2. Getting Started

2.1. Introduction

The first thing to understand about the OpenMM “application layer” is that it is not exactly an application in the traditional sense: there is no program called “OpenMM” that you run. Rather, it is a collection of libraries written in the Python programming language. Those libraries can easily be chained together to create Python programs that run simulations. But don’t worry! You don’t need to know anything about Python programming (or programming at all) to use it. Nearly all molecular simulation applications ask you to write some sort of “script” that specifies the details of the simulation to run. With OpenMM, that script happens to be written in Python. But it is no harder to write than those for most other applications, and this guide will teach you everything you need to know. There is even a graphical interface that can write the script for you based on a simple set of options (see Section 3.5), so you never need to type a single line of code!

On the other hand, if you don’t mind doing a little programming, this approach gives you enormous power and flexibility. Your script has complete access to the entire OpenMM application programming interface (API), as well as the full power of the Python language and libraries. You have complete control over every detail of the simulation, from defining the molecular system to analyzing the results.

2.2. Installing OpenMM

OpenMM is installed using the Conda package manager (https://docs.conda.io). Conda is included as part of the Anaconda Python distribution, which you can download from https://docs.continuum.io/anaconda/install. This is a Python distribution specifically designed for scientific applications, with many of the most popular mathematical and scientific packages preinstalled. Alternatively you can use Miniconda (available from https://docs.conda.io/en/latest/miniconda.html), which includes only Python itself, plus the Conda package manager. That offers a much smaller initial download, with the ability to then install only the packages you want.

(A third option is to compile OpenMM from source. This provides more flexibility, but it is much more work, and there is rarely a need for anyone but advanced users to compile from source. Detailed instruction are in Chapter 8.)

1. Begin by installing the most recent 64 bit, Python 3.x version of either Anaconda or Miniconda.

2. (Optional) If you want to run OpenMM on a GPU, make sure you have installed modern drivers from your vendor.

  • If you have an Nvidia GPU, download the latest drivers from https://www.nvidia.com/Download/index.aspx. CUDA itself will be provided by the cudatoolkit package when you install openmm in the next steps.

  • If you have an AMD GPU and are using Linux or Windows, download the latest version of the drivers from https://support.amd.com. On OS X, OpenCL is included with the operating system and is supported on OS X 10.10.3 or later.

3. Open a command line terminal and type the following command

conda install -c conda-forge openmm

With recent conda versions (v4.8.4+), this will install a version of OpenMM compiled with the latest version of CUDA supported by your drivers. Alternatively you can request a version that is compiled for a specific CUDA version with the command

conda install -c conda-forge openmm cudatoolkit=10.0

where 10.0 should be replaced with the particular CUDA version you want to target. We build packages for CUDA 9.2 and above on Linux, and CUDA 10.0 and above on Windows. Because different CUDA releases are not binary compatible with each other, OpenMM can only work with the particular CUDA version it was compiled with.

Note

Prior to v7.5, conda packages for OpenMM where distributed through the omnia channel (https://anaconda.org/omnia). Starting with v7.5, OpenMM will use the conda-forge channel. Check the documentation for previous versions in case you want to install older packages.

4. Verify your installation by typing the following command:

python -m openmm.testInstallation

This command confirms that OpenMM is installed, checks whether GPU acceleration is available (via the OpenCL and/or CUDA platforms), and verifies that all platforms produce consistent results.