ATOMDANCE: Kernel-based denoising and choreographic analysis for protein dynamic comparison.

Comparative methods in molecular evolution and structural biology rely heavily upon the site-wise analysis of DNA sequence and protein structure, both static forms of information. However, it is widely accepted that protein function results from nanoscale nonrandom machine-like motions induced by evolutionarily conserved molecular interactions. Comparisons of molecular dynamics (MD) simulations conducted between homologous sites representative of different functional or mutational states can potentially identify local effects on binding interaction and protein evolution. In addition, comparisons of different (i.e., nonhomologous) sites within MD simulations could be employed to identify functional shifts in local time-coordinated dynamics indicative of logic gating within proteins. However, comparative MD analysis is challenged by the large fraction of protein motion caused by random thermal noise in the surrounding solvent. Therefore, properly denoised MD comparisons could reveal functional sites involving these machine-like dynamics with good accuracy. Here, we introduce ATOMDANCE, a user-interfaced suite of comparative machine learning-based denoising tools designed for identifying functional sites and the patterns of coordinated motion they can create within MD simulations. ATOMDANCE-maxDemon4.0 employs Gaussian kernel functions to compute site-wise maximum mean discrepancy between learned features of motion, thereby assessing denoised differences in the nonrandom motions between functional or evolutionary states (e.g., ligand bound versus unbound, wild-type versus mutant). ATOMDANCE-maxDemon4.0 also employs maximum mean discrepancy to analyze potential random amino acid replacements allowing for a site-wise test of neutral versus nonneutral evolution on the divergence of dynamic function in protein homologs. Finally, ATOMDANCE-Choreograph2.0 employs mixed-model analysis of variance and graph network to detect regions where time-synchronized shifts in dynamics occur. Here, we demonstrate ATOMDANCE's utility for identifying key sites involved in dynamic responses during functional binding interactions involving DNA, small-molecule drugs, and virus-host recognition, as well as understanding shifts in global and local site coordination occurring during allosteric activation of a pathogenic protease.

Statistical machine learning for comparative protein dynamics with the DROIDS/maxDemon software pipeline.

Comparative analysis of protein structure or sequence alignments often ignores the protein dynamics and function. We offer a graphical user interface to a computing pipeline, complete with molecular visualization, enabling the biophysical simulation and statistical comparison of two-state functional protein dynamics (i.e., single unbound state vs. complex with a ligand, DNA, or protein). We utilize multi-agent machine learning classifiers to identify functionally conserved dynamic motions and compare them in genetic or drug-class variants. For complete details on the use and execution of this profile, please refer to Babbitt et al. (2020b, 2020a, 2018) and Rynkiewicz et al. (2021).

Function and evolution of B-Raf loop dynamics relevant to cancer recurrence under drug inhibition.

Oncogenic mutations in the kinase domain of the B-Raf protein have long been associated with cancers involving the MAPK pathway. One constitutive MAPK activating mutation in B-Raf, the V600E (valine to glutamate) replacement occurring adjacent to a site of threonine phosphorylation (T599) occurs in many types of cancer, and in a large percentage of certain cancers, such as melanoma. Because ATP binding activity and the V600E mutation are both known to alter the physical behavior of the activation loop in the B-Raf ATP binding domain, this system is especially amenable to comparative analyses of molecular dynamics simulations modeling various genetic and drug class variants. Here, we employ machine learning enabled identification of functionally conserved protein dynamics to compare how the binding interactions of four B-Raf inhibitors impact the functional loop dynamics controlling ATP activation. We demonstrate that drug development targeting B-Raf has progressively moved towards ATP competitive inhibitors that demonstrate less tendency to mimic the functionally conserved dynamic changes associated with ATP activation and leading to the side effect of hyperactivation (i.e. inducing MAPK activation in non-tumorous cells in the absence of secondary mutation). We compare the functional dynamic impacts of V600E and other sensitizing and drug resistance causing mutations in the regulatory loops of B-Raf, confirming sites of low mutational tolerance in these regions. Lastly, we investigate V600E sensitivity of B-Raf loop dynamics in an evolutionary context, demonstrating that while sensitivity has an ancient origin with primitive eukaryotes, it was also secondarily increased during early jawed vertebrate evolution.Communicated by Ramaswamy H. Sarma.

DROIDS 3.0-Detecting Genetic and Drug Class Variant Impact on Conserved Protein Binding Dynamics.

The application of statistical methods to comparatively framed questions about the molecular dynamics (MD) of proteins can potentially enable investigations of biomolecular function beyond the current sequence and structural methods in bioinformatics. However, the chaotic behavior in single MD trajectories requires statistical inference that is derived from large ensembles of simulations representing the comparative functional states of a protein under investigation. Meaningful interpretation of such complex forms of big data poses serious challenges to users of MD. Here, we announce Detecting Relative Outlier Impacts from Molecular Dynamic Simulation (DROIDS) 3.0, a method and software package for comparative protein dynamics that includes maxDemon 1.0, a multimethod machine learning application that trains on large ensemble comparisons of concerted protein motions in opposing functional states generated by DROIDS and deploys learned classifications of these states onto newly generated MD simulations. Local canonical correlations in learning patterns generated from independent, yet identically prepared, MD validation runs are used to identify regions of functionally conserved protein dynamics. The subsequent impacts of genetic and/or drug class variants on conserved dynamics can also be analyzed by deploying the classifiers on variant MD simulations and quantifying how often these altered protein systems display opposing functional states. Here, we present several case studies of complex changes in functional protein dynamics caused by temperature, genetic mutation, and binding interactions with nucleic acids and small molecules. We demonstrate that our machine learning algorithm can properly identify regions of functionally conserved dynamics in ubiquitin and TATA-binding protein (TBP). We quantify the impact of genetic variation in TBP and drug class variation targeting the ATP-binding region of Hsp90 on conserved dynamics. We identify regions of conserved dynamics in Hsp90 that connect the ATP binding pocket to other functional regions. We also demonstrate that dynamic impacts of various Hsp90 inhibitors rank accordingly with how closely they mimic natural ATP binding.