High throughput protein crystallization
EMBL Practical Course on Protein Expression, Purification and Crystallization
August 14th-20th, 2000 EMBL Outstation Hamburg, Germany
Bernhard Rupp
University of California, LLNL-BBRP, Livermore, CA 94551
Institut für Theoretische Chemie und Strukturbiologie der Universtät Wien, A 1090 Wien
© 2000 Bernhard Rupp
Crystallization space and screening
Global success rate and hot spots
Truncation and orthogonalization of
sample space
Correlations and variance analysis
Overview - cost and efficiency
considerations
Appendix – Is hanging drop better
than sitting drop?
One of the requirements for any structural genomics project will be a high throughput crystallization facility. In the absence of any predictive ab-initio algorithms or rules for crystallization, random screening for crystallization conditions provides a powerful approach by statistically valid sampling of the multidimensional parameter space of crystallization with the aim to establish correlations between crystallization probability and intrinsic parameters and properties of proteins. At the same time, once enough data are accumulated, orthogonalization of the sample space and trimming of the multidimensional parameter vertices become possible, and correlations between success rate and protein properties may emerge from correlation analysis. Such an optimized protocol maximizes the probability of success while minimizing the number of sample space dimensions, leading to increased throughput at reduced cost. The implementation through our CRYSTOOL program and the design of robotic systems at a reasonable cost from commercially available components will be described. The concepts of high throughput and efficiency in crystallization are applicable to any crystallization project even in smaller labs, and not limited to large-scale structural genomics efforts.
.
Crystallization screening is the search for conditions that allow controlled phase separation and crystal growth in a multi-component mixture or solution. The process takes place in (or needs to move into) the metastable region between binodal and spinodal phase boundary lines. In addition, kinetic parameters such as nucleation frequency, growth rate and diffusion rates should favor the development of few well ordered, well diffracting crystals. While there is agreement on the desirable result, the best way to achieve such results is however hotly debated. Much of the ‘knowledge’ is anecdotal, with little statistical evidence or control experiments to prove its general efficiency or usefulness.
In addition, the
need to rapidly and efficiently screen large numbers of very different proteins
requires a protocol and procedure that can be automatically applied, analyzed,
and implemented at a reasonable cost.
All screening protocols search the protein crystallization space for successes. We define success as the first appearance of a diffracting crystal[1]. The crystallization space is a multi-dimensional vertex with chemical components, general physical parameters, and method specific parameters as its components. The chemicals can be grouped in classes like precipitant, additive, buffer, and detergent. Physical parameters might be pH or temperature, and method specific parameters might be drop size, oil composition, or the phase of the moon. Protein specific parameters may include concentration, tags, cofactors, ATP, Mg, etc.
The dimensionality of a reasonable experimental crystallization space containing most of the above-mentioned parameters is about 100 (keep the monstrous polyhedral figure in mind for now). Systematically sampling, even rather coarsely at three data points per axis, low, medium, high; would require 5x1047 experiments. Not an option.
Our problem is that we are trying to optimize
something for which we have no target model, and therefore no target function
to minimize. Systematic sampling is out of the question, not only due to supply
and time constraints – also, very important, due the fact that there is
no guarantee that a particular protein will in fact ever crystallize[2].
At best, we have some ideas of what not to do to a specific protein – like
using phosphate when Ca is a metal cofactor.
The approach we are left with for now, is statistically valid sampling of the crystallization space. We want to achieve two objectives with our sampling: first, we want to use the least experiments to get to a first success, and secondly, we want to screen comprehensively, to be able to make statistical judgments about the sample space and the correlation of conditions of success with certain properties of proteins. While we cannot give any guarantee that condition x will work for specific protein y, we can say that the overall probability of success are higher for ‘protein’ under condition x, thus increasing throughput.
