Journal of Structural and Functional Genomics 6: 195–202, 2005.

Springer 2005
DOI 10.1007/s10969-005-5243-9

Automatic classification and pattern discovery in high-throughput protein

crystallization trials

Christian Cumbaa & Igor Jurisica*

Ontario Cancer Institute, 610 University Avenue, Toronto, Ontario, M5G 2M9, Canada; and Northeast
Structural Genomics Consortium. *Author for correspondence (e-mail: juris@ai.utoronto.ca)
Received 31 July 2004; accepted in revised form 04 May 2005

Key words: association-rule discovery, image analysis, protein crystallization

Abstract

Conceptually, protein crystallization can be divided into two phases search and optimization. Robotic
protein crystallization screening can speed up the search phase, and has a potential to increase process
quality. Automated image classification helps to increase throughput and consistently generate objective
results. Although the classification accuracy can always be improved, our image analysis system can classify
images from 1536-well plates with high classification accuracy (85%) and ROC score (0.87), as evaluated on
127 human-classified protein screens containing 5600 crystal images and 189472 non-crystal images. Data
mining can integrate results from high-throughput screens with information about crystallizing conditions,
intrinsic protein properties, and results from crystallization optimization. We apply association mining, a
data mining approach that identifies frequently occurring patterns among variables and their values. This
approach segregates proteins into groups based on how they react in a broad range of conditions, and
clusters cocktails to reflect their potential to achieve crystallization. These results may lead to crystallization
screen optimization, and reveal associations between protein properties and crystallization conditions. We
also postulate that past experience may lead us to the identification of initial conditions favorable to
crystallization for novel proteins.

Introduction cal and biological parameters), and the unknown

One of the fundamental challenges in modern
molecular biology is the elucidation and under-
standing of the rules by which proteins adopt
their three-dimensional structure. Currently, the
most powerful method for protein structure
determination is single crystal X-ray diffraction,
although new breakthroughs in NMR and in silico
approaches are growing in their importance.
A crystallography experiment begins with a
well-formed crystal that ideally diffracts X-rays
to high resolution. For proteins, this process is
often limited by the difficulty of growing crystals
suitable for diffraction. This is partially due to
the large number of parameters affecting the
crystallization outcome (e.g., purity of proteins,
intrinsic physico-chemical, biochemical, biophysi-
correlations between the variation of a parameter
and the propensity for a given macromolecule to
crystallize.
Conceptually, protein crystallization can be di-
vided into two phases: search and optimization.
Approximate crystallization conditions are identified
during the search phase, while the optimization
phase varies these conditions to ultimately yield
high quality crystals. There is yet no reliable
means of generating quality crystals from
arbitrary proteins, and thus automatic methods
are needed to conduct high-throughput protein
crystallization trials.
Robotic systems now routinely handle most
phases of these trials: cocktail preparation, cock-
tail and protein solution mixing in sub-microliter
volumes, and imaging of each experiment.
196
Typically, this pipeline ends with a human expert
responsible for manually assessing an image from
each experiment photograph for a positive
crystallization outcome. One of the current
challenges in the field is to replace the expert with
a high-accuracy automatic image-classification
system (e.g., Wilson, 2002; Bern et al., 2004).
Another challenge is to discover the factors that
effect crystallization of different proteins under
different chemical conditions.
We report here on parallel efforts on both
these fronts: (1) automated image classification
to determine, per plate, in what wells and at
what time points crystals form, and (2) pattern
discovery in the resulting matrix of experimental
outcomes, considering each protein, cocktail and
multiple time points, to identify what protein and
cocktail properties influence crystallization.
The main advantage of this approach is a sys-
tematic screening of many proteins under same
conditions (Luft et al., 2003). To date,
High-Throughput Screening (HTS) lab at the
Hauptman–Woodward Medical Research Institute
(HWI) screened over 4000 proteins, under 1961
conditions. Unlike other protein crystallization
databases, such as BMCD (Gilliland et al.,
1996), HWI data includes both successful and
unsuccessful crystallization trials. This created a
computational challenge of analyzing over
40 million images, but a great potential for data
mining.
Materials and methods
Data sources and preparation
We obtained image data from the HTS lab at the
Hauptman–Woodward Medical Research Insti-
tute (HWI), as time-stamped, per-plate archives
of 1536 images (one image per well for each time
point). Provided with each archive are the fol-
lowing records:

