ArticlePDF Available

Carbon Sequestration in Synechococcus Sp.: From Molecular Machines to Hierarchical Modeling



The U.S. Department of Energy recently announced the first five grants for the Genomes to Life (GTL) Program. The goal of this program is to "achieve the most far-reaching of all biological goals: a fundamental, comprehensive, and systematic understanding of life." While more information about the program can be found at the GTL website (, this paper provides an overview of one of the five GTL projects funded, "Carbon Sequestration in Synechococcus Sp.: From Molecular Machines to Hierarchical Modeling." This project is a combined experimental and computational effort emphasizing developing, prototyping, and applying new computational tools and methods to elucidate the biochemical mechanisms of the carbon sequestration of Synechococcus Sp., an abundant marine cyanobacteria known to play an important role in the global carbon cycle. Understanding, predicting, and perhaps manipulating carbon fixation in the oceans has long been a major focus of biological oceanography and has more recently been of interest to a broader audience of scientists and policy makers. It is clear that the oceanic sinks and sources of CO(2) are important terms in the global environmental response to anthropogenic atmospheric inputs of CO(2) and that oceanic microorganisms play a key role in this response. However, the relationship between this global phenomenon and the biochemical mechanisms of carbon fixation in these microorganisms is poorly understood. The project includes five subprojects: an experimental investigation, three computational biology efforts, and a fifth which deals with addressing computational infrastructure challenges of relevance to this project and the Genomes to Life program as a whole. Our experimental effort is designed to provide biology and data to drive the computational efforts and includes significant investment in developing new experimental methods for uncovering protein partners, characterizing protein complexes, identifying new binding domains. We will also develop and apply new data measurement and statistical methods for analyzing microarray experiments. Our computational efforts include coupling molecular simulation methods with knowledge discovery from diverse biological data sets for high-throughput discovery and characterization of protein-protein complexes and developing a set of novel capabilities for inference of regulatory pathways in microbial genomes across multiple sources of information through the integration of computational and experimental technologies. These capabilities will be applied to Synechococcus regulatory pathways to characterize their interaction map and identify component proteins in these pathways. We will also investigate methods for combining experimental and computational results with visualization and natural language tools to accelerate discovery of regulatory pathways. Furthermore, given that the ultimate goal of this effort is to develop a systems-level of understanding of how the Synechococcus genome affects carbon fixation at the global scale, we will develop and apply a set of tools for capturing the carbon fixation behavior of complex of Synechococcus at different levels of resolution. Finally, because the explosion of data being produced by high-throughput experiments requires data analysis and models which are more computationally complex, more heterogeneous, and require coupling to ever increasing amounts of experimentally obtained data in varying formats, we have also established a companion computational infrastructure to support this effort as well as the Genomes to Life program as a whole.
OMICS A Journal of Integrative Biology
Volume 6, Number 4, 2002
© Mary Ann Liebert, Inc.
Carbon Sequestration in Synechococcus Sp.:
From Molecular Machines to Hierarchical Modeling
and DONG XU3
The U.S. Department of Energy recently announced the first five grants for the Genomes to
Life (GTL) Program. The goal of this program is to achieve the most far-reaching of all
biological goals: a fundamental, comprehensive, and systematic understanding of life.” While
more information about the program can be found at the GTL website (www.doegenomesto-, this paper provides an overview of one of the five GTL projects funded, “Carbon
Sequestration in Synechococcus Sp.: From Molecular Machines to Hierarchical Modeling.”
This project is a combined experimental and computational effort emphasizing developing,
prototyping, and applying new computational tools and methods to ellucidate the biochem-
ical mechanisms of the carbon sequestration of Synechococcus Sp., an abundant marine
cyanobacteria known to play an important role in the global carbon cycle. Understanding,
predicting, and perhaps manipulating carbon fixation in the oceans has long been a major
focus of biological oceanography and has more recently been of interest to a broader audi-
ence of scientists and policy makers. It is clear that the oceanic sinks and sources of CO2
are important terms in the global environmental response to anthropogenic atmospheric in-
puts of CO2and that oceanic microorganisms play a key role in this response. However, the
relationship between this global phenomenon and the biochemical mechanisms of carbon
fixation in these microorganisms is poorly understood. The project includes five subprojects:
1Sandia National Laboratories, Albuquerque, New Mexico.
2Sandia National Laboratories, Livermore, California.
3Oak Ridge National Laboratory, Oak Ridge, Tennessee.
4Lawrence Berkeley National Laboratory, Berkeley, California.
5Los Alamos National Laboratory, Los Alamos, New Mexico.
6University of California, San Diego, California.
7University of Illinois, Urbana/Champaign, Illinois.
8University of Michigan, Ann Arbor, Michigan.
an experimental investigation, three computational biology efforts, and a fifth which deals
with addressing computational infrastructure challenges of relevance to this project and the
Genomes to Life program as a whole. Our experimental effort is designed to provide biol-
ogy and data to drive the computational efforts and includes significant investment in de-
veloping new experimental methods for uncovering protein partners, characterizing protein
complexes, identifying new binding domains. We will also develop and apply new data mea-
surement and statistical methods for analyzing microarray experiments. Our computational
efforts include coupling molecular simulation methods with knowledge discovery from di-
verse biological data sets for high-throughput discovery and characterization of protein-
protein complexes and developing a set of novel capabilities for inference of regulatory path-
ways in microbial genomes across multiple sources of information through the integration
of computational and experimental technologies. These capabilities will be applied to Syne-
chococcus regulatory pathways to characterize their interaction map and identify compo-
nent proteins in these pathways. We will also investigate methods for combining experi-
mental and computational results with visualization and natural language tools to accelerate
discovery of regulatory pathways. Furthermore, given that the ultimate goal of this effort is
to develop a systems-level of understanding of how the Synechococcus genome affects car-
bon fixation at the global scale, we will develop and apply a set of tools for capturing the
carbon fixation behavior of complex of Synechococcus at different levels of resolution. Fi-
nally, because the explosion of data being produced by high-throughput experiments re-
quires data analysis and models which are more computationally complex, more heteroge-
neous, and require coupling to ever increasing amounts of experimentally obtained data in
varying formats, we have also established a companion computational infrastructure to sup-
port this effort as well as the Genomes to Life program as a whole.
AS DESCR IBED IN TH E Genomes to Life (GTL) Program literature (, the goal
of this program is to “achieve the most far-reaching of all biological goals: a fundamental, compre-
hensive, and systematic understanding of life”:
DOE’s Genomes to Life program will make important contributions in the quest to venture beyond
characterizing such individual life components as genes and other DNA sequences toward a more
comprehensive, integrated view of biology at a whole-systems level. The DOE offices of Biological
and Environmental Research and Advanced Scientific Computing Research have formed a strategic
alliance to meet this grand challenge.
The plan for the 10-year program is to use DNA sequences from microbes and higher organisms,
including humans, as starting points for systematically tackling questions about the essential processes
of living systems. Advanced technological and computational resources will help to identify and un-
derstand the underlying mechanisms that enable organisms to develop, survive, carry out their nor-
mal functions, and reproduce under myriad environmental conditions.
This approach ultimately will foster an integrated and predictive understanding of biological sys-
tems and offer insights into how both microbial and human cells respond to environmental changes.
The applications of this level of knowledge will be extraordinary and will help DOE fulfill its broad
missions in energy, environmental remediation, and the protection of human health.
The Genomes to Life program has four stated goals:
1. Identify and characterize the molecular machines of life—the multiprotein complexes that execute cel-
lular functions and govern cell form
2. Characterize gene regulatory networks
3. Characterize the functional repertoire of complex microbial communities in their natural enviroments at
the molecular level
4. Develop the computational methods and capabilities to advance understanding of complex biological
systems and predict their behavior
As stated above, the effort discussed in this paper is focused on understanding the carbon sequestration
behavior of Synechococcus Sp. through experimental and computational methods with the major effort in
the development of computational methods and capabilities for application to Synechococcus. Thus, this
project can be thought of as a “Goal 4” project. To this end, the work has been divided into five subpro-
1. Experimental Elucidation of Molecular Machines and Regulatory Networks in Synechococcus Sp.
2. Computational Discovery and Functional Characterization of Synechococcus Sp. Molecular Machines
3. Computational Methods Towards the Genome-Scale Characterization of Synechococcus Sp. Regulatory
4. Systems Biology for Synechococcus Sp.
5. Computational Biology Work Environments and Infrastructure
These five subprojects are discussed individually in the sections that follow, 1.0 through 5.0, respectively.
The computational work in this proposal is captured in sections 2.0, 3.0, 4.0, and 5.0, while the experi-
mental biology, including experimental methods development, required to integrate and drive the compu-
tational methods development and application are discussed in 1.0. “Computational Discovery and Func-
tional Characterization of Synechococcus Sp. Molecular Machines,” discussed in section 2.0, is aimed
directly at the GTL Goal 1 in the context of Synechococcus and includes primarily computational molecu-
lar biophysics and biochemistry as well as bioinformatics. “Computational Methods Towards the Genome-
Scale Characterization of Synechococcus Sp. Regulatory Pathways,” section 3.0, is also highly computa-
tional, focused primarily on the development and application of bioinformatics and data mining methods
to elucidate and understand the regulatory networks of Synechococcus Sp. (GTL Goal 2).
In section 4.0, we discuss our planned efforts to integrate the efforts discussed in sections 1.0, 2.0, and
3.0 to enable a systems biology understanding of Synechococcus. This work will support GTL Goals 1 and
2 and is focused on developing the computational methods and capabilitities to advance understanding of
Synechococcus as a complex biological system. Given the available information and data on Synechococ-
cus, the effort discussed in section 4.0 will initially (year 1) employ other microbial data in order to ad-
vance the state of the art of computational systems biology for microorganisms. This will give our Syne-
chococcus experimental effort (section 1.0) time to ramp up and produce the data needed to drive this effort
in project years 2, and 3 (FY04 and FY05).
In section 5.0, “Computational Biology Work Environments and Infrastructure,” we discuss a number of
developments to enable the support of high-throughput experimental biology and systems biology for Syne-
chococcus including work environments and problem solving environments, as well as high performance
computational resources to support the data and modeling needs of GTL researchers.
Understanding, predicting, and perhaps manipulating carbon fixation in the oceans has long been a ma-
jor focus of biological oceanography and has more recently been of interest to a broader audience of sci-
entists and policy makers. It is clear that the oceanic sinks and sources of CO2are important terms in en-
vironmental response to anthropogenic inputs of CO2into the atmosphere and in global carbon modeling.
However, the actual biochemical mechanisms of carbon fixation and their genomic basis are poorly under-
stood for these organisms as is their relationship to important macroscopic phenomena. For example, we
still do not know what limits carbon fixation in many areas of the oceans. Linking an organism’s physiol-
ogy to its genetics is essential to understand the macroscopic implications of an organism’s genome (i.e.,
linking “genomes to life”).
The availability of Synechococcus’ complete genome allows such an effort to proceed for this organism.
Thus the major biological objective of this work is to elucidate the relationship of the Synechococcus’
genome to Synechococcus’ relevance to global carbon fixation through careful studies at various length
scales and levels of complexity. To this end, we will develop a fundamental understanding of binding, pro-
tein complexes, and protein expression regulation in order to provide a complete view of the protein bind-
ing domains that mediate the most relevant protein–protein interactions and their related regulatory net-
works. In addition, we will investigate molecular machines and regulatory networks within the
Synechococcus cell. Specifically, we will develop a fundamental understanding of the protein binding do-
mains that mediate protein–protein interactions and form the basis of the Synechococcus molecular ma-
chines most relevant to carbon fixation. In addition, we will investigate Synechococcus’ regulatory network
and study a few mo-lecular machine complexes in detail. Our goal will be to elucidate the fundamental in-
formation regarding binding, protein complexes, and protein expression regulation to enable a systems-level
understanding of carbon fixation in Synechococcus.
The subproject will focus on three protein families: the carboxysome, ATP-binding cassette (ABC) trans-
porters, and histidine kinase-response regulators. The carboxysome is a complex of proteins that converts
inorganic carbon (CO2) to sugars via the photosynthetic carbon reduction chain. It is the primary organelle
involved in carbon fixation, yet its structure and function are not well understood. ABC transporters and
histidine kinase-response regulators appear to be intricately linked. The kinase-regulator signal transduc-
tion system detects environmental changes and regulates cellular response by inducing appropriate genes.
Cellular response often includes the influx and efflux of carbon and nutrients; thus, the kinase-regulators
affect ABC transporters, proteins involved in active transport. The mechanisms by which these two sys-
tems are co-regulated are not understood.
In the text that follows, we outline some of the important characteristics and questions regarding car-
boxysomes, ABC transporters, and histidine kinase-response regulators, and discuss how we will analyze
these interdependent protein complexes using mass spectrometry, display techniques, and gene microarray
analysis. Finally, we discuss our plans for employing our experimental program to integrate and prototype
the computational capabilities discussed in the sections that follow.
