Advanced Computational Structural Genomics
1. Policy Needs for Computational Structural Genomics
1.1. Societal Needs for Advanced Simulation Research in Biology
The pace of extraordinary advances in molecular biology has advanced in the past
decade due in large part to discoveries coming from genome projects on human and model
organisms. The advances in the genome project so far, happening well ahead of schedule and
under budget, have exceeded any dreams by its protagonists, let alone formal expectations.
Biologists expect the next phase of the genome project to be even more startling in terms of
dramatic breakthroughs in our understanding of human biology, the biology of health and of
disease.
Why can we not wait for a computational science effort until the conclusion of the
genome project, and what are the broader implications? Significant personal and economic
costs will be borne by society for delays in our exploiting the discoveries of the genome
projects on behalf of the Nation. The importance of each individual, each human being on the
planet, the individual's productivity and contribution to society, the quality of that individual's
life including that of our shared environment, has transcendental value. A major break
through in technology, the human genome project, puts us on the brink of truly extraordinary
progress in realizing the transcendental goal of human well-being. Only today can biologists
begin to envision the necessary experimental, computational and theoretical steps necessary to
exploit genome sequence information for its medical impact, its contribution to biotechnology
and economic competitiveness, and its ultimate contribution to environmental quality.
The experimental (left) and computational (right) hierarchies will increasingly become
codependent as the research community models greater biological complexity.
Individualized medicine- the recognition of individual differences in drug and treatment
response, in disease development and progression, in the appearance and thus diagnosis of
disease- justifies a much more aggressive approach to utilizing DNA sequence information
and the introduction of simulation capabilities to extract the implicit information contained in
the human genome. Each year thousands die and a hundred times more people suffer adverse
reactions of various extents to drugs that are applied in perfectly correct usage, dose and
disease specificity. Cancer in particular stands out as a disease of the genes. Two patients with
what appear to be identical cancers (based on cellular pathology) at the same stage of
malignancy, growth and dissemination will have very different responses to the same therapy
regime; for example, one could rapidly fail to respond to treatment, become terminal in
weeks, and the other could respond immediately and ultimately recover fully.
Re-engineering microbes for bioremediation will depend directly on the understanding
derived from the simulation studies proposed here. The design of new macromolecules and
the redesign of microbe metabolism based on simulation studies, truly grand challenges for
applied biology, will contribute to a wide range of environmental missions for the
Department, including the environmental bioremediation of the Nation's most contaminated
sites, often mixed waste for which no economically feasible cleanup technology currently
exists. Many aspects of sustainable development, changing the nature of industrial processes
to use environmentally friendly processes, depend on the same kinds of advances. Similarly,
in the longer term the understanding of living systems will also contribute to environmental
research on carbon management.
To exploit the inherent genome information derived from knowing the DNA sequences,
computational advances, along with related experimental biotechnology, are essential.
Knowing the sequence of the DNA does not tell us about the function of the genes,
specifically the actions of their protein products - where, when, why, how the proteins act is
the essence of biological knowledge required. Encoded in the DNA sequence is a protein's
three-dimensional topography, which in turn determines function; uncovering this sequencestructure-function relationship is the core goal of modern structural biology today. The goal of
the Advanced computational structural genomics initiative is to link sequence, structure and
function, and to move from analysis of individual macromolecules to macromolecular
assemblies and complex oligomeric interactions that make up the complex processes within
the cell.
1.2 Elucidating the Path from Protein Structure, to Function, to Disease
A particular sequence derived from the genome encodes in a character string the threedimensional structure of a protein that performs a specific function in the cell. The primary
outcome of the Argonne Workshop on Structural Genomics emphasized the experimental goal
of determining the relationship between sequence and structure at a level unattainable before
the coordinated effort to map and sequence the genomes of many organisms. It is in fact a
logical extension of the genome effort to systematically elaborate DNA sequences into full
three dimensional structures and functional analysis; this new effort will require the same
level of cooperation and collaboration among scientists as was necessary in the original
genome project.
The Structural Genomics initiative is gaining momentum and will primarily focus initially on
the infrastructure and support necessary to realize high throughput experimental structures by
x-ray crystallography and high field NMR. Success of this effort is in some sense guaranteed,
and the computational challenges that are posed when a structural genomics initiative of this
scale is contemplated is already clear cut. It is the outgrowth from this direction, a
computational genomics initiative- that will capture everything from sequence, structure and
functional genomics to genetic networks and metabolic engineering to forward folding of
macromolecules and kinetics to cellular level interactions. This approach makes all the
modeling and simulation of macromolecules fair game and moves toward the complex
systems approach, the complicated level of real living systems rather than individual
macromolecules.
The sequencing of microbial genomes containing several thousand genes and the eventual
completion of human and other model organism genomes is the underlying driving force for
understanding biological systems at a whole new level of complexity. There is for the first
time the potential to understand living organisms as whole, complex dynamic systems and to
use large-scale computation to simulate their behavior. Modeling all levels of biological
complexity is well beyond even the next generation of teraFLOP computers, but each
increment in the computing infrastructure makes it possible to move up the biological
complexity ladder and solve previously unsolvable classes of problems.
2. Scientific Scope of Computational Structural Genomics
2.1 The First Step Beyond the Genome Project:
High-Throughput Genome Assembly, Modeling, and Annotation
The first step in the biological hierarchy is a comprehensive genome-based analysis of the
rapidly emerging genomic data. With changes in sequencing technology and methods, the rate
of acquisition of human and other genome data over the next few years will be ~100 times
higher than originally anticipated. Assembling and interpreting these data will require new
and emerging levels of coordination and collaboration in the genome research community to
develop the necessary computing algorithms, data management and visualization system.
