Acellera's Blog

Webinar on PlayMolecule (25/05/2021)

Webinar on PlayMolecule 25/05/2021 13:00 CEST

Discover PlayMolecule!

Join the CompBioMed webinar on May 25th focusing on our platform PlayMolecule. This e-Seminar is intended for scientists and researchers in the field of pharmaceutical sciences that are interested in learning new methods and technologies to improve high throughput virtual screening, computational hit finding, and molecular dynamics simulations applied to biological systems. In this e-Seminar, you will learn how to launch jobs on PlayMolecule , how to retrieve the results and create workflows connecting the different applications available on the platform.

Date: 25/05/2021
Time: 13:00 CET
You can read more about this seminar here
You can register here

alejandroWebinar on PlayMolecule (25/05/2021)
Read more

Release of ACEMD 3.4

We have recently released a new version of ACEMD, Acellera’s molecular dynamics simulation software.

This new version brings support to the latest NVIDIA GPUs, including the Ampere architecture, as well as performance improvements. The simulation speed has been benchmarked against several systems at typical production conditions on different GPU devices (including GTX 1080, GTX1080 Ti, RTX 2080 Ti and RTX 3090). For the DHFR benchmark, on RTX 3090, ACEMD achieves the speed of ~1.3 µs/day.

Check ACEMD documentation page for more details, including the new installation instructions.

ACEMD is distributed under two different licenses: ACEMD Basic, free of use, that allows to run simulations on one GPU, and ACEMD Pro for any number of GPUs.

ACEMD logo

alejandroRelease of ACEMD 3.4
Read more

TorchMD – A deep learning framework for molecular simulations

This week our paper on TorchMD has been published in the Journal of Chemical Theory and Computation. TorchMD is a framework for molecular simulations that enables users to do research faster in force-field development as well as integrate neural network potentials seamlessly into the dynamics with the simplicity and power of PyTorch. It is the result of the collaboration of Acellera with Universitat Pompeu Fabra, Freie Universität Berlin, Rice University, CNR-IBF- Italy, Università degli Studi di Milano and ICREA.

alejandroTorchMD – A deep learning framework for molecular simulations
Read more

A new tutorial on SkeleDock

If you have recently visited PlayMolecule, you might have noticed that there is a new app available: SkeleDock.

SkeleDock is a scaffold docking algorithm (also known as template-guided docking), which allows users to dock a series of compounds to a template with which they share some common scaffold or moiety.

We have written a short tutorial on how to use it here. We hope that you find it useful!

If you want to know more about the inner workings of this algorithm, you can read the publication here

alejandroA new tutorial on SkeleDock
Read more

Release of ACEMD 3.3

We are proud to announce the release of ACEMD 3.3. This new version adds key functionalities to Acellera’s MD software. One of them is the integration of PLUMED to bring state-of-the-art free energy calculation capabilities to ACEMD. PLUMED is an open-source library for free energy calculation, implementing several enhanced-sampling methods, such as metadynamics, and an extensive set of collective variables.

Other changes in this release include an update to the latest OpenMM version, an improvement of the PRMTOP file parser, among others. Check the ACEMD documentation page for more details, including our new PLUMED tutorial!

alejandroRelease of ACEMD 3.3
Read more

ACEMD and OpenMM unite to tackle new challenges in molecular simulations

First, a bit of history...

In 2008, ACEMD was the first molecular dynamics software using GPUs (graphical processing units) to accelerate bio-molecular simulations. ACEMD came from a GPU conversion of 2006 CellMD, the first MD code to run on commercially available consumer-grade accelerator processors. ACEMD allowed us to demonstrate unprecedented simulation speeds for a decade, which was unmatched by any CPU-only or GPU software. After more than a decade, we can say for certain that the decision to invest in GPU-based technology was a correct one. GPUs have become an integral part of scientific computing, new architectures are announced almost on an annual basis, and the performance just keeps growing.

Nevertheless, GPUs are not novel anymore, and enabling more science by just significantly improving the speed of MD engines is no longer a possibility. In the last years, our research and development team has focused on new challenges, with higher immediate impact such as machine learning, HTMD for handling high-throughput molecular simulations using adaptive sampling and Markov state models, and finally PlayMolecule, an infrastructure for molecular discovery.

