A Mechanistic Variant Effect Predictor for disentangling the biophysical mechanisms driving disease

Genome sequencing has grown exponentially, but determining the phenotypic consequences of mutations remains a key challenge in human genetics. Recently, unsupervised methods, which model the distribution of sequence variation across organisms, have emerged as promising tools for quantifying the clinical pathogenicity of variants, rivaling the accuracy of experimental approaches. However, these machine learning predictors fall short in discerning the effects induced by pathogenic mutations that impact on organismal fitness and drive disease. I propose to address this knowledge gap by combining the evolutionary-based variant tolerance predictors with protein folding models derived from energy landscape theory. In order to fold in biological timescales, an amino acid chain must minimize energy conflicts within it. In natural proteins, remaining conflicts can be measured with a transferable energy function, pointing out regions that did not evolve maintaining foldability, as active or binding sites. Incorporating this approach, I will focus on disentangling the biophysical ambiguity between the alteration of specific functions and local stability changes along proteins, firstly for single variants from which thermodynamic data is available for testing. Furthermore, I will integrate the folding and evolutionary models into a single algorithm to characterize and predict disease-driving mechanisms along the human proteome, a Mechanistic Variant Effect Predictor (mechVEP). Finally, I will apply this pipeline for investigating the epistatic and dominance effects of multiple mutations, paving the way for fine-tuning disease predictors for different ancestries. In summary, findings from this study will provide completely new insight into deciphering the underlying molecular mechanisms driving disease, opening new opportunities for diagnosis and therapy.