Annotation the elucidation and description of biologically relevant features in the sequence
is essential in order for genome data to be useful. The quality with which annotation is done
will have direct impact on the value of the sequence. At a minimum, the data must be
annotated to indicate the existence of gene coding regions and control regions. Further
annotation activities that add value to a genome include finding simple and complex repeats,
characterizing the organization of promoters and gene families, the distribution of G+C
content, and tying together evidence for functional motifs and homologs.
Table I shows many high-priority computational challenges associated with the analysis,
modeling and annotation of genome data. The first is the basic assembly and interpretation of
the sequence data itself (analysis and annotation) as it is produced at increasing rates over the
next five years. There will be a never-ending race to keep up with this flow, estimated at
about 200 million base pairs per day by some time in 1999, and to execute the required
computational codes to locate and understand the meaning of genes, motifs, proteins, and
genomes as a whole. Cataloging the flood of genes and proteins and understanding their
relationship to one another, their variation between individuals and organisms, and evolution
represents a number of very complex computational tasks. Many calculations such as multiple
sequence alignments, phylogenetic tree generation, and pedigree analysis are NP-hard
problems and cannot be currently done at the scale needed to understand the body of data
being produced, unless a variety of shortcuts is introduced. Related to this is the need to
compare whole genomes to each other on many levels both in terms of nucleic acid and
proteins and on different spatial scales. Large-scale genome comparisons will also permit
biological inference of structure and function.
Table I. Current and Expected Sustained Capability Requirements for Major Community
Genomics Codes
Problem Class
Sequence assembly
Binary sequence
comparison
Multiple sequence
comparison
Gene modeling
Phylogeny trees
Protein family
classification
Sustained Capability
1999
>1012 flops
1012 flops
Sustained Capability
2000
1014 flops
>1014 flops
1012 flops
>1014 flops
>1015 flops
1011 flops
>1010 flops
1017 flops
1013 flops
1012 flops
Sequence Comparisons Against Protein Families for Understanding Human Pathology
Searching a protein sequence database for homologues is a powerful tool for discovering the
structure and function of a sequence. Amongst the algorithms and tools available for this task,
Hidden Markov model (HMM)-based search methods improve both the sensitivity and
selectivity of database searches by employing position-dependent scores to characterize and
build a model for an entire family of sequences.
HMMs have been used to analyze proteins using two complementary strategies. In the first, a
sequence is used to a search a collection of protein families, such as Pfam, to find which of
the families it matches. In the second approach an HMM for a family is used to search a
primary sequence database to identify additional members of the family. The latter approach
has yielded insights into protein involved in both normal and abnormal human pathology such
as Fanconi Anaemia A, Gaucher disease, Krabbe disease, polymyositis scleroderma and
disaccharide intolerance II.
HMM-based analysis of the Werner Syndrome protein sequence (WRN) suggested it
possessed exonuclease activity, and subsequent experiments confirmed the prediction [1].
Like WRN, mutation of the protein encoded by the Klotho gene lead to a syndrome with
features resembling ageing. However, Klotho is predicted to be a member of the family 1
glycosidase (see Figure). Eventually, large-scale sequence comparisons against HMM models
for protein families will require enormous computational resources to find these sequencefunction correlations over genome-scale databases.
The similarities and differences between two plant and archeal members of a family of
glycosidases that includes a protein implicated in ageing. Ribbons correspond to the betastrands and alpha-helices of the underlying TIM barrel (red) and family 1 glycosidase domain
(cyan). Amino acid side chains drawn in magenta, yellow and green are important for
structure and/or function. The loop in yellow denotes a region proposed to be important for
substrate recognition. The 2-deoxy-2-fluorglucosyl substrate bound at the active site of one of
the enzymes is shown with carbon atoms in grey, oxygen in red and fluorine in green.
2.2 From Genome Annotation to Protein Folds: Comparative Modeling and Fold Assignment
The key to understanding the inner workings of cells is to learn the three-dimensional atomic
structures of the some 100,000 proteins that form their architecture and carry out their
metabolism. These three-dimensional (3D) structures are encoded in the blueprint of the DNA
genome. Within cells, the DNA blueprint is translated into protein structures through
exquisitely complex machinery- itself composed of proteins. The experimental process of
deciphering the atomic structures of the majority of cellular proteins is expected to take a
century at the present rate of work. New developments in comparative modeling and fold
recognition will short circuit this process, that is we can learn to translate the DNA message
by computer.
Protein fold assignment.
A genome-encoded amino acid sequence (center) is tested for compatibility with a library of
known 3D protein folds[2]. An actual library would contain of the order of 1000 folds; the
one shown here illustrates the most common protein folding motifs. There are two possible
outcomes of the compatibility test: that the sequence is most compatible with one of the
known folds or that the sequence is not compatible with any known fold. The second outcome
may mean either that the sequence belongs to one of the folds not yet discovered, or that the
compatibility measures are not fully developed to detect the distant relationship of sequence
to its structure.
The goal of fold assignment and comparative modeling is to assign each new genome
sequence to the known protein fold or structure that it most closely resembles, using
computational methods. Fold assignment and comparative modeling techniques can then be
helpful in proposing and testing hypotheses in molecular biology, such as in inferring
biological function, predicting the location and properties of ligand binding sites, in designing
drugs, testing remote protein-protein relationships. It can also provide starting models in Xray crystallography and NMR spectroscopy.
The success of these methods rests on a fundamental experimental discovery of structural
biology: the 3D structures of proteins have been better conserved during evolution than their
genome sequences. When the similarity of a target sequence to another sequence with known
structure is above a certain threshold, comparative modeling methods can often provide
quantitatively accurate protein structure predictions, since a small change in the protein
sequence usually results in a small change in its 3D structure. Even when the percentage
identity of a target sequence falls below this level, then at least qualitative information about
the overall fold topology can often be predicted.
Several fundamental issues remain to amplify the effectiveness of fold assignment and
comparative modeling. A primary issue in fold assignment is the determination of better
multipositional compatibility functions which will extend fold assignment into the "twilight
zone" of sequence homology. In both fold assignment and comparative modeling, better
alignment algorithms that deal with multipositional compatibility functions are needed. A
move toward detailed empirical energy functions and increasingly sophisticated optimization
approaches in comparative modeling will occur in the future.
