Andrej ali is with the Laboratories of Molecular Biophysics,
The Rockefeller University, New York, New York
10021, USA. sali@rockefeller.edu
Structural genomics promises to deliver experimentally determined three-dimensional
structures for many thousands of protein domains. These domains will be carefully
selected, so that the methods of fold assignment and comparative protein structure
modeling will result in useful models for most other protein sequences. The
impact on biology will be dramatic.
Recently, some 200 structural biologists from both academia and industry
gathered in Avalon, New Jersey to explore the science and organization of
the new field of structural genomics1. The aim of the structural
genomics project is to deliver structural information about most proteins.
While it is not feasible to determine the structure of every protein by experiment,
useful models can be obtained by fold assignment and comparative modeling
for those protein sequences that are related to at least one known protein
structure. Structure determination of 10,000 properly chosen proteins should
result in useful three-dimensional models for hundreds of thousands of other
protein sequences. In other words, structural genomics will put each protein
within a comparative modeling distance of a known protein structure (Eaton
Lattman, Johns Hopkins University).
Knowing the structures of biological macromolecules is almost always useful
for predicting, interpreting, modifying and designing their functions. The
high-throughput, coordinated structure determination efforts should be more
efficient for obtaining 10,000 structures than the current distributed and
largely uncoordinated efforts in individual structural biology laboratories.
Structural genomics will build on the human genome project and will be integrated
with the functional genomics project. Here, I review the recent meeting in
Avalon. Due to space restrictions, the review omits many of the presentations.
However, a complete list of speakers may be found on the meeting web site1.
The structural genomics project will deliver improved methods and a process
for high-throughput protein structure determination. The process involves:
(i) selection of the target proteins or domains, (ii) cloning, expression,
and purification of the targets, (iii) crystallization and structure determination
by X-ray crystallography or by nuclear magnetic resonance (NMR) spectroscopy,
and (iv) archiving and annotation of the new structures. Several pilot projects
in structural genomics that implement all of these steps are already underway
(Table 1 http://structbio.nature.com/web_specials/nsb1298_supp/). It is feasible to determine at least 10,000 protein structures within
the next five years. No new discoveries or methods are needed for this task.
The average cost per structure is expected to decrease significantly from $200,000
per structure to perhaps less than $20,000 per structure. These savings will
be achieved by the economies of scale, new or improved methods, and their
parallelization in most of the steps in the process.
Table 1. List of pilot projects in structural genomics1
Target selection The targets for structure determination generally will be individual domains
rather than multi-domain proteins (Chris Sander, Millenium Information). Such
targets will be easier to determine by X-ray crystallography or NMR spectroscopy
than the more flexible multi-domain proteins, although it would be beneficial
to determine whole proteins whenever possible. Target selection will begin
by pairwise comparison of all protein sequences, followed by clustering into
groups of domains. The targets for the structural genomics project will be
the representatives of the groups without structurally defined members. The
groups of domains should contain members that share at least 30% sequence
identity (30-seq families), rather than those that correspond to superfamiles
or fold families.
This relatively high granularity is needed because of the limitations in
the current methods for comparative modeling and for inferring function from
structure. Errors in the comparative models become relatively small as the
sequence identity increases above 30% (Andrej ali, Rockefeller University)
and function is generally conserved within families (John Moult, Center for
Advanced Research in Biotechnology, Rockville, Maryland); Christine Orengo,
University College, London). Given the clustering into the 30-seq families,
the total number of targets for experimental structure determination has been
estimated to be on the order of 10,000. As an alternative, it has also been
suggested that 100,000 protein structures need to be determined by experiment
(David Eisenberg, UCLA); this would allow calculation of models with higher
accuracy than is possible with only 10,000 known structures. After a list
of the 30-seq families is obtained, the representatives will be picked and
prioritized. A variety of different criteria likely will be used for this
task; for example, size of the family, biological knowledge about the family,
distribution of the members among various organisms, the pharmaceutical relevance,
the likelihood of a successful structure determination, and so forth.
