 |
Online ISSN: 1097-0134 Print ISSN: 0887-3585 Copyright © 2003 Wiley-Liss, Inc. |
|
| Data mining crystallization databases: Knowledge-based approaches to optimize protein crystal screens |
Matthew S. Kimber 1, François Vallee 1, Simon Houston 1, Alexander Ne akov 1, Tatiana Skarina 2, Elena Evdokimova 2, Steven Beasley 2, Dinesh Christendat 2, Alexei Savchenko 2, Cheryl H. Arrowsmith 1 2, Masoud Vedadi 1, Mark Gerstein 3, Aled M. Edwards 1 2 * |
| 1Affinium Pharmaceuticals Inc., Toronto, Ontario, Canada 2Division of Molecular and Structural Biology, Ontario Cancer Institute and Department of Medical Biophysics, University of Toronto, Toronto, Ontario, Canada 3Molecular Biophysics and Biochemistry, Yale University, New Haven, Connecticut |
| email: Aled M. Edwards (aled.edwards@utoronto.ca) |
*Correspondence to Aled M. Edwards, Affinium Pharmaceuticals Inc., 10th floor, South Tower, 100 University Avenue, Toronto, Ontario M5J 1V6, Canada
Funded by:
Ontario Research and Development Challenge Fund Best Foundation
| Keywords |
| crystal screening • crystallization • data mining • structural proteomics |
| Abstract |
 |  | | | | | |
| Protein crystallization is a major bottleneck in protein X-ray crystallography, the workhorse of most structural proteomics projects. Because the principles that govern protein crystallization are too poorly understood to allow them to be used in a strongly predictive sense, the most common crystallization strategy entails screening a wide variety of solution conditions to identify the small subset that will support crystal nucleation and growth. We tested the hypothesis that more efficient crystallization strategies could be formulated by extracting useful patterns and correlations from the large data sets of crystallization trials created in structural proteomics projects. A database of crystallization conditions was constructed for 755 different proteins purified and crystallized under uniform conditions. Forty-five percent of the proteins formed crystals. Data mining identified the conditions that crystallize the most proteins, revealed that many conditions are highly correlated in their behavior, and showed that the crystallization success rate is markedly dependent on the organism from which proteins derive. Of the proteins that crystallized in a 48-condition experiment, 60% could be crystallized in as few as 6 conditions and 94% in 24 conditions. Consideration of the full range of information coming from crystal screening trials allows one to design screens that are maximally productive while consuming minimal resources, and also suggests further useful conditions for extending existing screens. Proteins 2003;51:562-568. © 2003 Wiley-Liss, Inc. |
| Digital Object Identifier (DOI) |
| Article Text |
 | | | | | | |
The ultimate goal of structural proteomics is to obtain, through experimental or computational methods, 3D structural models for every protein in nature. Driving this ambitious program is the expectation that structural information will provide functional insights for many of the proteins predicted by genome-sequencing efforts that cannot be ascribed a function using current sequence homology-based approaches. The challenges in structural proteomics are significant; it has been estimated that some 16,000 structures will have to be determined using experimental approaches to obtain reasonable coverage of fold space.[1] Recent advances in X-ray crystallography methodology and associated technologies have made it possible, at least in ideal cases, to go from data collection to a refined structure in a matter of hours[2-6]; however, actually growing a diffraction-quality crystal is far more time and resource intensive.