Protein concentration and sample preparation
date;

The chemical composition of the cocktail used
in each well;

Binary classification of each image (crystal/no
crystal) generated by human experts.
We assembled cocktail information from 111
plates into a table of 1961 distinct cocktails
involving 152 distinct chemical compounds. We
compiled the classification data into an incom-
plete 111 · 1961 matrix of experimental out-
comes (the precipitation index).
Members of the Northeast Structural Genom-
ics Consortium (NESG) provided the protein for
the trials. The proteins were prepared from pro-
teins produced in E. coli, using pET vectors with
HexaHis tags, and purified by NiNTA affinity
purification followed by gel filtration purification
(Acton et al., 2005). Where identifying information
was available, additional protein information was
obtained from the following sources:

NESG SPINE database (Goh et al., 2003):
organism, SWISS-PROT/TrEMBL reference,
amino acid sequence;

UniProt (Apweiler et al., 2004): EC number,
Interpro domains;

The Protein Calculator (Putnam, 1999): molec-
ular weight, estimated pI, estimated charge at
pH 7.0, estimated wavelength extinction coeffi-
cients for 278, 279, 280 and 282 nm.
These data for the 111 proteins were assem-
bled into a third table.
Image classification
Image classification uses extracted features to la-
bel the image with one of the outcome classes –
unknown, clear, precipitate, crystal (we intend to
further subdivide precipitates and crystals, and
allow for multiple classes to be used simulta-
neously). Different algorithms can be used to
implement individual stages. To improve perfor-
mance, it may also be useful to apply a boosting
approach by combining results from multiple
techniques. A variety of standard classification
methods have been used for classifying crystalli-
zation images: Fisher linear discriminants
(Cumbaa et al., 2003), self-organizing maps
(Spraggon et al., 2002), naive Bayes classifiers
(Wilson, 2002), decision trees (Bern et al., 2004;
Zhu et al., 2004), and support vector machines
(Zhu et al., 2004).
At present, we use Linear Discriminant Analy-
sis (LDA) (Fisher, 1936) for classification,
and we perform image analysis in four stages:

registration (locating the well within the image),
segmentation (locating the droplet within the
well), feature extraction (computing 59 measures
of droplet texture), and finally classification. These
are the same stages described in (Cumbaa et al.,
2003), with the following improvements:

Image registration. To reduce image registra-
tion error, we now factor robotic camera mo-
tion and plate-to-plate variations in apparent
well diameter (due to adjusted camera height)
into our image registration algorithm. As the
imaging robot moves along its programmed
path, slight miscalibrations in its positioning
system will introduce a consistent drift in (x, y)
coordinates of consecutive wells. Registration
proceeds as follows:
(a) A first approximation of the coordinates of
all 1536 wells is simultaneously computed
at multiple well-diameter values.
(b) The well diameter that minimizes the vari-
ance in measured (Dx, Dy) values is se-
lected.
(c) The computed coordinates for each well
are corrected by imposing a uniform (Dx,
Dy) between adjacent wells.

