Acellera's Blog

Adaptive Molecular Dynamics on the Cloud with HTMD and AceCloud

From João M. Damas

Researchers have been looking beyond traditional Molecular Dynamics (MD) for a long time now. While many have been using many forms of biased MD (for example, metadynamics), we have alternatively proposed Adaptive Molecular Dynamics 1,2 as a way to unbiasedly enhance the exploration of the phase space through on-the-fly learning of the explored space.

Since the beginning, HTMD 2 has provided an Adaptive MD protocol, which automatically manages all the process of analyzing the explored space and spawning of new MD simulations. An explanation of how Adaptive MD works and examples on how to use it can be found in the cited papers and in the online documentation of HTMD. One of the limitations of Adaptive MD has been the access to computing resources to take advantage of its high-throughput nature.

However, with the advent of cloud computing, resources are more readily available in an elastic manner that suits different needs. We have developed AceCloud 3 to run MD simulations on AWS and, together with HTMD, it can be used to run Adaptive MD simulations on the cloud. By using a system example for the generators (initial simulations), one can use the following script to quickly run an Adaptive MD job using HTMD and AceCloud:


References:
1. S. Doerr and G. De Fabritiis. On-the-fly learning and sampling of ligand binding by high-throughput molecular simulations. Journal of Chemical Theory and Computation, 10(5):2064–2069, 2014.
2. S. Doerr, M. J. Harvey, Frank Noé, and G. De Fabritiis. HTMD: High-throughput molecular dynamics for molecular discovery. Journal of Chemical Theory and Computation, 12(4):1845–1852, 2016.
3. M. J. Harvey and G. De Fabritiis. AceCloud: Molecular dynamics simulations in the cloud. Journal of Chemical Information and Modeling, 55(5):909–914,2015.

alejandroAdaptive Molecular Dynamics on the Cloud with HTMD and AceCloud
read more

Acellera Flows

“A command-line interface designed to perform complex workflows with few inputs”

By Alberto Cuzzolin

Nowadays, Molecular Dynamics (MD) simulation is known to be an important tool in drug discovery process. Although its potential is clearly proved, it is still not trivial to handle all the steps necessary to retrieve results. To handle all these aspects, we developed HTMD, a python framework that manages all these aspects and allows an easily scale up to an high-throughput manner.

Despite HTMD provides all the necessary functionalities, we decided to develop automated workflows, called “protocols or flows” that will ease and standardize the use of HTMD features. A command-line interface was designed to perform complex workflows with few inputs, allowing the user to customize its own protocol.

Several protocols are available with their specific options that the user can play with.
While the interface is simple and fixed (few options are available), Acellera Flows provide enhanced flexibility through template generation, meaning that the entire workflow is saved in pure python script.

At the moment, 4 protocols are available:

  • Build Flow: A protocol to build systems with amber and charmm.
  • SimpleRun Flow: A protocol to equilibrate and run MD simulation
  • RunAdaptLig Flow: A protocol to run adaptive sampling for ligand-protein recognition
  • CheckAdaptLig FLow: A protocol to analyze the ligand-protein adaptive simulation

Coming soon:

  • RunAdaptProt Flow: A protocol to run adaptive sampling for protein conformation
  • MembraneBuilder Flow: A protocol to prepare phospholipid bilayer
  •  CheckAdaptProtFlow: A protocol to analyze the protein adaptive simulations

For more details, please visit our software website.

alejandroAcellera Flows
read more

Acellera’s new website

As you might have noticed, Acellera’s website has experienced a few changes: Our consultancy services have been included inside the products section so it is easier to find out what fits your needs best; we have also simplified the Science section, classifying our articles in categories in order to make it easier for you to find them. At the bottom of every page, you can see the latest posts in the blog, our recent activity on Twitter -we invite you to follow us 😉 – and the subscription form to our products’ newsletters.

All our pages follow a similar architecture so that it is easier to navigate trough our contents and find what you are looking for. We have also included some videos to better illustrate some possible uses of our products and/or software. Finally, the new Partnership page includes information on how Acellera manages relationships with companies and laboratories which want to work with us.

We hope you like the new website and we will be glad to hear your feedback in the comments section.

alejandroAcellera’s new website
read more

Automated Preparation and Simulation of Membrane Proteins with HTMD

By Stefan Doerr

Molecular dynamics has matured to the point where users can simulate multiple protein systems over timescales as large as milliseconds (see reference of HTMD and ACEMD at https://www.acellera.com/science/).
However, the preparation of the protein systems remains a complex step.
Tools already exist which allow the preparation of an MD system using visual GUIs or webservers.
However few, like HTMD, are built allowing scriptable system preparation for multiple hundreds of such systems.

