Equitable variant interpretation across the protein order-to-disorder continuum

Background

Computational variant effect predictors (VEPs) are increasingly used as evidence in clinical genetics, but they must perform reliably for all patients. Clinical reference data do not represent all ancestry groups equally, and we recently showed that this leads many clinically trained VEPs to perform unevenly across populations (Pathak et al., 2026). Population-free methods, based on evolutionary and protein language model signals, offer a promising alternative: they are considerably more consistent across ancestry groups while still performing well at identifying pathogenic variants.

However, these methods also have an important weakness. Intrinsically disordered regions (IDRs) are flexible protein regions that lack a stable three-dimensional structure; they make up around a third of the human proteome and mediate much of cellular signalling and regulation. Pathogenic IDR variants are enriched in dominant disease genes, including genes associated with cancer predisposition and neurodevelopmental disorders (e.g. BRCA1 and MECP2), yet we recently found that population-free VEPs agree poorly with one another in these regions and have reduced sensitivity for pathogenic variants (Fawzy and Marsh, 2025). Many are therefore left unresolved as variants of uncertain significance (VUS), delaying or preventing diagnosis.

A key reason is that these methods rely heavily on evolutionary conservation, which is weaker in IDRs. Improving interpretation therefore requires modelling the properties that determine how disordered regions function, such as their flexibility, short binding motifs and transient interactions with other proteins. Most analyses, including our own, have also treated order and disorder as discrete classes, but proteins span a continuum from stable domains to complete disorder, with corresponding differences in constraint and pathogenic mechanism. This project will determine how predictor reliability changes along this continuum and develop methods that capture these properties, making the most ancestry-robust predictors reliable across the whole proteome.

Aims

1. Where do existing predictions become unreliable as proteins become more flexible? The student will place residues on the order-disorder continuum using AlphaFold confidence, disorder predictors and conformational ensemble properties. They will then test whether the same predictor score provides equivalent evidence across it, comparing calibration strategies with and without structural context, and across ancestry groups.
2. Can sequence and interaction models better identify harmful changes in flexible regions? Building on the Glasgow supervisors' protein language models of protein interactions (Liu et al., 2025), the student will combine language model representations with local constraint, regulatory motifs and other features of IDR biology, prioritising approaches that do not rely on clinically labelled training variants. Since many IDRs mediate transient interactions or fold upon binding, AlphaFold 3 co-folding will test whether modelling interaction partners reveals pathogenic mechanisms missed by conventional predictors. Validation will use deep mutational scanning (DMS) data generated within the supervisory team (Çubuk et al., 2026) and from rapidly growing public resources.
3. Can the resulting predictions provide reliable evidence for clinical variant interpretation across diverse populations? Improved predictions will be translated into calibrated evidence within the ACMG/AMP framework that diagnostic laboratories use to classify variants. Using independent clinically classified variants, we will test whether this improves recognition of pathogenic variants while limiting false positives, and whether gains hold across ancestry groups. Validated methods will be incorporated into VUSterpret, a clinical variant interpretation platform we are developing with clinical genomics collaborators.

Supervisory team and training

Edinburgh brings clinical variant interpretation, calibration and experimental measurement of variant effects (Marsh, Kudla); Glasgow brings machine learning (Yuan) and protein evolution and interactions (Robertson). The student will gain training across all of these areas, and insight into how computational evidence is used in practice through NHS Lothian clinical genomics colleagues, with whom Marsh works closely on clinical variant interpretation.

References

• Çubuk H, Plech M, Aslanzadeh V, Zikánová M, Škopová V, Kmoch S, Shen Y, Marsh JA, Kudla G. Mechanistic modeling of recessive disease through allelic integration of variant effects. Cell Systems 2026;17(6):101595. doi:10.1016/j.cels.2026.101595.
• Fawzy M, Marsh JA. Assessing variant effect predictors and disease mechanisms in intrinsically disordered proteins. PLoS Computational Biology 2025;21(8):e1013400. doi:10.1371/journal.pcbi.1013400.
• Liu D, Young F, Lamb KD, Quiros AC, Pancheva A, Miller CJ, Macdonald C, Robertson DL, Yuan K. PLM-interact: extending protein language models to predict protein-protein interactions. Nature Communications 2025;16:9012. doi:10.1038/s41467-025-64512-w.
• Pathak AK, Bora N, Badonyi M, Livesey BJ, SG10K_Health Consortium, Ngeow J, Marsh JA. Ancestry bias in variant effect predictors. Nature Genetics. 2026. Accepted in principle. preprint doi:10.1101/2024.05.20.594987