Following the lead of ACEMD, many MD applications started to adapt to GPUs, and now it would be hard to find any popular MD software without GPU capabilities. Among them was OpenMM, an open-source project born in 2008 as a minimal but versatile C++ framework with a simple Python API enabling access to high performance GPU-accelerated MD capabilities, attracting developers and users from academia. Like ACEMD, OpenMM was designed and implemented with GPU computing in mind, and eventually exceeded our software in terms of functionality, and more recently, in performance.

In 2017, we made the decision to embrace OpenMM and started to develop the next generation ACEMD using several parts of OpenMM C++ backend, in the same way that we were already using CUDA libraries for FFT. In 2019, we released ACEMD v3, which merged the previous ACEMD codebase with OpenMM kernels. Among other things, this gave ACEMD multithreaded CPU support and potential access to more features. In the first months of 2020, we have already submitted a pull request to integrate ACEMD multi-time step integrators into OpenMM, a feature which currently gives a 10% improvement in speed to ACEMD. Other open source contributions will appear soon.

OpenMM and ACEMD remain different in several aspects. ACEMD is a simple, stand-alone MD executable, rather than a library. OpenMM provides a library of low-level molecular simulation capabilities (force field terms, integrators, thermostats, etc), while ACEMD relies on the Python framework HTMD and forcefield tools to build, manage and analyze simulations. ACEMD focuses on stability, reliability, performance, easy-to-use and professional support, in a similar way as RedHat provides a commercial package for Linux. ACEMD has also its own additional tests and integration with other tools developed by Accellera.

Our joint collaborative project...

Today (27/5/2020), thanks to a one-year seed grant from the Chan Zuckerberg Initiative (CZI) Acellera is joining the OpenMM development team, together with lead OpenMM developer Peter Eastman, Tom Markland from Stanford University (whose lab focuses on QM/MM and machine learning for quantum chemistry), John Chodera from the Sloan Kettering Institute (whose lab focuses on free energy calculations). This grant aims to support the continued development of OpenMM to better serve its broad biomolecular modeling community, and its extension to integrate machine learning to enable genomic-scale biomolecular modeling, simulation, and prediction. Our collaborative project aims to secure long-term sustainable federal funding for OpenMM from the National Institutes of Health in a proposal submitted earlier this year.

“Hundreds of thousands of scientists each day use open source software to carry out their research,” said CZI Head of Science Cori Bargmann. “Scientists deserve better tools, and we’re helping to meet that need by supporting open source projects that will advance biomedical science and foster greater access to critical software.”

This new series of year-long grants of the CZI’s Essential Open Source Software for Science program, aims to support open source software projects essential to biomedical research, enabling software maintenance, growth, development, and community engagement. View the full list of grantees. Open source software is crucial to modern scientific research, advancing biology and medicine while providing reproducibility and transparency. Yet even the most widely used research software often lacks dedicated funding.

Prof. Gianni De Fabritiis, head of the Computational Science Laboratory (Universitat Pompeu Fabra) and CEO/CSO at Acellera thinks that “this is the way forward. By joining forces we can have one of the largest development teams in molecular simulations and have the strength to tackle the most challenging research projects ahead to make ACEMD and OpenMM incredibly useful for the research community.”

alejandroACEMD and OpenMM unite to tackle new challenges in molecular simulations
Read more

Virtual Screening against SARS-CoV-2

Using SkeleDock and KDeep to rank compounds against SARS-CoV-2 main protease

Postera has launched an initiative that allows researchers across the globe to suggest compounds that they think might be active against the main protease of SARS-CoV-2 . Given that most of the submitted compounds are based on co-crystallized fragments, we decided to use our scaffold-docking algorithm (SkeleDock) to predict their binding mode, and KDeep, to get an estimate of the binding affinity. This same protocol was proven successful in a blind competition (D3R GC4), so we decided to try it against this target, hoping that it can help researchers prioritizing compounds into further validation steps.


We downloaded all non-covalent submitted compounds from this site , and we checked for large maximum common substructures (MCS) with the co-crystallized fragments. Those compounds which had a large MCS with a fragment were selected. For each selected compound, the fragment with largest MCS was used as template to guide its pose prediction, using SkeleDock. Finally, the predicted poses were passed to KDeep, which computed a binding affinity for each complex.


