Probing impact of allosteric (or distal) mutations on protein function through time-dependent linear response theory

Advanced sequencing technologies provide ever-increasing data about human genetic variation and viral evolution. However, predicting the outcomes of missense mutations in protein coding regions remains a challenge, creating a bottleneck in discriminating biomedically-relevant variants from neutral ones (with little or no effect on phenotype). Outcome predictions are particularly poor when a missense mutation alters amino acids that are located far from a protein’s functional/binding sites. These shortcomings also impair protein design. Here we implement a novel physics-based computational metric to predict the outcome of the distal mutations on protein function. In particular, we investigate TEM-1 beta-lactamase, the most widespread enzyme to confer beta-lactam antibiotic resistance to gram-negative bacteria. Through time-resolved responses to perturbations performed at critical active sites, we find that allosteric versus non-allosteric positions can be distinguished from one another based upon their Fourier-transformed perturbation response profiles. Further, we show that through Principle Component Analysis (PCA) it is possible to further separate allosteric from non-allosteric positions and that this analysis can be extended to generally capture per-position substitution behavior with respect to Ampicillin fitness. Finally, we implement a machine learning approach via logistic regression which shows that these time-dependent response metrics can act as features that are able to determine, with high accuracy, positions which result in increased fitness over the wild-type in the presence of antibiotics Cefotaxime or Ampicillin.