The availability of tera-scale computational power will further extend fold assignment in the
following ways: the alignment by dynamic programming is performed in order L2 time, where
L is the length of the sequence/structure. For 100,000's of sequences and 10,000's of
structures (each of order 102 amino acids long), an all to all comparison of sequence to
structure using dynamic programming would require on order of 1013 operations. The use of
multipositional functions that will scale as L3 or L4 depending on the complexity of the
compatibility function, and may require 1015 to 1017 FLOPS, and will likely need an effective
search strategy in addition. The next generation of 100 teraFLOP computers addresses both
fundamental issues in reliable fold assignment. First, an increase in complexity in alignment
algorithms for multipositional functions that scale beyond L2, and secondly the development
of new multipositional compatibility functions whose parameters are derived by training on
large databases with multiple iterations, resulting in an increase in sensitivity and specificity
of these new models.
Availability of tera-flop computing will greatly benefit comparative protein structure
modeling of both single proteins and the whole human genome. Once an alignment is
determined, comparative modeling is formally a problem in molecular structure optimization.
The objective function is similar in complexity to typical empirical protein force fields used in
ab initio global optimization prediction and protein folding, and scales as M2, where M is the
number of atoms in the model. A typical calculation for a medium sized protein takes in the
order of 1012 Flops. More specifically, the increased computing power is likely to improve the
accuracy of comparative models by alleviating the alignment problem and the loop modeling
problem, at the very least.
It is probable that the effect of alignment errors will decrease by performing independent
comparative modeling calculations for many different alignments. This would in essence
correspond to a conformational search with soft constraints imposed by the alignment
procedure. Such a procedure would increase the computational cost to 1015 Flops for a single
protein, and to 1020 Flops for the human genome. Specialized loop modeling procedures can
be used after the initial comparative models are obtained by standard techniques, typically
needing about 5000 energy function evaluations to obtain one loop conformation. It is
standard to calculate an "ensemble" of 25-400 loop conformations for each loop sequence.
Thus, the total number of function evaluations is on the order of 106 for prediction of one loop
sequence. Applying these procedures to the whole human genome would take on the order of
1018 Flops.
Large-Scale Comparative Modeling of Protein Structures of the Yeast Genome
Recently, a large-scale comparative protein structure modeling of the yeast genome was
performed[3]. Fold assignment, comparative protein structure modeling, and model
evaluation were completely automated. As an illustration, the method was applied to the
proteins in the Saccharomyces cerevisiae (baker's yeast) genome. It resulted in all-atom 3D
models for substantial segments of 1071 (17%) of the yeast proteins, only 40 of which have
had their 3D structure determined experimentally. Of the 1071 modeled yeast proteins, 236
were related clearly to a protein of known structure for the first time; 41 of these have not
been previously characterized at all. Many of the models are sufficiently accurate to facilitate
interpretation of the existing functional data as well as to aid in the construction of mutants
and chimeric proteins for testing new functional hypotheses. This study shows that
comparative modeling efficiently increases the value of sequence information from the
genome projects, although it is not yet possible to model all proteins with useful accuracy.
The main bottlenecks are the absence of structurally defined members in many protein
families and the difficulties in detection of weak similarities, both for fold recognition and
sequence-structure alignment. However, while only 400 out of a few thousand domain folds
are known, the structure of most globular folds is likely to be determined in less than ten
years. Thus, comparative modeling will conceivably be applicable to most of the globular
protein domains close to the completion of the human genome project.
Large-scale protein structure modeling. A small sample of the 1,100 comparative models
calculated for the proteins in the yeast genome is displayed over an image of a yeast cell.
1.3 Low Resolution Folds to Structures with Biochemical Relevance: Toward Accurate
Structure, Dynamics, and Thermodynamics
As we move to an era of genetic information at the level of complete genomes, classifying the
fold topology of each sequence in the genome is a vital first step toward understanding gene
function. However, the ultimate limitation in fold recognition is that these algorithms only
provide "low-resolution" structures. It is crucial to enhance and develop methods that permit a
quantitative description of protein structure, dynamics and thermodynamics, in order relate
specific sequence changes to structural changes, and structural changes to associated
functional/phenotypic change.
These more accurate approaches will greatly improve our ability to modify proteins for novel
uses such as to change the catalytic specificity of enzymes and have them degrade harmful
waste products. While often tertiary fold of proteins involved in disease change dramatically
on mutation, those whose fold remains invariant, may have quantitative differences in
structure that can have important macroscopic effects on function that can be manifested as
disease. More accurate screening for new drug targets that bind tightly to specific protein
receptors for inhibition will require quantitative modeling of protein/drug interactions.
Therefore the next step is the quantitative determination of protein structure starting from the
fold prediction, and ultimately directly from sequence.
Estimate of the number of updates of the molecular dynamics algorithmic kernal that can be
accomplished for a 100 amino acid protein and 3000 water molecules with 100 tflop
resources.
The effort possible from a Scientific Simulation Initiative of 100 terascale means that time
and size scales become accessible for the first time for predicting and folding protein
structure, and simulating protein function and thermodynamics. The simulation of a protein
folding trajectory of a 100 amino acid protein and 3000 water molecules would require 10121015 updates of an empirical energy function (and its derivatives) that scales as N2, where N
will range from 10,000 to 100,000. Many folding trajectories will likely be needed to
understand the process on how proteins self-assemble. Optimization approaches for
generating a diversity in biochemically relevant structures and misfolded conformers will take
on the order of 108-1010 updates depending on the number of conformations and
improvements in algorithms. The simulations of relative free energies important for
determining drug binding affinities can take on the order of 106 steps, and the thermodynamic
derivatives >106 updates.