Some compounds might generate poses with high predicted affinity, however, they might not be synthetically feasible, hence, we visually inspect the best compounds (according to predicted ligand efficiency) to evaluate their medchem profile, which resulted in a list of 79 compounds. Predicted poses and affinities for each compound are available here.

Predicted pose for compound MAK-UNK-105-15, shown together with the fragment used to guide the docking (Mpro-x1093, glass texture).

If you want to contact us for more information on this project, feel free to email us at!


alejandroVirtual Screening against SARS-CoV-2
Read more

In silico binding assay against SARS-CoV-2 main protease

Melatonin as a potential compound against SARS-CoV-2

In silico binding assay suggests melatonin can bind to SARS-CoV-2 main protease

After the effort of Diamond team to crystallize several fragments on the main protease of SARS-CoV-2 (PDB ID: 6YB7), we decided to screen FDA approved drugs for similarity with these fragments. We found out that melatonin shares common moieties with fragment x0104, which could make it a good candidate to fight this disease.

In order to evaluate the capacity of melatonin to bind to this target we ran an in silico binding assay using ACEMD platform, with the AdaptiveGoal protocol.

As shown in the video, we were able to see the binding event and the behaviour of melatonin (shown in green) in the cavity of the main protease of SARS-CoV-2, which looks promising, as it partially overlaps with the crystallized fragment, shown in grey. Trajectories are available here.

We have submitted this compound for further validation here. A similar submission has been previously proposed by John Chodera here.



A library of FDA approved drugs was downloaded from Enamine Store. Fingeprints were computed with RdKit's function Chem.RDKFingerprint for all the molecules in the library, as well as for the fragments that had crystallized to the main protease of the virus, available here. Finally, we computed similarity for each possible pair and the we decided to focus on melatonin, given its similarity to fragment x0104.

Simulation preparation

The protein (Mpro-x0104) was protonated at pH 7.0 using ProteinPrepare, to be then modeled using the AMBER99SB force field. The ligand (melatonin) was modeled with the general amber force field (GAFF) using Parameterize. The ligand was placed in the bulk in a random orientation. The complex was then solvated with TIP3P water molecules and neutralized.

3 nanoseconds of equilibration were run at 310K using ACEMD. Constraints on the heavy atoms of the protein were applied during the first half of the equilibration, 1 kcal * mol−1 * Å−2 for the alpha carbons and 0.1 for all others. This equilibration step was followed by 10 ns of simulation under production conditions (no constraints). This production simulation was used as the generator of our AdaptiveGoal protocol.


AdaptiveGoal protocol runs several simulations in parallel, and when a given number of simulations finishes, new simulations are launched. The key is that this new round (or epoch) of simulations starts from one of the frames of the previoulsy finished simulations. The protocol assigns a score to each frame, and the new round of simulations starts from the frames with highest score. The score of each frame is computed by an arbitrary, user-defined function. In this case, we gave a score of 1.0 to every frame where the ligand was closer than 20.0 Å to the binding site (defined by the location of Histidine 41 in the original structure) and the value of that distance multiplied by -1 if it was further away. Hence, once the ligand enters the sphere of 20 Å centered in the binding site, subsequent rounds of simulations will focus on exploring that area. This helps saving GPU time.

Markov State Model analysis

In the picture below, the 5 macrostates identified by our analysis are shown. Fragment x0104 is shown in bold Licorice for reference. Different lag times and number of macrostates were tried when computing the markov model. In all combinations tried, the green macrostate overlapping with the fragment was always present, which is, again, a good sign that suggests that this compound could bind in that area of the protein. The other macrostates shown are also resilient to changes in the aforementioned parameters, particularly, the red one.

All the code is available in our github repository.

We will keep posting updates in this page on the status of this and other proposals. If you want to contact us for more information on this project, feel free to email us at!


  • 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). pdf
  • M. Harvey and G. De Fabritiis, An implementation of the smooth particle mesh Ewald method on GPU hardware, J. Chem. Theory and Comput. 5 (9), 2371-2377 (2009). Link
  • I. Buch, T. Giorgino and G. De Fabritiis, Complete reconstruction of an enzyme-inhibitor binding process by molecular dynamics simulations, PNAS, 108, 10184-10189, (2011). Link
  • 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.
alejandroIn silico binding assay against SARS-CoV-2 main protease
Read more