Adapted with permission from J. Chem. Theory Comput., 2017, 13 (9), pp 4003–4011. Copyright 2017 American Chemical Society.

 

The purpose of HTMD is to provide all tools necessary for an integrated Molecular Dynamics simulation based discovery pipeline.
Therefore, in our most recent published work (S. Doerr, T. Giorgino, G. Martinez-Rosell, J.M. Damas, G. de Fabritiis High-Throughput Automated Preparation and Simulation of Membrane Proteins with HTMD in J. Chem. Theory Comput. 2017,) we present the system building and preparation tools of HTMD and it’s application on a complex use case which is the preparation of protein-membrane systems.
We apply a single building and equilibration protocol on all eukaryotic membrane proteins of the OPM database (Orientations of Proteins in Membranes)
and perform a short equilibration runs to test for the stability of the systems.
All data from the building and equilibration runs of the OPM systems is provided to the users through a webservice which also allows direct view of the trajectories, various plots and download options for further inspection.

We believe that this and upcoming advances of HTMD will help simplifying the process of performing MD-based experiments and will further broaden the user-base and popularity of Molecular Dynamics simulations.

alejandroAutomated Preparation and Simulation of Membrane Proteins with HTMD
read more

Preparing a Molecular System for MD with PlayMolecule

By Gerard Martínez-Rosell

PlayMolecule Introduction

One step of the simulation workflow has typically remained relatively under addressed, namely, the preparation steps before building a molecular system. These preparation steps aim to make a protein structure, usually extracted from the PDB database, ready for system building. In particular, there’s two main points that need to be addressed: (a) titration of the residues and (b) optimization of hydrogen bond (H-bond) network. Resolved protein structures usually don’t contain hydrogen atoms and therefore need to be added by the researcher before running a molecular dynamics (MD) simulation. While most hydrogens can be easily guessed, some protein residues prove to be more challenging as they co-exist in different protonation states. While in a constant-pH simulation the residues would be free to switch among the different protonation states, in a classic MD simulation protonation states are fixed and therefore must be decided beforehand. These protonation states greatly depend on the local environment of the residue and the simulation pH.

PlayMolecule ProteinPrepare

For this reason, we devised ProteinPrepare, a protein preparation web application that allows the estimation and visualization of charge states, optimization of the hydrogen bonding network of protonated structures, thorough visual inspection of the results, and rapid iteration of changes. The application provides an evaluation of the titration states of a target protein’s residues on the basis of their local environment and the optimization of its hydrogen-bonding network through the placement of missing hydrogen atoms and flipping of side chains. Special consideration is given to residues whose pKa is close to the solvent pH because they are more prone to be misclassified by the customary binary (protonated vs unprotonated) assignments; the user can interact with the results, force their chosen titration states, and have the application reoptimize the structure. The computation is executed on the server taking advantage of High-Throughput Molecular Dynamics (HTMD) a Python framework for simple molecular-simulation-based discovery. In particular, we used the proteinPrepare() functionality of HTMD, currently based on PROPKA 3.1 and PDB2PQR 2.1. As such, in contrast to most other graphical tools, this web application also provides the short HTMD Python code required to perform the same task offline for the specific structure. The web application is publicly available at www. playmolecule.org for use on the web as part of the PlayMolecule web platform.

Conclusion

The protein preparation tool presented offers three main features: (1) help in deciding the protonation states and charges of residues while still retaining the capability to change them, (2) optimization of the hydrogen-bonding network as it might have been approximative from the crystal structure, and (3) both an interactive and scriptable way to perform these tasks. Deciding on protonation states in classical molecular dynamics is critical because a change in charge can have drastic effects and invalidate all of the results. The hydrogen-bonding network optimization is important to help the protein retain the original state of the crystal structure, especially when losing an active or inactive state can take milliseconds or more to recover. The third aspect is purely one of convenience: the use of the HTMD framework allows the reproduction of any result obtained through the ProteinPrepare web interface on local computing resources and the possibility to automate the preparation steps, e.g., repeating them for a large set of structures in a high-throughput context. Coding of HTMD based analysis on local resources takes place through the Python language and can thus take advantage of any of its numerous libraries and facilities for reproducible research.

