Making it rain: cloud-based molecular

simulations for everyone

Pablo R. Arantes,∗,†,⊥ Marcelo D. Polêto,∗,‡ Conrado Pedebos,∗,¶,§ and Rodrigo

Ligabue-Braun∗,k

†Department of Bioengineering, University of California, Riverside, CA 92521, United

States
‡Department of Biochemistry, Virginia Tech, Blacksburg, VA 24061, United States
¶School of Chemistry, University of Southampton, Highfield Campus, Southampton SO17
1BJ, United Kingdom
§Present Address: Department of Biochemistry, University of Oxford, Oxford OX1 3QU,
United Kingdom
kDepartment of Pharmacosciences, Federal University of Health Sciences of Proto ALegre
(UFCSPA), Porto Alegre 90050-170, RS, Brazil.
⊥Corresponding author

E-mail: [email protected]; [email protected]; [email protected];

[email protected]

Abstract

We present a user-friendly front-end for running molecular dynamics (MD) simula-

tions using OpenMM toolkit on the Google Colab framework. Our goals are: 1) to high-

light the usage of a cloud-computing scheme for educational purposes for a hands-on

approach when learning MD simulations and 2) to exemplify how low-income research

1
 groups can perform MD simulations in the microsecond timescale. We hope this work

facilitates teaching and learning of molecular simulation throughout the community.

Introduction

Reliable structural depictions of biomolecules are essential for understanding how the nat-
ural world works. 1 Experimental techniques for obtaining such structures are mainly repre-
sented by X-ray crystallography, nuclear magnetic resonance spectroscopy, and cryo-electron
microscopy, with the latter facing an explosive increase in application. 2,3 In addition, com-
putational methods for biomolecular modeling (particularly proteins) have reached an un-
precedented quality, with AlphaFold2 4 and RoseTTAFold 5 being the most recent examples.
All these advances, combined with the availability of the human structural proteome 6 usher
the demand for dynamical, time-dependent methods for biological structure analysis that
are essential for holistic inspections of protein function in the organism-level context. 7,8
Molecular dynamics (MD) simulations, a classic physics approximation for describing
atomic trajectories, is widespread among the methods for describing biomolecular motion. 9,10
These simulations, however, require progressively more computation power, and efforts are
increasing to allow massive runs of MD. 10 Thus, one alternative to in-house computing
facilities is the use of cloud computing for molecular modeling projects, especially due to its
scalability, reliability, and lower costs. 11 However, the use of cloud computing requires from
the user the ability to perform technical or tedious tasks, such as configuring remote computer
nodes, installing the required software under specific conditions, and running calculations (by
launching, monitoring, and terminating them). 12 Some propositions for facilitating the use
of cloud computing for MD simulations have been made, including variations on AMBER, 13
CHARMM-GUI, 14 QwikMD, 15 and NAMD. 16 All these implementations, however, leave
some intermediate task to be performed or need to be set up on a cloud structure by the
end user. An exception is described by Engelberger et al. 17 which reported a collection of

2
Jupyter notebooks for educational purposes that performs phylogenetic analysis, molecular
modeling, molecular docking, molecular dynamics, and coevolutionary analysis. Although
those protocols were not focused on MD calculations, they greatly inspired the present work.
Considering the ever-growing need for faster, affordable computer resources for novice
users, here we present a user-friendly front-end for running molecular dynamics simulations
using OpenMM toolkit 18 on the Google Colab framework. 19,20 For educational purposes,
the user is only required to provide minimal input files, such as an initial biomolecular
tridimensional structure (a PDB file), and the following MD simulation tasks are facilitated
from there on. Usual simulation analyses (e.g. RMSD, RMSF, PCA) are also implemented
in an automated fashion. Despite its click-and-go appeal, the protocols presented here can be
customized for other specific cases, while also allowing students to have a hands-on approach
when learning MD simulations.