Carboxysomes and inorganic carbon fixation
Carboxysomes are polyhedral-shaped inclusion bodies approximately 100 nm in diameter and present in
cyanobacteria and many chemoautotrophs (Cannon et al., 2001). They consist of a protein shell and are
filled with the enzyme ribulose 1,5 bisphosphate carboxylase/oxygenase (RuBisCO). Carboxysomes main-
tain two known functions. They act in a carbon-concentrating mechanism (CCM) to increase the concen-
tration of CO2from the relatively low value of 20 mM in seawater, and they convert CO2to sugars. In fact,
RuBisCO catalyzes the photosynthetic conversion.
The functions of the carboxysome are dependent on the protein composition and structure of the com-
plex. In H. neapolitanus,SDS-PAGE analysis indicates the carboxysome consists of nine to fifteen polypep-
tides of which at least six are associated with the shell (Cannon et al., 1983; Holthuijzen et al., 1986). In
the freshwater strain Synechococcus PCC7942, up to 30 different polypeptides are evident, but of the 30
only 10 have similar molecular weights as those found in other organisms (Price et al., 1992). It is likely
that many polypeptides represent contaminants, as protein purifications are difficult in cyanobacterium due
to high membrane content. Genetic analysis has provided some information as to the likely nature of the
shell composition. In H. neopolitanus,csoS1 encodes a shell protein that has been duplicated twice (csoS1B
and csoS1C) (English et al., 1994) and is highly homologous to the Synechococcus PCC7942 genes ccmK
and ccmO (Shively et al., 1996). CsoS1 and ccmK and -O genes exist in cso (carboxysome) and ccm
(carbon-concentrating mechanisms) gene clusters that, in their respective organisms, are proximal to the
genes that code for RuBisCO large and small subunits. Functions for most of the genes within the clusters,
though, remain unknown.
The protein composition of carboxysomes is especially important in their role in the CCM. The CCM
provides concentrated CO2, the substrate for RuBisCO, through dehydration of HCO3
2, a relatively high
concentration (2 mM) source of inorganic carbon in seawater. HCO3
2is converted to CO2by the enzy-
matic activity of carbonic anhydrase (CA). It is not clear how high a concentration CO2is made available
to RuBisCO. It is possible that CA is present in the carboxysome. It has been suggested that CA is se-
questered with RuBisCO in the carboxysome in some systems (Fridlyand et al., 1996; Kaplan et al., 1989),
but not in others (Bedu et al., 1992; Lanaras et al., 1985). To date, a direct biochemical link between CA
and the carboxysome in the CCM has not been established. Additionally, CA genes have not been found
to map to carboxysome genome regions.
A number of questions with regard to carboxysome structure and function remain open. The number of
proteins, the composition of the shell, and the localization of proteins encoded by functionally important
gene clusters are not known. It appears that more proteins are present than can be accounted for by Ru-
BisCO and the shell, and it is possible that CA is present acting independently or in synergy with other
components. Finally, molecular interactions between protein components are not known.
The ABC transporter and histidine kinase-response regulator systems
ABC transporters are a superfamily of proteins that transport a wide variety of solutes including amino
acids, ions, sugars, and polysaccharides. These transporters have four domains, two hydrophobic integral mem-
brane protein domains and two hydrophilic ATP-binding domains that are thought to couple ATP hydrolysis
to transport (Chang et al., 2001). The domains can be found as separate proteins or as a single fused protein.
In bacterial systems involved in solute uptake, there is a separate solute binding protein that binds the solute
being transported, docks with the other components, and allows the solute to diffuse through a channel before
disassociating. Clearly, ABC transporters are sophisticated protein machines that carefully recognize particu-
lar compounds and move them into and out of the cell. In Synechococcus WH8102, there are about 80 genes
that encode for ABC transporters, including about eighteen specific to substrate-binding proteins.
The regulation of transport is a complex multi-level process. The two-component histidine kinase-
response regulator signal transduction system appears to play a critical role (Hoch et al., 1995) in this
process. Histidine kinase proteins or protein domains sense some property and activate or repress through
phosphorylation a second protein called a response regulator. The response regulator acts as a transcription
factor and regulates gene induction levels. For instance, sensing increased phosphate levels could lead
through the two-component system to increased induction of genes that code for affinity phosphate trans-
porter or binding proteins. In fact, two component regulatory systems have been linked to phosphate trans-
port, nitrogen transport, and porin regulation (Ninfa et al., 1995; Pratt et al., 1995). In Synechococcus
WH8102, there are six histidine kinases and nine response regulators, and a number of these genes are lo-
cated physically adjacent to transporters. The genetic information supports biochemical links between reg-
ulatory and transport systems.
There are a number of questions pertinent to the level of interaction between the signal transduction compo-
nents and ABC transporters. It is interesting to consider what histidine kinases and response regulators affect
specific ABC transporters and what is the level of regulation in transporter function. Also to be considered is
whether multiple transporter functions are effected concurrently and whether there is cross talk between signals.
Protein complex analysis by mass spectrometry
Protein complex compositions and structures, particularly in the carboxysome, will be explored using a
couple of techniques that make use of protein mass spectrometry. Novel methods to analyze proteome-scale
protein complexes use an affinity purification technique combined with protein identification by mass spec-
trometry (Gavin et al, 2002; Ho et al., 2002). Cassettes containing tags (poly-His or Protein A) will be in-
serted at the 39end of the genes encoding proteins central to the carboxysome in Synechococcus. After
clonal selection, cultures will be grown under various conditions of light flux, CO2levels, and nutrient con-
centrations. Cellular lysates will be subjected to tandem affinity purification utilizing low-pressure columns
in order to fish out” bait proteins and proteins complexed to them (Puig et al., 2001). Proteins will be
eluted off the column and separated by SDS-PAGE, and individual protein bands will be excised, digested,
and subjected to mass spectrometer analysis. We will conduct protein mass spectrometry with a Bruker
Apex FTICR instrument with an electrospray ionization source and a Thermoquest LCQ ion trap interfaced
with micro-HPLC. This strategy has the advantages that it is adaptable to high-throughput techniques and
does not require the use of antibodies for protein identification (a situation where a priori information as
to the identity of the protein is needed).
Mass spectrometry will also be used to isolate protein-protein interactions. Protein complexes will be col-
lected as before with tandem affinity purification. In a new step, the protein complexes bound to the col-
umn will be chemically crosslinked with amine- or sulfhydryl-specific crosslinkers. The chemically cross-
linked complexes will be digested with trypsin, and the peptides separated by capillary reversed-phase HPLC
and analyzed by FTICR-MS. Comparing crosslinked data sets to individual protein analysis provides shifts
in mass spectra that can be compared to theoretical distance constraints to calculate three dimensional pro-
tein arrangements and likely protein–protein interactions.
Protein complexes and networks by phage display
In addition to mass spectrometry, protein complexes can be characterized by using phage display tech-
niques to identify protein-binding domains. Phage display techniques refer to methods by which random
samplings of peptide fragments are presented on the surface of viral particles (phage) and are screened for
peptide-protein interactions with a probe protein (Smith et al., 1997). A library of degenerate oligonucleotide
inserts is cloned into the coding sequence of phage coat proteins so that each viral-containing clone dis-
plays a peptide on its surface corresponding to a specific sequence within the library. A probe protein is
fixed to any number of possible substrates, and peptide-protein interactions are elucidated by mixing
library-containing clones with the probe, selecting and amplifying positives, and repeating the process 2–3
times in a process called panning. Advantages of phage display include the ability to immediately isolate
molecular recognition domains and ligand sequences, the ability to construct different libraries to study par-
ticular problems, and the ability to screen up to 1010 different peptides at once.
We will use phage display using key carboxysome constituents as probe proteins. Libraries of random
peptides and libraries based on gene clusters known to be associated with carboxysomes will provide two
techniques to establish protein-peptide interactions. In the case of random peptide libraries, the genome will
be searched for regions that code for positives indicating potentially new protein-protein interaction pairs.
Peptide libraries developed from carboxysome gene clusters represent targeted efforts. Cognate binding
partners will be verified by enzyme-linked immunosorbent assay (ELISA) or by yeast 2-hybrid screening.
In addition to protein complex characterization, phage display is a powerful technique that has been used
to predict entire protein networks (Tong et al., 2002). As protein binding domains mediate protein-protein
interactions, characterizing binding domains can be used to infer networks by developing lists of probable
protein interactions. Similar strategies will be employed here. Binding domains are defined in families ac-
cording to similarities in sequence, structure, and binding ligands. Each member of a family binds a simi-
lar sequence, but binding for any particular family member is specified by variances within the core bind-
ing domain. Using phage display, we will study known prokaryotic binding domain families in an effort to
develop recognition rules. Protein binding domains known to exist in Synechococcus include SH3, PDZ,
CBS, and TPR domains (Falzone et al., 1994; Bateman et al., 2002).
Gene and protein regulatory networks by microarray analysis
As discussed, the WH8102 genome has six histidine kinases and nine response regulators that are major
components in its regulatory network. They are thought to control the sensing of the major nutrients of
phosphate, nitrogen, and light. The immediate objective of our experimental investigations is to define the
web of interactions controlled by these components with the ultimate goal to developing a systems-level
understanding of Synechococcus. Our efforts will focus on characterizing the regulation of the transport
genes, two component systems, and some stress-associated genes using a DNA microarray containing the
entire genome. We will use a gene knockout strategy whereby all histidine kinases and many of the re-
sponse regulators will be inactivated in turn, and the effect of the gene knockouts on the induction of ABC
transporters will be measured as a function of nutrient and light levels. Microarray analysis will be com-
pleted using state-of-the-art hyperspectral scanning. Using bioinformatic analyses to characterize the up-
stream regions of the genes regulated by a particular stress, common cis-regulatory sites potentially used
by the response regulators will be predicted. Then, the complete genome can be searched for other puta-
tive sites with these sequences (see below). In all cases, binding to DNA will be verified experimentally.
One of the advantages of Synechococcus as a model system is that bioinformatic analyses can incorpo-
rate the data for the complete genomes of the related cyanobacteria Prochlorococcus in both the sequence
definition phase and the sequence scanning phase. For example, if a sequence motif is found upstream of
a gene in all three genomes during genome scanning, a logical prediction is that these genes are regulated
in similar ways.
ABC transporter function and regulation will also be analyzed by studying protein expression levels. We
will follow the predicted eighteen substrate binding proteins involved in the ABC transporter system using
polyclonal antibodies to each protein. Antibodies will be produced in chickens and rabbits from proteins
purified from E. coli expression of PCR amplified gene segments. Protein expression will be analyzed by
SDS-PAGE and Western blotting. Eventually, protein arrays will be developed. Whole cells will be labeled
with an amine reactive fluorophore. The fluorophore migrates into the periplasmic space (but not inside the
cell) and will label the lysines of proteins in this region and the cell surface. We will then solubilize the
cells and incubate the labeled proteins with the array of antibodies and scan the membrane to quantity
the amount of each protein. As the signal should be proportional to the number of accessible lysine-residues
on each protein, we will carefully calibrate the signal from each protein against known amounts of protein
obtained from the expressed protein in E. coli.
The experimental design with respect to computational biology
The comprehensive experimental program is designed to integrate and prototype a number of computa-
tional capabilities with the goal of developing iterative and synergistic efforts. Whole protein complex char-
acterization supports Rosetta-like efforts to analyze protein structure, and breaking the complexes down
into the fundamental binding units supports molecular biophysical calculations. Furthermore, experimental
and theoretical results describing binding domains and recognition rules are critical in developing predicted
protein interaction maps. The analysis of gene and protein regulatory networks supports a number of ef-
forts from the acquisition, storage, and analysis of microarray data to the development of supervised and
unsupervised clustering techniques to predict cis-regulatory sites, regulatory networks, and protein func-
tion. These and other computational efforts are outlined below.
Our computational discovery and characterization of molecular machines effort is designed to develop
high-performance computational tools for high-throughput discovery and characterization of protein
protein complexes. This will be achieved primarily through coupling of molecular simulation methods with
knowledge discovery from diverse biological data sets which can then be applied, in conjunction with
experimental data, to the Synechococcus proteome to enable discovery and functional annotation of protein
complexes. Our efforts will be pursued through several synergetic approaches: low-resolution high-through-
put Rosetta-type algorithms, high performance all-atom modeling tools, and knowledge-based algorithms
for functional characterization and prediction of the recognition motifs. These are discussed individually in
the text that follows.
Rosetta-type algorithms
There are currently no highly reliable tools for modeling of protein-protein complexes. Building upon
proven methods for ab initio protein modeling, we will develop and apply Rosetta-like algorithms for fast
characterization of protein-protein complexes complexes with two approaches: (1) for cases where struc-
tures of unbound members are known, the Rosetta potential will be used to dock them together while per-
mitting conformational changes of the components, and (2) if experimental data are available, sparse con-
straints will be incorporated (from NMR and mass-spectroscopy experiments). Both approaches will help
achieve the goal of developing high-throughput methods of characterizing protein–protein complexes.
All-atom simulations
Our existing parallel codes for biomolecular-scale modeling will be extended as necessary to model
protein–protein complexes in Synechococcus. All-atom simulations will be initially focused on two prob-
lems: (1) interpretation of the phage display data and (2) investigation of the functional properties of Syne-
chococcus membrane transporters. The computational algorithms and software developed in this effort will
be applicable broadly to molecular machines in other organisms to provide understanding of protein inter-
actions in general.