A preliminary analysis suggests that among recent protein structures determined
in the hope of discovering a new fold, more than two-thirds actually have
exhibited an already known fold (Steven Brenner, Stanford). While this unveils
weaknesses in the fold assignment methods, it is not a waste of resources.
In practical terms, the structure of a protein is worth determining if we
do not know it, for whatever reason. In fact, it may even be an advantage
when the newly determined protein is related to other known structures because
such structure is likely to be more informative about its function than in
the case of structural 'orphans'.
Given a fixed number of experimentally determined protein structures, the
quality and the number of models for the rest of the proteins will be influenced
strongly by the methods for fold assignment and comparative protein structure
modeling. The impact of expected improvements in these methods may be equivalent
to several years of the experimental protein structure output (Andrej ali).
A number of different approaches to fold assignment have been described, applied,
and evaluated (Ron Levy, Rutgers University; George Rose, John Hopkins University;
Barry Honig, Columbia University; Manfred Sippl, CAME, Salzburg; David Eisenberg;
Eugene Koonin, NCBI; Steven Brenner). The methods that rely on multiple sequence
information, such as PSI-BLAST, appear surprisingly sensitive in detecting
remote relationships (Liisa Holm, EMBL-EBI, Cambridge; Eugene Koonin; Steven
Brenner).
Cloning, expression and purification Obtaining protein samples for crystallization trials and NMR measurements
likely will be the bottleneck in the structural genomics process. Although
the current methods are sufficiently powerful to begin the project, advances
are expected in parallelization and automation of the existing methods for
cloning, expression, and purification of proteins. In addition, entirely new
technologies may arise, such as an in vitro system that already has
allowed researchers to produce proteins up to concentrations of 6 mg ml
−1 (Shigeyuki Yokoyama, Tokyo University).
Crystallization and X-ray crystallography Stephen Burley (Rockefeller University) described how to test protein samples
for their propensity to form useful crystals. Crystals have a higher chance
of being formed when the protein species in solution is conformationally homogeneous.
This can be evaluated by dynamic light scattering, as well as by limited proteolysis
combined with mass spectroscopy. For example, candidates that are monodisperse
under standard aqueous conditions have a probability of at least 75% to yield
crystals suitable for X-ray studies, in comparison to 10% for poly-disperse
samples.
Wayne Hendrickson (Columbia University) described several pivotal advances
that have matured only recently and form the basis of crystallographic analysis
at a markedly accelerated level. These include: undulator sources at synchrotrons,
CCD detectors, cryo-crystallography, MAD phasing analysis, and selenomethionyl
proteins2. It is now possible to sustain a data collection throughput
of four structures per day on a single undulator beam line. Assuming 180 operational
days per year and that only one out of three data sets results in a useful
final structure, this gives a conservative estimate of approximately 360 structures
per year per single undulator beam line. This means that only several dedicated
undulated beam lines are needed over the course of five years to achieve the
goal of determining 10,000 new protein structures.
Tom Terwilliger (LANL) has developed the SOLVE computer program for obtaining
electron density maps of proteins directly from multiwavelength or multiple
isomorphous replacement X-ray diffraction data. An optimization of a scoring
function replaces a complex rule-based process followed manually by crystallographers.
This system has already allowed crystallographers to obtain images of proteins
within hours of the end of X-ray data collection. The procedure has been successfully
applied to MAD data with up to 26 selenium sites per derivative.
Sung-Hou Kim (University of California, Berkeley) described results from
the first structural genomics pilot project, focusing on the Methanococcus
janaschii genome (Table 1http://www.nature.com/nsmb/index.html). He presented the structures of two proteins, an ATP dependent molecular
switch with a new ATP binding motif, and a nucleotide pyrophosphatase, with
a new fold. For each protein, the type of a ligand bound to it, and thus its
biochemical function, were predicted from the protein structure alone. The
predictions were subsequently confirmed by direct experiment. These results
support the idea that protein function can be anticipated from structural
information alone and thus that high-throughput structure determination will
make a valuable contribution to biology.
