X-ray crystallography has been used as a tool for studying the structure of proteins since the Nobel Prize (1962) winning work of Max Perutz and Sir John Chowdry. The technique has been especially useful in studying the three dimensional interaction of potential drug molecules with target molecules. X-rays bounce off the crystal (diffraction) and the intensity of the diffracted beams are measured using film, X-ray sensitive plates, CCD cameras, or the latest innovation, pixel-array detectors. The diffraction data is transformed into an electron density map which the skilled crystallographer uses to build a three-dimensional model of the atomic structure It has been a methodology incorporating as much art as science, based on the experience of the experimenter, and has involved a great deal of tedious and laborious experimental design and execution. As soon as automated robotic manipulation and microliter scale liquid handling entered the laboratory realm, crystallographers became keenly interested in the possibilities such technology offered to their discipline. The Lab Man asked Eric Baldwin, Director of Protein Crystallography at Bristol-Myers Squibb (BMS) in Princeton N.J. to give us an update on automation in this interesting field.

Eric indicates that the X-ray method requires the production of a crystal. Protein samples need to be prepared and those protein samples combined with the drug molecule to be studied. The complex of protein and drug has to be induced to form a crystal. This is a largely empirical art. The crystallizer will often try many hundreds of different experiments to cause the protein-drug complex to crystallize. In the late 1970?s it was common for crystallizers to exchange lists of recipes for growing protein crystals that had worked for them in the past. By the early 1980?s many were experimenting with the application of experimental design methods that produced novel combinations of reagents that might produce crystals. These new recipe lists were adopted by a commercial enterprise which began to sell crystal screening kits. These kits were very popular and now 1000?s of solutions are available for sale. Traditionally all crystallization experiments were set up by hand. A few micro-liters of the protein solution and an equal amount of one crystal screening solution would be mixed together and the experiment observed for several days to see if crystals would grow. This would be repeated for 100?s or 1000?s of different solutions, trying to find a combination of reagents that would grow useful crystals. This was obviously very labor intensive and an excellent opportunity for automation to improve efficiency and eliminate errors caused by human fatigue.

In the mid-1990?s the first commercial crystallization robots appeared. These were widely purchased in pharmaceutical crystallography labs and were successfully used to screen the many hundreds to thousands of crystallization conditions needed to find those that grew crystals. Many of the early attempts tried to directly replicate the manual process, the hanging drop method for growing protein crystals. This was quite slow and introduced less-reliable steps that the researcher ?had to watch? or would do by hand to avoid a robot error. The most error prone event was the final step when the crystallization experiment was sealed for vapor equilibration. The hanging drop method mixes the protein solution and the crystallization screening solution on a surface. The surface which could hold only one crystallization drop or as many as 96 drops needs to be flipped over and suspended above a reservoir to create the crystallization environment. Automated cover slip flippers were generally not reliable enough for routine un-supervised operation.

Other automation difficulties included the large range in reagent viscosity and volumes that the robot needed to be able to dispense during the typical experiment. Protein crystallization experiments often use polymers of polyethylene glycol (PEG) to facilitate the crystallization process. When these polymers are mixed with proteins they compete with the protein for water of solvation. This action promotes the assembly of proteins into aggregates that might cause crystal nucleation. PEG solutions are quite viscous. Hence, the automation application would need to be able to accurately dispense low viscosity solutions like water and high viscosity solutions like PEG. The dispensed volumes are also quite variable. Crystallization drops may be a combination of 1 microliter of protein solution and 1 microliter of crystallization screening solution. This 2 microliter drop is suspended over a 100-500 microliter well solution. Hence, a 100- to 500-fold range in volume dispensing needs to be done accurately with the same apparatus. A robotic system that performs well with these requirements is difficult to implement by merely automating the manual process. New robotic systems requiring new workflows have been designed to work in an unsupervised mode, account for viscosity changes, and dispense low or high volumes with task specific hardware.