The most common sample protocols compared for efficiency by Segelke, 1995, include Grid Screen (McPherson, 1992), Footprint screening (Strua E, in Ducruix and Giege, 1999) and random (stochastic) CRYSTOOL sampling (Segelke and Rupp, 1998). The sparse matrix protocol by Jankaric and Kim (1991) is essentially a random protocol that overpopulates certain few (and not necessarily successful, as Hampton Crystal Screen II shows) ‘random’ points in space. Incomplete factorial screens (Carter JW, in Ducruix and Giege, 1999) are well suited for optimization purposes, and for screening experiments factorials can be treated as sparse random points. The important advantage of CRYSTOOL screening is that it samples the selected space – and not just a few points in the space – thoroughly, which is crucial for data mining and statistically valid correlation analysis, orthogonalization, and truncation of the sample space (see next chapter). CRYSTOOL also has the greatest chance of finding different crystal forms – a useful feature commonly undervalued. Much time has been spent optimizing unsuitable crystal forms when other forms were readily obtained, much better diffracting, from extensive screening.
Figure 2 shows a presentation of sampling protocols in schematic crystallization space (Segelke, 2000). In the 2-dimensional case, the analogy to the ‘battleship’ game is obvious: What would be the best strategy to strike the first hit on your opponents ships with the least trials? It turns out, even by simple trial, that it is random. Segelke 1995 has rigorously derived that the best efficiency for random sampling holds also in n-dimensional space.
The experimental support for this conclusion comes from the following experiment:
Five proteins (four known to crystallize and a new one) were subjected to
different sampling protocols (this in fact did keep quite a number of graduate
students busy, Segelke 1995). On average, the random protocol indeed has the
highest success rate. The notable exception is Catalase, which crystallizes
under pretty much every condition, so there is no significant advantage to
either sampling strategy. The more infrequent the conditions under which the
protein crystallizes, the more advantageous in fact becomes random sampling.
This has been proven in mathematical terms, again by Segelke, 1995.
Random sampling to set up the above experiments is implemented in the CRYSTOOL program at https://www.ruppweb.org/crystool/
The potentially most crucial question to ask is when to stop crystallization trials and to consider modifying the protein. We attempted to estimate the cutoff based on the few cases where proteins were crystallized and success rates were given or could be extracted from the publications. Data are from Jankaric and Kim (1991), the PA consortium (Eisenberg et al.), and our own collection of crystal growing efforts (graph from Segelke, 2000).
The graph allows an important conclusion: The success rate of crystallization appears bimodal. There is a majority of proteins, ~65%, that crystallized rarely (or not at all), but of those which were crystallized successfully, the success rates are rather high. We recognize that the statistical basis for this conclusion is still weak, but with upcoming data from the TB structural genomics consortium we expect to increase statistical significance.
The practical implication of the relatively high success rate for ‘well behaved’ or ‘crystallizable’ is that it is more efficient to stop screening and revert to protein engineering once a certain number of trials have failed to yield a success. Again, using data from known and newly crystallized proteins, we have estimated, based on the cumulative probability distribution, a reasonable number of experiments to be about 300. As a reminder, this is to achieve best throughput with a large number of proteins – your commitment to an individual protein may lure you into screening on and on. Just make sure you realize that at this point you do have a problem and may have to do smarter things that brute force screening. What these smarter things might be is now an individual problem of a specific protein (preparation) – and not a matter of screening statistics.
The graph above (Segelke, 2000) shows how many trials need to be attempted to have a certain chance of success in obtaining a crystal. If a protein had a success rate of 2%, after conducting 300 random experiments, you would be almost certain to obtain a crystal. Even in case of a ‘rarely’ crystallizing protein (0.2%), you still would have a nearly 50% chance of finding a crystal. The table below shows actual success rates of a variety of proteins crystallized using CRYSTOOL screens in my lab and at UCSD (Table 1):
Protein |
Source |
Percent success rate |
Lysozyme |
LLNL |
25 |
XBRCTb |
LLNL |
17 |
Catalase |
UCSD |
10 |
Calmodulin |
LLNL |
8 |
Phospholipase |
UCSD |
8 |
Rnase B |
UCSD |
6 |
Subtilisin |
LLNL |
4 |
Thaumatin |
UCSD |
3 |
We thus believe that a starting cutoff, subject to refinement by inclusion of new data, of about 300 experiments (288, twelve 6x4 plate equivalents) gives a rather good chance of obtaining a crystal – if the protein has a tendency to crystallize well, as about 35% of (soluble) proteins appear to do in fact. Random sampling up to this number of experiments in addition often yields multiple crystal forms, some with superior quality to others.