Nuclear magnetic resonance spectroscopy Kurt Wüthrich (ETH, Zürich), Gaetano Monte-lione and Gerhardt
Wagner (Harvard Medical School) described recent and prospective advances
in NMR spectroscopy that will allow structure determination of larger proteins
as well as increase the accuracy and speed of the analysis. These advances
include partial deuteration of protein samples, new magnets (up to 1 GHz),
superconducting probes, the TROSY detection method, triple resonance methods
for resonance assignment, spatial information from residual dipolar coupling,
and software for resonance assignments and structure determination. For example,
Montelione and colleagues have developed the program AUTOASSIGN to perform
automated backbone assignments in a few minutes of CPU time for proteins smaller
than 200 residues. This tool, integrated with the program AUTO_STRUCTURE for
automatic analysis of NOESY data and structure generation, decreased the time
needed for structure refinement of a Z-domain from several months to several
hours. If the NMR data collection for a 120 residue protein takes 4−5
weeks on a 600 MHz spectrometer and if automated data processing takes 1 day,
120 structures can be determined in 1 year by 12 machines (Montelione), approximately
at a cost of using one synchrotron beam line. By using higher field 800 MHz
to 1 GHz magnets and other sensitivity-enhancement methods, this throughput
could be significantly increased.
Anticipating the usefulness of NMR for structural genomics, Shigeyuki Yokoyama
of Tokyo University announced the Japanese Genomic Sciences Center that will
include a large NMR facility. The center will also support computational analysis
of the human genome, sequencing and mapping of the mouse genome, and collaborations
with high-throughput protein crystallography efforts elsewhere in Japan. The
center will have 20 NMR instruments by 2001 and will be open to international
collaboration. A 1 GHz magnet is being constructed in collaboration with the
Tsukuba laboratory, the engineers of the magnets used for the bullet train.
NMR spectroscopy has become increasingly more useful and efficient. However,
given the synchrotrons, X-ray crystallography is likely to retain an edge
in throughput, accuracy, maximal protein size, and possibly cost per protein
structure determination. Nevertheless, NMR spectroscopy will be indispensable
for maximizing the benefits of structural genomics because of its unique ability
to obtain information about the internal dynamics of a protein on multiple
time scales as well as about its ligand binding properties. For example, Kurt
Wüthrich described the importance of flexible tails for the function
of several proteins. Another example is the routine use of chemical shift
perturbation to enumerate binding sites on a protein for a thousand ligands
per day from a library of 15,000 compounds, with up to 1 mM sensitivity in
the dissociation constant (Steven Fesik, Abbott Laboratories). Also potentially
feasible is NMR structure determination of membrane protein structures, which
are exceptionally difficult to crystallize.
Databases Several different databases will be useful in facilitating the structural
genomics project (Table 2http://www.nature.com/nsmb/index.html). These databases will be used for target selection, coordination of
the experimental efforts, archiving of the new structures, dissemination of
the results, and for performing new research with the old and new data. Helen
Berman of Rutgers University described a new macromolecular structure database
that will replace the Brookhaven Protein Data Bank3. Liisa Holm,
Christine Orengo, and Steven Brenner described CATH, FSSP, and SCOP databases
respectively. These databases classify protein structures in a hierarchical
manner, generally consisting of families (proteins that share at least 30%
sequence identity), superfamilies (proteins that are probably related by divergent
evolution but can have very little sequence similarity), and fold families
(proteins that share similar folds). Orengo pointed out a great need for an
exhaustive, systematic and computer friendly database of protein functional
annotations. Such a database is necessary for the development and use of the
methods that correlate structure to function. Andrej ali described
ModBase, a large database of comparative protein structure models. This database
is calculated with program Modeller and currently contains protein models
for five genomes. It will continue to grow with the availability of new protein
structures and sequences, as well as the improvements in the modeling software.