Additional image features. The 23 original im-
age features (of which 20 measure microcrystal
correlations, 2 measure straight-lines, and 1 mea-
sures image smoothness) have been augmented
by 36 additional features: 2 Euler-number
measures, 2 quadtree-decomposition statistics,
10 measures of local extrema, 19 additional
straight-line measures, and 3 additional image
energy measures. The combined 26 straight-line
and image energy features constitute a multi-
scale analysis of image texture missing from our
earlier work.
We have evaluated performance of the
augmented classifier using a set of 127 plates
containing a total of 5600 crystal images and
189472 non-crystal images. Computing image
feature vectors using unoptimized Matlab code
compiled to a Linux binary took an average of
12.27 s per image, or 5.23 h per plate, per CPU
on 3.0 GHz Xeon processors. We tested the
accuracy of our LDA classifier using 10-fold
cross-validation. We also tested k-nearest-neigh-
197
bor (KNN) classification accuracies for all values
of k between 1 and 50 neighbors.
Pattern discovery in the precipitation index
Our precipitation index database may be
regarded as a transaction database and is thus
subject to frequent-itemset and association-rule
discovery. Transactions in this case are sets of
facts about proteins, cocktails, or particular
experiments, e.g.:
{Cocktail #222, pH < 6.0, contains: PEG 20000,
CAPS, crystallizes proteins: glucose isomerase,
Q8U2K6 (N-type ATP pyrophosphatase)}.
An itemset is a subset of a transaction. Fre-
quent-itemset discovery (Goethals, 2003) finds all
itemsets in a database whose frequency, or sup-
port, exceeds some minimum. An association rule
describes the co-occurrence of two disjoint
itemsets (i.e., antecedent  consequent) in the
transactions in the database, such as:
{protein concentration >10 mg/ml, medium molecular
weight, low pI, organism: A. aeolicus, cocktail contains
CaCl2*2H2O}  {crystal}
This rule’s consequent is supported by 20 of
the 25 transactions supporting the antecedent in
our database, and thus we say the rule has
support 20 and confidence 0.8 (20/25). Association-
rule discovery (Agrawal et al., 1993) finds all
association rules in a database whose support
and confidence exceed minimum values.
We factored our database three different ways:
(1) As a set of 111 transactions on proteins, each
transaction containing protein properties and
the set of all cocktails that produced crystalli-
zation result. Using this database, we per-
formed frequent itemset mining (using
support ‡8) in order to discover clusters of
protein properties and crystallizing cocktails.
(2) As a set of 1961 transactions on cocktails,
each transaction containing a cocktail’s
chemical additives and the set of all protein
that crystallized with this cocktail. This
database was subjected to frequent itemset
mining (using support ‡8) in order to
discover clusters of cocktail properties and
crystallizing proteins.
(3) As a set of 168588 transactions on individual
experiments, each transaction containing
198

both protein and cocktail information, and
the crystallization outcome for a single
protein/cocktail pair. To this database, we
performed association-rule discovery (using
support ‡8 and confidence ‡0.45) to find pro-
tein and cocktail properties associated with a
positive crystal outcome.
Due to the requirements of the data mining
tools, we quantized the real-valued variables
(e.g., pH, chemical concentration) in our
database prior to itemset mining. Chemical
concentrations were reported only as present or
not present. All other real-valued variables were
split into equal-sized bins (low, medium, and
high) according to computed 33rd and 67th per-
centiles.
Results
We initially counted all crystal occurrences. It
is also important to test on how many plates the
system misses all existing crystals. We have
found that 2 out of 75 plates missed all crystals
(4 nice crystals and 8 murky ones), i.e., we have
missed 4 good crystal hits for 2 out of 75 pro-
teins. The misses can be partially explained by
the registration error, which the updated image
classification system diminishes.
Precipitation index mining
Frequent itemset mining produced 2455 sup-
ported itemsets on the protein transactions, and
31682 supported itemsets on the cocktail transac-
tions. Association-rule mining on experiment
transactions revealed 5487 supported association
rules predicting crystal outcome. Examples of all
three results are shown in Tables 1–3.

