From aspirin to the most sophisticated and specialized drugs, it is difficult to overstate the impact that pharmaceutical chemistry has made on modern medicine.

Less widely known is that a drug’s operation is dictated by a host of properties that depend on the drug’s crystal structure. A new drug might be the silver bullet for a killer virus, but if it dissolves at the wrong rate in the human bloodstream, it may be useless - or dangerous. In legal matters, patents apply only to a single crystalline form.

It’s therefore no surprise that big pharmaceutical companies and other researchers are eager for a computational technique that can predict the possible crystal structures of a
molecule. Crystal-structure prediction aims to provide that technique.

The hitch has been harnessing the computing resources to make those predictions.

Computational crystal structure prediction typically requires the processing of an enormous computational workload in the form of thousands of small jobs.  At University College London, Sally Price’s group, working in collaboration with OMII-UK and the university-wide Condor pool, can draw upon an integrated distributed-computing system.

This gave them the computing power they need to predict crystal structure.

Why model a crystal?

Crystal-structure prediction answers the question ‘what crystal structures will an organic molecule adopt?’

Different researchers have united to break that question down into two smaller steps ‘what crystal structures could this organic molecule adopt,’ and ‘which structure is the molecule most likely to adopt?’ These two questions naturally separate the computational procedure into two steps.

The first step is the generation of a bank of structures that the molecule could adopt (usually based on very simple-to-evaluate criteria). The second step is the calculation of a lattice energy for each structure, which can be used to estimate the structure’s stability.

Determining the lattice energy of a structure presents a computational challenge, since it requires energy calculations for thousands, or tens of thousands, of possible crystal structures. One solution is to exploit a Condor pool and run individual energy calculations on the unused cycles of desktop machines — for example, unused undergraduate PCs.

The spoiler is that a copious series of complications arises from the first step of the crystal-structure prediction: generating thousands of structures, distributing thousands
of jobs across a Condor pool, and collecting the results as they are generated.

At UCL, OMII-UK helped Sally Price’s group by providing a set of tools and a custom-made system that allows the group’s prediction procedure to fully exploit the university’s Condor pool. OMII-BPEL orchestrates the workflow through which the various programs involved in crystal-structure prediction are executed and coordinated, and the GridSAM job submission interface handles the submission of thousands of energy-evaluation jobs to the Condor pool and the subsequent collection of the results.

Meanwhile, the PlotWS webservice allows users to monitor the progress of the calculations.
The basic elements of this system have been in place since 2004, but as new workflows have come online, the spare cycles of more desktop machines can be be put to work predicting the crystal structures of organic molecules supplied from across the country by experimental and industrial collaborators.

—Matthew Habgood, University College London. A version of this article originally appeared in the OMII-UK Newsletter.