Our paper on variant classification studies is published in Frontiers in Genetics

Our paper on real-world performance studies of variant classification platforms is now published in Frontiers in Genetics, in the Human and Medical Genomics section. We are pleased to share this work with the researchers, laboratory teams, and software developers who face the questions it addresses.

Written by Vladimir Mitev at Helena Bioinformatics, the paper proposes a method for investigating how a classification platform behaves when it encounters real cases. It brings together a structured review process, worked examples, and a specification that other teams can use and adapt.

Why a disagreement needs an explanation

A genetic variant is a change in a DNA sequence. Classifying it involves assessing the evidence for whether it contributes to disease. That assessment can depend on population frequency, experimental findings, inheritance, family observations, and the rules used to combine those sources.

When software and a reference laboratory assign different classifications to the same variant, an agreement percentage leaves much of the story unresolved. The software may contain a defect. The laboratory may have family evidence that was never supplied to the software. The two may be applying different versions of interpretation guidelines. Some cases remain uncertain even after the available evidence has been reviewed.

Those situations call for different actions. A rule implemented incorrectly needs correction. Missing family evidence needs expert review and, where appropriate, further investigation. Uncertainty needs to remain visible in the record. Treating every disagreement as the same event makes it harder to identify what should happen next.

This is particularly difficult for small development teams. They need a disciplined way to document early findings while they build the independent review arrangements and organizational resources required for larger studies. Our paper sets out a proposed method for that stage.

Following a result through three layers

The framework separates analytical performance, classification performance, and clinical performance. These layers help a reviewer locate where an observation arose.

Figure 1. The three layers help locate whether an observation concerns data processing, classification, or clinical reporting. Reproduced from Mitev, Frontiers in Genetics (2026), under the Creative Commons Attribution license. Select the figure to view it at full size.

At the analytical layer, the question is whether variants were detected, annotated, and processed correctly within the stated scope. A variant missing from the supplied input file requires a different investigation from a variant received by the platform and then lost during processing.

At the classification layer, the reviewer examines how evidence and interpretation rules produced the assigned class. The framework uses several reference sources, with their weight determined by evidence strength. An expert-panel assessment, a published functional study, and a single laboratory report each have a defined role in that comparison.

At the clinical layer, the focus includes prioritization and the information presented for professional review. The paper distinguishes the quality of a narrative report from its factual faithfulness. A report can be incomplete or poorly organized; it can also make an unsupported claim. Recording these separately helps identify the appropriate corrective work.

Human experts remain responsible for evaluating conflicting evidence, explaining disagreements, and completing clinical sign-off before results are used in patient care. Software can organize the evidence and preserve the record, but the proposed protocol requires documented professional judgment.

Figure 2. Automated steps prepare the evidence package; expert reviewers assess disagreements and document the resulting decisions. Reproduced from Mitev, Frontiers in Genetics (2026), under the Creative Commons Attribution license. Select the figure to view it at full size.

What the demonstration shows

The paper illustrates the framework across three cohorts comprising 97 scored cases, processed with three versions of the Helena classifier. CellGenetics, the clinical partner reference laboratory in Sofia, supplied the comparator classifications used in the demonstration.

One example concerns a target variant that was absent from the input supplied to the platform. The investigation checked the relevant coordinates and followed the data through processing before attributing the absence to the upstream pipeline. That sequence matters: an upstream explanation should only be accepted after checking that the platform preserved the input correctly.

Another example identified a defect in how a classification rule was applied across inheritance patterns. The paper records the defect and the required remediation. Other disagreements involved evidence that could not be obtained from the variant file alone, or documented limitations in the automated workflow.

The framework gives each reviewed outcome a primary category, records additional contributing factors, and connects the finding to an action. It also preserves the classifier version, evidence sources, and review decisions so that later work can examine the basis of an earlier result.

The cohorts use different classifier versions and, in some comparisons, different units of analysis. Their agreement percentages therefore cannot be treated as a simple trend in platform performance. The examples illustrate how the method is applied; they do not establish clinical accuracy or fitness for use.

Making the method available for independent use

This remains an early demonstration on one platform with one peer laboratory comparator. The categories were developed using the same cohorts that illustrate them, and the paper describes the associated conflicts of interest. Independent application is needed to establish whether different reviewers can use the method consistently in other settings.

The paper proposes a concrete next test: independent reviewers score the same cases using a fixed specification, without seeing each other's decisions, and report their agreement before resolving differences. That would provide evidence about reproducibility that the current demonstration cannot supply.

We are glad to have this work published and available for examination beyond our own team. The framework is non-proprietary, and its supplementary specification is free to adopt and adapt. Researchers and developers can read the decision rules, examine the examples, and assess how the method could apply to their own studies.

Thank you to our partners

We thank our clinical partners at CellGenetics and their team in Sofia. Their clinical reports provided the reference laboratory comparisons used in the demonstration cohorts and made it possible to examine the framework against real laboratory interpretations.

We also thank ELTA 90M, our distribution partner in Bulgaria, for sponsoring the publication fee in full. Their support made the open-access publication of this paper possible. As disclosed in the article, ELTA 90M had no role in the study design, data collection, analysis, interpretation, manuscript preparation, or decision to submit.

We are grateful to both teams for their contributions to this work and its publication.

Read the full paper in Frontiers in Genetics

Vladimir Mitev. A methodological framework for real-world performance studies of clinical variant classification platforms at early organizational stages. Published 14 September 2026. doi:10.3389/fgene.2026.1925492