Simulation of the Molecular Mechanism of Membrane Binding of Viral Envelope
Proteins
The first step of infection by enveloped virus proteins involves the attachment of a viral
envelope glycoprotein to the membrane of the host cell. This attachment leads to fusion of
viral and host cell membranes, and subsequent deposition of viral genetic material into the
cell. It is known that the envelope protein, hemagglutinin, exists under normal conditions in a
"retracted" state where the peptide that actually binds to the sialylated cell surface receptor is
buried some 100Å from the distal tip of the protein, and is therefore not capable of binding to
the cell membrane. However, at low pH a major conformational change occurs whereby the
fusion peptide is delivered (~100Å !) via a "spring-loaded mechanism" to the distal tip where
it is available for binding to the cell membrane.
While the crystal structures of the inactive (neutral pH) and active (low pH) forms are known,
the molecular details of the large-scale conformational change that produces the activated
hemagglutinin, and of the binding of the fusion peptide to the cell surface receptors, are not
known. As many aspects of this process are likely prototypical of a class of enveloped viruses
including HIV, a detailed understanding of the molecular mechanism would be beneficial in
guiding efforts to intervene before the viral infection is completed.
Atomistic simulations of the process of viral binding to cell membranes are extremely
demanding computationally [4]. Taken together, the hemagglutinin, lipid membrane, and
sufficient water molecules to solvate the system in a periodically replicable box adds up to
more than 100,000 atoms, an order of magnitude larger than what is presently considered a
large biomolecular system. Furthermore, the large scale conformational changes in the
protein, and displacements of lipids as the protein binds the membrane involve severe time
scale bottlenecks as well.
X-ray crystal structure of the fusion-activated (low pH) form of the influenza virus
hemagglutinin poised above a small (256 lipids) patch of a lipid membrane from an MD
simulation. The N-terminal peptides that actually bind to the membrane are not present in this
hemagglutinin structure. An all-atom model of this complex plus solvent in a periodically
replicated box would contain >100,000 atoms, and long time scale motions would need to be
simulated to understand this step in the mechanism of viral infection.
2.4 Advancing Biotechnology Research: In Silico Drug Design
The robust prediction of protein structure ultimately ties directly into our ability to rationally
develop highly specific therapeutic drugs and to understand enzymatic mechanisms.
Calculations of affinities with which drug molecules bind to proteins that are important in
metabolic reactions and molecular signaling processes, and the mechanisms of catalytic
function of certain key enzymes, can be obtained with computational approaches described
here. However, there are many challenges to the simulations of these biological processes that
arise from the large molecular sizes and long time scales relevant for biological function, or
the subtle energetics and complex milieu of biochemical reactions. Some of these same
computational methods and bottlenecks arise in the evaluation of the binding affinity of drugs
to specific targets, but the task of the drug designer is further complicated by the need to
identify a small group of compounds out of a virtually limitless universe of combinatorial
possibilities.
Only a small percentage of a pool of viable drug candidates actually lead to the identification
of a clinically useful compound, with typically over $200 million spent in research costs to
successfully bring it to market. On average, a period of 12 years elapses between the
identification and FDA approval of a successful drug, with the major bottleneck being the
generation of novel, high-quality drug candidates. While rational, computer-based methods
represent a quantum leap forward for identifying drug candidates, substantial increases in
compute power are needed to allow for both greater selection sensitivity and genome-scale
modeling of future drugs.
Table 2 below gives an estimate of the method and current computational requirements to
complete a binding affinity calculation for a given drug library size. Going down the table for
a given model complexity is a level of computational accuracy. The benefits of improved
model accuracy must be offset against the cost of evaluating a model over the ever increasing
size of the drug compound library brought about by combinatorial synthesis, and further
exacerbated by the high throughput efforts of the genome project and structural annotation of
new protein targets.
Table 2. Estimates of current computational requirements to complete a
binding affinity calculation for a given drug library size.
Modeling complexity
Method
Size of
library
Required
computing
time
Molecular Mechanics SPECITOPE 140,000
~1 hour
Rigid ligand/target
LUDI
30,000
1-4 hours
CLIX
30,000
33 hours
Molecular Mechanics Hammerhead 80,000
3-4 days
Partially flexible
ligand
DOCK
17,000
3-4 days
Rigid target
DOCK
53,000
14 days
50,000
21 days
1
~several
days
1
>several
weeks
Molecular Mechanics ICM
Fully flexible ligand,
Rigid target
Molecular Mechanics AMBER
Free energy perturbation
CHARMM
QM Active site and
Gaussian,
MM protein
Q-Chem
State-of-the-art quantum chemistry algorithms can expand the applicability of QM/MM
methods to simulate a greatly expanded QM subsystem for enzymatic studies, or even
estimation of drug binding affinities. An estimate of the cost of using QM methods for
evaluating a single energy and force evaluation system of 104 heavy atoms, would require
resources that can handle ~1016 FLOPS; on a 100 teraFLOP machine this would require five
minutes. Parallel versions of these methods have been implemented and are available at a
number of universities and DOE laboratories. On current generation teraFLOP platforms, QM
calculations of unprecedented size are now possible, allowing HF optimizations on systems
with over 1000 atoms and MP2 energies on hundreds of atoms. The presently available
traditional-scaling (~N3) first principles molecular dynamics code is running efficiently on
serial platforms, including high-end workstations and vector supercomputers. The ultimate
goal is to develop linear scaling quantum molecular dynamics code. It will be important to
adapt these codes to the parallel architectures, requiring rewriting parts of existing programs,
and developing linear scaling algorithms
2.5 Linking Structural Genomics to Systems Modeling: Modeling the Cellular Program
The cellular program that governs the growth, development, environmental response, and
evolutionary context of an organism does so robustly in the face of a fluctuating environment
and energy sources. It integrates numerous signals about events the cell must track in order to
determine which reactions to turn on, off, or slow down and speed up. These signals, which
are derived both from internal processes, other cells, and changes in the extracellular medium,
arrive asynchronously, and are multivalued in meaning. The cellular program also has
memory of its own particular history as written in the complement and concentrations of
chemicals contained in the cell at any instant. The circuitry that implements the working of a
cell and/or collection of cells is a network of interconnected biochemical, genetic reactions,
and other reaction types.