Google Colab framework and MD simulations

Google Colaboratory, or Colab for short, is a product from Google Research organization. 19,20
According to their own website, 19 “Colab allows anybody to write and execute arbitrary
python code through the browser, and is especially well suited to machine learning, data
analysis and education.” In practice, Colab uses virtual machines to execute Jupyter note-
book codes on its server using GPUs and each notebook connection (a session) can run up
to 12 hours in Colab free version, while the Colab paid version (Colab Pro) has a time limit
of 24 hours per session. Currently, only the United States, Canada, Japan, Brazil, Germany,
France, India, United Kingdom, and Thailand have access to Colab Pro. Moreover, Colab
is equipped with Nvidia K80, T4, P4 and P100 GPUs and Colab Pro users have access to
their fastest GPUs.
In terms of MD simulation, trajectories files can be very large and proper storage space
is a fundamental feature for computational resources. While the temporary storage system

3
can be limited when connected to a Colab computing node, users are allowed to mount their
personal Google Drive account so all output files are directly written to a more permanent
storage system. Conversely, users can also use the Google Drive system to upload input files
that can be used in their notebooks, which gives much flexibility to Colab computational
structure. Google Drive free version allows users to store up to 15GB of data, while paid
versions can store more than 2TB of data.
Understanding the potential use of such a computing environment and the educational
importance of introducing MD simulation concepts to new users, we combined Google Colab
framework and the OpenMM MD toolkit 18 in order to create notebooks containing entire
workflows for MD simulations. Here, we report 3 notebooks (Figure 1): A) exemplifying
a complete pipeline for MD simulations, from structure preparation to MD simulation and
basic analyses; B) Running MD simulations using input files imported from a Google Drive
folder; and C) combining molecular modeling techniques and MD simulations.

General Notebook structures

All workflows presented here follow the Jupyter Notebook structure 21 and have 2 initial
configuration steps that should take less than 7 minutes. The first step is for software and
dependency installation, which includes the OpenMM engine, 18 MDanalysis, 22 PyTraj, 23
Numpy, 24 Ambertools 25 and others. The second step is the mounting of the user’s Google
Drive filesystem in order to store output files. The remaining tasks are the usual ones
observed in any MD simulation protocol, namely: topology generation, solvation and neu-
tralization of a simulation box, equilibration run, production run and trajectories analyses.
Users are allowed to modify the OpenMM code to adapt each of these steps as needed.
The main difference between the 3 notebooks presented here is the origin of the input file
(i.e. the solute molecule). Notebook A) was designed to fetch a protein structure directly
from the Protein Data Bank, to create its topology using AmberTools along with the actual
simulation box itself containing solvent and neutralizing ions. Notebook B) was designed to

4
Figure 1: Workflow of the 3 presented Jupyter notebooks. Blue boxes are configuration steps
required by the Colab environment, red boxes are actual MD related tasks and yellow boxes
represent tasks that are external to the Jupyter notebook. During every MD related task,
output files are directly written at the user’s Google Drive storage filesystem.

5
show how users can provide their own input files for the OpenMM engine via Google Drive
storage system. As demonstrated, inputs generated by CHARMM-GUI 14 server can be
easily uploaded to a Google Drive folder, allowing the user to use them to start a simulation.
Moreover, notebook C) was designed to demonstrate how molecular modeling techniques
and MD simulations can be combined to generate structurally relevant information. In this
case, we combined our MD notebook with the AlphaFold2 algorithm that was recently ported
to Google Colab. 4,26 Together, these protocols are meant to cover most of the basic needs of
MD approaches.

Usage examples

In order to show the feasibility of using such a cloud-based computational resource for
research purposes, we carried out a 1 microsecond long simulation of lysozyme using the
Amber FF19sb force field in explicit OPC water, totalizing 26,033 atoms in the simulation
box. For this, we used a Colab Pro version, which allowed us to use P100 GPUs to yield
an average simulation speed of ≈230 ns/day. Due to the limiting 12/24h job limit in Colab,
jobs are expected to be terminated when they reach that time. In order to circumvent that
limitation, we divided the entire trajectory in 50 strides of 20 ns each. At the end of each
stride, OpenMM generates a .xml state file containing all atom’s positions and velocities,
which are read at the beginning of the next stride. All these steps can be easily defined by
the user in all notebooks reported here and restart routines are automatically set up. Thus,
in order to complete our 1 microsecond trajectory, we simply ran notebook A) 5 times, which
took us 6 days of work. Structural analyses of our lysozyme simulation are shown in Figure
2 below.

