Estimating the Functional Impact of Single Amino Acid Changes Using Machine Learning

One of the great challenges in genetics is to accurately separate functional from neutral variation in human genome sequence. Towards more accurate pathogenicity prediction of human genetic variants, it has become possible to experimentally map the fitness landscape of missense mutations for specific targeted human disease-associated proteins. Coupling exhaustive mutagenesis with a deep sequencing read-out of multiplexed yeast functional complementation assays can yield a “variant effect map” covering most of possible amino acid changes in these proteins. The first challenge I wished to address is that for each variant effect map, a substantial minority of missense mutations remain unmeasured. The second challenge is that variant effect maps are available for less than 1% of disease-associated proteins, with a high-quality fully comprehensive experimental atlas of functional missense variation being likely decades in the future. Towards the first challenge, I imputed the functional effect of unmeasured variants using machine learning model trained on variants with experimental measurement, each of which is represented by a list of features including conservation scores, chemical physical proprieties of the residues, and engineered feature derived from the functional effect of similar amino acid substitutions in the same map. Towards the second challenge, I built a computational model VARITY to predict the pathogenicity of all possible missense variants for ~18,000 human proteins. I exploited a larger reservoir of training examples with uncertain accuracy and representativity, each assigned with a differential weight determined via hyperparameter tunning that optimizes predictive performance on a small high-quality set of annotated variants. Performance of the final model (VARITY) exceeds that of all analyzed previously described computational methods.Ph.D