The experimental task of mapping genetic regulatory networks using genetic footprinting and
two-hybrid techniques is well underway, and the kinetics of these networks is being generated
at an astounding rate. Technology derivatives of genome data such as gene expression microarrays and in vivo fluorescent tagging of proteins through genetic fusion with the GFP protein
can be used as a probe for network interaction and dynamics. If the promise of the genome
projects and the structural genomics effort is to be fully realized, then predictive simulation
methods must be developed to make sense of this emerging experimental data. First is the
problem of modeling the network structure, i.e. the nodes and connectivity defined by sets of
reactions among proteins, small molecules and DNA. Second is the functional analysis of a
network using simulation models built up from "functional units" describing the kinetics of
the interactions.
Prediction of networks from genomic data can be approached from several directions. If the
function of a gene can be predicted from homology, then prior knowledge of the pathways in
which that function is found in other organisms can be used to predict the possible
biochemical networks in which the protein participates. Homology approaches based on
protein structural data or functional data for a protein previously characterized can be used to
predict the type of kinetic behavior of a new enzyme. Thus structural prediction methods that
can predict the fold of the protein product of a given gene are fundamental to the deduction of
the network structure.
The non-linearity of the biochemical and genetic reactions, along with the high degree of
connection (sharing of substrates, products and effectors) among these reactions, make the
qualitative analysis of their behavior as a network difficult. Furthermore, the small numbers of
molecules involved in biochemical reactions (typical concentrations of 100 molecules/cell)
ensure that thermal fluctuations in reaction rates are expected to become significant compared
to the average behavior at such low concentrations. Since genetic control generally involves
only one or two copies of the relevant promoters and genes per cell, this noise is expected to
be even worse for genetic reactions. The inherent randomness and discreteness of these
reactions can have significant macroscopic consequences such as that common inside living
cells.
There are three bottlenecks in the numerical analysis of biochemical reaction networks. The
first is the multiple time scales involved. Since the time between biochemical reactions
decreases exponentially with the total probability of a reaction per unit time, the number of
computational steps to simulate a unit of biological time increases roughly exponentially as
reactions are added to the system or rate constants are increased. The second bottleneck
derives from the necessity to collect sufficient statistics from many runs of the Monte-Carlo
simulation to predict the phenomenon of interest. The third bottleneck is a practical one of
model building and testing: hypothesis exploration, sensitivity analyses, and back
calculations, will also be computationally intensive.
Modeling the Reaction Pathway for the Glycolytic Biochemical System
A novel gene expression time series analysis algorithm known as the Correlation Metric
Construction uses a time lagged correlation metric as a measure of distance between reacting
species. The constructed matrix R is then converted to a Euclidean distance matrix D, and
multidimensional scaling, MDS, is used to allow the visualization of the configuration of
points in high dimensional space as a two dimensional stick and ball diagram. The goal of this
algorithm is to deduce the reaction pathway underlying the response dynamics, and was used
on the first few steps of the glycolytic pathway determined by experiment[5].
The reconstituted reaction system of the glycolytic pathway, containing eight enzymes and 14
metabolic intermediates, was kept away from equilibrium in a continuous-flow, stirred-tank
reactor. Input concentrations of adenosine monophosphate and citrate were externally varied
over time, and their concentrations in the reactor and the response of eight other species were
measured. The CMC algorithm showed a good prediction of the reaction pathway from the
measurements in this much-studied biochemical system. Both the MDS diagram itself and
the predicted reaction pathway resemble the classically determined reaction pathway. In
addition, CMC measurements yield information about the underlying kinetics of the network.
For example, species connected by small numbers of fast reactions were predicted to have
smaller distances between them than species connected by a slow reaction.
Plot of the time-lagged correlation function of G6P with all other species. The experimentally
determined lagged correlation functions. The graph clarifies the temporal ordering data
inherent in the correlation functions. (B) The 2D projection of the MDS diagram. Each point
represents the calculated time series of a given species. The closer two points are, the higher
the correlation between the respective time series. Black (gray) lines indicate negative
(positive) correlation between the respective species. (C) Predicted reaction pathway derived
from the CMC diagram. Its correspondence to the known mechanism is high.
3. Implicit Collaborations Across the DOE Mission Sciences
Computational modeling plays a key role in all scientific understanding. Therefore it is
probably not surprising that disparate scientific disciplines share common modeling strategies,
algorithmic bottlenecks, and information technology issues, and use computer scientists and
applied mathematicians as a common resource. In this Section we emphasize what are the
cross-cutting computational problems whose solutions could benefit a large number of
scientific applications, or where technologies should be transferred between the computational
biology domain and other disciplines.
3.1 Computer Hardware
The advanced computational biology simulations described in this white paper will require
computer performance well beyond what is currently available, but computational speed alone
will not ensure that the computer is useful for any specific simulation method. Several other
factors are critical including the size of primary system memory (RAM), the size and speed of
secondary storage, and the overall architecture of the computer. This will undoubtedly be true
for global climate, combustion, and other basis energy science areas.
Many parallel processing computer architectures have been developed over the years, but the
dominant parallel architecture that has emerged is the distributed memory, multiple
instruction-multiple data architecture (MIMD), which consists of a set of independent
processors with their own local RAM memory interconnected by some sort of communication
channel. Such an architecture is characterized by the topology and speed of the
interconnection network, and by the speed and memory size of the individual processors. All
of the current generation of teraFLOP-class computers, including ASCI Red and Blue are of
this design.