6
Figure 2: Structural analyses of MD simulations using PyTraj embedded in the notebooks:
RMSD values over time (A), its distribution (B) and a 2D RMSD plot (C); radius of gyration
over time (D) and its distribution (E); Cross-correlation analysis (F); RMSF per residue (G);
Principal Component Analysis (H) and its eigenvector distribution (I).

7
Applicability

The approach shown here is specially suited for educational purposes. Professors can share
the notebook with students and each one can run an entire MD simulation, from system
preparation to production runs and analyses in one day. Such hands-on classes do not
require local compilation of MD software since all dependencies are installed directly at
the computing node via Anaconda module. The usage of OpenMM engine and its well
known compatibility with GPU processing is especially well suited for the Colab computing
infrastructure.
Another benefit is that students are not required to have much previous computational
background. In fact, using the notebooks presented here would require only a Google Drive
account for each student and the appropriate storage space on it. An important feature
we made sure was present is the structure/trajectory visualization that is embedded in the
notebooks. Students can also choose to see, study and edit the OpenMM code behind each
task.
Moreover, considering that we were able to use the Colab structure to produce 1 microsec-
ond long MD simulations at a relatively low cost, this approach could be used for under-
privileged investigators around the world to conduct their research. In order to demonstrate
this, we evaluated both Pro and Free versions of Google Colab in terms of MD simulation
speed of system of different sizes, namely PDB ID 7CI3 (18K atoms), 6ZCT (25K atoms),
7AEH (46K atoms), 6VYO (66K atoms) and 7BTF (195K atoms). For these simulations,
we solvated the systems using explicit OPC water model and used a timestep of 2 femtosec-
ond for production simulations. We used a Tesla T4 GPU for our Free version calculations
and a Tesla P100-PCIE GPU for our Pro Calculations. Performances for each system are
shown in Figure 3. Assuming no data loss due to resubmissions and only one active session
per user, one could produce nearly 1 microsecond long trajectories for systems with ≈200K
using the Pro version for ≈1 month. A cost-free alternative would require ≈50 days for the
same system. Despite the necessity of ressubmitting a job 1-2 times a day, we believe useful

8
research can be done using this approach.

Figure 3: Simulation speed of multiple systems using Colab Pro (magenta) and Free (green)
versions. Fully solvated systems simulated, from left to right: PDB ID 7BTF (195K atoms),
6VYO (66K atoms), 7AEH (46K atoms), 6ZCT (25K atoms) and 7CI3 (18K atoms)

Conclusions

Here we present an affordable (cost-free) strategy for performing molecular dynamics sim-
ulations using cloud computing resources. It involves shareable, ready-to-use, customizable
Jupyter notebooks that guide the end-user in running their calculations using the Google
Colab services. We hope this work facilitates teaching and learning of molecular simula-
tions, as well as allowing for low-income research groups to perform MD in the microsecond
timescale.

9
Acknowledgement

We would like to thank the OpenMM team for developing an excellent and open source engine
and the AlphaFold team for developing an excellent model and open sourcing the software.
We also would like to thank Sergey Ovchinnikov, Milot Mirdita and Martin Steinegger for
their fantastic ColabFold and David Koes for the excellent py3Dmol plugin.

Supporting Information Available

All notebooks presented here are freely and publicly available at https://github.com/
pablo-arantes/Making-it-rain.

References

(1) Boodhun, N. Seeing is believing: structures and functions of biological molecules.

https://doi.org/10.2144/btn-2017-0123 2018, 64, 143–146.

(2) Rout, M. P.; Sali, A. Principles for Integrative Structural Biology Studies. Cell 2019,
177, 1384–1403.

(3) Callaway, E. Revolutionary cryo-EM is taking over structural biology. Nature 2020,
578, 201.

(4) Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature
2021 2021, 1–11.

(5) Baek, M. et al. Accurate prediction of protein structures and interactions using a three-
track neural network. Science 2021, eabj8754.

(6) Tunyasuvunakool, K. et al. Highly accurate protein structure prediction for the human
proteome. Nature 2021 2021, 1–9.

10
 (7) Bershtein, S.; Kleiner, D.; Mishmar, D. Predicting 3D protein structures in light of
evolution. Nature Ecology & Evolution 2021 2021, 1–4.

