🧬 biology

Intrinsic sequence evidence cannot reach a pathogenic classification for missense variants under a partitioned ACMG framework

This study demonstrates that under a partitioned ACMG framework preventing evidence reuse, intrinsic sequence data alone is mathematically insufficient to classify missense variants as pathogenic, revealing that reducing the burden of uncertain significance requires diverse evidence types rather than improved computational predictors.

Original authors: Anees Ahmed Mahaboob Ali, Radhakrishnan Delhibabu, Everette Jacob Remington Nelson

Published 2026-09-08

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Anees Ahmed Mahaboob Ali, Radhakrishnan Delhibabu, Everette Jacob Remington Nelson

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the hidden world of human genetics, a single letter change in our DNA code can sometimes spell the difference between health and a rare, life-altering disease. When doctors sequence a patient's genome to find the cause of a mysterious illness, they often encounter a specific type of change called a missense variant. This is a mutation where one building block of a protein is swapped for another, like replacing a brick in a wall with a slightly different one. The problem is that for the vast majority of these changes, science cannot yet say with certainty if they are harmless or dangerous. In medical reports, these are labeled as "variants of uncertain significance," leaving clinicians and families in a state of limbo, unable to make definitive treatment decisions. The standard rules used by geneticists to classify these variants rely on a point system, where different types of evidence are added up to reach a verdict. However, a critical gap has emerged: the evidence that can be gathered just by looking at the DNA sequence itself is often too weak to cross the threshold needed to declare a variant dangerous.

A team of researchers at the Vellore Institute of Technology in India has now measured exactly how wide that gap is, and they have found a surprising, structural limit to what DNA analysis alone can achieve. They built a sophisticated computer system to test the rules used by geneticists, focusing on a group of rare bleeding and platelet disorders where getting the diagnosis wrong can lead to dangerous or even fatal treatments. Their work reveals that no matter how advanced a computer predictor becomes, it cannot classify a missense variant as pathogenic based solely on the DNA sequence and its annotations. The rules of the game simply do not allow it. The researchers discovered that to solve these cases, doctors must look beyond the sequence to other kinds of evidence, such as how the patient's body actually functions or how the disease runs in a family.

The researchers developed a tool called DISCERN to explore this problem. Unlike previous tools that tried to guess the answer by looking only at the genetic code, DISCERN was designed to act like a careful auditor. It was built to ensure that every piece of evidence used to make a decision was counted exactly once, preventing the kind of double-counting that can accidentally inflate the importance of a finding. The system was tested against a massive database of expert classifications for bleeding disorders, a field where the stakes are incredibly high. For instance, confusing two similar-looking bleeding disorders can lead a doctor to prescribe a drug that is harmless for one but causes severe blood clots for the other, or to recommend a surgery that is unnecessary and dangerous. The researchers wanted to see if their system, which strictly followed the rules of evidence, could solve these difficult cases using only the genetic data available in a standard lab report.

When they ran the system on 425 missense variants that experts had already labeled as either clearly harmful or clearly harmless, the results were stark. Using only the intrinsic evidence found in the DNA sequence, the system assigned a "pathogenic" or "likely pathogenic" classification to zero of the 425 variants. It simply could not reach the necessary score. The system was able to identify harmless variants, but the rules of the evidence framework meant that the points available from the sequence alone were insufficient to cross the line into a dangerous classification. Even when the researchers tried to restore every possible piece of information that the strict rules had withheld, the system could only recover a small fraction of the cases that experts had solved. The study showed that the barrier was not a flaw in the computer program or a lack of computing power; it was a fundamental property of the guidelines themselves. The evidence required to prove a missense variant is dangerous simply does not exist in the sequence data alone.

This finding challenges the common hope that a better computer algorithm will eventually solve the problem of uncertain genetic results. The researchers demonstrated that even the most advanced existing predictors, which are often used as a shortcut in diagnosis, hit the same invisible ceiling. The issue is not that the computers are bad at reading the code, but that the code itself does not contain enough information to make the final call. To move a variant from "uncertain" to "pathogenic," a doctor needs a different kind of proof: a functional test showing how the protein behaves, data showing the disease running through a family tree, or a specific clinical symptom that points to only one disease. The study highlights that the solution to the uncertainty crisis lies not in better software, but in gathering these other, more complex types of evidence that connect the genetic change to the actual experience of the patient.