Steven Brenner introduced PRESAGE, a database that stands to become the central
nervous system for coordinating the efforts in structural genomics. PRESAGE
includes entries for proteins; each entry contains information about the protein's
homologs, current status of structure determination, links to three-dimensional
models, and so forth. The entries in PRESAGE will be contributed by the community
at large. Biologists will use it as a portal for accessing the most up-to-date
information about proteins of interest. For example, PRESAGE would have eliminated
the duplicated structure determination of elongation factor 5A by two structural
genomics pilot projects (Sung-Hou Kim; Tom Terwilliger).
Using the structures Development of the databases and software will be necessary to integrate
raw structural information and structure-derived results with other information,
such as genetic and biochemical analysis, gene expression patterns, mutational
analysis, binding data from protein chips, interaction data from two-hybrid
systems, metabolic pathway information, and so forth (Chris Sander). These
tools will then allow researchers to address efficiently a host of new questions.
Some such studies, illustrating but the tip of the iceberg, have already been
performed and were described at the meeting. For example, questions about
the distribution of different structures among the genomes, and the implications
for models of evolution, have been asked by Mark Gerstein (Yale University)
and Eugene Koonin. In addition, David Eisenberg focused on understanding the
extreme thermostability of proteins in thermophilic organisms in terms of
the amino acid composition and structural features of the proteins. The study
was based on a comparison of proteins that have homologs in both thermophilic
and mesophilic organisms and which have homologs of known structure. It appears
that thermostability of proteins originates in part from salt bridges and
from deletions of loop regions.
Jeff Skolnick (Scripps Research Institute) described a method for predicting
protein function based on structure prediction and three-dimensional motif
identification. This method could push the detection of protein function much
further into the twilight zone of sequence identity. First, the tertiary structures
of the sequences of interest are predicted either ab initio or by threading.
Then, the resulting models are screened using a library of three-dimensional
descriptors of protein active sites, termed 'fuzzy functional forms'. If the
geometry and residue types in the predicted structures match a three-dimensional
motif, then the protein is predicted to have the specified molecular function.
The idea and its power relative to local sequence pattern searches were illustrated
by searching for proteins with a disulfide oxidoreductase active site in three
bacterial genomes.
Impact of structural genomics Suggestions about how best to proceed with the structural genomics project
were discussed. Ideas ranging from continuing 'research as usual' to establishing
large centers for structural genomics were explored. The pilot projects will
help shape the full scale effort. It is likely that relatively large centers
will provide the infrastructure for X-ray crystallography and NMR spectroscopy,
allow the economy of scale, foster collaboration, and also ensure that the
technical and repetitive jobs are not performed by students and postdocs in
the academic laboratories. Centers will make it possible for academic and
pharmaceutical researchers to obtain protein samples and research information,
such as purification protocols and crystallization conditions. Such an infrastructure
will enable the individual laboratories to focus on more challenging research
problems, including functional characterization of their favorite proteins,
structure determination of disease targets, large proteins and protein complexes,
study of post-translational modifications, and systemic questions involving
networks of proteins. Structural biologists in the pharmaceutical industry
will increase their impact on the drug discovery process because they will
be in a position to determine the structures of many more protein−ligand
complexes, at a lower cost and faster, using the structures, samples and protocols
from the high-throughput structure determination. The impact of structural
genomics on the structural and functional characterization of proteins, as
well as on our understanding of the machinery of life and its evolution, will
surely result in profusion of new drug targets and better drugs. Structural
genomics will refocus biology, just as the advent of of high throughput DNA
sequencing machines freed researchers to focus on more elaborate experiments
rather than on accumulating sequence data.
Structure-based functional genomics, October 4−7, Avalon, New Jersey (1998). http://lion.cabm.rutgers.edu/bioinformatics_meeting/. The meeting was organized by Gaetano Montelione, Stephen Anderson, Edward Arnold, and Ann Stock of the Center for Advanced Biotechnology and Medicine at Rutgers University (CABM), with the backing of the New Jersey Commission on Science and Technology, CABM, the Merck Genome Research Institute, and the Burroughs Wellcome Fund.
ali
$200,000
per structure to perhaps less than $20,000 per structure. These savings will
be achieved by the economies of scale, new or improved methods, and their
parallelization in most of the steps in the process.