Just as with the computer hardware, there have been a large number of software programming
paradigms developed for parallel computers. A great deal of research has gone into
developing software tools to assist in parallel programming or even to automatically
parallelize existing single-processor software. Selected parallelization and debugging tools
can assist and new programming models such as Object-Oriented programming (using C++ or
FORTRAN90) can help hide the details of the underlying computer architecture. At the
current time, however, efficiently using massively parallel computers primarily involves
redesigning and rewriting software by hand. This is complicated by the facts that the best
serial (single processor) algorithms are often not the best suited for parallel computers and the
optimal choice of algorithm often depends on the details of the computer hardware.
The different simulation methods presently used have different requirements of parallel
computer hardware. Simulation methods which involve calculating averaged properties from
a large number of smaller calculations that can be individually run on gigaFLOP class
processors are most ideally suited for parallelism. These methods include classical and
quantum Monte Carlo. In these simulations a minimal amount of initial data can be sent to
each processor which then independently calculates a result that is communicated back to a
single processor. By choosing an appropriate size of problem for each single processor
(problem granularity), these algorithms will work efficiently on virtually any MIMD
computer, including separate computer workstations linked by local-area networks.
The quantum chemical and molecular dynamics methods, or certain optimization algorithms,
in which all processors are applied to the calculation of a single chemical wave function or
trajectory, involve much greater challenges to parallelization and involve greater constraints
on the parallel computer architecture. Since all processors are involved in a single
computation, interprocessor communication must occur. It is the rate of this communication,
characterized in term of raw speed (bandwidth) and initialization time (latency) that usually
limits the efficient use of parallel computers. The minimal necessary communication rate
depends exquisitely on the simulation type, choice of algorithm and problem size. Generally,
it is essential software design criteria that as the problem scales to larger size, the ratio of
computational operations per communication decrease (or at least remain constant), so that for
some problem size, the communication rate will not constitute a bottleneck. Moreover, it is
important that the work per processor, or "load balance", scale evenly so that no processors
end up with much larger computational loads and become bottlenecks. In a broad sense, the
nature of computational biology simulations–in particular the physical principle that
interactions attenuate with distance–will ensure that scalable parallel algorithms can be
developed, albeit at some effort.
Even given the very broad range of simulation methods required by computational biology, it
is possible to provide some guidelines for the most efficient computer architectures.
Regarding the size of primary memory, it is usually most efficient if a copy of the (6xN) set of
coordinates describing a time step of a molecular dynamics simulation or the (NxN) matrices
describing the quantum chemical wavefunction, can be stored on each processing element.
For the biological systems of the sort described in this white paper, this corresponds to a
minimum of several hundred megabytes of RAM per processor. Moreover, since many of the
simulation methods involve the repeated calculation of quantities that could be stored and
reused (e.g. two electron integrals in quantum chemistry or interaction lists in molecular
dynamics), memory can often be traded for computer operations so that larger memory size
will permit even larger simulations.
Similarly, general estimates can be made for the minimal interprocessor communication rates.
Since the goal of parallel processing is to distribute the effort of a calculation, for tightly
coupled methods such as quantum chemical simulations, it is essential that the time to
communicating a partial result be less than the time to simply recalculate it. For example, the
quantum chemical two-electron integrals require 10-100 floating point operations to calculate,
so that they can be usefully sent to other processors only if that requires less than ~100 cycles
to communicate to send the 8 or 16 byte result. Assuming gigaFLOP speeds for individual
processing elements in the parallel computers, this translates roughly to gigabyte/sec
interprocessor communication speeds. (Note that many partial results involve vastly more
operations, so that they place a much weaker constraint on the communication rate.
3.2 Information Technologies and Database Management
"Biology is an Information Science" [6] and the field is poised to put into practice new
information science and data management technologies directly. Two major conferences are
emerging within the field of computational biology (ISMB - Intelligent Systems for
Molecular Biology; and RECOMB - Research in Computational Biology). Each year,
associated workshops focus on how to push new techniques from computer science into use in
computational biology. For example, at ISMB-94 a workshop focused on problems involved
in integrating biological databases; follow-up workshops in 1995 and 1996 explored CORBA
and java as methods toward integration solutions. In 1997, a post-conference workshop
focused on issues in accurate, usable annotations of genomes. This year, a pre-conference
workshop will explore how to use text-processing and machine-translation methods for
building ontologies to support cross-linking between databases about organisms. All of these
criteria require ultra-high-speed networks to interconnect students, experimental biologists,
and computational biologists and publicly funded data repositories. This community will, for
example, benefit directly from every new distributed networking data exchange tool that
develops as a result of Internet-II and the high-speed Energy Sciences network.
Data warehousing addresses a fundamental data management issue: the need to transparently
query and analyze data from multiple heterogeneous sources distributed across an enterprise
or a scientific community. Warehousing techniques have been successfully applied to a
multitude of business applications in the commercial world. Although the need for this
capability is as vital in the sciences as in business, functional warehouses tailored for specific
scientific needs are few and far between. A key technical reason for this discrepancy is that
our understanding of the concepts being explored in an evolving scientific domain change
constantly, leading to rapid changes in data representation. When the format of source data
changes, the warehouse must be updated to read that source or it will not function properly.
The bulk of these modifications involve extremely tedious, low-level translation and
integration tasks which typically require the full attention of both database and domain
experts. Given the lack of the ability to automate this work, warehouse maintenance costs are
prohibitive, and warehouse "up-times" severely restricted. This is the major roadblock to a
successful warehouse to for scientific data domains. Regardless of whether the scientific
domain is genome, combustion, high energy physics, or climate modeling, the underlying
challenges for data management are similar, and present in varying degrees for any
warehouse. We need to move towards the automation of these scientific tasks. Research will
play a vital role in achieving that goal and in scaling warehousing approaches to dynamic
scientific domains. Warehouse implementations have an equally important role; they allow
one to exercise design decisions, and provide a test-bed that stimulates the research to follow
more functional and robust paths.