References

Martínez-Rosell, G.; Giorgino, T.; De Fabritiis, G.; J. Chem. Inf. Model., 2017, 57 (7), pp 1511–1516
Doerr, S.; Harvey, M. J.; Noe, F.; De Fabritiis, G.; J. Chem. Theory Comput. 2016 , 12, 1845 −1852.
Søndergaard, C. R.; Olsson, M. H. M.; Rostkowski, M.; Jensen, J. H. J. Chem. Theory Comput. 2011 , 7, 2284 −2295.
Dolinsky, T. J.; Czodrowski, P.; Li, H.; Nielsen, J. E.; Jensen, J. H.; Klebe, G.; Baker, N. A. Nucleic Acids Res. 2007 , 35, W522 −W525.

alejandroPreparing a Molecular System for MD with PlayMolecule
read more

Complete protein–protein kinetics by molecular dynamics

From Stefan Doerr

Introduction
Proteins are essential in the regulation of most of the processes in a cell. They do so by interacting with other molecules, including other proteins.
Protein-protein interactions have been the subject of study for a long time using both experimental and computational methods, however they have eluded the atomic-level analysis and accuracy that unbiased molecular dynamics simulations can provide.

While protein-ligand interactions have been simulated already in many studies including our own (http://www.pnas.org/content/108/25/10184.full, http://pubs.acs.org/doi/full/10.1021/ci4006063), these interactions are relatively simple compared to two proteins interacting. Due to the size of proteins and their potential interaction sites, protein-protein interactions happen on timescales that used to be out of our computational reach until now.

In this study, Frank Noe’s team in collaboration with our group has been able to produce simulations that investigate both the binding and unbinding processes of two proteins called Barnase/Barstar. This was only possible through the use of GPU based ACEMD simulations (https://www.acellera.com/products/molecular-dynamics-software-gpu-acemd/) and adaptive sampling methods such as those implemented in the HTMD software (https://www.acellera.com/products/high-throughput-molecular-dynamics/) and the analysis of powerful Markov model methods.

Thus we were able to obtain an atomic level description of interaction states, equilibrium populations and kinetics of the binding process of the two proteins. Additionally, investigations on the effect of mutations on this process were reported. All results were validated with reference to experimental studies previously performed.

This study is of great interest as it demonstrates the possibility of protein-protein investigation using MD and opens the gates to the investigation of other very important protein-protein interaction processes.

Reference:
Plattner N. et al. Complete protein–protein association kinetics in atomic detail revealed by molecular dynamics simulations and Markov modelling. Nat. Chem. (2017).
doi:10.1038/nchem.2785

Franck ChevalierComplete protein–protein kinetics by molecular dynamics
read more

A Decade of GPUs based MD

By Gerard Martinez-Rosell
Introduction

ACEMD was introduced in the field of Molecular Dynamics (MD) back in 2008. Its release, together with other MD engines, revolutionized the simulation throughput by leveraging the intrinsic parallelism present in a specific computer processor termed GPU (Graphic Processor Unit). Using a cluster of GPUs or a distributed network like GPUGRID one is able to reach performances of HPC-class computers using commodity software, yielding a significant (x10) reduction in the Euro/simulated time cost. Furthermore, the evolution of GPUs performance, which is mainly driven by a market demanding more graphic processing power (e.g. in videogames), has doubled approximately every 3 years and with it, since 2008, ACEMD performance has increased 2.1x. To illustrate this fact, we have compiled the maximum simulated time per year published by prof. Gianni de Fabritiis using ACEMD (which yields the oldest historic data) and we extrapolate that the MD field will reach the second aggregated time timescale by 2022.

Challenges

It is really interesting to envision how things will have changed by then and how the increased performance will help to overcome current challenges. For instance, two of the present issues in MD are polarizability and fix protonation. These limitations are a model simplification necessary to reduce the computational cost that could be overcome in exchange of more computational power. One could easily picture that, in a context of computer power abundance, polarizable force fields and constant-pH simulations will become much more popular and the main choices to address the study of systems especially sensitive to those effects such as certain ligand-protein binding complexes or pH-sensitive proteins. Furthermore, access to increased computational power will probably motivate the community to tackle currently impossible approaches such as mesoscale simulations (e.g. whole protein cascade pathway, full virus, etc.).