According to Eric, at BMS the ability to prepare 1000?s of crystallization experiments using automation has been very useful for quickly finding the conditions that produce crystals for many proteins. But the ability to set up these experiments created new bottlenecks that needed to be addressed. Each crystallization experiment needs to be visually examined under a microscope and repeatedly checked over the course of a week or more to determine if the experiment produced crystals or a result that might be close to conditions that would give crystals. To address this bottleneck, automation experts at BMS built and deployed a crystal imaging system that takes high-magnification digital pictures of crystallization experiments. This has dramatically reduced the need for scientists to examine experiments under the microscope and provides a time-lapsed series of images to help understand how the experiment evolved over time.

A camera with a motor driven zoom lens is mounted over a controllable X/Y stage. A robotic arm moves crystallization experimental trays to/from storage and the imaging stage. Crystallization experiments evolve over many days and so multiple images are captured over the course of each experiment on a user-defined schedule. A database allows association of protein information and crystallization conditions with images. An extension of this custom software, Image Viewer, is used for displaying and scoring crystallization images. This tool is used externally by the BMS outsourcing partner in India to manually annotate all crystallization experiments, and is used internally by BMS crystallizers to rapidly visualize crystallization ?hits? and follow up with new experiments. During the night-time crystallization images are scored overseas and all of the interesting results are emailed to the crystallizers a report that they receive at 8 AM in the morning. "We have set up 100?s or 1000?s of crystallization experiments on a Friday afternoon, and the crystallization scientist has had an email on Monday morning when he returned to work, indicating that a big crystal was ready for data collection", says Eric. The scientist could go right to the correct drop and harvest the crystal and data collection could begin immediately. The team has also integrated RockMaker, a key third-party software package, to design crystallization experiments and dispense them on Tecan liquid handlers. Scientists can drag the conditions from Image Viewer back to RockMaker to initiate the design of a focused experiment, allowing for very efficient experiment design flow. Finally, a fully integrated robot-ready inventory of bar-coded solutions for rapid custom crystallization experimentation has been created.

When asked to speculate on future directions for automation, Eric noted that one obvious extension would be to replace manual scoring of crystallization experiment images with a computational scoring process. As far back as 25 years ago scientists at the Naval Research Labs in Washington experimented with using defense image analysis technology to accurately score the crystalline and non-crystalline results of crystallization experiments. Because crystals have edges, detecting them is relatively easy with modern algorithms. However, non-crystalline material in the experiment, precipitate, looks like snow and can hide crystals. More difficult is the fact that snowy precipitate can be ?good? or ?bad?. The experienced human eye can tell the difference between ?good shiny precipitate? that might lead to crystals from the bad brown precipitate that is just denatured proteins. Distinguishing these different precipitates with image analysis tools will be a challenge for many years ahead, but because we have millions of images that have been manually annotated we may now have a superior basis set for a machine learning experiments.

Another area of developing technology is the field of data collection automation. Robots have been available for the past 5-7 years that will automatically mount crystals for X-ray data collection. However, the latest generation of detectors threatens to make these robots obsolete. The new pixel-array detectors are so fast that the rate-limiting step in data collection is not exposure of the crystals to the X-rays but movement of the sample to and from the liquid nitrogen storage Dewar to the X-ray beam. New robots will be needed to rapidly select the crystals and place them in the X-ray beam every few seconds.

Keeping track of all of these samples is also becoming a problem. One instrument company has introduced RFID technology that works at liquid nitrogen temperature. The crystals are mounted on RFID encoded pins and stored in a liquid nitrogen bath until they can be exposed to the X-ray beam. Sample information is loaded onto the RFID. When the pin is selected for data collection and moved into position, the RFID reader in the robot arm downloads the pin information to connect the sample information with the resulting diffraction data set.

In The Lab Man's opinion, the automation of X-ray crystallography is fascinating microcosm of how laboratory automation evolves. First we attempt to directly automate the manual procedure, only to discover that it doesn't always translate well. We re-engineer the process to better take advantage of automation, and then realize we have created new bottlenecks, some manipulative in nature, others perhaps data related. As those eventually get addressed, the bottlenecks shift. Mix in globalization developments and what we have is a complex evolution that has a lot of potential blind alleys, which can consume both funds and time. To gain the best productivity from such evolutions, long-term vision and guidance is essential, which at BMS is readily available from an excellent team of internal automation experts.

Until next time,

Domo Arigato, Mr. Roboto !