Image classification
Discussion
The following table compares the performance of
Image classification
Fisher LDA on our 59-feature dataset against
the original 23-feature dataset, introduced in
Reducing drop volumes to the sub-microliter,
(Cumbaa et al., 2003), and the KNN classifica-
increasing number of cocktails, changing optical
tion (for k=50):
setup for image capture and modifying optical
Precision, recall, and accuracy are based on
qualities of crystallization plates creates chal-
the classifiers’ confusion matrices:
lenges that require specialization and integration
where Precision = TP/(TP + FP), recall = TP/
of image analysis methods. Evaluating the aug-
(TP + FN), and accuracy = (TP + TN)/(TP
mented image classification system shows a
+ TN + FN + FP). The ROC scores measure
marked improvement in classification accuracy,
the area under the respective classifiers’ Receiver
by every measure. Precision is still low in abso-
Operating Characteristic curves (Figure 1).
lute terms, made so in part by the proportional
skew of the data towards non-crystal images. Fu-
Precision Recall Accuracy ROC score ture work will focus on finer-grained classification
of outcomes by adding new feature-extraction
23 features 0.10 0.62 0.83 0.80
algorithms.
59 features 0.12 0.68 0.85 0.85
59 features (KNN) 0.13 0.76 0.85 0.87
Most complex domains may suffer from a po-
tential disconnect between what experts perceive
as important features and what is objectively a
computationally viable image feature with high
59 features 59 features 23 features discriminatory potential. For that reason, it is
(KNN) useful to approach the problem in iterations,
taking multiple approaches to compute diverse
True False 4248 1352 3788 1812 3500 2100
features, collecting the data, and then thoroughly
positives negatives
evaluating which combination of features
False True 28881 160591 27139 162333 31557 157915
positives negatives
provides the best accuracy on testing results.
Importantly, specific features may discriminate
 199

0.9

0.8

0.7

False-positive rate
0.6

0.5

0.4
59 features (KNN)
0.3
59 features (LDA)
0.2 23 features (LDA)
0.1

0
0 0.2 0.4 0.6 0.8 1
False-negative rate
Figure 1. Receiver operating characteristic curves of the three classifiers.

Table 1. Sample frequent itemsets discovered in the database of protein transactions.

Pattern Support

Cocktail #1456 (HR Crystal Screen-1-20: Ammonium Sulfate 0.2 M, Sodium Acetate trihydrate 0.1 M, 8
pH 4.6, PEG 4000 25.0%)
Cocktail #1499 (HR Crystal Screen-2-13: Ammonium Sulfate 0.2 M, Sodium Acetate trihydrate 0.1 M,
pH 4.6, Polyethylene Glycol Monomethyl Ether 2000 30.0%)
Cocktail #1018(KBr 0.1 M, MES 0.1 M, pH 6.0, PEG 400 40.0%), 8
Cocktail #1024(KCl 0.1 M, MES 0.1 M, pH 6.0, PEG 400 40.0%)
Cocktail #1018, Cocktail #1042 (NaNO 30.1 M, MES 0.1 M, pH 6.0, PEG 400 40.0%) 8
low pI, Cocktail #1445 (HR Crystal Screen-1-9: Ammonium Acetate 0.2 M, Tri-sodium Citrate dihydrate) 10
Charge at pH 7.0£)7.8, Cocktail #1445 11
Bacterial protein source, Cocktail #1445 12
Bacterial protein source, low pI, Charge at pH 7.0£)7.8, Cocktail #1445 10
The first three illustrate patterns of co-crystallization. The next three associate cocktails with protein properties (isoelectric point,
charge at pH 7.0, source organism) in successful crystallization experiments..

Table 2. Sample frequent itemsets discovered in the database of cocktail transactions.

Pattern Support

Glucose isomerase, YB61_HAEIN (Hypothetical UPF0152 protein HI1161), Q8U2K6 (N-type ATP pyrophosphatase) 108
Glucose isomerase, Q8U189 (putative Transcriptional activator), Q8U2K6 125
Basic pH, PEG 4000, Q8U2K6 47
PEG 8000, glucose isomerase, Q8U189 39
Neutral pH, HEPES, glucose isomerase, Q8U2K6 48
Basic pH, TRIS, Q8U189, Q8U2K6 32
Acidic pH, PEG 4000, sodium acetate, glucose isomerase, YB61_HAEIN 8
Acidic pH, PEG 8000, sodium acetate, glucose isomerase, YB61_HAEIN 14
The first two illustrate clusters of co-crystallizing proteins. The third illustrates discovered crystallization trends of a single protein. The
rest are examples of co-crystallization trends of pairs of proteins.
200

Table 3. Sample antecedents of association rules predicting crystallization discovered in the experiment transactions.