Knowledge fusion algorithms
Because existing data mining algorithms for identification and characterization of protein complexes are
not sufficiently accurate, nor do they scale well for genome-wide studies, we will extend or develop new
algorithms to improve predictive strength and allow new types of predictions to be made. Our approach
will involve (1) developing “knowledge fusion” algorithms that combine many sources of experimental,
genomic and structural information, (2) coupling these algorithms with modeling and simulation methods,
(3) implementing high performance, optimized versions of our algorithms. Specifically algorithms for three
interrelated problems will be investigated: (1) identification of pair-wise protein interactions, (2) construc-
tion of protein-protein interaction maps, and (3) functional characterization of the identified complexes.
Together these three elements yield a synergistic computational approach to discovering and character-
izing molecular machines. Thus, for example, protein pair identification tools will be used to provide the
initial sets of putative pairs of interacting proteins, either by filtering our experimental data or bioinfor-
matics leads (from efforts described in section 3.0) for specific metabolic subsystems of Synechococcus.
This initial set of targets and the available experimental constraints will be investigated further through the
use of the Rosetta-like algorithms and all-atom methods. The resulting information will then be used to re-
fine the knowledge fusion algorithms as well as applied for the functional characterization of the verified
protein assemblies.
Genome-scale techniques for measuring, detecting, mining, and simulating protein-protein interactions
will be critical for transforming the wealth of information currently being generated about individual gene
products into a comprehensive understanding of the complex processes underlying cell physiology. Current
approaches for accomplishing this formidable task include direct experimentation, genome mining, and com-
putational modeling. This effort will exploit all three approaches. Below we briefly discuss the current state-
of-art and existing limitations of these approaches.
The leading experimental genome-wide high-throughput methods for characterizing protein-protein in-
terations include the two-hybrid system (Fields et al., 1989; Uetz et al., 2000), protein arrays (Finley et al.,
1994), and the phage display (Rodi et al., 1999). Although direct identification methods provide wide genome
coverage, they have a number of limitations intrinsic to their experimental design. First, a protein must pre-
serve a correct fold while attached to the chip surface (or linked to the hybrid domain). Otherwise, the
method can capture non-native interactions. Second, the binary nature of these approaches is even more re-
strictive because many of the cellular machines are multiprotein complexes, which may not be fully char-
acterized by pair wise interactions. Finally, short-living protein complexes are a tremendous problem for
all of these methods. Transient protein–protein complexes are thought to comprise a significant fraction
of all regulatory interactions in the cell and may need additional stabilization for the detection.
Over the last 5 years, experimental approaches have been supplemented by bioinformatics methods based
on genome context information. These methods explore correlations of various types of gene contexts and
functional interactions between corresponding encoded proteins. Several types of genomic context have
been utilized including: fusion of genes (Marcotte et al., 1999; Enright et al., 1999), also called the Rosetta
Stone approach; co-occurrence of genes in potential operons (Overbeek et al., 2000; Dandekar et al., 1998),
and co-occurrence of genes across genomes (Pellegrini et al., 1999) which is based on an assumption that
proteins having similar phylogenetic profiles (strings that encode the presence or absence of a protein in
every known genome) tend to be functionally linked or to operate together.
Unfortunately the usefulness of many valuable bioinformatics methods are seriously limited due to (1)
high loads in false negatives (resulting from incomplete coverage) and false positives (resulting from indi-
rect interference detection), (2) low genome coverage due to a low percentage of genes that meet underly-
ing assumptions (e.g., in a comparative study by Huynen (Huynen et al., 2000), the conservation of gene
order for Mycoplasma genitalium had the highest coverage, 37%, among all available genomes and all con-
sidered methods), and (3) predictions that are mostly derived from sequence analysis and do not incorpo-
rate any information about the structure of the interacting proteins.
For these reasons, the full power of bioinformatics approaches realized only in close integration with ex-
perimental and/or other computational methods. We will use such a collaborative approach in this effort as
we develop new identification algorithms, combining information from several heterogeneous sources.
Development of rosetta-based computational methods for characterization of
protein–protein complexes
The computer program “ROSETTA” is currently the leading program for protein structure prediction
(rated first in CASP-2001), and its approach represents a powerful foundation for applying computational
methods for characterization of protein-protein complexes. In this work, we will create a tool that will en-
able the assessment of the probability of candidate pairs of proteins forming a complex, assuming known
structures. This tool will be immensely useful for many applications aimed at genome-level categorization.
One example is the prediction of putative binding partners given a set of proteins with known structures.
As the number of known protein structures grows exponentially, such a question will soon embrace a sig-
nificant part of the bacterial proteome. Ultimately, we hope to obtain a detailed characterization of the ob-
tained molecular machine (e.g., footprint interaction areas and spatial parameters of the molecular machines)
leading to accelerated functional characterization and a better prediction of the recognition motifs. Below
we describe our research strategy in detail: how we will tune the program to protein complex characteri-
zation, our plans for integrating experimental constraints, and several possible strategies for increasing
Rosetta performance.
In the past, Rosetta has mainly been applied to the task of predicting the entire, end-to-end, structures of
protein domains starting from just the primary amino acid sequence. The main limitation of this ab initio
protein structure prediction approach has been protein size, given the computational difficulties of the
method. In this work, we will employ the Rosetta method in a different fashion. Starting from a set of pro-
teins whose unbound structure is mostly known from experimental measurements, but whose arrangement
in a complex is the subject of interest, we will tune the Rosetta algorithm to the protein complex charac-
terization problem. This will require several innovations of the method: (1) introducing a probability dis-
tribution for the distance between two centers of mass for individual chains, (2) adapting the Rosetta po-
tential functions to identify the optimal docking arrangement between two proteins in a complex, and (3)
adapting Rosetta to allow for conformational changes induced by the binding event.
Two primary questions will be addressed: (1) which proteins are binding partners and (2) how, physi-
cally, do they bind? The first question can be answered by trying all combinations of putative partners in
a set and assessing the relative quality of the fit or misfit for the binding configurations predicted by Rosetta;
from this a probability of binding can be estimated. The second question can be addressed by reporting the
optimal configuration found by Rosetta for any given pair. In practice, one would report not a single con-
formation but rather multiple plausible configurations, and then corroborate these by attempting to refine
the interaction configuration by molecular dynamic (MD) simulation and other all-atom modeling methods
in a high throughput implementation.
We will develop our technique in stages, moving from small, static protein–protein models to large com-
plex ones, in the following timeline:
1. Simulate the protein–protein complex available from PDB
2. Simulate the subdomain-peptide recognition, comparing with data from phage display experimental
3. Study the assembly of small ribosomal proteins where RDC constraints will guide the assembly
4. Carry out simulations without using experimental constraints
5. Conduct studies on a large set of proteins from Synechococcus identifying targets with tools developed
in cooperation the experimental group and the bioinformatics group
We will progress to systems involving multiprotein complexes, protein–protein complexes involving dy-
namic behavior of the components, and systems with a large number of decoys (proteins that are proven
not to be putative binding partners). It is anticipated that some of the most interesting biological systems
in the foreseeable future will be reachable only for “low resolution” technologies such as Rosetta. For one
example, we may consider multiprotein complexes involving 40–60 proteins (e.g., cellular membrane
Mass spectroscopy coupled with cross-linking (MS/CL) and NMR experiments are very attractive tech-
nologies for characterizing the molecular machinery in the cell. Integration of such experimental informa-
tion with the Rosetta method is a centerpiece of our strategy, as constraint-supported docking is likely to
have a much higher success rate for correct identification of the complexes. This experimentally obtained
information will be used to train our knowledge-based prediction methods tools, improve the Rosetta po-
tentials and ultimately enable us to reduce our dependence on constraints and successfully perform uncon-
strained simulations/docking of selected complexes.
High-performance all-atom modeling of protein machines
We will model two “molecular machine” problems in Synechococcus. In the first effort, we interpret data
from phage display experiments (section 1.0); in the second we will deduce functional properties of Syne-
chococcus membrane transporters.
The phage display library screens discussed in section 1.0 for Synechococcus proteins will yield ligands
that bind to specific proteins. Due to uncertainties (e.g., counts of expressed ligands on phage surfaces, al-
teration in binding strength due to ligand tethering, calibration of fluorescence measurements, etc), these ex-
periments will provide only a qualitative measure of binding affinity. Thus the relative binding strength of
an individual ligand/protein pair cannot be accurately compared to other pairings. We will use molecular-
scale calculations to compute relative rankings of affinities for the ligands found to bind to each probe pro-
tein in the phage library screens. These rankings will be used in the protein/protein interaction models dis-
cussed in section 4.0. Additionally, we will identify mutated ligand sequences with likely binding affinity
that can be searched for within the Synechococcus proteome to infer protein/protein pairings beyond those
indicated by the phage experiments. This work will proceed in 2 stages: we will first compute ligand con-
formations, then perform flexible docking of ligands to the known binding domains of the target proteins.
In phage display experiments a short peptide chain (ligand) is expressed on a phage surface where it po-
tentially binds to a protein (the probe or target) in the surrounding solution. The ligand is fused to coat pro-
teins (typically pVIII or pIII proteins) on the phage surface. We will model ligand conformation and ori-
entation (relative to the phage surface) for representative ligands found as hits in the library scans performed
experimentally, and thus inferred to bind to specific prokaryotic protein motifs in Synechococcus. Because
the ligands are short (9-mers to 20-mers), we anticipate being able to compute their structure “de novo,”
using a combination of computational approaches: Monte Carlo, molecular dynamics and parallel temper-
ing. In all of these methods, water can be explicitly treated, which is a critical contributor to the native
structure of the ligand in an aqueous solution. The tethering of the ligand to the phage surface can also be
naturally included in the models, as can the presence of the phage surface, which affects the energetics of
the ligand conformation and the ligand/water interactions.
Atomistic Monte Carlo (MC) techniques will first be used for this problem. Sandia’s Towhee MC code
(Martin et al., 2000) can perform simple MC moves as well as protein-specific moves (e.g., torsional rota-
tions of backbone bonds, side-chain regrowths) to sample a peptide chain’s conformational space. Towhee
also has a configurational-bias (CB) MC capability (Martin et al., 1999) so that a peptide chain can be
“grown,” atom by atom, into a surrounding medium in a manner that preferentially generates lower energy
conformations. Towhee can then solvate the system with water using similar CB-MC insertions.
We will also investigate the use of a new method, parallel tempering (or replica-exchange) (Mitsutake
et al., 2001), to generate low-energy ligand conformations. In parallel tempering, multiple copies of a mol-
ecular-scale simulation are created and simulated at different temperatures using traditional MD. Periodi-
cally, the respective temperatures of a pair of ensembles are swapped according to Monte Carlo rules. The
method is highly parallel since individual replicas run with little communication between them. Parallel
tempering can find low-energy conformations much more quickly than a standard MD simulation. Garcia
et al. (2001) used these methods to find the native beta-hairpin conformational state of an 18-mer peptide
fragment of protein G in explicit solvent within a few nanoseconds of MD simulation time, starting from
a denatured conformation. Similar work predicted alpha-helical structures in short peptide chains (Sanbon-
matsu et al., 2002). We propose to enhance our LAMMPS MD code to include a replica-exchange capa-
bility whereby P 5M3N processors can run M replicas, each on N processors. This will enable us to ef-
ficiently apply all the LAMMPS features (particle-mesh Ewald, rRESPA, force fields) to computing ligand
All-atom methods will also be applied to investigate ligand conformations for docking. Both high-accuracy
conformations and low-accuracy conformations will be investigated to maximize the information yielded
given the computational requirements of the flexible docking methods we will employ.
Transport proteins found in cell membranes are known to be important to the functioning of Syne-
chococcus, as to all microbes and all cells. Yet much about these molecular machines is unknown, from
the function and regulation of individual transporters to the interaction and cross-talk between multiple
transporters. We will model three types of transporters in Synechococcus: ion channels, small multi-drug
resistance (SMR) transporters, and ATP binding cassette (ABC) transporters. The goal of this effort is to
uncover the physical basis for the function of these transporters. These studies will provide molecular in-
sight for the system-level cell models developed in this effort (section 1.0), for example, as boundary con-
ditions on the cell as it interacts with its extracellular environment.
“Knowledge fusion” based characterization of biomolecular machines
Developing new data mining and statistical analysis methods for the identification of protein-protein in-
teractions is an urgent need for several reasons. First, interactions can be deduced in unusual ways from
many very diverse data sources (for example, from the fact that genes from one genome are fused in genome
of another organism). Second, information is accumulating in databases of all kinds (complete genomes,
expression arrays, proteomics, structural) at an unprecedented rate, resulting in a landslide of unanalyzed
data. Furthermore, taking advantage of the information in these structural and biophysical databases pre-
sents a special challenge as many existing data mining techniques were developed for sequence databases
and conceptually new approaches will be needed for the structural domain.