The purpose of protein crystallization trials is to efficiently find useful lead conditions from which crystal size and morphology can be optimized. Commonly, one explores a wide range of different solutions in which the salt concentration and type, pH, additive type, temperature, and precipitant type and concentration are varied. The precipitant is usually a long-chain polymer [poly-ethylene glycol (PEG), jeffamine], an inorganic or organic salt, or some small organic molecule (MPD, isopropanol, ethanol).[2]
There are two general approaches to the design of crystallization screens. One strategy aims to blanket potentially useful crystallization space with screens of a few hundreds to over a thousand conditions (see, e.g., the JBScreen at www.jenabioscience.com/jbscreen.html). The other strategy is to use smaller, more efficient screens based on previously successful conditions.[3-8] The most widely used variant of this second strategy, developed by Jancarik and Kim, is based on an incomplete factorial approach, which explores a range of conditions biased toward previously successful crystallization conditions.[8] The popularity of this screening strategy can be ascribed to many reasons, including its economy (1-2 mg of protein are needed), ease of use, manageable size, and convenience (it is sold in preformulated kit form by various companies). As originally formulated, this screen was intended to be modified reiteratively as the experiences of users were incorporated.[3] In practice, however, this has not happened. Partly this is due to a lack of a clear metric by which to decide which of a set of potential screening conditions is better, but also, more fundamentally, a lack of sufficient, standardized experimental data by which to evaluate a screen. While the optimal conditions for crystallizing a particular protein crystal form are often collected and archived in a database,[9] the detailed results of each screen, including partial successes or failures, are not. As a result, while individual crystallization conditions can perhaps be shown to be more or less successful, researchers have tended to supplement the common factorial screens with additional screens rather than attempting to systematically optimize any given one. This may be a reasonable strategy when trying to crystallize one or two proteins, in which the effort and expense involved in making the additional protein, setting up the screen, and evaluating the results is manageable. However, in the context of structural proteomics this extra effort and expense is multiplied over hundreds or even thousands of proteins, and thus the desirability of screens that minimize the number of experiments while maximizing the probability of success becomes far more pronounced. Here, by collating data collected for 755 proteins from 6 organisms we show that it is possible to use the information gleaned from previous screens to improve screening strategies using objective empirical criteria.
| | | | | | |
Proteins predicted not to have membrane-spanning domains from six organisms - the prokaryotes Staphylococcus aureus, Escherichia coli K12, Pseudomonas aeruginosa, Heliobacter pylori, and Thermotoga maritima and the archaeote Methanobacterium thermoautotrophicum - were amplified by PCR, cloned into E. coli expression vectors, overexpressed, and purified using His6 technology, as described elsewhere (see [[10]] for a general overview and [[11][12]]
for typical examples of procedures). Proteins were typically stored at
4°C, in 20 mM HEPES, pH 7.5, and 500 mM NaCl. The solutions for the
initial screen were purchased from Hampton Research. Proteins were
screened in 24-well Lindbro plates, using a 2-
l + 2-l drop size and 700 l
in the well. For some proteins both the his-tagged and non-his-tagged
sample were both screened, and these were scored as separate samples.
Samples also included separate domains of multidomain proteins. For
most samples, two to four protein concentrations, typically ranging
from 5-40 mg ml-1, were screened in parallel; in almost all
cases this included at least one experiment in the 10- to 15-mg/ml
range. The data for these multiple experiments were pooled. All
crystallization experiments were performed at ambient temperature
(approximately 293 K). In total, over 35,000 experiments were performed.
Screening
results were scored by eye after approximately 1 day, 1 week, 1 month,
and 3 months. For proteins where screening multiple samples at
different protein concentrations yielded different outcomes, only the
most favorable outcome was scored and reported. To minimize
subjectivity in scoring, results for each experiment were reduced to
one of three assessments - clear, precipitate, or crystalline. Samples
were required to have at least two conditions in which they are soluble
and two where they are not and no more than five conditions for which
data was missing (if, e.g., the condition was not set or the drop dried
out before it could be scored). These criteria reject 6 E. coli proteins, 3 T. martima proteins, 6 M. thermoautotrophicum proteins, and 55 S. aureus proteins. Minimal screen 6 was derived by sequentially searching all combinations of conditions (48!/6! 42! 
1.2 × 107
combinations) for the one that crystallized the most proteins. Minimal
screen 12 was derived by using minimal screen 6 as a seed and searching
all combinations of the remaining conditions for the six that best
complemented minimal screen 6. Minimal screen 24 was found by repeating
this condition twice more. Although this procedure is not guaranteed to
find the globally optimal subsets, this limitation is likely far less
serious than the one imposed by the limited amount of data available.
