Development Of Theoretical Methods For Studying Biological Macromolecules

National Heart, Lung, And Blood Institute

Investigators

Linked publications, trials & patents

Abstract

GraphVAMPNet, using graph neural networks and variational approach to Markov processes for dynamical modeling of biomolecules Finding a low dimensional representation of data from long-timescale trajectories of biomolecular processes, such as protein folding or ligandreceptor binding, is of fundamental importance, and kinetic models, such as Markov modeling, have proven useful in describing the kinetics of these systems. We combine VAMPNet and graph neural networks to generate an end-to-end framework to efficiently learn high-level dynamics and metastable states from the long-timescale molecular dynamics trajectories. This method bears the advantages of graph representation learning and uses graph message passing operations to generate an embedding for each datapoint, which is used in the VAMPNet to generate a coarse-grained dynamical model. This type of molecular representation results in a higher resolution and a more interpretable Markov model than the standard VAMPNet, enabling a more detailed kinetic study of the biomolecular processes. Deep attention based variational autoencoder for antimicrobial peptide discovery Antimicrobial peptides (AMPs) have been proposed as a potential solution against multiresistant pathogens. Designing novel AMPs requires exploration of a vast chemical space which makes it a challenging problem. Recently natural language processing and generative deep learning have shown great promise in exploring the vast chemical space and generating new chemicals with desired properties. We leverage a variational attention mechanism in the generative variational autoencoder where attention vector is also modeled as a latent vector. Variational attention helps with the diversity and quality of the generated AMPs. The generated AMPs from this model are novel, have high statistical fidelity and have similar physicochemical properties such as charge, hydrophobicity and hydrophobic moment to the real to the real antimicrobial peptides. pKa prediction by machine learning Machine learning techniques are developing rapidly in recent years and have been applied to numerous scientific fields. We have presented four tree-based machine learning models for protein pKa prediction. The four models, Random Forest, Extra Trees, eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), were trained on three experimental PDB and pKa datasets, two of which included a notable portion of internal residues. We observed similar performance among the four machine learning algorithms. The best model trained on the largest dataset performs 37% better than the widely used empirical pKa prediction tool PROPKA and 15% better than the published result from the pKa prediction method DelPhiPKa. Equivariant graph neural based electrostatic embedding in QM/MM simulations MM empirical force field calculations on GPUs are much faster than QM calculations on an even small region. One solution to this problem is to replace the expensive QM calculations with neural networks. However, this is challenging because the model is not transferable and has to be trained on the bulk region data for every simulation. The main contribution of this work is to develop a novel equivariant neural network where the intra-QM interactions are modeled with a GNN and the MM to QM region is modeled with a sparse network. The E(3)-equivariant convolutions over higher-order tensors ensure the features at each layer remain equivariant. Hence, this gives up to 2 orders of magnitudes higher sampling efficiency while training, thus making a NN based QM/MM simulation feasible. Automatic differentiation for Particle Mesh Ewald There has been recent interest in using automatic differentiation for simulations using end-to-end differentiable implementation of molecular dynamics in packages like JaxMD, TorchMD, DIffTaichi. These methods utilize the progress made in automatic differentiation for neural networks for empirical and machine learned force fields. However, these methods lack a long range electrostatics and rely only on cutoff schemes. In this work, we are adding long range ewald terms to the package JaxMD. We plan to extend the approach to higher order multipoles in the future as well since it will provide the complicated hessian terms automatically. Evaluation of Binding Free Energies The ability to accurately predict binding free energies is a cornerstone of rational drug design. The SAMPL challenges propose each year several sets of host and guests pairing for which no experimental data is available yet. It is designed as a test for the molecular simulations community to assess the capability and robustness of free energies estimation methods. Our submission to the drugs-of-abuse section of the SAMPL8 challenge yielded excellent performances by bridging the gap between Quantum Mechanical (QM) and Molecular Mechanical (MM) methods, which we further analyzed to offer some deeper insights on the methods, demonstrating e.g. the adequacy of semi-empirical methods for this type of work. Hessians for Permanent Electrostatics and Polarizability models Based on an implementation of the first derivative of the electrostatic interaction terms for multipole terms, we pursued our efforts to propose an implementation of the second degree derivative, namely the Hessian terms. This second order term is the core of tools such as Normal Mode Analysis and are key to predicting structural properties, amongst other. Hessian terms for polarizable force fields (classical force fields including a representation of the electronic mobility) have also been implemented. Enhancements and Extensions of Grid Inhomogeneous Solvation Theory Calculations Grid Inhomogeneous Solvation Theory (GIST) provides a statistical mechanical formalism for determining the thermodynamics of water in a region of interest by mapping various solvation properties onto a grid. This helps to identify thermodynamic hot spots for solvation, indicating where solvent binding may be favorable or unfavorable, thus helping to e.g. guide rational drug design. Enthalpy-related properties requires the calculation of water-water and water-solute energy on this grid, which is usually the most time-consuming part of the calculation. We have greatly increased the speed of this calculation by parallelizing it with MPI. In addition, we have both improved the speed of the entropy calculation and its convergence via the introduction of a correction term for the "reference volume" in the nearest neighbor entropy and a new version of the nearest neighbor search. The new search allows a user to choose how many layers of neighboring voxels should be used, with higher values improving convergence of the entropy of bulk solvent. Finally, we have extended GIST to be able to handle solvents other than water and including ions. Evaluating the Effect of Positional Restraints on GIST Calculations In order to ensure that global rotations and translations of the solute of interest are removed, previous GIST calculations have required either rotating the entire system to fit the grid or running simulations with positional restraints on the solute to keep it oriented in the grid. The former makes it difficult to do energy calculations with a long cutoff (since the coordinates are rotated out of the system reference unit cell), and the latter can potentially introduce some bias into the results from the use of restraints. We have developed a third option where the grid itself is rotated to match the solute of interest. This allows GIST to be used on simulations with no positional restraints, and facilitates the use of GIST with the particle mesh Ewald method for calculating electrostatics, which improves its convergence. Work is currently underway to evaluate what affect past methods.