Antecedent Support Crystal

confidence

Protein: Archaea, high WME282, unknown function; cocktail: PEG 4000, magnesium sulfate heptahydrate 10/10 1
Protein: medium MW, unknown function, P. furiosus; cocktail: PEG 8000, ammonium phosphate-dibasic 8/8 1
Protein: 12<mg/mL, Archaea, charge>)3.3, high WME282; cocktail: PEG 8000, TAPS 15/16 0.94
Protein: medium MW, high pI, P. furiosus; cocktail contains PEG 4000, TAPS 15/16 0.94
Protein: medium MW, charge >)3.3, P. furiosus; cocktail: PEG 4000, TAPS 15/16 0.94
Protein: medium MW, high pI; cocktail: PEG 8000, TAPS, P. furiosus 15/16 0.94
Protein: medium MW, charge>)3.3; cocktail: 8000, TAPS, P. furiosus 15/16 0.94
Protein: Archaea, medium MW, high pI; cocktail: PEG 4000, TAPS 15/16 0.94
Protein: Archaea, medium MW, high pI; cocktail: PEG 8000, TAPS 15/16 0.94
Protein: medium MW, low pI, )7.8<charge£)3.3; cocktail: acidic pH, MPD 10/22 0.45
Protein: mg/mL£10, medium pI, E. faecalis; cocktail: acidic pH, 20000, MES 11/20 0.55
Protein: Archaea, medium MW, high pI; cocktail: basic pH, lithium chloride 17/21 0.81
Protein: medium MW, low pI, )7.8< CHARGE £)3.3; cocktail: acidic pH, MPD, sodium cacodylate 9/15 0.6
Protein: medium pI, acidic pH, IPR011060; cocktail: MPD, sodium cacodylate 10/15 0.67
Protein: medium pI, IPR005627; cocktail: acidic pH, MPD 11/22 0.5
Protein: 12<mg/mL, medium pI, IPR005627; cocktail: acidic pH, MPD 11/22 0.5
Protein: acidic, IPR005627, IPR011060; cocktail: spermine tetra-hcl 8/9 0.89
Protein: IPR005627, IPR011060; cocktail: manganese sulfate monohydrate 9/18 0.5
Protein: IPR005627, IPR011060; cocktail: PEG 3350, bis-tris 9/13 0.69
Protein: acidic pH, IPR005627, IPR011060; cocktail: MPD, sodium cacodylate 10/15 0.67
Protein: high WME278, EC4.2.-.-; cocktail: PEG 1000, MES 15/19 0.79
Protein: Bacteria, medium pI, EC4.2.-.-; cocktail: PEG 1000, MES 15/19 0.79
Protein: low MW, low pI, EC3.5.-.-; cocktail: PEG 400, MES 11/19 0.58
Specific protein properties (source organism, EC numbers, Interpro domains, molecular weight, isoelectric point, wavelength
extinction coefficients, charge at pH 7.0, and concentration) combine with specific cocktail properties (pH, additives) to predict a
crystallization result. Confidence of the crystallization result is indicated in the right column; support for the rule is indicated in the
middle column.

