The application of statistical methods to comparatively framed questions about protein dynamics can potentially enable investigations of biomolecular function beyond the current sequence and structural methods in bioinformatics. However, chaotic behavior in single protein trajectories requires statistical inference be obtained from large ensembles of molecular dynamic (MD) simulations representing the comparative functional states of a given protein. Meaningful interpretation of such a complex form of big data poses serious challenges to users of MD. Here, we announce DROIDS v3.0, a molecular dynamic (MD) method + software package for comparative protein dynamics, incorporating many new features including maxDemon v1.0, a multi-method machine learning application that trains on large ensemble comparisons of concerted protein motions in opposing functional states and deploys learned classifications of these states onto newly generated protein dynamic simulations. Local canonical correlations in learning patterns generated from self-similar MD runs are used to identify regions of functionally conserved protein dynamics. Subsequent impacts of genetic and drug class variants on conserved dynamics can also be analyzed by deploying the classifiers on variant MD runs and quantifying how often these altered protein systems display the opposing functional states. Here, we present several case studies of complex changes in functional protein dynamics caused by temperature, genetic mutation, and binding interaction with nucleic acids and small molecules. We studied the impact of genetic variation on functionally conserved protein dynamics in ubiquitin and TATA binding protein and demonstrate that our learning algorithm can properly identify regions of conserved dynamics. We also report impacts to dynamics that correspond well with predicted disruptive effects of a variety of genetic mutations. In addition, we studied the impact of drug class variation on the ATP binding region of Hsp90, similarly identifying conserved dynamics and impacts that rank accordingly with how closely various Hsp90 inhibitors mimic natural ATP binding.Statement of significance We propose a statistical method as well as offer a user-friendly graphical interfaced software pipeline for comparing simulations of the complex motions (i.e. dynamics) of proteins in different functional states. We also provide both method and software to apply artificial intelligence (i.e. machine learning methods) that enable the computer to recognize complex functional differences in protein dynamics on new simulations and report them to the user. This method can identify dynamics important for protein function, as well as to quantify how the motions of molecular variants differ from these important functional dynamic states. For the first time, this method of analysis allows the impacts of different genetic backgrounds or drug classes to be examined within the context of functional motions of the specific protein system under investigation.

Machine learning based detection of genetic and drug class variant impact on functionally conserved protein binding dynamics