The focus of this task is to develop advanced data mining algorithms to elucidate (1) which proteins in
a cell interact both directly (via multiprotein complex) and indirectly (via biochemical process, metabolic
or regulatory pathway), (2) where on the protein surface the interaction occurs, and (3) what biological
functions the protein complex performs. Existing data mining tools for making such inferences have low
predictive accuracy and do not scale for genome-wide studies. This is largely due to incorporation of data
at a single or very few levels, lack of sufficient data, and/or computational intractability of exact algorithms.
We will improve the predictive accuracy of such methods through several approaches: developing “knowl-
edge fusion” based algorithms that make predictions by fusing knowledge extracted from various sources
of bioinformatics, simulation, and experimental data, coupling these algorithms with modeling and simu-
lation methods for approximating structure-related missing data, and extending the applicability of these al-
gorithms to the genome-scale through developing their high performance optimized versions suited for
Terascale computers.
Our development strategy will involve three parts: identification of pair-wise protein interactions, con-
struction of protein interaction maps of these complexes, and functional characterization of identified com-
plexes. These tools will be prototyped with application to the Synechococcus proteome in coordination with
our regulatory pathway mining effort and used to obtain information necessary for our systems biology
Discovery and characterization of Synechococcus molecular machines
Our aim is to apply the computational methods developed in this effort to Synechococcus. We will ini-
tially verify known molecular machines in Synechococcus to prototype our methods. Ultimately we expect
that these methods will enable us to (1) discover novel multiprotein complexes and protein binding domains
that mediate the protein-protein interactions in Synechococcus, and (2) better understand the functional
mechanisms involved in carbon fixation and environmental responses to carbon dioxide levels. In particu-
lar we will characterize Synechococcus protein-protein interactions that contain known interaction domains
such as SH3 domains, and LRRs, as well as the Synechococcus protein complexes related to carboxyso-
mal, ABC transporter systems and also protein-protein interactions involved into circadian system and light-
signal transduction pathways.
About 10% of the genes of bacterial genomes are dedicated to transport and there are approximately 200
transporter families. Validations and applications of our biomolecular machines characterization pipeline
methods will be tested by focusing on elucidating the functional mechanisms of protein complexes related
to carboxysomal and ABC transporter systems in Synechococcus. The categorical data analysis based pre-
diction methods will be applied to all amino-acid sequences of interest in Synechococcus genome. This will
generate a probability matrix with a probability of interaction assigned to each protein pair. Rosetta-based
modeling methods will be applied to a selected set of more likely interacting protein pairs. This will pro-
vide a basis for determining putative structural properties of selected proteins and give hints about poten-
tial protein-protein interaction residue sites. The identified structural properties will be used by prediction
methods to further validate and/or refine a set of interacting protein pairs. Thus, these knowledge-based
prediction methods coupled with modeling and simulation will determine a set of possible protein pairs in-
volved in the carboxysomal and ABC transporter complexes and a set of their putative binding sites.
In living systems, control of biological function occurs at the cellular and molecular levels. These con-
trols are implemented by the regulation of activities and concentrations of species taking part in biochem-
ical reactions (Stephanopoulos et al., 1998). The complex machinery for transmitting and implementing the
regulatory signals is made of a network of interacting proteins, called regulatory networks. Characteriza-
tion of these regulatory networks or pathways is essential to our understanding of biological functions at
both molecular and cellular levels. Traditionally, study of regulatory pathways is done on individual basis
through ad hoc approaches. With the advent of high-throughput measurement technologies, e.g., microarray
chips for gene/protein expression and two-hybrid systems for protein-protein interactions, and bioinfor-
matics, it is now feasible and essential to develop new and effective protocols for tackling the challenge of
systematic characterization of regulatory pathways. The impact of these new high-throughput methods can
be greatly leveraged by carefully integrating new information with the existing (and evolving) literature on
regulatory pathways in all organisms. The main focus of this subproject is to develop a suite of computa-
tional capabilities for (1) information extraction from large-scale biological data, including microarray gene
expression data, genomic sequence data, protein-protein interaction data, and (2) systematic construction of
biological pathways through fully utilizing information extracted from large-scale biological data mining
and analysis. These capabilities will improve the state of the art in biological pathway inference, in terms
of both prediction reliability and application scope. Initially, these capabilities will be applied to a set of
selected pathways in Synechococcus, including ABC transporter pathway. As the technologies mature, they
will be ported on to high-performance supercomputers to provide community-wide services for general bi-
ological pathway constructions.
Derivation of regulatory pathways through combining multiple sources of information:
our vision
Given the complexity of regulatory pathways and incomplete nature of existing high-throughput exper-
imental data, it is unlikely that one could develop a computational capability for accurate and automated
derivation of novel biological pathways in the near future. It will be more realistic if we set our goal to
identify which parts of a target pathway are inferable from the publicly available data/information and to
construct them; and then to identify what data we may need, specific to a particular pathway, for a full
characterization of the pathway. Through rational design of experiments for further data collection to fill
in the “gaps,” we can significantly reduce the cost and time needed to fully characterize a pathway. To
make the experimental data more useful, we will first develop a number of improved capabilities for gen-
eration and interpretation of data. Initially these data include (a) microarray gene-expression data, (b) ge-
nomic sequence data, and (c) protein–protein interaction data. Then we will investigate an inference frame-
work for pathways, which can make use of all of these data. This inference framework will be able to pull
together pathway information from our own work and from earlier relevant investigations. With that cor-
pus we will be able to (1) to assign weights to each data item, based on our assessment on the quality of
each data source and the cross-validation information from other sources, (2) identify components of a tar-
get pathway and their interaction map to the extent possible, and (3) identify the parts of a target pathway
that are not inferable from the available data. This framework will be organized such that new sources of
information or analysis tools can be easily added as they become available, without affecting the other parts
of the framework. We envision that we can quickly generate a set of possible candidate pathway models,
possibly with certain parts missing or uncertain, using this inference framework. An iterative process will
then follow to design and conduct experiments through rational design and then feed the new and more
specific data to this inference framework to refine the models. As more high-throughput genomic/proteomic
data become available, we can expect that fewer pathway-specific experiments will be needed for complete
characterization of a pathway and hence the more automated this inference framework will become.
Improved experimental and computational capabilities for high-throughput biological data
generation and interpretation
High-throughput biological data generation capabilities, like microarray gene expression chips or two hy-
brid systems for protein-protein interactions, provide powerful tools for scientists to investigate the inter-
nal structures and functional mechanisms of various biological processes at the molecular level. However
it is well known that these high-throughput technologies, at the current stage, could be quite noisy, mak-
ing it a highly challenging problem to interpret the data in a biologically meaningful and systematic man-
ner. Compounding to the stochastic nature of the underlying biological processes under investigation, many
non-biological technical factors could contribute to the readings of the biological data collected, which could
significantly affect our data interpretation. For example, in the case of microarray chips, the readings of
microarray signal intensity could be highly sensitive to small differences in local surface properties of a
micrarray chip, making it even difficult to reproduce the same data under the same controllable experi-
mental conditions. Our early focus will be to improve experimental designs for microarray experiments for
more reliable data collection.
In this project, we will develop improved experimental processes in order to reduce microarray expres-
sion variability. We propose to perform a variety of statistically designed experiments to elucidate the er-
ror sources for microarray experiments. Yeast microarrays will be used in these initial experiments since
most experience with microarrays has resulted from experiments with yeast microarrays. The results from
the final optimized microarray experiments will generate information about the error structure of the micro-
array data. This information will be used to evaluate bioinformatics algorithms by providing a realistic er-
ror structure. In addition, this information will facilitate the use of improved algorithms that require knowl-
edge of the covariance structure of the noise. This will result in improved data quality and reduced depen-
dence on data normalization and preprocessing. As a result, it will be possible to place more confidence in
the assumption that the observed variations in the data could be directly attributed to actual biological vari-
ation in the sample rather than on the experimental variability as has often been the case with current micro-
array experiments. Similar studies will be extended to other high-throughput biological data production ex-
periments, to improve the overall quality of our initial data for information extraction.
Improved data mining capabilities will then be developed for various biological data sources. Our initial
focus will still be on microarray chip data. Data clustering is a powerful technique for grouping unstruc-
tured data sharing common or similar features. It facilitates recognition of patterns shared by some subsets
of the data, and identification of significant hidden signals. Data clustering is often used as the first step in
mining a large quantity of data. The basic idea of clustering can be stated as follows. For a given set of
data points (e.g., multi-dimensional vectors), group the data into clusters so that (1) data of the same clus-
ter have similar (specified) features, and (2) data of different clusters have dissimilar features. We will de-
velop significantly improved capabilities for data clustering, based on the concept of minimum spanning
trees (Prim, 1957). Our preliminary studies have established rigorous relationship between data clustering
and searching for substrings in a string with certain properties (Xu et al., 2002; Olman et al., 2002). Based
on this fundamental relationship, we will develop a suite of data clustering tools capable of solving data
clustering problems that are not easily solvable using any existing clustering tools, including problems of
identifying/extracting data clusters from a very noisy data set, etc. Such improved data mining capabilities
should allow us to more effectively extract significant patterns and co-relations from a large quantity of un-
structured data, to be used for pathway inference.
Parallel efforts will be given to developments of more effective capabilities for mining genomic sequence
data for inference of operon structures, identification of orthologs, and identification of protein binding sites,
for the purpose of deriving possible parts list (which genes may be involved in a pathway) of a target path-
way. The “parts” information, along with the possible protein-protein interaction information, causality in-
formation in a gene network, and identified co-regulated gene list, will provide a set of basic components
for computational capabilities for pathway inference.
Improved computational capabilities for inference of biological pathways
Our proposed pathway inference framework will consist of the following main components:
1. Prediction of potential genes/proteins involved in a specific pathway
2. Function assignment of a protein
3. Identification of co-regulated genes and interacting proteins
4. Mapping proteins to a known biological pathway
5. Inference of a specific pathway consistent with available information
Prediction of potential genes/proteins involved in a specific pathway. The first step towards pathway
construction is to select a relatively small set (,100) of genes/proteins (list A) that are potentially involved
in the target pathway. To do this, we will first use gene expression profiles to find highly induced genes
under the condition that activates the pathway. For example, in studying leucine transport pathway in yeast,
we can monitor the up-regulated genes under leucine rich environment. Some of the induced genes are
probably involved directly in the pathway; others may be indirectly involved (e.g., general energy pathway)
or not relevant at all (due to noise in experimental data). Then we will use protein-protein interaction/
co-regulated gene information (see the following) to add more genes to this list. Any proteins that interact
directly with a protein in list A will be selected. In addition, proteins that are connected through a series of
protein-protein interactions in list A will also be selected. This will form the list of the potential components
for the target pathway. We will also use homologous information to build this list. If the homologs of two
proteins in another organism interact with each other, the two proteins may interact with each other in Syne-
chococcus as well. A user can review this list and add/remove the genes based on his/her own knowledge.
Function assignment of a protein. Some genes in the selected list may not have function assignment. We
will use gene expression data, protein-protein interaction data, and co-regulation/regulon data to infer a pos-
sible functional role of a hypothetical protein. The idea is “guilt-by-association”: if a hypothetical protein
interacts with two or more proteins with the same function, it is likely that the hypothetical protein also has
this function. If we have determined a hypothetical protein in a regulon and other genes in the regulon have
the same known cellular role, then this hypothetical protein may have the same cellular role as well. In ad-
dition, we can study genes with similar expression profiles to the hypothetical protein. If there is a con-
sensus among the functions of these genes, it may suggest the function of the hypothetical protein.
Identification of co-regulated genes and interacting proteins. To determine co-regulated genes, we will
first cluster gene expression profiles using the selected list. Then we will retrieve upstream regions of these
genes. If the co-expressed genes share similar sequence-motifs in their upstream regions, we will assign
them as co-regulated genes. Also genes predicted to be in the same operon will also be added to the list of
co-regulated genes. After identifying co-regulated genes, we will determine which protein–protein inter-
actions determined above are true interactions (it is expected some of them are false interactions). If two
proteins are co-regulated, their interaction determined above is probably true. If a protein–protein interac-
tion is predicted from multiple sources, this interaction pair should have higher likelihood to be true.
Mapping proteins to a known biological pathway. We will search the selected list of proteins against all
known biological pathways. This effort will benefit from the literature mining tools also developed in this
subproject. If a significant fraction of the proteins can map to a known pathway of a different organism
(through homology), a pathway can be constructed for these proteins using the known pathway as a tem-
plate. The unmapped parts of the pathway may suggest what other proteins we should look in Synechococcus
For each of the identified component protein or protein-protein interaction, using our own or others’ tools,
we will first thoroughly evaluate the reliability of the prediction, based on well-characterized gene/protein
functions, protein–protein interactions, operons/regulons. These reliability values will be used in our infer-
ence framework when putting all pieces together.