Clustering was performed in ClustalX using pairwise identity scores
once screening results had been 
encoded
as amino acid sequences.[13] Cladogram was produced in Phylodraw.[14]
| | | | | | |
A total of 755 protein samples from T. maritima, E. coli, M. thermoautotrophicum, S. aureus, P. aeruginosa, and H. pylori were screened against conditions 1-48 of the Jancarik and Kim screen. The success rates for different genomes (Table I) ranged from 67.6% (for T. maritima, albeit with the smallest sample size) to 36.7% in the case of H. pylori. The differences may reflect differences in the intrinsic properties of proteins from different organisms, perhaps influenced by the intracellular environment of the natural host, or may simply reflect the quality of the protein produced by the E. coli host. Overall, 338 protein samples (45%) yielded at least 1 crystallization lead. For each protein that crystallized, crystals were observed, on average, in 4.7/48 conditions.
| a The
average number of conditions under which crystals were obtained,
considering only those samples for which at least one crystal was
obtained. |
There were large differences in the number of proteins that crystallized in each condition (Table II),
ranging from 76 for condition 9 to 4 for condition 27. Surprisingly,
for 99 of the 338 proteins crystallized (29.3%) crystals were obtained
in only one condition; this is approximately 10-fold higher than might
be expected if crystallization in different conditions were behaving as
independent random variables (48 × 0.1 × 0.947 
3.4%).
Table II. Overview of Results Produced by the Jancarik and Kim Screen | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Buffers when present were at 0.1 M. a pH is experimentally measured pH, performed in duplicate on two different batches of the screen. b Number of proteins for which this condition yielded the only crystal lead. |
Efficacy of Different Precipitants
In 17 of the 48 conditions, salt was the major precipitant. These conditions together crystallized a total of 200 of the 338 proteins (59.2%) [Fig. 1(a)]. There were significant differences in the effectiveness of different salts. Sodium citrate, for example, is represented by a single condition in the screen, 39, but was the eighth most productive condition in the screen. Moreover, this condition was by far the most likely condition to yield the only crystal for a given protein. This unusual behavior may be related to citrate's strong metal chelating abilities, something that suggests further investigation.[2] Ammonium sulphate, long considered an exceptionally good salt for crystallizing proteins, was the major precipitant in four conditions (4, 32, 39, and 47), all of which yield substantial numbers (50, 36, 63, and 30) of crystals. Lithium sulphate (16) and formate salts (33, 34, and 44) were also successful. Citrate, sulphate, and formate salts formed a cluster that affected protein solubility in a similar fashion; they tended to crystallize the same subset of proteins (Fig. 2). The other salts used, namely, phosphate (3, 11, 35, and 48), acetate (7 and 25), and tartrate (2 and 29), were much less successful precipitants; in the case of phosphate, this likely reflected the fact that all of these conditions are acidic. Only 69 proteins crystallized in these 8 conditions.
 | Figure 1. (a)
Venn diagram showing the number of proteins for which crystals were
obtained in conditions where salt (17 screening conditions, red), PEG
(22 conditions, blue), or an organic molecule (9 conditions, green) was
the major precipitant. (b) Number of proteins with a given
number of successful screening conditions. While some exceptional
samples crystallize in up to two-thirds of all conditions, most
proteins crystallize in relatively few conditions. (c) Number of
crystals contained for a given number of selected screening conditions.