The researchers also tested their system's ability to distinguish between diseases that look almost identical. In the world of bleeding disorders, different genetic causes can produce the same symptoms, making it easy to misdiagnose a patient. The system successfully used the combination of the genetic finding and the patient's specific symptoms to narrow down the correct disease, but it did so with a crucial safety feature. If the system was not confident enough to make a call, or if a proposed treatment was dangerous for the suspected condition, it would stop and recommend a specific test to get more information. This "safety interlock" worked perfectly in their tests, flagging every scenario where a treatment would be harmful and never flagging a harmless one. This proves that the system can act as a reliable guide, not by guessing the answer, but by knowing exactly when it needs more information to be safe.

Ultimately, this work provides a clear map of where genetic medicine currently stands and where it must go. The researchers showed that for missense variants, the path to a definitive diagnosis is blocked if one relies only on the genetic sequence. The "ceiling" they found is a structural limit of the current guidelines, meaning that no amount of computational power can break through it without new data. The path forward requires a shift in how genetic evidence is gathered, moving toward a model where the genetic finding is coupled with the patient's specific clinical picture and functional test results. By measuring the gap so precisely, the study offers a practical roadmap for laboratories: stop trying to force a verdict from the sequence alone, and instead focus on collecting the specific, non-genetic evidence that the rules require to make a safe and accurate diagnosis.

Technical Summary: Intrinsic Sequence Evidence Cannot Reach a Pathogenic Classification for Missense Variants Under a Partitioned ACMG Framework

Problem Statement
Missense variants in rare-disease genes are predominantly classified as Variants of Uncertain Significance (VUS). Under the American College of Medical Genetics and Genomics (ACMG) and Association for Molecular Pathology (AMP) framework, the criteria available to a laboratory based solely on sequence and annotations are few and individually weak (e.g., absence from population databases and computational predictions of deleteriousness). Closing the gap to a "pathogenic" classification requires evidence types not inherent to the sequence, such as functional assays, segregation data, case-control counts, or phenotype-specific evidence (PP4).

The core problem addressed is the magnitude of this evidence shortfall. Existing automated classifiers often fail to account for the structural limitation that intrinsic sequence evidence alone cannot mathematically reach the threshold for pathogenicity under a framework that prohibits double-counting evidence. Furthermore, current tools struggle to integrate disease models (required for PP4) without reusing evidence or falling into circular reasoning (e.g., relying on ClinVar labels). In inherited bleeding and platelet disorders, misclassification has severe clinical consequences, such as administering contraindicated treatments (e.g., desmopressin in Type 2B von Willebrand disease) or missing definitive treatments (e.g., stem cell transplantation for Leukocyte Adhesion Deficiency Type 3).

Methodology: The DISCERN Framework
The authors introduce DISCERN, a system comprising three coupled components and a trustworthiness layer, designed to operate within a partitioned ACMG framework:

Variant Classification Engine: Implements ACMG/AMP criteria using a ClinGen point-based Bayesian formulation. It assigns calibrated posterior probabilities of pathogenicity.
Disease Discrimination Engine: Operates over ten clusters of confusable diseases (spanning 31 diseases and 29 genes). It computes a posterior over disease identity using likelihood ratios derived from specific laboratory observations (e.g., RIPA mixing studies, flow cytometry) that distinguish clinically similar conditions.
The Joint Model and Evidence Partition: The system's core innovation is a joint posterior $P(D, V | E)$ $P (D, V ∣ E)$ that couples disease identity ( $D$ $D$ ) and variant pathogenicity ( $V$ $V$ ). Crucially, the ACMG evidence space is partitioned so that each criterion is owned by exactly one factor (variant, disease, functional, segregation, phasing, or de novo).
- Variant Factor: Owns intrinsic criteria (e.g., PVS1, PP3, PM2).
- Disease Factor: Owns PP4 (phenotypic specificity).
- Functional Factor: Owns PS3/BS3.
- This structural partition prevents the double-counting of evidence, a risk inherent in coupled models where a phenotype-aware classifier might inadvertently re-add evidence already used to establish the disease.
Trustworthiness Layer: Includes isotonic regression for calibration, an abstention rule (withholding calls when posterior separation is insufficient), and a safety interlock that flags contraindicated management plans.