Inference of a specific pathway consistent with available information. Our initial objective is to work on
regulatory pathways that may have similar or related known pathways in other genomes. This way, we will
have at least a partial template to work with. As our capabilities mature, we will gradually move to more
ab initio constructions of pathways. In the template-based case, the key issue is to identify all the corre-
sponding parts (components and interactions) in the template pathway and the missing parts in Syne-
chococcus. Then the system will ask for either new data to be generated by further experiment or/and for
a revised template pathway by the user. In the more ab inito construction of a pathway, the system will
need some general guiding information from the user like what functional classes of proteins may be in-
volved, how large the target pathway is approximately, which known genes/proteins are probably involved,
etc. Then inference framework should suggest a list of possible models based on the general information
extracted from various data sources and specific information about the target pathway. We will explore a
number of possible computational approaches as the main tool for constructing such frameworks, includ-
ing Bayesian networks (Friedman et al., 2000), expert system techniques (Takai-Igarashi and Kaminuma,
1999), and combinatorial optimization approaches (Kyoda et al., 2000). These methods have been applied
to infer regulatory networks using single source of information. They were only successful on artificial sim-
ulated data but not on real biological data, since a single source of data is insufficient to derive regulatory
networks as discussed above. Here, by combining different sources of information, we expect to make sig-
nificant advances in reference of regulatory networks.
Ultimately, all of the data that is generated from experiment must be interpreted in the context of a model
system. Individual measurements can be related to a very specific pathway within a cell, but the real goal
is a systems understanding of the cell. Given the complexity and volume of experimental data as well as
the physical and chemical models that can be brought to bear on subcellular processes, systems biology or
cell models hold the best hope for relating a large and varied number of measurements to explain and pre-
dict cellular response. Clearly, cells fit the working scientific definition of a complex system: a system
where a number of simple parts combine to form a larger system whose behavior is much harder to un-
derstand. The primary goal of this subproject is to integrate the data generated from the overall project’s
experiments and lower level simulations, along with data from the existing body of literature, into a whole
cell model that captures the interactions between all of the individual parts. It is important to note here that
all of the information that is obtained from the previously described efforts in this project is vital to the
work here. In a sense, this is the “Life” of the “Genomes to Life” theme of this project.
The precise mechanism of carbon sequestration in Synechococcus is poorly understood. There is much
unknown about the complicated pathway by which inorganic carbon is transferred into the cytoplasm and
then converted to organic carbon. While work has been carried out on many of the individual steps of this
process, the finer points are lacking, as is an understanding of the relationships between the different steps
and processes. Understanding the response of Synechococcus to different levels of CO2in the atmosphere
will require a detailed understanding of how the carbon concentrating mechanisms in Synechococcus work
together. This will require looking these pathways as a system.
The aims of this section are to develop and apply a set of tools for capturing the behavior of complex
systems at different levels of resolution for the carbon fixation behavior of Synechococcus. We briefly de-
scribe here those 4 objectives.
Protein interaction network inference and analysis using large-scale experimental data and
simulation results
Experimentally, regulatory networks are generally probed with microarray experiments, and protein net-
work interactions have been investigated with 2-hybrid screening. All the inference computational techniques
have so far been based on probabilistic frameworks that search the space of all possible labeled graphs. Our
aim is to infer networks from multiple sources including phage display experimental data and simulation re-
sults from subprojects 2 and 3. Furthermore, instead of searching networks in the space of all labeled graphs
we proposed to search networks in the space of scale-free graphs. The scale-free nature of protein networks
was first discovered by Jeong et al. (Jeong, 2000) and independently verified by Gomez et al. (Gomez, 2001).
Since the number of scale-free graphs is many order of magnitudes smaller than the number of labeled graphs
we expect to develop a method far more efficient than the current state-of-the art.
It is well established that proteins interact through specific domains. While many proteins are composed
of only one domain, multiple domains are also present and must be considered when reconstructing net-
works (Uetz, 2000). Probabilities of attraction between protein domains have been derived from phage dis-
play data (Tong, 2002), and protein–protein interaction databases. Note that probability of attraction be-
tween domains can be calculated from binding energies computed through molecular simulations in
subproject 2. Considering two multi-domains proteins iand j, one can then define a probability (pij) of at-
traction between these proteins as (Gomez, 2002):
pij 5
where vi(vj) is the domain set of protein i(j), and p(dm,dn) is the probability of attraction between domains
dmand dn. Thus, the problem of inferring a protein-protein interaction network from domain–domain in-
teraction probabilities reduces to finding a graph G5(V,E) where the vertices of Vare proteins and the
edges of Eare protein-protein interactions that maximizes the probability:
eij[Epij P
eklÒE(1 2pkl ) (2)
The trivial solution to this problem, which consists of selecting only the edges with probability of .0.5
is not appropriate because protein–protein interaction networks are generally scale-free networks [11], which
is an additional constraint not captured in Eq. 4-2.
Our proposed work is composed of the four following steps:
1. Develop methodology to characterize and analyze scale-free networks and protein interaction networks
2. Compute domain–domain attraction probabilities from phage display data, molecular simulations, and
protein–protein interaction databases
3. Sample scale-free networks that maximize P(E) computed in step 2 using labeled graph sampling algo-
rithm and characteristics developed in step 1
4. Compare predicted networks with experimentally derived two-hybrid networks. Adjust domain–domain
attraction probabilities and repeat steps 2–4 until agreement between predicted and two-hyrid networks
is reached.
The above four tasks will be tested with the yeast proteome for which there is already ample data and
then will be applied to Synechococcus when experimental and simulation data become available.
Discrete component simulation model of the inorganic carbon to organic carbon process
Once protein networks have been inferred, one can then study their dynamics. While even the simplest
prokaryotic cells are extremely complex, this complexity is generally driven by a relatively small number
of unique cellular components. One of the consequences of this is that many important processes in cells
can be controlled by the interaction of a very small number of individual reactants. This can lead to a wide
range of different behaviors associated with cells of identical type due to the fluctuations in the number and
position of its reactants. In many cases, it is important to understand how this randomness affects cell be-
havior through computer modeling.
There are two different ways in which the individual particle method can be implemented. In the first
model, “reactions” are calculated by a stochastic method such as that described by Gillespie (Gillespie,
1976) with recent developments by Gibson and Bruck (Gibson, 2000). In this method, there is a set of pos-
sible reactions that can occur given the products that exist. There are also reaction rates that are associated
with any of these events occurring. For calculations where spatial details are more important, a second
model is used that is a little more sophisticated. In this model, each of the objects is modeled separately
and its spatial position tracked separately, in the spirit of the code by Stiles and Bartol (Stiles,
2001). (We note here for clarity that the “particles” described in this section are not atoms or even neces-
sarily molecules, but simply individual objects in the cell that must be tracked separately.)
There are two primary tasks associated with this objective:
Stochastic method. We will first build a serial version of the code, based on the work that has already
been done by Lok and Brent at tMSI. We will test this code on yeast data, and Synechococcus data from
other subprojects when it becomes available. In this serial version we will address the event scheduling is-
sues related to the sub-volume partitioning so that the debugging processes will be more straightforward
than it would be on the parallel version. After the serial code is working, we will begin to develop a mas-
sively parallel version of this code based on domain decomposition.
Individual particle method. This method will begin by adapting Sandia’s existing particle code (ICARUS)
to work on biological systems. Boundary conditions will be implemented that allow reactions on the inter-
faces. This will model biological processes that occur on the cell membrane and the surfaces of internal
structures. The ultimate goal is to be able to handle more than 107individual particles using hundreds of
In both models we can start with a higher concentration of inorganic carbon near the membrane and then
run the model forward in time to generate a simulation of how the inorganic and organic carbon (in the car-
boxysomes) coexist inside the cell. Once the network is set up, one can then change individual reactant
amounts or reaction rates and test to see how this affects the results. Finally, these techniques can be used
to help determine unknown variables in the network by comparing the results against experimentally de-
terminable quantities.
Continuous species simulation of ionic concentrations
While a discrete particle simulation is useful for situations where there is a relatively small number of
particles, once the concentration of a particular species becomes large enough the discrete method becomes
impractical and unnecessary. In this case, the particle number is large enough that the overall behavior is
better understood as a continuous phenomenon, where the particle concentration is modeled as a continu-
ous function of space and time. The interactions between various species are described in terms of partial
differential equations, and the resulting formulae belong to a general class of equations known as reaction/
diffusion equations.
One code used to solve the reaction/diffusion equations essential for this effort is a widely used produc-
tion code at Sandia called (Shadid, 1997). This code has been shown to successfully scale to
more than 1,000 processors with very little loss of speed. We plan on using a version of MPSalsa to per-
form much of the work proposed here.
Despite much research, there is still not a clear consensus on the mechanism by which inorganic carbon
is transported across the cell membrane (Kaplan, 1999). There are many mechanisms that are being con-
sidered. The simplest is that it passes through the cell membrane as CO2, and this behavior has been well
documented in many microbes. It is also believed that HCO3
2is actively transported across the membrane
via either an ion gradient or by an ATP fueled pump. There is now increasing belief that there may be mul-
tiple mechanisms for getting inorganic carbon into the cytoplasm. Some of the CO2that exists in the cy-
toplasm is converted into HCO3
2. When the HCO3
2reaches the carboxylation site, it is converted to CO2,
which is then used by RuBisCO to form 3-phosphoglycerate (PGA).
The specific goal in this aim is to study interplay between CO2and HCO3. The first work will be done
making minor modifications to the existing code ( ) that allow for species to be created at in-
terfaces to help model specific biological mechanisms, such as membrane transport. Eric Jakobsson and his
co-workers at UIUC have done extensive modeling of membranes and ion channels. They will be provid-
ing support to this project by modeling proposed ion channel structures based on sequence data to help for-
mulate the boundary conditions for the for inorganic carbon species formulation. The boundary conditions
on the simulation can be set to represent both the steady diffusion of CO2across the membrane, and point
sources of HCO3
2related to specific pumps located in the membrane. The carboxylation site could also be
modeled as a sink for HCO3
2and a source for CO2and PGA.
Once the work gets done obtaining all of the boundary conditions regarding inorganic carbon transport,
the simulation will be used to study what concentrations of carbon could be sequestered given various known
kinetic constants associated with RuBisCO (as discussed in section 1.0) and membrane transport. We will
then compare our results to experimental measurements obtained in this proposal and elsewhere, and use
this to drive the direction of future experiments.
Synechococcus carboxysomes and carbon sequestration in bio-feedback, hierarchical modeling
This aim answers the questions, “How does one utilitize genomic and proteomic information in prob-
lems manifest at the ecosystem level?” To begin, consider a conceptual organization that can be modeled
via a hierarchical, object-oriented design, whereby conceptually discrete systems are linked by levels of in-
teraction. Details of each level are handled within a “black box,” communicating to levels above and be-
low by specified rules based on scientifically known or hypothesized mechanisms of interaction. The model
allows the connection of levels by the imposition of de novo laws as discovered in the respective disci-
plines, their rules of interaction and axiomatic behavior, as well as an actual examination of the state of the
As a demonstrative example of the utility of a hierarchical, feedback model, we have tested the prelimi-
nary implementation in the complex scenario of the genetic basis of flu pandemics. Influenza is a negative-
stranded RNA virus of the family Orthomyxoviridae. Importantly, each pandemic has been associated with
the discovery of a new serotype for the virus’ hemagglutinin (HA) protein. Swine, and particularly birds,
serve as reservoirs for the HA subtypes. With this basic knowledge of genetic factors underlying influenza’s
virulence, we now seek factors that create HA variation. RNA-RNA recombination is known in numerous
viruses, including influenza (for review, see Worobey and Holmes 1999). Using our hierarchical model, we
demonstrated the greatly increased morbidity associated with the RNA-RNA recombination model (as op-
posed to no RNA-RNA recombination).
To investigate the importance of Synechococcusin carbon cycling using a data-driven, hierarchical model,
we seek to directly incorporate genomic and proteomic knowledge of Synechococcus to understand how
conditions, such as a 1°C increase in ambient temperature, affect carbon fixation of important and ubiqui-
tous marine populations (Fig. 1). We propose to do this by underlaying the carboxysome of Figure 2 with
known carbon fixation metabolic pathway information such as that available at
keggmaps/syn_wh/07sep00/html/map00710.html. The network dynamics of the previous sections of this
proposal give us a model of carbon fixation dependent on a variety of external parameterizations, such as
ambient water temperature, CO2diffusion rates, and Synechococcus growth rates.
FIG. 1. Hierarchical model relating pathways to carbon cycling.
FIG. 2. Carbon concentrating mechanism (from Kaplan and Reinhold, 1999).
A broader result of this work on Synechococcus is to help us understand how biological reactions to en-
vironmental conditions feedback onto the environmental conditions themselves: thus the loop back in Fig-
ure 1 between CO2affecting growth rates and marine biomass, which in turn affect carbon sequestration.
The strains in Figure 1 each encapsulate a variant in CO2fixation pathways as similarly used in the previ-
ous worked example.
The explosion of data being produce by high-throughput experiments will require data analysis tools and
models which are more computationally complex, more heterogeneous, and require coupling to enormous
amounts of experimentally obtained data in archived ever changing formats. Such problems are unprece-
dented in high performance scientific computing and will easily exceed the capabilities of the next gener-
ation (PetaFlop) supercomputers.
The principal finding of a recent DOE Genomes to Life (GTL) workshop was that only through com-
putational infrastructure dedicated to the needs of biologists coupled with new enabling technologies and
applications will it be possible “to move up the biological complexity ladder” and tackle the next genera-
tion of challenges. This section discusses the development of a number of such capabilities including work
environments such as electronic notebooks and workflow environments, and high performance computa-
tional systems to support the data and modeling needs of GTL researchers, particularly those involved in
this Synechococcus study.