3.3 Ensuring Scalability on Parallel Architectures
In all of the research areas, demonstrably successful parallel implementations must be able to
exploit each new generation of computer architectures that will rely on increased number of
processors to realize multiple teraFLOP computing. We see the use of the paradigm/software
support tools as an important component of developing effective parallelization strategies that
can fully exploit the increased number of processors that will comprise a 100 teraFLOP
computing resource. However, these paradigm/software support libraries have largely been
developed on model problems at a finer level of granularity than "real life" computational
problems. But it is the "real life" problems that involve complexity in length scales, time
scales, and severe scalings of algorithmic kernals that are in need of the next and future
generations of multiple teraFLOP computing. The problems described in this initiative
provide for a more realistic level of granularity to investigate the improved use of software
support tools for parallel implementations, and may improve the use of high computing
resources by the other scientific disciplines, such as climate, combustion, and materials, that
need scalable algorithms.
Even when kernals can be identified and parallel algorithms can be designed to solve them,
often the implementation does not scale well with the number of processors. Since 100
teraFLOP computing will only be possible on parallel architectures, problems in scalability is
a severe limitation in realizing the goals of Advanced Computational Structural Genomics.
Lack of scalability often arises from straightforward parallel implementations where the
algorithm is controlled by a central scheduler and for which communication among
processors is wired to be synchronous. In the computer science community, various
paradigms and software library support modules exist for exploring better parallel
implementations. These can decompose the requirements of a problem domain into high-level
modules, each of which is efficiently and portably supported in a software library that can
address issues involving communication, embedding, mapping, etc for a scientific application
of interest.
These tools allow different decompositions of a parallel implementation to be rapidly
explored, by handling all of the low-level communication for a given platform. However,
these paradigms and their supporting software libraries usually are developed in the context of
model mathematical applications, which are at a finer level of granularity than "real life"
computational science problems. A unique opportunity exists to use some of the
computational science problems described in the previous chapters to refine or redefine the
current parallel paradigms currently in use. The outcome of this direction could be broader
than the particular scientific application, and may provide insight in how to improve the use
of parallel computing resources in general.
3.4 Meta Problem Solving Environments
In order to take advantage of teraFLOP computing power, biological software design must
move towards an efficient geographically distributed software paradigm. One novel paradigm
is the "Meta Problem Solving Environment", a software system based on the "plug and play"
paradigm. Algorithms are encapsulated into specially designed software modules called
components which the user can link together produce a simulation. Software components can
exist anywhere on a large computing network and yet can be employed to perform a specified
simulation on any geographically distributed parallel computing environment to which the
user has access. In order to satisfy this design requirement, software components must run as
separate threads, respond to signals of other components and, in general, behave quite unlike a
standard subroutine library. Therefore, the software architecture must specify the rules for a
component to identify standard communication interfaces as well the rules that govern how a
component responds to signals. Also, components must be permitted to specify resources and
obtain communication lines.
Although the component based "Meta Problem Solving Environment" software paradigm is
far more complex than any current software system employed by the biological community
today, it has many advantages. By coming together to produce uniform component design
protocols, computational biologists could quickly apply the latest algorithmic advances to
novel new systems. Also, simulations could be constructed using the best available
components. For example, real space force component from one researcher can be combined
with a reciprocal space component from another and an integration component from another
group, and could significantly increase computational efficiency compared to any single
biological computing package. In addition, with good design protocols, all the components
employed in a given simulation need not be built by computational biologists. For instance,
parallel FFT component software already exists that could be used to improve significantly
SPME performance. Finally, the biological community will not be acting alone in adopting a
component based approach. The DOE ASCI project is currently using using tools such as
CORBA (Common Object Request Broker Architecture), and Los Alamos National
Laboratory's PAWS (Parallel Application Workspace) to produce component based software
products. Therefore, with appropriate communication within the community and with others,
biological computational scientists can begin to produce a new software standard that will
function well in large geographically distributed environment and very quickly lead to
teraFLOP computing capability.
3.5 Usability Evaluation and Benchmarking
In the last five years, there has been major progress in benchmarking massively and highly
parallel machines for scientific applications. For example, the NAS Parallel Benchmarks
(NPB) were a major step toward assessing the computational performance of early MPPs, and
served well to compare MPPs with parallel vector machines. Based on numerical algorithms
derived from CFD and aerospace applications, the NPB soon were widely adopted as an
intercomparison metric of parallel machine performance. Subsequently the NPB were
included in the international benchmarking effort, PARKBENCH (PARallel Kernels and
BENCHmarks). In 1996, portable HPF and MPI versions of the NPB were made available.
With increasing industry emphasis on parallelism, a number of other organizations, as part of
their procurement process, are designing benchmark suites to evaluate the computational
performance of highly parallel systems.
The Advanced Computational Structural Genomics program will involve both research and
development activities and implementation of comprehensive biological modeling
simulations. Therefore, its success will depend not only on raw computational performance
but also on an effective environment for code development and analysis of model results.
These issues suggest that a new benchmarking approach is needed, one that evaluates not only
raw computational performance but also system usability. In addition, there should be formal
tracking of benchmark results throughout the course of the program to provide an objective
measure of progress. To meet these objectives, the Advanced Computational Structural
Genomics will support benchmarking and performance evaluation in four areas:
Identify or create scalable, computationally intensive benchmarks that reflect the production
computing needs of the particular scientific community or common algorithm, and evaluate
the scalability and capability of high performance systems.
Establish system performance metrics appropriate for scientific modeling.
Establish usability metrics for assessing the entire research/ development/ analysis
environment.
Establish a repository of benchmark results for measuring progress in the initiative and for
international comparison.
Identify or create scalable, computationally intensive benchmarks that reflect the needs of the
scientific community. The purpose of this activity is to develop benchmarks that can be
shared by the community to gain insight into the computational performance of machines that
will be proposed by vendors in response to the Scientific Simulation Initiative. It is well
known that the GFlops performance level achieved on an application is not a reliable measure
of scientific throughput. For example, an algorithm may achieve a very high floating point
rate on a certain computer architecture, but may be inefficient at actually solving the problem.