Benchmarking and Evaluation
The system was evaluated on the ClinGen Evidence Repository (eRepo) expert panel surface for inherited bleeding and platelet disorders, specifically focusing on 425 missense variants with expert pathogenic/benign labels.

Blinding: Evaluations were ClinVar-blinded and used a time-split protocol to prevent label leakage.
Comparators: DISCERN was compared against REVEL, AlphaMissense, InterVar, and GeneBe.
Differential Diagnosis: A curated benchmark of 42 published cases was used to test disease discrimination, comparing against phenotype-blind baselines and the LIRICAL tool.

Key Results

The Intrinsic Evidence Ceiling:
The most significant finding is a structural limitation: On intrinsic sequence evidence alone, the framework assigned a pathogenic classification to 0 of 425 missense variants.
- The maximum points achievable from intrinsic evidence (PM2 + PP3) is 3.0, falling short of the 6 points required for "likely pathogenic."
- Restoring all withheld criteria (functional, segregation, phenotype) recovered only 89 of the 316 variants the expert panel called pathogenic.
- This confirms that the "uncertain fraction" is not a failure of prediction quality but a consequence of the evidence framework's arithmetic.
Performance Metrics:
- DISCERN achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.939 (95% CI: 0.906–0.967) on the eRepo surface, statistically indistinguishable from REVEL (0.954).
- Under a time-split evaluation, AUROC was 0.927, again statistically indistinguishable from REVEL.
- Calibration: DISCERN demonstrated excellent out-of-sample Expected Calibration Error (ECE) of 0.017, significantly better than raw REVEL (0.092) and AlphaMissense (0.167). Unlike other tools, DISCERN provides a calibrated probability as delivered, without requiring post-hoc fitting to expert labels.
Evidence Partition and Double-Counting:
- The partition successfully prevented evidence reuse. In a genome-wide analysis of 12,240 variants, the partition prevented pathogenicity inflation in 33.2% of variants that would have been misclassified by a standard scorer summing all criteria into a single factor.
- The system reproduced expert reasoning with high agreement for criteria it could derive (PVS1 $\kappa=0.98$ , PP3 $\kappa=0.95$ ) and correctly withheld criteria where input data was missing (e.g., PM2 where gnomAD frequency was undetermined).
Safety and Differential Diagnosis:
- The safety interlock achieved 100% sensitivity and specificity in flagging contraindicated management plans across five treatment-divergent scenarios.
- In differential diagnosis, DISCERN achieved 100% Top-1 accuracy on the 42-case benchmark. However, the authors note that in this specific benchmark, the gene identity largely determined the diagnosis (93% accuracy for a phenotype-blind gene lookup), and the "informative subset" where phenotype was required to break ties was small ( $n=3$ ).
- When compared to LIRICAL on phenotype-only inputs (withholding the gene), DISCERN showed a non-significant lead (+17%), but when allowed to use non-HPO laboratory findings (35 features), it significantly outperformed LIRICAL (+35%, $p=0.02$ ).

Significance and Claims
The paper claims that the inability of intrinsic sequence evidence to classify missense variants as pathogenic is a property of the ACMG point arithmetic, not a defect of any specific implementation.

The Ceiling is Structural: No better in silico predictor can solve the VUS problem for missense variants because the shortfall lies in the type of evidence available (functional/phenotypic), not the quality of the sequence prediction.
The Role of Coupling: The study demonstrates that a disease-variant coupling is necessary to supply the missing evidence (specifically PP4 and functional PS3) from clinical information a service already holds.
Modesty on Coupling Validation: The authors explicitly state that the disease-variant coupling itself has not been proven in real data within this study. Public data lacks the necessary paired phenotype-genotype cohorts with independent ground truth to validate that using the disease posterior to resolve a VUS actually moves the variant toward its true state. This evaluation is pre-registered for future cohort data.
Practical Implication: For diagnostic laboratories, the study concludes that reducing the VUS fraction requires acquiring different kinds of evidence (functional, segregation, or specific phenotypes), not merely adopting stronger sequence predictors. The DISCERN framework provides a mechanism to auditably integrate these evidence streams while preventing double-counting.

Technical Summary: Intrinsic Sequence Evidence Cannot Reach a Pathogenic Classification for Missense Variants Under a Partitioned ACMG Framework

More like this