Once we have accumulated sufficient data covering our parameter space, we can begin to analyze a variety of things. I will give a few examples, many more are conceivable and it will depend on data range, user input, and our ingenuity what we actually can extract from our data.
Let us consider the crystallization space again as an n-dimensional vertex, with n largely defined by the number of components used to create our screening solutions. As a first step, we can populate the n-dimensional vertex with global success rate, i.e., the ‘hot spots’ for all of the proteins we tried. At this point we do not know with any statistical certainty whether such hot spots (Na-Malonate at high molarities, Mc Pherson, 2000) – or dead spots (other than Crystal Screen II) – exist (figure from Segelke, 2000).
Considering that the vortex dimensionality is largely defined by the number of components used to create our screening solutions, we can refine our analysis. CRYSTOOL for example, uses 85 different components. It is conceivable that certain extreme combinations, located close to the center, surface or edges of the n-dimensional vertex, never yield any crystals. Such combinations can be easily excluded from the CRYSTOOL random screening setup. Example: High ionic strength and high large Mw PEG concentration, protein concentration below 1 mg/ml (such conditions are not to be confused with inherently impossible ones, like phosphate and Ca++, which are excluded already by CRYSTOOL).
Components may also be highly correlated, similar to linear dependence in mathematical terms. Redundant salts can be eliminated (or the frequency of Na-Malonate increased, McPherson, 2000). Such is equivalent to orthogonalization in a mathematical sense (we are looking to find the principal components or axes of the vertex). Expensive but ineffective detergents could be removed from the sample space – again reducing dimensionality and cost while increasing efficiency.
We can take the analysis further and correlate success rate with certain properties of ‘protein’. Examples are PI vs. pH of success, His6–tag vs. precipitant type, etc. To find out whatever parameters in fact do correlate is the goal of such analysis, aiming to find unexpected, highly significant and probably multidimensional correlations. In fact, with sufficiently dense population of the sampling space, a global variance and multiple regression analysis in the sense of CW Carter can be attempted, although not for individual protein optimization, but for ‘protein’ in general, with possible subdivision of ‘protein’ in protein classes identified by certain common properties or qualities. I would expect that in few years from now we will have a wealth of statistically significant correlations which to rationally explain will give ample opportunity for further in-depth analysis.
How are we going to create these data-to-be-mined efficiently? For the NIH TB structural genomics consortium we have proposed and are currently implementing a modest cost design based on commercially available components, which will provide sufficient throughput to obtain quality crystals of about 400 target proteins in selected pathways of Mycobacterium tuberculosis.
The computational backbone of our setup is a Windows NT based system, with SQL server as an open database, and applications linked and controlled by VBX (web) scripts. Main components are a Packard Instruments Multiprobe II (MPII) liquid handling station for screen preparation and tray setup, Velocity11 PlateSeal plate sealer, and an optical recognition system (likely Emerald Biosciences), arranged around a central Hudson Controls PlateCrane (TM) plate handler and stacker. The need for a seamless interface between various applications requires open software architecture, essentially precluding the use of our Cyberlab200 robot or any other closed proprietary applications (DSI CrystalScore system). The C200 is also very slow and would need major modifications. Plates will be standard footprint plates (which is smaller than the Linbro plates) in sitting drop setup (siliconized cover slides, in addition to requiring complicated handling, are prohibitively expensive). Greased trays are avoided by tape sealing. The following table of expenses clearly shows the dramatic drop in cost when disposing of cover slides and grease. In addition, none of the existing plates are optimal for robotic dispensing, and a new plate design is in the making.
