University of Edinburgh
Harnessing the human mutational spectrum to advance missense variant effect prediction
Abstract
dc:description.abstractGenetic mutations provide the raw material for evolution. However, they do not arise evenly across the genome. Human germline mutation rates are shaped by local DNA sequence, as well as by the processes of replication and repair. This creates a characteristic mutational spectrum in which some nucleotide changes occur far more frequently than others. Because natural selection can act only on the variants that arise, observed patterns of human variation reflect both mutational opportunity and the removal of deleterious variants through purifying selection. Separating these two influences is therefore essential for understanding genetic constraint. The first chapter frames missense variation as a central challenge at this interface between mutation and selection. It explains why the consequences of amino acid substitutions are strongly dependent on biological context and how they can disrupt protein function through distinct molecular mechanisms. It reviews the biological signals used to predict these effects, particularly evolutionary conservation, protein structure, and human genetic constraint, and shows how these signals underpin modern variant effect predictors, while examining the strengths, limitations, and biases of these tools. The chapter then turns to the processes that shape the human germline mutation spectrum and the models used to estimate variation in mutation rates across the genome. These foundations motivate the central premise of the thesis, that more realistic models of mutational opportunity can improve estimates of missense constraint and, in turn, strengthen variant effect prediction. Building on this premise, the second chapter maps missense constraint across the human proteome by comparing variation observed in ~731k human exomes with expectations from a basepair resolution germline mutation model. Coding single nucleotide variants (SNVs) expected under neutrality are simulated and contrasted with the observed burden of rare variants, allowing constraint to be quantified as observed-to-expected depletion in the human population. This produces a fine-grained tolerability map of amino acid substitutions accessible through SNVs and shows that missense constraint varies systematically with nucleotide context and codon position, and becomes stronger in predictable structural environments within proteins. At the gene level, the resulting estimates provide clearer separation between selective regimes, suggesting that a more accurate expected burden reduces baseline noise and allows human genetic constraint to be measured more precisely. Broad classes of substitutions can reveal general patterns of pathogenicity, however, interpreting individual missense variants remains challenging because their effects depend strongly on biological context. Although variant effect predictors are widely used for this purpose, their scores are often difficult to interpret biologically and are not consistently calibrated across genes. As a result, the same numerical score may imply different levels of impact in different proteins. The third chapter addresses this by introducing a universal gene-level calibration framework that applies the same observed-to-expected principle to individual predictor scores. By comparing the predicted burden of damaging variation with expectations under a neutral mutation model, the framework converts raw scores into biologically interpretable observed-to-expected ratios that can be compared across the proteome. In developmental disorder cohorts, this calibration substantially improves the ability of all tested predictors to distinguish de novo variants in affected individuals from those in controls. Proteome-wide analyses further show that calibration strengthens the average pathogenicity evidence provided by these predictors, extending their reliable application to genes with limited previous clinical annotation. Most variant effect predictors are trained on datasets of clinically curated or population-derived variants. Because these resources disproportionately represent individuals of European descent, they may introduce systematic differences in predictor performance across populations. The fourth chapter investigates this possibility by comparing the scoring patterns of 55 variant effect predictors for missense variants from 14 ancestry groups. The analysis reveals striking disparities, where several predictors classify a markedly greater proportion of variants as damaging in underrepresented populations, such as individuals of Malay ancestry, than in those of European ancestry. In contrast, predictors that do not rely on clinical or population-based training data produce more consistent estimates of pathogenicity burden across ancestry groups. These differences appear to be closely associated with biases in the datasets used to train the models, highlighting the urgent need to adopt tools that minimise ancestry bias. This thesis demonstrates that the human mutational spectrum can be used to construct more accurate neutral expectations for coding variation, sharpen estimates of missense constraint, and place variant effect predictor scores on a consistent and biologically meaningful scale. It also shows that widely used prediction tools can carry substantial ancestry biases. These findings provide practical approaches for improving the accuracy, interpretability, and fairness of missense variant interpretation in human genetics and clinical diagnosis.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Pathak, Ankit Kumar
- Advisors dc:contributor.advisor
-
- Marsh, Joseph
- Semple, Colin
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier URI
- https://doi.org/10.7488/era/7794