Important to enabling technologies is the issue of ease of use and coupling between geographically and
organizationally distributed people, data, software, and hardware. Today most analysis and modeling is done
on desktop systems, but it is also true that most of these are greatly simplified problems compared to the
needs of GTL. Thus an important consideration in the GTL computing infrastructure is how to link the GTL
researchers and their desktop systems to the high performance computers and diverse databases in a seam-
less and transparent way. We believe that this link can be accomplished through work environments that
have simple web or desktop based user interfaces on the front-end and tie to large supercomputers and data
analysis engines on the back-end.
These work environments have to be more than simple store and query tools. They have be conceptu-
ally integrated “knowledge enabling” environments that couple vast amounts of distributed data, advanced
informatics methods, experiments, and modeling and simulation. Work environment tools such as the elec-
tronic notebooks have already shown their utility in providing timely access to experimental data, discov-
ery resources and interactive teamwork, but much needs to be done to develop integrated methods that al-
low the researcher to discover relationships and ultimately knowledge of the workings of microbes. With
large, complex biological databases and a diversity of data types, the methods for accessing, transforming,
modeling, and evaluating these massive datasets will be critical. Research groups must interact with these
data sources in many ways. In this effort, we will develop a problem solving environment with tools to sup-
port the management, analysis, and display of these datasets. We will also develop new software tech-
nologies including “Mathematica-type toolkits for molecular, cellular and systems biology with highly op-
timized life science library modules embedded into script-driven environments for rapid prototyping. These
modules will easily interface with database systems, high-end simulations, and collaborative workflow tools
for collaboration and teaching.
Working environments: the lab benches of the future
This project will result in the development of new methods and software tools to help both experi-
mental and computational efforts characterize protein complexes and regulatory networks in Syne-
chococcus. The integration of such computational tools will be essential to enable a systems-level un-
derstanding of the carbon fixation behavior of Synechococcus. Computational working environments
will be an essential part of our strategy to achieve the necessary level of integration of such computa-
tional methods and tools.
Because there is such diversity among computational life science applications in the amount and type of
their computational requirements, the user interface designed in this effort will be designed to support three
motifs. The first is a biology web portal. These have become popular over the past three years because of
their easy access and transparent use of high performance computing. One such popular web portal is
ORNL’s Genome Channel. The Genome Channel is a high-throughput distributed computational environ-
ment providing the genome community with various services, tools, and infrastructure for high quality analy-
sis and annotation of large-scale genome sequence data. We plan to leverage this existing framework, which
is based on the Genomes Integrated Supercomputer Toolkit (GIST), to create a web portal for the applica-
tions developed in this proposal.
The second motif is an electronic notebook. This electronic equivalent of the paper lab notebook is in
use by thousands of researchers across the nation. Biology and Pharma labs have shown the most interest
in this collaboration and data management tool. Because of its familiar interface and ease of use, this mo-
tif provides a way to expose reluctant biologists to the use of software tools as a way to improve their re-
search. The most popular of the electronic notebooks is the ORNL enote software. This package provides
a very generic interface that we propose to make much more biology centric by integrating the advanced
bioinformatics methods described in this proposal into the interface. In out years we plan to incorporate
metadata management into the electronic notebook to allow for tracking of data pedigree, etc.
The third motif will be a Matlab-like toolkit whose purpose would be fast prototyping of new computa-
tional biology ideas and allow for a fast transition of algorithms from papers into tools that can be made
available to an average person sitting in the lab. No such tool exists today for biology.
For all three of the working environment motifs we will build an underlying infrastructure to: (1) sup-
port new core data types that are natural to life science, (2) allow for new operations on those data types,
(3) support much richer features, and (4) provide reasonable performance on typical life science data. The
types of data supported by electronic notebooks and problem solving environments should go beyond se-
quences and strings and include trees and clusters, networks and pathways, time series and sets, 3D mod-
els of molecules or other objects, shapes generator functions, deep images, etc. Research is needed to al-
low for storing, indexing, querying, retrieving, comparing, and transforming those new data types. For
example, such tools should be able to index metabolic pathways and apply a comparison operator to re-
trieve all metabolic pathways that are similar to a queried metabolic pathway.
Creating new GTL-specific functionality for the work environments
There are a large number of GTL-specific functionalities that could be added to the work environments.
For this effort we have selected three that have wide applicability across computational biology to illustrate
how the work environments can be extended. These are as follows:
1. Graph data management for biological network data
2. Efficient data organization and processing of microarray databases
3. High performance clustering methods
Graph data management for biological network data. In this first area, we will develop general-purpose
graph-based data management capabilities for biological network data produced by this Synechococcus ef-
fort as well as from other similar efforts. Our system will include an expressive query language capable of
encoding select-project queries, graph template queries, regular expressions over paths in the network, as
well as subgraph homomorphism queries (e.g. find all of examples of pathway templates in which the en-
zyme specification is a class of enzymes). Such subgraph homomorphism queries arise whenever the con-
straints on the nodes of the query template are framed in terms of generic classes (abstract noun phrases)
from a concept lattice (such as the Gene Ontology), whereas the graph database contents refer to specific
enzymes, reactants, etc. Graph homomorphism queries are known to be NP-hard and require specialize tech-
niques that cannot be supported by translating them into queries supported by conventional database man-
agement systems.
This work on graph databases is based on the premise that such biological network data can be effectively
modeled in terms of labeled directed graphs. This observation is neither novel nor controversial: a number
of other investigators have made similar observations (e.g. the AMAZE database, VNM00). Other investi-
gators have suggested the use of stochastic Petri Nets (generally described by Directed Labeled Graphs) to
model signaling networks. Some nodes represent biochemical entities (reactants, proteins, enzymes, etc.) or
processes (e.g. chemical reactions, catalysis, inhibition, promotion, gene expression, input-to-reaction, out-
put-from-reaction, etc.). Directed edges connect chemical entities and biochemical processes to other bio-
chemical processes or chemical entities. Undirected edges can be used to indicate protein interactions.
Current systems for managing such network data offer limited query facilities, or resort to ad hoc pro-
cedural programs to answer more complex or unconventional queries, which the underlying (usually rela-
tional) DBMSs can not answer. The absence of general purpose query languages for such graph databases
either constrains the sorts of queries biologists may ask, or forces them to engage in tedious programming
whenever they need to answer such queries. For these reasons, we will focus our efforts on the develop-
ment of the graph query language and a main memory query processor. We plan to use a conventional re-
lational DBMS for the persistent storage (perhaps DB2, which supports some recursive query processing).
The main memory graph query processor will directly call the relational database management system (i.e.,
both will reside on the server). The query results will be encoded (serialized) into XML and a SOAP-based
query API will be provided, to permit applications or user interfaces to run remotely.
We will also explore the use of graph grammars for describing query languages, network data, and the
evolution of biological networks. Graph grammars are the graph analog of conventional string grammars.
Thus the left hand side of a GG rule is generally a small graph, whereas the right hand side of the GG rule
would be a larger (sub-) graph. Graph grammars can be used for graph generation (e.g. to model network
evolution) and graph parsing. They have been used to describe various sorts of graph languages. Graph
grammars could be useful for specifying the graph query language powerful enough for the graph opera-
tions described above.
Efficient data organization and processing of microarray databases. Microarray experiments have proven
very useful to functional genomics and the data generated by such experiments is growing at a rapid rate.
While initial experiments were constrained by the cost of microarrays, the speed with which they could be
constructed, and occasionally by the sample generation rates, many of these constraints have been or are
being overcome. One could readily envision that data production rates might increase another factor of 5
or 10. Note that we are concerned here with the processed data, not the much larger raw image data, which
we assume will likely not be kept in a DBMS. Datasets of 50 or 100 GB/year 33 or 4 years exceed likely
main memory configurations. This does not even account for record overhead, or indices. It is likely that
most of this data will be kept on disk. Thus we will need efficient database designs, indices and query pro-
cessing algorithms to retrieve and process these large datasets from disk.
It will be necessary to support queries over these relations in combination with the spot data in order to
permit queries that are meaningful to the biologists. Note that the space described above represents the
Cartesian product of the experimental conditions and the genes. However, we can expect replication of spots
and experiments, since replication is essential to reliable statistical analysis of this very noisy data. In ad-
dition, it will be necessary to support ontology of the various genes, gene products and a biological net-
work database describing various cellular processes (metabolic, signal transduction, gene regulation).
A common query might ask (over some subset of the experimental design) which genes are overexpressed
relative to their expression for standard experimental conditions. Other queries might request restricting the
set of genes considered to certain pathways or retrieving pathways in addition to genes. To support such
queries, it is necessary to join the results of conditions on the experimental design with the microarray spot
data in order to identify the genes that are overexpressed. This implies the capability of searching over one
or more of the spot attributes.
Indexing over a billion or more elements is a daunting task. Conventional indexing techniques provided
by commercial database systems, such as B-trees, do not scale. One of the reasons for this is that general-
purpose indexing techniques are designed for data that can be updated over time. Recognizing this prob-
lem, other indexing techniques have been proposed, notably techniques that take advantage of the static na-
ture of the data, as is the case with much of scientific data resulting from experiments or simulations.
One of the most effective methods of dealing with large static data is called “bitmap indexing” (Wu et al.,
2001; Wu et al., 2002). The main idea for bitmap indexing is to partition each attribute into some number
of bins (such as 100 bins over the range of data values), and to construct bitmaps for each bin. Then one
can compressed the bitmaps and perform logical operations to achieve a great degree of efficiency.
LBNL has developed highly efficient bitmap indexing techniques that were shown to perform one to
two orders of magnitude better than commercial software, and where the size of the indexes are only
20–30% the size of the original vertical partition (Wu et al., 2001; Wu et al., 2002). To achieve this we
have developed specialized compression techniques and encoding methods that permit the logical oper-
ation to be performed directly on the compressed data. We have deployed this technique in a couple of
scientific applications, where the number of elements per attribute vector reaches hundreds of millions
to a billion elements. We propose here to use this base of software to the problem of indexing micro-
array spot data.
High performance clustering methods. In this third area, we will be implementing a clustering algorithm
named RACHET into our work environments. RACHET builds a global hierarchy by merging clustering
hierarchies generated locally at each of the distributed data sites and is especially suitable for very large,
high-dimensional, and horizontally distributed datasets. Its time, space, and transmission costs are at most
linear in the size of the dataset. (This includes only the complexity of the transmission and agglomeration
phases and does not include the complexity of generating local clustering hierarchies.)
Clustering of multidimensional data is a critical step in many fields including data mining, statistical data
analysis, pattern recognition and image processing. Current popular clustering approaches do not offer a
solution to the distributed hierarchical clustering problem that meets all these requirements. Most cluster-
ing (Murtagh, 1983; Day and Edelsbrunner, 1984; Jain et al., 1999) are restricted to the centralized data sit-
uation that requires bringing all the data together in a single, centralized warehouse. For large datasets, the
transmission cost becomes prohibitive. If centralized, clustering massive centralized data is not feasible in
practice using existing algorithms and hardware. RACHET makes the scalability problem more tractable.
This is achieved by generating local clustering hierarchies on smaller data subsets and using condensed
cluster summaries for the consecutive agglomeration of these hierarchies while maintaining the clustering
quality. Moreover, RACHET has significantly lower (linear) communication costs than traditional central-
ized approaches.
In summary, this project will require an infrastructure that enables easy integration of new methods and
ideas and supports biology collaborators at multiple sites so they can interact as well as access to data, high
performance computation, and storage resources.
We are grateful to the vision and support of the DOE Office of Science, the sponsor of the Genomes to
Life program. The DOE Genomes to Life (GTL) program is unique in that it calls for “well-integrated, mul-
tidisciplinary (e.g. biology, computer science, mathematics, engineering, informatics, biphysics, biochem-
istry) research teams,” with strong encouragement to “include, where appropriate, partners from more than
one national laboratory and from universities, private research institutions, and companies.” Such guidance
is essential to the success of the GTL program in meeting its four ambitious goals. To this end, our effort
includes participants from four DOE laboratories (Sandia National Laboratories, Oak Ridge National Lab-
oratory, Lawrence Berkley National Laboratory, and Los Alamos National Laboratory), four universities
(California/San Diego, Michigan, California/Santa Barbara, and Illinois/Urbana/Champaign), and three in-
stitutes (The National Center for Genomic Resources, The Molecular Science Institute, and the Joint Insti-
tute for Computational Science).
BATEMAN, A., BIRNEY, E., CERRUTI, L., et al. (2002). The Pfam protein families database. Nucleic Acids Res.
30, 276–280.
BEDU, S., LAURENT, B., and JOSET, F. (ed.) (1992). Membranes and Soluble Carbonic Anhydrase in a Cyanobac-
terium. Synechocystis PCC6803, Vol. III. (Kluwer Academic Publishers, Dordrecht, The Netherlands).