Conditions were added one at a time, where the condition added was the
one that most increased the number of crystals obtained relative to the
previously chosen subset. Note that almost all of the crystals obtained
can be obtained from approximately half of the screening conditions and
that the last nine conditions could be omitted without affecting the
number of samples crystallized. [Normal View 32K | Magnified View 106K] |
 | Figure 2. (a)
Subset of the overall data, representing the 338 proteins successfully
crystallized. Different conditions are arrayed vertically and different
proteins horizontally. Yellow squares represent crystals, dark blue
squares represent precipitated proteins, and cyan squares represent
soluble proteins. Proteins and conditions were sequentially clustered
using ClustalX using identity matrices for scoring. (b) Distance
matrix of clustered conditions. Condition numbers are noted along the
diagonal; off-diagonal elements represent the number of instances in
which proteins were found to crystallize in both the condition to the
right and the one above it. Note the strong correlation between the PEG
conditions (bottom right corner) as well as the
citrate/sulphate/formate salts (top left corner). The scale is
logarithmic. On the left is a Cladogram
showing the degree of relatedness of crystal screen conditions as
inferred from the degree to which they crystallize the same protein
samples. Note that the crystallization conditions cluster primarily on
the basis of the chemical nature of the major precipitant and
secondarily on the basis of pH. [Normal View 40K | Magnified View 147K] |
High-molecular-weight PEGs are widely considered the most successful protein crystallization agents.[2][9] Of the 338 total proteins, 229 crystallized (67.7%) in the 14 conditions that consist of PEG 4000 or 8000 at concentrations greater than 18%, possibly with some buffer and inert salt (excluding condition 45, which contains zinc), for an average of 54 crystals per condition. The six most productive conditions overall were all in this group. However, there was, in general, a great deal of redundancy among the PEG conditions [Fig. 2(b)]. This region of parameter space appears to be heavily oversampled, a consequence of strong bias toward previously successful experiments built into the original design of the screen. Low concentrations of PEG (36, 37) appeared to be less effective at obtaining crystals but crystallized a different set of proteins than those obtained at higher concentrations. PEG 400 (13, 14, and 23) appeared to affect the solubility of proteins in a manner that more resembled the action of a salt than that of a high-molecular-weight PEG. This may reflect the fact that smaller PEGs probably precipitate proteins more by a solvent competition effect than a volume exclusion effect. Also, zinc in conjunction with PEG yielded different solubility behavior than other salt/PEG combinations, likely reflecting the ability of transition metals to effect strong interactions between proteins by virtue of their ligand chemistry.
Organic precipitants formed a minor part of the screen. There are four MPD conditions (1, 5, 21, and 26) that yielded a modest numbers of crystals (39 proteins crystallized, average of 14 crystals per condition). The 5 conditions in which isopropanol was the primary precipitant (8, 12, 19, 24, and 27) yielded few crystals (average 9 crystals per condition, 39 proteins crystallized) unless combined with 20% PEG (40 and 41). In vapor diffusion experiments, the slow diffusion of water from the drop, where vapor pressure is higher, to the well, where vapor pressure is lower, drives a gradual increase in protein and precipitant concentration in the drop that may eventually lead to the crystallization of the protein. Volatile precipitants, on the other hand, tend to diffuse in the opposite direction, decreasing protein concentration. In the case of isopropanol this happens rapidly, making this precipitant perhaps better suited to batch experiments than vapor diffusion. Overall, the 9 conditions that contain organic precipitants crystallized 66 proteins, 10 of which grew only in these conditions.
Mining the Information to Identify Minimal Screens
The strong interdependence of various conditions of the Jancarik and Kim screen implies that it is, in its present formulation, less than ideal. Several conditions produced few crystals, while other groups of conditions, such as those based on high-molecular-weight PEGs, were too highly correlated. We set out to identify reduced condition sets that minimally compromise the chances of getting at least one crystal. Sequential searches of all combinations of conditions yielded a series of minimal screens (Table III) that comprised sets of conditions optimized for maximal probability of successfully obtaining a crystal [Fig. 1(c)]. A potential drawback of optimizing coverage at the expense of redundancy was that alternate crystal forms with different diffraction qualities might be missed; in those instances where a different crystal form would prove desirable, a second round of screening could then be initiated.
| Table III. Minimal Screens | ||||||||||||||||||||||||||||||
| ||||||||||||||||||||||||||||||
| These
screens are not only potentially useful in their own right (e.g., in
cases where material is limited) but also may potentially serve as the
nucleus of a new, more efficient screen. |
Six conditions - 6, 10, 18, 38, 39, and 43 (referred to hereafter as minimal screen 6) - yielded crystals for 205 of the 338 proteins (60.6%) successfully crystallized by the full screen. It is interesting to note the dispersion of these conditions - high-molecular-weight PEG at acid, neutral, and basic pH (10, 18, and 6), low-molecular-weight PEG (43), and two different salts (38 and 39). Augmenting this set with conditions 4, 17, 30, 36, 41, and 45 (minimal screen 12) yielded 268 of the 338 crystals (79.3%), and adding a further 12 conditions - 1, 11, 13, 14, 16, 20, 21, 28, 33, 34, 35, and 42 (minimal screen 24) - yielded 318 of the 338 crystals (94.1%). In addition to the conditions defined as minimal screen 24, the omission of conditions 22, 23, 27, 32, and 46 from the screen would have each resulted in 2 fewer proteins being crystallized, while the omission of conditions 8, 9, 12, 15, 25, 26, 37, 44, 47, and 48 would have each resulted in 1 fewer protein being crystallized. Nine conditions - 2, 3, 5, 7, 19, 24, 29, 31, and 40 - could have been omitted from the screen entirely without losing a single crystal from 755 samples.