(8) Fleishman, S. J.; Horovitz, A. Extending the New Generation of Structure Predictors
to Account for Dynamics and Allostery. Journal of Molecular Biology 2021, 167007.

(9) Liu, X.; Shi, D.; Zhou, S.; Liu, H.; Liu, H.; Yao, X. Molecular dynamics simulations
and novel drug discovery. https://doi.org/10.1080/17460441.2018.1403419 2017, 13,
23–37.

(10) Hollingsworth, S. A.; Dror, R. O. Molecular Dynamics Simulation for All. Neuron 2018,
99, 1129–1143.

(11) Ebejer, J.-P.; Fulle, S.; Morris, G. M.; Finn, P. W. The emerging role of cloud computing
in molecular modelling. Journal of Molecular Graphics and Modelling 2013, 44, 177–
187.

(12) Wong, A. K. L.; Goscinski, A. M. The Design and Implementation of the VMD Plugin
for NAMD Simulations on the Amazon Cloud. International Journal of Cloud Com-
puting and Services Science (IJ-CLOSER) 2012, 1 .

(13) Purawat, S.; Ieong, P. U.; Malmstrom, R. D.; Chan, G. J.; Yeung, A. K.; Walker, R. C.;
Altintas, I.; Amaro, R. E. A Kepler Workflow Tool for Reproducible AMBER GPU
Molecular Dynamics. Biophysical Journal 2017, 112, 2469–2474.

(14) Jo, S.; Kim, T.; Iyer, V. G.; Im, W. CHARMM-GUI: A web-based graphical user
interface for CHARMM. Journal of Computational Chemistry 2008, 29, 1859–1865.

(15) Ribeiro, J. V.; Bernardi, R. C.; Rudack, T.; Stone, J. E.; Phillips, J. C.; Fred-
dolino, P. L.; Schulten, K. QwikMD - Integrative Molecular Dynamics Toolkit for
Novices and Experts. Scientific Reports 2016 6:1 2016, 6, 1–14.

11
(16) Nicolas-Barreales, G.; Sujar, A.; Sanchez, A. A Web-Based Tool for Simulating Molec-
ular Dynamics in Cloud Environments. Electronics 2021, Vol. 10, Page 185 2021, 10,
185.

(17) Engelberger, F.; Galaz-Davison, P.; Bravo, G.; Rivera, M.; Ramı́rez-Sarmiento, C. A.
Developing and Implementing Cloud-Based Tutorials That Combine Bioinformatics
Software, Interactive Coding, and Visualization Exercises for Distance Learning on
Structural Bioinformatics. Journal of Chemical Education 2021, 98, 1801–1807.

(18) Eastman, P.; Swails, J.; Chodera, J. D.; McGibbon, R. T.; Zhao, Y.; Beauchamp, K. A.;
Wang, L.-P.; Simmonett, A. C.; Harrigan, M. P.; Stern, C. D.; Wiewiora, R. P.;
Brooks, B. R.; Pande, V. S. OpenMM 7: Rapid development of high performance
algorithms for molecular dynamics. PLOS Computational Biology 2017, 13, e1005659.

(19) Google Colaboratory. https://colab.research.google.com/notebooks/intro.

ipynb.

(20) Bisong, E. Google Colaboratory. Building Machine Learning and Deep Learning Models
on Google Cloud Platform 2019, 59–64.

(21) Project Jupyter. https://jupyter.org/.

(22) Michaud-Agrawal, N.; Denning, E. J.; Woolf, T. B.; Beckstein, O. MDAnalysis: A

toolkit for the analysis of molecular dynamics simulations. Journal of Computational
Chemistry 2011, 32, 2319–2327.

(23) Roe, D. R.; Thomas E. Cheatham, I. PTRAJ and CPPTRAJ: Software for Processing
and Analysis of Molecular Dynamics Trajectory Data. Journal of Chemical Theory and
Computation 2013, 9, 3084–3095.

(24) Harris, C. R. et al. Array programming with NumPy. Nature 2020, 585, 357–362.

(25) Case, D. A.; et al., https://ambermd.org/AmberTools.php.

12
(26) Ovchinnikov, S.; Mirdita, M.; Steinegger, M. ColabFold - Making Protein folding ac-
cessible to all via Google Colab. 2021,

13