CANNON, G.C., and SHIVELY, J.M. (1983). Characterization of a homogeneous preparation of carboxysome from
Thiobacillus neapolitanus.Arch. Microbiol. 134, 52–59.
CANNON, G.C., BRADBURNE, C.E., ALDRICH, H.C., et al. (2001). Microcomartments in prokaryotes: carboxysomes
and related polyhedra. Appl. Env. Microbiol. 67, 5351–5361.
CHANG, G., and ROTH (C.B.) 2001, Structure of MSBA from E. coli: a homolog of the multidrug resistance ATP
binding cassette (ABC) transporters. Science 293, 1793–1800.
DANDEKAR, T., SNEL, B., HUYNEN, M., et al. (1998). Conservation of gene order: a fingerprint of proteins that
physically interact. Trends Biochem. Sci. 23, 324–328.
DAY, W., and EDELSBRUNNER, H. (1984). Efficient algorithms for agglomerative hierarchical clustering methods.
J. Classificiation 1, 7–24.
ENGLISH, R.S., LORBACH, S.C., QIN, X., et al. (1994). Isolation and characterization of a carboxysome shell gene
from Thiobacillus neapolitanus.Mol. Microbiol. 12, 647–654.
ENRIGHT, A.J., ILIOPOULOS, I., KYRPIDES, N.C., et al. (1999). Protein interaction maps for complete genomes
based on gene fusion events. Nature 402, 86–90.
FALZONE, C.J., KAO, Y.H., ZHAO, J.D., et al. (1994). Three-dimensional solution structure of PSAE form the
cyanobacterium Synechococcus sp. strain PCC-7002, a photosystem-i protein that shows structural homology with
SH3 domains. Biochemistry 33, 6052–6062.
FIELDS, S., and SONG, O. (1989). A novel genetic system to detect protein–protein interactions. Nature 340, 245–246.
FINLEY, R.L., and BRENT, R. (1994). Interaction mating reveals binary and ternary connections between drosophila
cell cycle regulators. Proc. Natl. Acad. Sci. USA 91, 12980–12984.
FRIDLYAND, L., KAPLAN, A., and REINHOLD, L. (1996). Quantitative evaluation of the role of a putative CO2-
scavenging entity in the cyanobacterial CO2-concentrating mechanism. Biosystems 37, 229–238.
FRIEDMAN, N., LINIAL, M., NACHMAN, I., et al. (2000). Using Bayesian networks to analyze expression data.
J. Comput. Biol. 7, 601–620.
GARCIA, A.E., and SANBONMATSU, K.Y. (2001). Exploring the energy landscape of a beta hairpin in explicit sol-
vent. Proteins–Structure Funct. Genet. 42, 345–354.
GAVIN, A.C., BOSCHE M., KRAUSE R., et al. (2002). Functional organization of the yeast proteome by systematic
analysis of protein complexes. Nature 415, 141–147.
GIBSON, M.A., and BRUCK, J. (2000). Efficient exact stochastic simulation of chemical systems with many species
and many channels. J. Phys. Chem. 104, 1876–1889.
GILLESPIE, D. (1976). A general method for numerically simulating the stochastic time evolution of coupled chemi-
cal reactions. J. Comp. Phys. 22, 403–434.
GOMEZ, S.M., LO S.-H., and RZHETSKY A. (2001). Probabilistic prediction of unknown metabolic and signal-
transduction network. Genetics 159, 1291–1298.
GOMEZ, S.M., and RZHETSKY, A. (2002). Towards the prediction of complete protein–protein interaction networks.
R.B. Altman, A.K. Dunker, L. Hunter, K. Lauderdale, and T.E. Klein, eds. Pacific Symposium on Biocomputing (Li-
hue, Hawaii, World Scientific), 413–424.
HU, Y., GRUHLER, A., HEILBUT, A., et al. (2002). Systematic identification of protein complexes in Saccharomyces
cerevisiae by mass spectrometry. Nature 415, 180–183.
HOCH, J.A., and SILHAVY T.J. (1995). Two-component signal transduction
. (Washington, DC, ASM Press).
HOLTHUIJZEN, Y.A., VANBREEMEN, J.F.L., KUENEN, J.G., et al. (1986). Protein composition of the carboxysomes
of Thiobacillus neapolitanus.Arch. Microbiol. 144, 398–404.
HUYNEN, M., SNEL, B., LATHE, W., et al. (2000). Predicting protein function by genomic context: quantitative eval-
uation and qualitative inferences. Genome Res. 10, 1204–1210.
JAIN, A.K., MURTY, M.N., and FLYNN, P.J., (1999). Data clustering: a review. ACM Computing Surveys 31, 264–323.
JEONG H., TOMBOR B., ALBERT, R., et al. (2000). The large-scale organization of metabolic networks. Science
407, 651–654.
KAPLAN, A., FRIEDBERG, R., SCHWARZ, R., et al. (1989). The “CO2concentrating mechanism” of cyanobacte-
ria: physiological, molecular and theoretical studies. Photosynth. Res. 17, 243–255.
KAPLAN, A., and REINHOLD, L. (1999). CO2concentrating mechanisms in photosynthetic microorganisms. Annu.
Rev. Plant Physiol. Plant Mol. Biol. 50, 539–570.
KYODA, K.M., MOROHASHI, M., ONAMI, S., et al. (2000). A gene network inference method from continuous-
value gene expression data of wild-type and mutants. Genome Inform. 11, 196–204.
LANARAS, T., HAWTHORNTHWAITE, A.M., and CODD, A. (1985). Localization of carbonic anhydrase in the
cyanobacterium Chlorogloeopsis fritschii. FEMS Microbiol. Lett. 26, 285–288.
MARCOTTE, E.M., PELLEGRINI, M., THOMPSON, M.J., et al. (1999). A combined algorithm for genome-wide pre-
diction of protein function. Nature 402, 83–86.
MARTIN, M.G. (2000). Available:
MARTIN, M.G., and SIEPMANN, J.I. (1999). Novel configurational-bias Monte Carlo method for branched mole-
cule—transferable potentials for phase equilibria—2. United-atom description of branched alkanes. J. Phys. Chem.
B 103, 4508–4517.
MITSUTAKE, A., SUGITA, Y., and OKAMOTO, Y. (2001). Generalized-ensemble algorithms for molecular simula-
tions of biopolymers. Biopolymers Pept. Sci. 60, 96–123.
MURTAGH, F. (1983). A survey of recent advances in hierarchical clustering algorithms. Compu. J. 26, 354–359.
NINFA, A.J., and ATKINSON, M.R. (1995). Control of nitrogen assimilation by the NRI-NRII two component sys-
tem of enteric bacteria. Two-Component Signal Transduction. J.A. Hoch and T.J. Silhavy, eds. (ASM Press, Wash-
ington, D.C.).
OLMAN, V., XU, D., and XU, Y. (2002). A new framework for biological data clustering using minimum spanning
trees. Proceedings of the 7t h Pacific Symposium on Biocomputing (in press).
OVERBEEK R., LARSEN N., PUSCH G.D., et al. (2000). WIT: integrated system for high-throughput genome se-
quence analysis and metabolic reconstruction. Nucleic Acids Res. 28, 123–125.
PELLEGRINI, M., MARCOTTE, E.M., THOMPSON, M.J., et al. (1999). Assigning protein functions by comparative
genome analysis: protein phylogenetic profiles. Proc. Natl. Acad. Sci. USA 96, 4285–4288.
PRATT, L.A., and SILHAVY, T.J. (1995). Porin regulon in Esherichia coli. Two-Component Signal Transduction. J.A.
Hoch and T.J. Silhavy, eds. (ASM Press, Washington, D.C.).
PRICE, G.D., COLEMAN, J.R., and BADGER, M.R. (1992). Association of carbonic anhydrase activity with car-
boxysomes isolated from the cyanobacterium Synechococcus PCC7942. Plant Physiol. 100, 784–793.
PRIM, R.C. (1957). Shortest connection networks and some generalizations. Bell System Tech. J. 36, 1389–1401.
PUIG, O., CASPARY, F., RIGAUT, G., et al. (2001). The tandem affinity purification method: a general procedure of
protein complex purification. Methods 24, 218–229.
RODI, D.J., JANES, R.W., SANGANEE, H.J., et al. (1999). Screening of a library of phage-displayed peptides iden-
tifies human Bcl-2 as a taxol-binding protein. J. Mol. Biol. 285, 197–203.
SANBONMATSU, K.Y., and GARCIA, A.E. (2002). Structure of Met-enkephalin in explicit aqueous solution using
replica exchange molecular dynamics. Proteins–Structure Funct. Genet. 46, 225–234.
SHADID, J., HUTCHINSON, S., HENNIGAN, G., et al. (1997). Efficient parallel computation of unstructured finite
element reacting flow solutions. Parallel Comput. 23, 1307–1325.
SHIVELY, J.M., LORBACH, S.C., JIN, S., et al. (1996). Carboxysomes: the genes of Thiobacillus neapolitanus. Lid-
strom, M.E., and Tabita, F.R., eds. Microbial Growth on C1 Compounds. (Kluwer, Dordrecht, The Netherlands), 56–63.
SMITH, G.P., and PETRENKO, V.A. (1997). Phage display. Chem. Rev. 97, 391–410.
STEPHANOPOULOS, G., ARISTIDOU, A.A., and NIELSEN, J. 1998. Metabolic engineering. Biotechnol. Bioeng.
58, 119–120.
STILES, J., and BARTOL, T. (2001). Monte Carlo methods for simulating realistic synaptic microphysiology using
Mcell. De Schutter, E., eds. Computational Neuroscience: Realistic Modeling for Experimentalists(CRC Press, Boca
Raton), 87–127.
TAKAI-IGARASHI, T., and KAMINUMA, T. (1999). A pathway finding system for the cell signaling networks data-
base. In Silico Biol. 1, 129–146.
TONG, A.H.Y., DREES, B., NARDELLI, G., et al. (2002). A combined experimental and computational strategy to
define protein interaction networks for peptide recognition modules. Science 295, 321–324.
UETZ, P., GIOT, L., CAGNEY, G., et al. (2000). A comprehensive analysis of protein–protein interactions in Sac-
charomyces cerevisiae. Nature 403, 623–627.
VAN HELDEN, J., NAIM, A., MANCUSO, R., et al. (2000). Representing and analysing molecular and cellular func-
tion using the computer. Biol. Chem. 381, 921–935.
WOROBEY, M., and HOLMES, E.C. (1999). Evolutionary aspects of recombination in RNA viruses. J. Gen. Virol.
80, 2535–2543.
WU, K., OTOO, E.J., and SHOSHANI, A. (2001). A performance comparision of bitmap indexes. ACM Int. Conf. In-
formation Knowledge Management 559–561.
WU, K., OTOO, E.J., and SHOSHANI, A. (2002). Compressed Bitmap Indexes for Faster Search Operations. LBNL
Technical Report, LBNL-49627.
XU, Y., OLMAN, V., and XU, D. (2002). A graph-theoretical approach for gene expression data analysis: an applica-
tion of minimum spanning trees. Bioinformatics 18, 526–535.
Address reprint requests to:
Dr. Grant S. Heffelfinger
Sandia National Laboratories
Bldg 701/2101, MS-0885
1515 Eubank SE
Albuquerque, NM 87123
This article has been cited by:
1. Shouqiang Cheng, Yu Liu, Christopher S. Crowley, Todd O. Yeates, Thomas A. Bobik. 2008. Bacterial
microcompartments: their properties and paradoxes. BioEssays 30:11-12, 1084-1095. [CrossRef]
2. Dylan M. Morris, Grant J. Jensen. 2008. Toward a Biomechanical Understanding of Whole Bacterial
Cells. Annual Review of Biochemistry 77:1, 583-613. [CrossRef]
3. G.X. Yu, E.M. Glass, N.T. Karonis, N. Maltsev. 2006. Knowledge-based voting algorithm for
automated protein functional annotation. Proteins: Structure, Function, and Bioinformatics 61:4,
907-917. [CrossRef]
... Additional 522 prokaryotic genomes and 436 eukaryotic genomes are listed as ongoing sequence projects. Knowledge about protein components, functional capacities, and overall metabolic potentials of these genomes will advance progress toward a comprehensive understanding of the genetic mechanisms of diverse biochemical processes [2]. The challenge is that the experiments needed to determine biological functionalities for the composed gene components in the sequenced genomes are labor-intensive and very expensive. ...
Automated annotation of high-throughput genome sequences is one of the earliest steps toward a comprehensive understanding of the dynamic behavior of living organisms. However, the step is often error-prone because of its underlying algorithms, which rely mainly on a simple similarity analysis, and lack of guidance from biological rules. We present herein a knowledge-based protein annotation algorithm. Our objectives are to reduce errors and to improve annotation confidences. This algorithm consists of two major components: a knowledge system, called “RuleMiner,” and a voting procedure. The knowledge system, which includes biological rules and functional profiles for each function, provides a platform for seamless integration of multiple sequence analysis tools and guidance for function annotation. The voting procedure, which relies on the knowledge system, is designed to make (possibly) unbiased judgments in functional assignments among complicated, sometimes conflicting, information. We have applied this algorithm to 10 prokaryotic bacterial genomes and observed a significant improvement in annotation confidences. We also discuss the current limitations of the algorithm and the potential for future improvement. Proteins 2005. © 2005 Wiley-Liss, Inc.