Analysis and Perspective

We have analyzed a large number of MD studies carried out since the release of ACEMD engine in 2008. One important lesson we can learn from these studies is that, while an initial phase of proof of concept and technique validation is necessary to evolve and mature MD simulations as any other technique, the field is getting closer to the inflexion point where both accuracy (i.e. force field improvement) and precision (i.e. higher throughput) together with the introduction of tools that decrease the learning curve and entry-point to the technology (e.g. HTMD software package) will allow MD to become a mainstream technique routinely used in drug discovery. This inflexion point will have several practical implications, including a shift of the current paradigm to a new one in which MD is not only used in retrospective and validation studies but into prospective searches where MD leverages its full predictive potential decipher the mechanism of unknown biological processes and bring new drugs into the market

References:

M. Harvey, G. Giupponi and G. De Fabritiis, ACEMD: Accelerated molecular dynamics simulations in the microseconds timescale, J. Chem. Theory and Comput. 5, 1632 (2009)
G Martinez-Rosell, T Giorgino, MJ Harvey, G de Fabritiis, Drug Discovery and Molecular Dynamics: Methods, Applications and Perspective Beyond the Second Timescale in Curr Top Med Chem. 2017 Apr 14.
S. Doerr, MJ. Harvey, F. Noé and G. De Fabritiis. HTMD: High-Throughput Molecular Dynamics for Molecular Discovery. J. Chem. Theory Comput. (2016) 12 (4), pp 1845-1852

Franck ChevalierA Decade of GPUs based MD
read more

Helical unwinding and side-chain unlocking unravel the outward open conformation of the melibiose transporter

By Alex Perálvarez-Marín


Introduction:

In Wang et al. we have studied and described an intermediate step of the molecular mechanism of sugar transport thanks to the possibility to run several MD simulation replicates in the 100 ns range using a single GPU-based desktop running ACEMD. Our solvated membrane protein system embedded in the lipid bilayer consisted of ca. 100,000 atoms, and the combination between ACEMD and a single GTX-780 helped us to understand discrete events happening between two metastable intermediates. The combination of MD simulations and experimental data was key to achieve our goal to understand the conformational changes to allow the entrance of the substrates within the transporter. For the full story, refer to Wang et al Sci Rep 2016(https://www.ncbi.nlm.nih.gov/pubmed/27658476).

References
For the full story, refer to Wang et al Sci Rep 2016(https://www.ncbi.nlm.nih.gov/pubmed/27658476)

Franck ChevalierHelical unwinding and side-chain unlocking unravel the outward open conformation of the melibiose transporter
read more

Acellera, celebrating the 10th Anniversary

What has been achieved in Molecular Dynamics and Acellera contribution.

Acellera celebrates its 10th anniversary. A decade focusing on molecular dynamics (MD) and dedicated to improve this technique as a strong partner of biophysical approaches. Objectively, molecular dynamics is the only technique allowing to study conformational changes and interaction at atomistic level, to quantify binding energies, binding kinetics for any kind of molecules since a force field has been defined for these. It represents an adequate combination with NMR, Crystallography, ITC, Surface plasmon resonance, Thermal shift that are already widely used in Drug Discovery.

What makes MD so interesting now?

In 10 years, we have seen how calculations done on cluster of CPU can now be done now on GPU cluster or in the Cloud. The first GPU offered already a large improvement on yield (ns/day) obtained. Where ns was the goal, nowadays it is a simple unit. In 2009, we ran microsecond, last year, 2 milliseconds (cumulated time) were obtained when performing a fragment screening (Ferruz et al, J. Chem. Inf. Model. 2015). If this rhythm is followed within the next 6 years, a second will probably be attained and maybe required for high index journal publication. Basically, we can simulate long process from side chain flip to domain motion and binding of fast Kon molecules for example.

The speed of GPU was undeniable in the achievement of long MD but they also generate higher amount of data to analyze. Where we needed Hard disk to store Gigabytes of data, we can count in terabyte…generated weekly. Storage is maybe the next challenge.

So it is possible to simulate the binding of benzamidine to Trypsin from scratch (Buch et al, PNAS 2011), to perform fragment screening, to study conformational change of antibodies, to analyze multibody cofactor and substrate molecular recognition…This has been possible not only thanks to the GPU development but also thanks to a huge work that contributed to innovative approach to sample the conformational space and energy of the system (Doerr et al, J. Chem. Theory Comput. 2014) and to the correct analysis of the trajectories.

