Lab assays and AI predictors join forces to classify cancer variants fairly across ancestries

Every year, thousands of people who undergo genetic testing for hereditary cancer risk receive a result that is neither reassuring nor actionable: a variant of uncertain significance, or VUS. These are changes in the DNA sequence of a cancer gene whose effect on the encoded protein simply cannot be determined with the evidence currently available. For patients and their families, a VUS result can mean months or years of limbo, with screening decisions postponed and relatives left wondering whether they carry a hidden risk. Now, a study published in Genome Medicine by Nikita Bora of the Lee Kong Chian School of Medicine, Mihaly Badonyi and Ankit K. Pathak of the University of Edinburgh’s MRC Human Genetics Unit, Joseph A. Marsh, and Joanne Ngeow, together with colleagues and the SG10K_Health Consortium, offers a systematic way out of that limbo — and, crucially, a way that works fairly regardless of a patient’s ancestry.

The core problem the team set out to address is one of the most quietly consequential issues in clinical genomics: ancestry bias. Many widely used computational tools for predicting whether a genetic variant is harmful are trained on clinical databases and population datasets that heavily under-represent global diversity. European populations dominate these resources, so the statistical patterns the tools learn reflect the variants and genetic backgrounds of European-ancestry individuals. When the same tools are applied to variants found in African, East Asian, South Asian, Latino, or Middle Eastern populations, their predictions can become unreliable — and, in a clinical setting, that unreliability translates directly into uneven diagnostic outcomes. A variant might be confidently classified as benign in one population and left indeterminate, or worse, misclassified, in another.

Bora and colleagues turned to two families of methods that are, by design, free of this ancestry bias. The first is experimental: multiplexed assays of variant effect, or MAVEs. These are high-throughput laboratory techniques that measure the functional consequence of thousands of variants simultaneously, typically by synthesising all possible mutations in a gene region and scoring how each one affects protein function or abundance in a pooled cell population. Methods such as variant abundance by massively parallel sequencing (VAMP-seq) quantify how stable each variant protein is inside cells, while deep mutational scanning approaches measure effects on activity or interaction. Because MAVEs measure function directly, their evidence does not depend on which human populations have been sequenced or clinically characterised.

The second family is computational: population-free variant effect predictors, or VEPs. The team focused on three leading examples — GEMME, EVE, and CPT-1. These tools infer the likely impact of a variant without relying on human clinical datasets. GEMME works by tracing the evolutionary conservation of each position and substitution across thousands of homologous protein sequences, reasoning that positions and amino acid changes conserved through evolution are less tolerant to mutation. EVE takes a different route, using a statistical model — a variational inference approach — trained on the vast evolutionary record contained in protein sequence families to estimate how likely a given amino acid is at a given position. CPT-1 incorporates structural biophysics, estimating the change in the Gibbs free energy (ΔΔG) that a variant induces in the folded protein, a quantity that reflects how strongly the mutation destabilises the protein’s three-dimensional structure. All three draw on evolutionary and physical principles rather than on human population data, making them inherently unbiased with respect to ancestry.

The study evaluated MAVE- and VEP-based classification across five cancer-associated genes for which high-quality MAVE datasets exist. The genes include well-known hereditary cancer players such as PTEN, a tumour suppressor whose loss predisposes carriers to breast, thyroid, and endometrial cancers, and TP53, the guardian of the genome, mutated in Li-Fraumeni syndrome. The team calibrated the experimental and computational evidence to the ACMG/AMP guidelines — the framework issued by the American College of Medical Genetics and Genomics and the Association for Molecular Pathology that clinicians worldwide use to combine evidence into five-tier variant classifications ranging from benign to pathogenic. This calibration step is technically demanding: it requires translating raw assay scores or predictor outputs into the standardised evidence strengths (strong, moderate, supporting, and so on) that the guidelines accept.

The results reveal a nuanced division of labour between the two approaches. MAVEs proved to be more conservative and more decisive classifiers. They assigned fewer variants to the pathogenic category, but they also produced fewer indeterminate classifications — meaning that when a MAVE does speak, it tends to speak clearly. The assays showed heightened sensitivity in the regions that had actually been functionally assayed, which makes sense: within the windows of the gene that the experiment covers, the direct measurement of function is the most reliable evidence available. VEPs, by contrast, identified a broader range of pathogenic variants overall, because they can score every variant in the protein, not just those covered by the assay. Each approach, in other words, has blind spots the other can fill.

The headline finding is what happens when the two are combined. By merging the clinical evidence strengths derived from MAVEs and from VEPs, the team reclassified over 90 percent of variants of uncertain significance across two reference datasets — the SG10K_Health cohort from Singapore’s National Precision Medicine programme, which includes Chinese, Malay, Indian, and other Southeast Asian ancestries, and the Mexico City Prospective Study, which captures Latino and admixed American populations. That figure is remarkable: nine out of ten variants that would otherwise sit in diagnostic limbo were moved to at least likely benign or likely pathogenic. For a patient carrying one of these variants, the difference between a VUS and a likely benign result can be the difference between indefinite anxiety and a clear answer.

The team also isolated variants where the MAVE-based and VEP-based classifications disagreed, and these discordances are technically informative. They highlight method-specific limitations: a MAVE may miss variants whose effects are context-dependent or outside the assayed region, while a VEP may misjudge variants in positions where evolutionary conservation is a poor guide to human disease relevance, or where structural destabilisation does not translate into loss of function. Understanding where and why the two approaches diverge is essential for building robust combined classifiers — and for knowing when a single line of evidence should be trusted.

The broader significance of the work lies in its implications for equity in clinical genomics. Under-represented populations have long borne the brunt of the VUS problem: because most clinical variant databases are European-dominated, variants in non-European patients are more likely to be left unclassified, delaying diagnosis and preventive care. By drawing on evidence sources that are ancestry-blind by construction — direct functional measurement and evolutionary or biophysical inference — the combined MAVE-plus-VEP framework offers a path to variant interpretation that generalises across populations. The authors suggest this could help mitigate disparities in hereditary cancer diagnosis across ancestries, a goal that becomes increasingly urgent as genome sequencing expands globally.

There are, of course, practical hurdles ahead. High-quality MAVE datasets exist for only a fraction of clinically relevant genes, and scaling the assays to the full exome remains a formidable experimental challenge. The calibration of assay scores and predictor outputs to ACMG/AMP evidence strengths must be standardised and validated across laboratories before widespread clinical adoption. But the study demonstrates the principle clearly: experimental and computational evidence, each ancestry-unbiased, each with complementary strengths, can be combined to resolve the vast majority of uncertain variants in hereditary cancer genes. As genome medicine matures, this kind of evidence integration may become the standard by which the promise of precision oncology is extended — fairly — to everyone.

}

Subject of Research: Combining MAVEs and computational predictors improves variant classification across ancestries in hereditary cancer genes

Article Title: Combining MAVEs and computational predictors improves variant classification across ancestries in hereditary cancer genes

Article References: Bora, N., Badonyi, M., Pathak, A. K., Manisekaran, R. G., Marsh, J. A., Ngeow, J., & SG10K_Health Consortium (2026). Combining MAVEs and computational predictors improves variant classification across ancestries in hereditary cancer genes. Genome Medicine. https://doi.org/10.1186/s13073-026-01799-4

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01799-4

Keywords: Combining, MAVEs, computational, predictors, improves, variant, classification, across, ancestries, hereditary, cancer, genes