... (7,10) Overall, carboxysomes are widely distributed among bacteria, which play a major role in the global carbon cycle, thus underscoring the importance of the carboxysome and the bacterial CCM. (7,87) Ecology of organisms having Pdu microcompartments Genes for B 12 -dependent 1,2-propanediol degradation are found in a number of bacterial genera including Salmonella, Klebsiella, Shigella, Yersinia, Listeria, Lactobacillus, Lactococcus and at least one species of Escherichia coli (E24377A). In all cases examined, pdu gene clusters are highly conserved indicating that microcompartments are essential to the physiology of 1,2-propanediol degradation. ...
Many bacteria conditionally express proteinaceous organelles referred to here as microcompartments (Fig. 1). These microcompartments are thought to be involved in a least seven different metabolic processes and the number is growing. Microcompartments are very large and structurally sophisticated. They are usually about 100-150 nm in cross section and consist of 10,000-20,000 polypeptides of 10-20 types. Their unifying feature is a solid shell constructed from proteins having bacterial microcompartment (BMC) domains. In the examples that have been studied, the microcompartment shell encases sequentially acting metabolic enzymes that catalyze a reaction sequence having a toxic or volatile intermediate product. It is thought that the shell of the microcompartment confines such intermediates, thereby enhancing metabolic efficiency and/or protecting cytoplasmic components. Mechanistically, however, this creates a paradox. How do microcompartments allow enzyme substrates, products and cofactors to pass while confining metabolic intermediates in the absence of a selectively permeable membrane? We suggest that the answer to this paradox may have broad implications with respect to our understanding of the fundamental properties of biological protein sheets including microcompartment shells, S-layers and viral capsids.
... As a consequence, host organisms can survive and prosper in a diversity of hostile environments. Identification of " key " genes that are responsible for these processes is an essential step toward understanding the genetic and biochemical basis of entire life processes in general, and biologically important biochemical processes related to sustainable sources of energy, environmental management, carbon cycle and sequestration, and human health protection in particular [4]. However, such a task presents a tremendous challenge for both experimental and computational scientists because of the complexity of gene networks involved in these processes. ...
This paper presents a novel algorithm for identification and functional characterization of "key" genome features responsible for a particular biochemical process of interest. The central idea is that individual genome features are identified as "key" features if the discrimination accuracy between two classes of genomes with respect to a given biochemical process is sufficiently affected by the inclusion or exclusion of these features. In this paper, genome features are defined by high-resolution gene functions. The discrimination procedure utilizes the Support Vector Machine classification technique. The application to the oxygenic photosynthetic process resulted in 126 highly confident candidate genome features. While many of these features are well-known components in the oxygenic photosynthetic process, others are completely unknown, even including some hypothetical proteins. It is obvious that our algorithm is capable of discovering features related to a targeted biochemical process.
The history of the Monte Carlo for complex chemical systems Towhee open source Monte Carlo molecular simulation tool is discussed. A proof is given that the Widom insertion method for computing the chemical potential is formally correct even when combined with the most general version of arbitrary trial distribution configurational-bias Monte Carlo. A simulation strategy for computing single component vapour–liquid phase coexistence curves is presented as a guide for inexperienced practitioners of Monte Carlo simulations. A review of papers that cite the Towhee code is presented. The paper concludes with a discussion about releasing and sustaining a simulation package that uses an open source software license.
Conference Paper
Oak Ridge National Laboratory (ORNL) leads two of the five big Genomes-to-Life projects funded in the USA. As a part of these projects researchers at ORNL have been using PVM to build a computational biology grid that spans the USA. This talk will describe this effort, how it is built, and the unique features in PVM that led the researchers to choose PVM as their framework. The computations such as parallel BLAST are run on individual supercomputers or clusters within this P2P grid and are themselves written in PVM to exploit PVM’s fault tolerant capabilities. We will then describe our recent progress in building an even more adaptable distributed virtual machine package called Harness. The Harness project includes research on a scalable, self-adapting core called H2O, and research on fault tolerant MPI. Harness software framework provides parallel software “plug-ins” that adapt the run-time system to changing application needs in real time. This past year we have demonstrated Harness’ ability to self-assemble into a virtual machine specifically tailored for particular applications. Finally we will describe DOE’s plan to create a National Leadership Computing Facility, which will house a 100 TF Cray X2 system, and a Cray Red Storm at ORNL, and an IBM Blue Gene system at Argonne National Lab. We will describe the scientific missions of this facility and the new concept of “computational end stations” being pioneered by the Facility.
Conference Paper
Petascale Virtual Machines (PVM) continues to be a popular software package both for creating personal grids and for building adaptable, fault tolerant applications. We will illustrate this by describing a computational biology environment built on top of PVM that is used by researchers around the world. We will then describe or recent progress in building an even more adaptable distributed virtual machine package called Harness. The Harness project includes research on a scalable, self- adapting core called H2O, and research on fault tolerant Message Passing Interface (MPI). The H2O core can be configured to support distributed web computing like SETI@home, parallel virtual machines like PVM, and OGSA compliant grid environments. The Harness software is being designed to scale to petascale virtual machines. We will describe work at Oak Ridge National Laboratory (ORNL) on developing algorithms for such petascale virtual machines and the development of a simulator to test these algorithms on simulated 100,000 processor systems.
The genomic era brought with it the capacity to unlock complex interactions in organisms and biological systems. Currently, by exploiting genomic and associated protein information through in silico analyses, postgenomic research is developing rapidly. This field, which encompasses functional genomics, structural genomics, transcriptomics, pharmacogenomics, proteomics and metabolomics, allows for a systems-wide approach to biological studies. To date, bacterial postgenomic research has focused mainly on a few representative pathogenic species, leaving the vast majority of the microbial community relatively overlooked. Amongst the under-represented microorganisms are the cyanobacteria, which are important for their beneficial natural product production, bioremediation and energy applications. Here, we highlight the current status of cyanobacterial postgenomic research and assess the potential for future metabolic engineering and "cell factory" or "microbial cell" development.
Full-text available
Phosphorus is an essential element for all life forms. However, it is limiting in most ecological environments where cyanobacteria inhabit. Elucidation of the phosphorus assimilation pathways in cyanobacteria will further our understanding of the physiology and ecology of this important group of microorganisms. However, a systematic study of the Pho regulon, the core of the phosphorus assimilation pathway in a cyanobacterium, is hitherto lacking. We have predicted and analyzed the Pho regulons in 19 sequenced cyanobacterial genomes using a highly effective scanning algorithm that we have previously developed. Our results show that different cyanobacterial species/ecotypes may encode diverse sets of genes responsible for the utilization of various sources of phosphorus, ranging from inorganic phosphate, phosphodiester, to phosphonates. Unlike in E. coli, some cyanobacterial genes that are directly involved in phosphorus assimilation seem to not be under the regulation of the regulator SphR (orthologue of PhoB in E coli.) in some species/ecotypes. On the other hand, SphR binding sites are found for genes known to play important roles in other biological processes. These genes might serve as bridging points to coordinate the phosphorus assimilation and other biological processes. More interestingly, in three cyanobacterial genomes where no sphR gene is encoded, our results show that there is virtually no functional SphR binding site, suggesting that transcription regulators probably play an important role in retaining their binding sites. The Pho regulons in cyanobacteria are highly diversified to accommodate to their respective living environments. The phosphorus assimilation pathways in cyanobacteria are probably tightly coupled to a number of other important biological processes. The loss of a regulator may lead to the rapid loss of its binding sites in a genome.
Carboxysomes are organelle-like polyhedral bodies found in cyanobacteria and many chemoautotrophic bacteria that are thought to facilitate carbon fixation. Carboxysomes are bounded by a proteinaceous outer shell and filled with ribulose 1,5-bisphosphate carboxylase/oxygenase (RuBisCO), the first enzyme in the CO(2) fixation pathway, but exactly how they enhance carbon fixation is unclear. Here we report the three-dimensional structure of purified carboxysomes from Synechococcus species strain WH8102 as revealed by electron cryotomography. We found that while the sizes of individual carboxysomes in this organism varied from 114 nm to 137 nm, surprisingly, all were approximately icosahedral. There were on average approximately 250 RuBisCOs per carboxysome, organized into three to four concentric layers. Some models of carboxysome function depend on specific contacts between individual RuBisCOs and the shell, but no evidence of such contacts was found: no systematic patterns of connecting densities or RuBisCO positions against the shell's presumed hexagonal lattice could be discerned, and simulations showed that packing forces alone could account for the layered organization of RuBisCOs.
Full-text available
Many heterotrophic bacteria have the ability to make polyhedral structures containing metabolic enzymes that are bounded by a unilamellar protein shell (metabolosomes or enterosomes). These bacterial organelles contain enzymes associated with a specific metabolic process (e.g. 1,2-propanediol or ethanolamine utilization). We show that the 21 gene regulon specifying the pdu organelle and propanediol utilization enzymes from Citrobacter freundii is fully functional when cloned in Escherichia coli, both producing metabolosomes and allowing propanediol utilization. Genetic manipulation of the level of specific shell proteins resulted in the formation of aberrantly shaped metabolosomes, providing evidence for their involvement as delimiting entities in the organelle. This is the first demonstration of complete recombinant metabolosome activity transferred in a single step and supports phylogenetic evidence that the pdu genes are readily horizontally transmissible. One of the predicted shell proteins (PduT) was found to have a novel Fe-S center formed between four protein subunits. The recombinant model will facilitate future experiments establishing the structure and assembly of these multiprotein assemblages and their fate when the specific metabolic function is no longer required.
Full-text available
Many microorganisms possess inducible mechanisms that concentrate CO2 at the carboxylation site, compensating for the relatively low affinity of Rubisco for its substrate, and allowing acclimation to a wide range of CO2 concentrations. The organization of the carboxysomes in prokaryotes and of the pyrenoids in eukaryotes, and the presence of membrane mechanisms for inorganic carbon (Ci) transport, are central to the concentrating mechanism. The presence of multiple Ci transporting systems in cyanobacteria has been indicated. Certain genes involved in structural organization, Ci transport and the energization of the latter have been identified. Massive Ci fluxes associated with the CO2-concentrating mechanism have wide-reaching ecological and geochemical implications.
In complex systems with many degrees of freedom such as peptides and proteins, there exists a huge number of local-minimum-energy states. Conventional simulations in the canonical ensemble are of little use, because they tend to get trapped in states of these energy local minima. A simulation in generalized ensemble performs a random walk in potential energy space and can overcome this difficulty. From only one simulation run, one can obtain canonical-ensemble averages of physical quantities as functions of temperature by the single-histogram and/or multiple-histogram reweighting techniques. In this article we review uses of the generalized-ensemble algorithms in biomolecular systems. Three well-known methods, namely, multicanonical algorithm, simulated tempering, and replica-exchange method, are described first. Both Monte Carlo and molecular dynamics versions of the algorithms are given. We then present three new generalized-ensemble algorithms that combine the merits of the above methods. The effectiveness of the methods for molecular simulations in the protein folding problem is tested with short peptide systems. © 2001 John Wiley & Sons, Inc. Biopolymers (Pept Sci) 60: 96–123, 2001
Polyhedral inclusion bodies have been observed in all of the cyanobacteria thus far examined and in many, but not all, chemoautotrophic bacteria (Codd 1988, Shively 1974, Shively et al. 1988, Shively,English 1991). The bodies were first isolated from Thiobacillus neapolitanus and shown to consist of a monolayer shell surrounding multiple molecules of the enzyme ribulose bisphosphate carboxylase/oxygenase, RuBisCO (Shively et al. 1973a,b). The inclusions were subsequently named carboxysomes. Presumably, all of the polyhedral inclusion bodies of similar appearance are carboxysomes. To date, evidence has been gathered which identifies the bodies of Nitrosomonas sp., several Nitrobacter sp., Anabaena cylindrica, Chlorogleopsis fritschii, Anacystis nidulans, Synechococcus sp., Synechocystis sp., Prochloron, and Prochlorothrix as carboxysomes (Shively,English 1991).
Pfam is a large collection of protein multiple sequence alignments and profile hidden Markov models. Pfam is available on the World Wide Web in the UK at, in Sweden at, in France at and in the US at The latest version (6.6) of Pfam contains 3071 families, which match 69% of proteins in SWISS-PROT 39 and TrEMBL 14. Structural data, where available, have been utilised to ensure that Pfam families correspond with structural domains, and to improve domain-based annotation. Predictions of non-domain regions are now also included. In addition to secondary structure, Pfam multiple sequence alignments now contain active site residue mark-up. New search tools, including taxonomy search and domain query, greatly add to the functionality and usability of the Pfam resource.
90% of the carbonic anhydrase (CA) activity recovered from Chlorogloeopsis fritschii cells, when broken under conditions which favour the isolation of carboxysomes, is particulate. Subsequent sucrose density gradient centrifugation of the carboxysome-containing pellet produced a sharp band of CA, well separated from the carboxysomes and thylakoids. The implications of these findings for the possible functions of carboxysomes and location of CA are discussed