Unified Pathogenicity Scales for Genetic Variant Interpretation

This direction is merged into another direction: variants of uncertain significance. Its last version stays readable and citable here.

Version 1 of 15 papers16 conceptsfrozen 2026-10-04

In the effort to identify the genetic causes of rare diseases, researchers face a fundamental "apples and oranges" problem. A mutation that breaks a protein involved in smelling a specific flower might be biologically "damaging" to that protein, but it is unlikely to cause a severe developmental disorder. Conversely, a seemingly minor change in a core developmental gene could be catastrophic.

Historically, computational tools have struggled to account for this context. They often produce scores that are either specific to a single type of mutation (like a single-letter swap) or specific to a single protein, making it nearly impossible to compare the "risk" of different variants across the entire genome. Recent work is shifting toward unified pathogenicity scales—frameworks that calibrate these scores so that a 0.9 in one gene or mutation type means exactly the same thing as a 0.9 in another.

The Challenge of Incommensurate Scores

To understand why calibration is necessary, consider two distinct gaps in traditional genetic modeling:

  1. The Structural Gap: Single Nucleotide Variants (SNVs) and small Insertions/Deletions (InDels) are physically different. Most models are trained on one or the other, leading to scoring distributions that cannot be compared. If an SNV gets a "70" and an InDel gets a "0.4," a clinician cannot easily decide which one is the more likely culprit.
  2. The Biological Gap: Evolutionary models can tell us if a protein sequence is "constrained" (meaning it hasn't changed much in millions of years). However, high constraint doesn't always equal human disease risk. Without calibration, a mutation in a non-essential gene might appear more "pathogenic" than a mutation in a critical gene simply because the model lacks organism-level context.

Bridging Mutation Types with Shared Architectures

To solve the structural gap, Chang Li, Kai Wang, and Xiaoming Liu developed MetaRNN. Their insight was that by using a consistent neural architecture—specifically Recurrent Neural Networks (RNNs)—and identical feature sets for both SNVs and InDels, the model could be forced to project both types of mutations into the same mathematical space.

In MetaRNN: differentiating rare pathogenic and rare benign missense SNVs and InDels using deep learning, the researchers demonstrate that this unified representational strategy produces directly comparable score distributions. This is not just a matter of convenience; it enables burden tests. In a burden test, researchers aggregate all the "risky" variants in a gene to see if they appear more often in sick patients than healthy ones. Without a unified scale, you couldn't easily "add up" the combined risk of an SNV and an InDel in the same calculation.

Calibrating Across the Proteome

Even if we can compare different mutations within a gene, we still need to compare them across different genes. A team including Rose Orenbuch, Thomas A. Hopf, and Debora Susan Marks addressed this by creating popEVE.

The core mechanism of popEVE, detailed in Proteome-wide model for human disease genetics, is the use of human population variation as a "ground truth" for calibration. While evolutionary history (how a protein changed across species) provides a raw score of biological damage, the researchers map these scores onto a unified, organism-level scale.

They do this by looking at how much variation humans actually tolerate in the wild. If a gene is highly conserved across mammals but humans carry many "damaging" mutations in it without getting sick, the model down-weights the pathogenicity of mutations in that gene. This ensures that a high score represents a consistent level of biological risk, whether the mutation occurs in a metabolic enzyme or a specialized sensory receptor.

Why Unified Scales Work

The intuition behind this research direction is that pathogenicity is a relative property, not an absolute one.

By standardizing these metrics, these models allow for:

  • Cross-Gene Prioritization: Clinicians can look at a list of 10 candidate mutations across 10 different genes and trust that the highest score is truly the most likely cause of the patient's symptoms.
  • Integrated Statistical Power: By merging SNVs, InDels, and even nonsense variants into a single metric, researchers can identify "disease genes" that were previously hidden because no single mutation type reached statistical significance on its own.

Ultimately, by moving from gene-specific rankings to a global, calibrated scale, tools like MetaRNN and popEVE transform genetic interpretation from a series of isolated guesses into a unified, proteome-wide map of human health and disease.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_unified_pathogenicity_scales_for_genetic_a11d4d56,
  title = {Unified Pathogenicity Scales for Genetic Variant Interpretation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unified-pathogenicity-scales-for-genetic-variant-interpretation/rdr_a11d4d561c554aa8a441c5f54e81fa8d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_unified_pathogenicity_scales_for_genetic_a11d4d56_v1,
  title = {Unified Pathogenicity Scales for Genetic Variant Interpretation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unified-pathogenicity-scales-for-genetic-variant-interpretation/rdr_a11d4d561c554aa8a441c5f54e81fa8d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}