, combined with rigorous statistical calibration, can resolve the majority of current VUS and prevent future ones, fundamentally advancing the utility of genomic medicine.