What can be expected now from MD in Drug Discovery?

As solvation/desolvation processes are critical parameters for interaction studies, all atoms simulations represent an adequate method. During the simulations, at each step, the forces on each atom are computed and the atomic position and velocity are updated according to Newton’s laws of motion. This means millions of calculations magnified by the number of atoms composing your system. A 25000 atoms system can be simulated at 500 ns/day. If we take 100 GPU, after 20 days we have a millisecond of calculation which may be enough for micro molar affinity ligand. When studying larger system, the time needed to sample the interaction increase and the throughput we can expect decreases.

A category of molecules enters this affinity range and offers advantage when calculating compound parameters: fragment. MD is well suited for fragment interaction and low to medium throughput screening. Brute throughput will depend on the resources assigned, but in a single analysis binding kinetics, energy and pathway can be determined. The accuracy of low energy pose description reaches crystallography resolution.

What can be improved?

Improvements are needed in the timescale MD can reach, in reproducing experimental binding and throughput, meaning the number of molecules we can test. At the end of the day, we want to explore the chemical space and we are still far from scratching its surface. So, Speed of calculations, sampling and analysis are the main parameters where we can still gain time and open the way to build and test molecules newly designed and not yet synthesized.
Blue Waters, Anton, MDGRAPE are synonyms of (impressive) fast calculations. A more affordable solution has been developed at Acellera: the workstation Metrocubo equipped with 4 GPU; it allows to simulate a 2 microsecond/day (based on DHFR benchmark).

Sampling approaches are also being optimized (metadynamics, replica exchange, umbrella sampling, accelerated MD). At Acellera, we use Adaptive Sampling with ACEMD. Finally, analysis protocols have been largely improved and is still an active field of development.

In the last 10 years, Acellera developed ACEMD, enhancing the execution of MD simulations. These can now be run on GPU (workstation like Metrocubo or cluster) or in the cloud. For this purpose, we developed AceCloud, a tool allowing to run calculations on AWS. Finally, on top of all, a new interface providing a unified environment for molecular discovery has been designed for both computational and medicinal chemistry groups: HTMD. This tool allows encoding best practices, hiding complexity, from start to end: PDB to free energy calculations, to solve biological problem.

Fragment screening, GPCR ligand binding, protein folding, nucleic acids complex interaction are only few of the topics studied successfully by high throughput molecular dynamics and we hope its use will be generalized within the next decade.

FC

 

Franck ChevalierAcellera, celebrating the 10th Anniversary
read more

Multibody cofactor and substrate molecular recognition

By Franck Chevalier


Introduction:

Molecular recognition is rarely a two-body protein-ligand problem, as it often involves the dynamic interplay of multiple molecules that together control the binding process. Myo-inositol monophosphatase (IMPase), a drug target for bipolar disorder, depends on 3 Mg2+ ions as cofactor for its catalytic activity. Although the crystallographic pose of the pre-catalytic complex is well characterized, the binding process by which substrate, cofactor and protein cooperate is essentially unknown. Here, we have characterized cofactor and substrate cooperative binding by means of large-scale molecular dynamics. Our study showed the first and second Mg2+ ions identify the binding pocket with fast kinetics whereas the third ion presents a much higher energy barrier. Substrate binding can occur in cooperation with cofactor, or alone to a binary or ternary cofactor-IMPase complex, although the last scenario occurs several orders of magnitude faster. Our atomic description of the three-body mechanism offers a particularly challenging example of pathway reconstruction, and may prove particularly useful in realistic contexts where water, ions, cofactors or other entities cooperate and modulate the binding process.

In all in-silico binding analyses, full kinetic and thermodynamic data were obtained by performing free-ligand binding, all-atom molecular dynamics simulations with the ACEMD molecular dynamics software and anlyzed with HTMD.

Result
We have fully characterized substrate and cofactor binding prior to the catalytic event.
We have also provided an atomic-level description of substrate and cofactor cooperation and binding.
Large-scale HTMD has been used and been able to recapitulate the binding events of Mg ions and natural substrate at IMPase, and we identified structures close to the X-ray solutions.
References
N Ferruz, G Tresadern, A Pineda-Lucena, G De Fabritiis, Multibody cofactor and substrate molecular recognition in the myo-inositol monophosphatase enzyme, Scientific Reports 2016

Franck ChevalierMultibody cofactor and substrate molecular recognition
read more