Table 2: major cost factors in crystallization setup
Another most time consuming step is the repeated inspection of the droplets to check for occurrence of crystals. Current detection software is limited to a binary decision crystal – no crystal, and will need improvement to ultimately achieve a finer score scale. In the first step, the automated analysis will provide a binary decision, with a minimum of false positives, whether a crystal-like object is present or not. Positives are tagged, and a manual inspection will follow.
Due to open architecture and compatibility with proven and customizable image recognition software, Emerald CrystalMonitor will be used as a data base front end driving the automated video stage and also interface to the setup robots. Each plates in a stack will be moved to the inspection table with the CCD microscope by the Plate Crane and shelved into another rack. A score list will be created and the operator notified by e-mail to inspect the video images in CrystalMonitor. Provided that false positives are infrequent, one can only hope that the visual inspection of crystals became a throughput-limiting factor!
We have provisions for one single, combined harvesting and optimization step in one to two plates with a mini-crystool screen around the initial conditions. From this point on, we operate essentially manually. The crystals will be conventionally harvested in standardized cryo-loops, cryo-protected, preliminary low temperature data screened, and suitable crystals cryo-shipped to the data collection facility (BNL, ALS) for automated mounting and data collection. The same procedure holds for Se-Met material, which may be used ab-initio, or after the first screening was successful. The actual variation of the procedure will largely depend on the experience gained through the expression stage, and subject to revision. We also expect to derive a successful set of cryo-protection and cooling procedures, with the analysis performed in a similar way as described for the crystallization experiments.
Automated crystallization design, setup, and analysis increases throughput, reduces cost, improves reproducibility, and minimizes errors. In addition, the comprehensive database created can be analysed to establish correlation between material properties and success rate. Complete mapping of the orthogonalized experimental parameter space may define regions with statistically significant higher probability of crystallization success. A automated system with a throughput of several hundert targets per year can be designed from commercially available components for about $200-$250k.
Anecdotal rumors exist that give preference to one or the other setup technique, for example, sitting vs. hanging drop. Again, no statistical validity is attached to these rumors. We therefore did a brief experiment testing ConA and Lysozyme in four different setups. We recognize that both fall into the category of ‘easy-to-crystallize’ proteins, and will follow up with other proteins.
We conducted 4 sets of 96 experiments (using the same array of 96 CRYSTOOL solutions), each either with siliconized slides in hanging drops, in sitting drops on Hampton plastic plates, on Hampton glass posts, and, to have the exactly same surface, broke siliconized cover slides into pieces and used the slide pieces to support sitting drops. The result can be summarized as follows: No significant differences in the success rates of either method exist for our test set. Graph from Schick, Segelke and Rupp, unpublished.
George A, Wilson WW. 1994. Predicting protein crystallization from a dilute solution property. Acta Crystallogr D 50:361-365
Giege R, Ducruix A. Eds. 1999. Crystallization of Nucleic Acids and Proteins : A Practical Approach (Practical Approach Series, 210) 2nd ed. Oxford University Press, New York.
Jancarik J, Kim SH. 1991. Sparse
Matrix Sampling : A Screening Method for Crystallization of Proteins., Journal of Applied Crystallography. 24,
409-411
McPherson A. 1982. Preparation and Analysis of Protein Crystals. Krieger, Malabar, FL.
McPherson A. 2000. Growth and Handling of Protein Crystals in High Throughput Production. ACA Meeting Series 27, 43
Segelke BW. 1995. Efficiency analysis of sampling protocols used in protein crystallization screening and crystal structure from two novel crystal forms of PLA2, , PhD Thesis, University of California, San Diego.
Segelke BW, Rupp B. 1998. Beyond the Sparse Matrix Screen: A Web Service for Randomly Generating Crystallization Experiments., American Crystallographic Association Annual Meeting, July 18-23, 1998, Arlington, VA. ACA Meeting Series 25, 78
Segelke BW. 2000. High throughput crystallization screening for structural genomics and data mining. American Crystallographic Association Annual Meeting, July 22-27, 2000, St.Paul, MN. ACA Meeting Series 27, 43