Note that had the conditions for the minimal screens been chosen by considering only the total number of crystals produced per condition significantly less productive screens would have resulted. For example, the 6 individually most productive conditions crystallized 180 proteins compared to 205 for the optimal 6.
Screening data can also be mined for trends in precipitation. For example, the conditions employing tartrate and acetate salts as the primary precipitant were not only among the poorest crystal producers but also among those least likely to precipitate a protein. Because for the majority of proteins supersaturation was never reached with these precipitants, a substantially higher concentration may be predicted to be more effective.
Rational Strategy for Producing Maximally Productive Screens
Although the Jancarik and Kim screen has proved a useful tool for a generation of crystallographers, it is clear that a more efficient screen of a similar size could be derived from it by substituting some of its less productive conditions with ones chosen that demonstrably complement its more productive conditions. With sufficient data it is a relatively straightforward procedure to eliminate those conditions that contribute little, and iterative cycles of further additions, testing, and elimination should allow the eventual optimization of the screen. Analysis of the data generated may also help suggest suitable candidate conditions for expanding the screen. Conditions that show a strong tendency to uniquely crystallize proteins are likely in regions of parameter space that are undersampled and could therefore yield more crystals. For example, of the 62 proteins crystallized by condition 38, 11 are uniquely crystallized by this condition, implying that citrate salts have some unique properties whose potential for crystallization is underexploited by the present screen. Similarly, further conditions with transition metal ions (such as condition 45) and intermediate-molecular-weight PEGs (such as in condition 43) might also prove useful additions.
Employing such a strategy, one can
experiment with a wide variety of conditions without increasing the
amount of work to unmanageable proportions and without sacrificing the
proven productivity of a core set of conditions - important
considerations given that the only practical manner to obtain
sufficient samples and data to implement this strategy is to
incorporate it into ongoing structural proteomics efforts. Also,
because of the tentative nature of additions, this strategy should
encourage the exploration of chemically diverse conditions that might
otherwise not be thought sufficiently safe to include in a fixed, generally used screen.
In this study, a crystallization database was created and mined to find the most productive screening conditions, highlighting the value of such efforts within structural proteomics projects. Clearly, not only this sort of information can be extracted from the data, and there is more analysis to perform. For example, whereas this data set comprises the best conditions to identify initial crystals, most of these crystals have not been optimized to form diffraction-quality crystals. It will be interesting to learn whether there are differences in the suitability of different conditions as a starting point for producing large, well-diffracting crystals. The data could also be analyzed for solubility properties rather than crystallization. In this way, one may be able to identify a set of solution conditions that most often yield soluble protein - something potentially useful for NMR experiments, for example. Both solubility data and crystallization data should ultimately be linked to the biophysical properties of the proteins, such as the isoelectric point or amino acid content. Clearly, the crystallization databases resulting from structural proteomics projects can yield important information, and efforts should be expended to ensure that they are routinely collected in a consistent, machine-readable format.