only certain subclasses of all images, and thus a
combination of features is necessary.
In protein crystallization, the most important
images represent crystals. Protein crystals come
in diverse forms: micro-crystals, micro-needles,
needle crystals, fan-shaped crystals, and larger
faceted crystals. Although most of the crystals
exhibit straight edges in a drop, this important
feature to discern in general may not be
detectable in micro-crystals in a high-throughput
setup, due to insufficient drop size. To diminish
this problem, we need to augment a set of
image features specifically for detecting micro-
crystals. In addition, many protein crystallization
outcomes can occur simultaneously in a single
experiment: precipitation, crystal formation,
microcrystal formation, skin effect, and phase
separation. Thus, extending classification of
precipitates by characterizing image smoothness
can also contribute to improving classification
accuracy.
Precipitation index mining
Data mining can integrate results from
high-throughput screens with information about
crystallizing conditions, intrinsic protein proper-
ties, and results from crystallization optimization.
Association mining helps segregate proteins into
groups based on how they react in a broad range
of conditions, and group cocktails to reflect their
potential to achieve crystallization. These results
may lead to crystallization screen optimization,
and reveal associations between protein proper-
ties and crystallization conditions. Clustering can
be used on the precipitation index, to identify
cocktails or proteins with similar crystallization
potential.

Preliminary results suggest relationships that
group together certain proteins, protein
characteristics, or chemical properties based on
crystallization outcomes. These patterns and
others discovered by frequent itemset and associ-
ation-rule mining are numerous. A filtering step
is required in order to separate the few, interest-
ing rules from the many uninteresting rules
discovered by a basic association-rule-mining
algorithm. For example,

We already limit our association-rule miner to
finding rules with crystal as the consequent.
More generally, interesting rules adhere to a
[{independent variables}  dependent variables]
format

A rule A B is uninteresting if a more general
rule A¢  B (i.e., A¢  A) has a higher confi-
dence.

An itemset A=B [ C is uninteresting if A oc-
curs no more frequently than if B and C were
independent (i.e., P(A) £ P(B)P(C)).

