This document is designed to help you set up an OpenFF QM+MM benchmark using your own set of molecules.
If you experience any issues during any stage of the procedure, please raise issues here. You can also receive support on the Open Force Field Slack channel benchmarks-partners
NOTE: all discussions are public, so please do not disclose any molecule information you wish to remain private.
Installation
conda
environments have been created which are fully self-consistent and contain at a minimum all of the required dependencies to run the calculations and generate the final analysis.
Create a new conda environment on any machines that will hold your local QCFractal server or run compute managers using the command:
conda env create openforcefield/openff-benchmark-optimization
Activate the environment using:
conda activate openff-benchmark-optimization
Any update to the workflow will be distributed via an update to this environment.
Software Updates
Whenever we need to update any of the software with bug fixes or improvements we will update the conda environment above. To install these updates it is recommended that you remove the current working environment with conda env remove -n openff-benchmark-optimization
and install again using the instructions above.
Compute Approaches
Multiple compute options have been created to allow users to generate the QM and MM optimizations required for the benchmark. The main difference between these is the use of a long-running self-hosted QCFractal server instance to store and collect all data during the benchmark.
If you have have chosen a compute option that requires a self-hosted QCFractal server instance, follow the instructions below.
Setting up a QCFractal Server
In order to be able to use a local QCFractal server instance, you will need to be able to host the instance on a machine with a port (e.g. 7777) that is network-accessible from either the head node or compute nodes of your cluster. In the case where you do not have access to an HPC cluster and are using a single machine to host the server and perform compute, all network communication will occur in-host via localhost
.
There are several benefits to using a self-hosted server:
There is a long-running database of results which can be queried at any time during execution for progress.
QCFractal will automatically handle the parallelization and distribution of all tasks between connected managers.
Error reporting is more verbose as it is captured into the database on failed tasks.
Server setup
Initialize your server instance with:
qcfractal-server init
See qcfractal-server init --help
for additional options. This will by default place the server data in ~/.qca/qcfractal
.
Start up the server instance with:
qcfractal-server start
See qcfractal-server start --help
for additional options.
Manager setup
We will submit molecules and computation requests to the server, but compute of those computation requests must be performed by one or more QCFractal manager(s) that can communicate with that server.
Generally, you can start a manager with:
qcfractal-manager -v --verify False --config-file <config.yaml>
The content of your manager’s config.yaml
, as well as whether you will launch it as a long-running process or as part of a batch job, depends on your cluster setup and the compute option you have chosen to use on it.
Troubleshooting
Where is the local QCArchive database stored
you can find the database in a path like
~/.qca/qcfractal/
on the machine where you activated the server.
Why do I see stereochemistry issues in the initial step in the workflow?
All file reading and writing is done via the Open Force Field Toolkit, with many cheminformatics operations powered by RDKit. When we read in a molecule we check with RDKit if the molecule has any missing stereochemistry information. We recommend using 3D SDF files where possible so that all stereochemistry is fully defined.
Missing molecules?
During the workflow, things can go wrong involving undefined stereochemistry when this happens a warning is shown about the molecule is moved to an error directory to keep track of.
During the analysis, you may notice some molecules are missing or the number of conformers of the molecule is not as expected. During optimization, we sometimes see proton transfer and the resulting optimized geometry is actually a tautomer of the input. In this case, the optimization is excluded from the analysis. This should not affect the results.