| | | | | | |
Initial
crystallization conditions for a novel macromolecular sample are
obtained by screening the sample against a wide variety of chemical cocktails,
in general a generic set preselected for their historically proven
efficacy in producing crystals. Despite the critical nature of this
step, however, no systematic effort has been made to optimize the set
of conditions to be used. Here, we used data generated by subjecting a
large set of proteins against a commonly used, commercially available
screen to obtain a clearer picture of the overall efficacy of the
present screening strategies and see if there are obvious ways to
improve them. This data leads to several nontrivial conclusions: (1)
Among archaeal and bacterial genomes, there appear to be large
differences in the degree to which proteins are tractable to
crystallization; (2) a small subset of the conditions, even in the
relatively small Jancarik and Kim screen, are responsible for a large
proportion of crystals obtained overall; (3) as a corollary to this,
screening hundreds of conditions, as advocated in some screening
protocols, is little more likely to yield a crystal than searching a
few tens of well-chosen conditions. The results of this experiment
suggest that iteratively adding new conditions, testing against a large
set of proteins, and rejecting those conditions that contribute least
should allow the fine-tuning of existing screens while still having a
useful screen in place at all times. Ultimately this will be of great
benefit to structural proteomics efforts as an efficient, optimized
screen will give maximal samples for structure solution while
minimizing the amount of time and material wasted on unneeded
experiments.
| | | | | | |
This work was supported in part by the Ontario Research and Development Challenge Fund. A.M.E. and C.H.A. are CIHR Investigators. D.C. was supported by a fellowship from the Best Foundation.
| References |
| | | | | | |
| 1 | Vitkup D, Melamud E, Moult J, Sander C. Completeness in structural genomics. Nat Struct Biol 2001; 8: 559-566. Links |
| 2 | McPherson A. Crystallization of Biological Macromolecules. Cold Spring Harbor, ME: Cold Spring Harbor Laboratory Press; 1999. p. 586. |
| 3 | Jankarik J, Kim SH. Sparse matrix sampling: a screening method for crystallization of proteins. J Appl Crystallogr 1991; 24: 409-411. Links |
| 4 | Kingston R, Baker H, Baker E. Search designs for protein crystallization based on orthogonal arrays. Acta Crystallogr D 1994; 50: 429-440. Links |
| 5 | Saridakis E, Chayen N. Improving protein crystal quality by decoupling nucleation and growth in vapor diffusion. Protein Sci 2000; 9: 755-757. Links |
| 6 | Scott W, Finch J, Grenfell R, Fogg J, Smith T, Gait M, Klug A. Rapid crystallization of chemically synthesized hammerhead RNAs using a double screening procedure. J Mol Biol 1995; 250: 327-332. Links |
| 7 | Cudney B, Patel S, Weisgraber K, Newhouse Y, McPherson A. Screening and optimization strategies for macromolecular crystal growth. Acta Crystallogr D 1994; 50: 414-423. Links |
| 8 | Carter CJ, Carter C. Protein crystallization using incomplete factorial experiments. J Biol Chem 1979; 254: 12219-12223. Links |
| 9 | Gilliland G, Tung M, Blakeslee D, Ladner J. Biological Macromolecule Crystallization Database, Version 3.0: new features, data and the NASA archive for protein crystal growth data. Acta Crystallogr D 1994; 50: 408-413. Links |
| 10 | Christendat D, Yee A, Dharamsi A, Kluger Y, Savchenko A, Cort JR, Booth V, Mackereth CD, Saridakis V, Ekiel I, Kozlov G, Maxwell KL, Wu N, McIntosh LP, Gehring K, Kennedy MA, Davidson AR, Pai EF, Gerstein M, Edwards AM, Arrowsmith CH. Structural proteomics of an archaeon. Nat Struct Biol 2000; 7: 903-909. Links |
| 11 |
Wu N,
Christendat D,
Dharamsi A,
Pai EF.
Purification, crystallization and preliminary X-ray study of orotidine 5 -monophosphate decarboxylase.
Acta Crystallogr D Biol Crystallogr
2000;
56
(7):
912-914. Links |
| 12 | Zhang R, Skarina T, Katz J, Beasley S, Khachatryan A, Vyas S, Arrowsmith C, Clarke S, Edwards A, Joachimiak A, Savchenko A. Structure of Thermotoga maritima stationary phase survival protein SurE: a novel acid phosphatase. Structure 2001; 9: 1095-1106. Links |
| 13 | Jeanmougin F, Thompson J, Gouy M, Higgins D, Gibson T. Multiple sequence alignment with Clustal X. Trends Biochem Sci 1998; 23: 403-405. Links |
| 14 | Choi J, Jung H, Kim H, Cho H. PhyloDraw: a phylogenetic tree drawing system. Bioinformatics 2000; 16: 1056-1058. Links |