Reported rules must satisfy some test of statis-
tical significance, e.g., a p-value.
Gopalakrishnan et al. (2004) discuss similar
challenges and describe a semi-manual process
for filtering their discovered rules for interesting-
ness and novelty.
Further, the quantization step should be
incorporated into the mining algorithm, i.e.,
quantitative association rule mining (Srikant and
Agrawal, 1996). Additionally, any reasoning
about crystallization conditions is limited by the
incomplete data in the database, e.g., buffers in
proteins, pH of cocktails, other characteristics of
proteins.
Case-based reasoning
We postulate that past experience will lead us to
the identification of initial conditions favorable
to crystallization. Moreover, it is believed that
solubility experiments can provide a quantitative
measure of similarity among proteins. If we con-
sider only 15 possible conditions, each having 15
possible values, the result would be 4.3789e +
017 possible experiments; impossible to test
exhaustively. Assuming that two proteins react
similarly when tested against a large set (over
1500) of precipitating agents in the search phase
201
of crystallization, then the planning strategies
successfully employed for the one may be profit-
ably applied to the other, while failed crystalliza-
tions can be avoided. Thus, it is important to
identify a suitable set of precipitating agents to
sort the outcomes of reactions for a relatively
large group of proteins. New crystallization chal-
lenges are then approached by the execution and
analysis of a set of precipitation reactions, fol-
lowed by an automated identification of similar
proteins and an analysis of the recipes used to
crystallize them, i.e., crystal growth method,
temperature and pH ranges, concentration of a
protein, and crystallization agent. Importantly,
this does not require a generic strategy that
would apply for all proteins – we treat each
protein individually using a KNN algorithm.
Further, a crystal result is not required either, as
different patterns of clear drops and precipitates
can be equally informative for the protein
similarity measure.
Acknowledgements
The authors would like to thank the Hauptman–
Woodward HTS lab for providing image data
and especially Angela Lauricella for her meticu-
lous hand-scoring of the images, George T. deT-
itta and Joe R. Luft for long term collaboration
and many fruitful discussions that lead to the
development of this system. We are grateful to
Thomas Acton, Rong Xiao, Gaetano Monteli-
one, and other members of the NESG Protein
Production team for providing the many protein
samples that have made this work possible. We
also thank Max Kotlyar for the use of his
association-mining software. This research was
supported by the Natural Science and Engineering
Research Council of Canada (#RGPIN 203833–
02), National Institutes of Health (#P50 GM-
62413), and IBM Shared University Research
grant.
References
1. Acton, T.B., Gunsalus, K., Xiao, R., Ma, L., Aramini, J.,
Baron, M.C., Chiang, Y., Clement, T., Cooper, B., De-
nissova, N., Douglas, S., Everett, J.K., Palacios, D.,
Paranji, R.H., Shastry, R., Wu, M., Ho, C.-H., Shih, L.,
Swapna, G.V.T., Wilson, M., Gerstein, M., Inouye, M.,
202
Hunt, J.F. and Montelione, G.T. (2005). Meth. Enzymol.
394, 210–243.
2. Agrawal, R., Imielinski, T. and Swami, A.N. (1993). In
Mining Association Rules between Sets of Items in Large
Databases. (Eds., Buneman, P., Jajodia, S.), Proceedings
of the 1993 ACM SIGMOD International Conference on
Management of Data, ACM Press, 207–216.
3. Apweiler, R., Bairoch, A., Wu, C.H., Barker, W.C., Bo-
eckmann, B., Ferro, S., Gasteiger, E., Huang, H., Lopez,
R., Magrane, M., Martin, M.J., Natale, D.A., O’Dono-
van, C., Redaschi, N. and Yeh, L.S. (2004) Nucleic Acids
Res. 32, D115–D119.
4. Bern, M., Goldberg, D., Stevens, R.C. and Kuhn, P.
(2004) J. Appl. Cryst. 37, 279–287.
5. Cumbaa, C., Lauricella, A.M., Fehrman, N.A., Veatch,
C.K., Collins, R.J., Luft, J.R., DeTitta, G.T. and Jurisica,
I. (2003) Acta Cryst., D 59, 1619–1627.
6. Fisher, R. (1936) Ann. Eugenics 7, 179–188.
7. Gilliland, G.L., Tung, M. and Ladner, J. (1996) J. Res.
Natl. Inst. Stand. Technol. 101(3), 309–320.
8. Goethals, B. (2003). Survey on Frequent Pattern Mining.
http://www.cs.helsinki.fi/u/goethals/publications/sur-
vey.pdf. Manuscript.
9. Goh, C.S., Lan, N., Echols, N., Douglas, S.M., Milburn,
D., Bertone, P., Xiao, R., Ma, L.C., Zheng, D., Wun-
derlich, Z., Acton, T., Montelione, G.T. and Gerstein, M.
(2003) Nucleic Acids Res. 31(11), 2833–2838.
10. Gopalakrishnan, V., Livingston, G., Hennessy, D.,
Buchanan, B. and Rosenberg, J.M. (2004) Acta Crystal-
logr. D. Biol. Crystallogr. 60, 1705–1716.
11. Luft, J.R., Collins, R.J., Fehrman, N.A., Lauricella, A.M.,
Veatch, C.K. and DeTitta, G.T. (2003) J. Struct. Biol.
142(1), 170–1799.
12. Putnam C. (1999). The Protein Calculator. http://
www.scripps.edu/cdputnam/protcalc.html. Scripps Re-
search Institute.
13. Spraggon, G., Lesley, S.A., Kreusch, A. and Priestle, J.P.
(2002) Acta Crystallogr. D Biol. Crystallogr. 58(11), 1915–
1923.
14. Srikant, R. and Agrawal, R. (1996). In Mining Quantita-
tive Association Rules in Large Relational Tables. (Eds.,
Jagadish, H.V., Mumick, I.S.). Proceedings of the 1996
ACM SIGMOD International Conference on Manage-
ment of Data, ACM Press, 1–12.
15. Wilson, J. (2002) Acta Cryst. D. 58, 1907–1914.
16. Zhu, X., Sun, S., Cheng, S.E. and Bern, M. (2004). Clas-
sification of Protein Crystallization Imagery. 26th Annual
International Conference of IEEE Engineering in Medi-
cine and Biology Society, IEEE Press.