Therefore, accurate and useful benchmarks must be based on a variety of numerical methods
that have proven themselves efficient for various modeling problems. Benchmarks for
scientific computing applications in general have traditionally used either a functional
approach or a performance-oriented approach. In the functional approach, a benchmark
represents a workload well if it performs the function as the workload, e.g., a modeling
workload is represented by a subset of computational biology codes. In the performanceoriented approach, a benchmark represents a workload well if it exhibits the same
performance characteristics as the workload, e.g., a structural genomics modeling workload
would be represented by a set of kernels, code fragments, and a communication and I/O test.
The assumption is that the system performance can be predicted from an aggregate model of
the performance kernels.
We propose that Advanced Computational Structural Genomics use mainly a functional
approach and identify or develop a set of benchmark models. The benchmark suite also must
include a range of simulation models typically found in the computational biological
community. These benchmarks should be developed in the spirit of the NAS Parallel
Benchmarks (the CFD applications benchmarks): the models should exhibit all the important
floating point, memory access, communication, and I/O characteristics of a realistic model,
yet the code should be scalable, portable, compact, and relatively easy to implement. As
mentioned above, this is a substantial effort, because many extant benchmark codes for some
scientific modeling are not readily portable, or are not currently well suited or even capable of
running on more than a few tens of processors.
Establish system performance metrics appropriate for scientific modeling. We propose that
Advanced Computational Structural Genomics select a few full-scale scientific applications,
and intercompare parallel systems using the wall clock seconds per model application.
Establish usability metrics and system attributes for assessing the entire
research/development/analysis environment. The benchmarks in (1) focus on PE performance
alone and do not measure any of the important system features that contribute to the overall
utility of a parallel system. These features include code development tools, batch and
interactive management capabilities, I/O performance, performance of the memory hierarchy,
compiler speed and availability, operating system speed and stability, networking interface
performance, mean time between failures or interrupts, etc. Many of these features have been
quite weak on current highly parallel computers as compared to vector supercomputers.
We propose that the Computational Structural Genomics area develop a new set of utility
benchmarks and attributes for highly and massively parallel machines as well as clusters of
SMPs. The goal of these benchmarks is to derive a set of metrics that measure the overall
utility of machines and systems for a productive scientific modeling environment. Such an
environment would emphasize model flexibility, rapid job turnaround, and effective
capabilities for analysis and data access. Also, the system attributes will specify a set of
desired capabilities necessary to administer and manage the system and its resources. At this
time there is no industry standard benchmark which addresses these issues. All scientific
disciplines coupled to the SSI can drive the state of the art in performance evaluation by
initiating such an effort.
Establish a repository of benchmark results for measuring progress in the initiative and for
international comparison. Benchmarking codes, metrics, and the detailed benchmark results
from the activities in (1), (2), and (3) must be available to the public and the high performance
computing community through frequent publication of reports, Web pages, etc., with the goal
of influencing the computer industry. The Basic will immediately benefit from using these
benchmarks for the next-generation procurement of production machines. will create a central
repository of benchmark codes and performance data. It will also publish annual reports about
the measured performance gains and accomplishments of the initiative, as well as providing
international comparison data.
We expect that the rapid development of new high-performance computing technology will
continue, with rapidly changing architectures and new software and tools becoming available.
Hence a benchmark must be designed to be able to evaluate systems across several hardware
generations, diverse architectures, and changing software environments. In return, we believe
that a well-designed benchmark can drive the HPC industry and create a better focus in the
industry on high-end computing.
Acknowledgements
A web-based version of this document, and the more comprehensive white paper on
computational biology can be found at:
http://cbcg.lbl.gov/ssi-csb/
We would like to thank the following individuals who contributed to the intellectual content of
this document:
2.1. The First Step Beyond the Genome Project: High-Throughput Genome Assembly,
Modeling, and Annotation
Phil LaCascio, Oakridge National Laboratory
Saira Mian, Lawrence Berkeley National Laboratory
Richard Mural, Oakridge National Laboratory
Frank Olken, Lawrence Berkeley National Laboratory
Jay Snoddy, Oakridge National Laboratory
Sylvia Spengler, Lawrence Berkeley National Laboratory
David States, Washington University, Institute for Biomedical Computing
Ed Uberbacher, Oakridge National Laboratory
Manfred Zorn , Lawrence Berkeley National Laboratory
2.2. From Genome Annotation to Protein Folds: Comparative Modeling and Fold Assignment
David Eisenberg, University of California, Los Angeles
Barry Honig, Columbia University
Alan Lapedes, Los Alamos National Laboratory
Andrej Sali, Rockefeller University
2.3. Low Resolution Folds to Protein Structures with Biochemical Relevance: Toward
Quantitative
Structure, Dynamics, and Thermodynamics
Charles Brooks, Scripps Research Institute
Yong Duan, University of California, San Francisco
Teresa Head-Gordon, Lawrence Berkeley National Laboratory
Volkhard Helms, University of California, San Diego
Peter Kollman, University of California, San Francisco
Glenn Martyna, Indiana University
Andy McCammon, University of California, San Diego
Jeff Skolnick, Scripps Research Institute
Doug Tobias, University of California, Irvine
2.4. Biotechnology Advances from Computational Structural Genomics: In Silico Drug
Design
Ruben Abagyan, NYU, Skirball Institute
Jeff Blaney, Metaphorics, Inc.
Fred Cohen, University of California, San Francisco
Tack Kuntz, University of California, San Francisco
2.5. Linking Structural Genomics to Systems Modeling: Modeling the Cellular Program
Adam Arkin, Lawrence Berkeley National Laboratory
Denise Wolf, Lawrence Berkeley National Laboratory
Peter Karp, Pangea
3.1. Computer Hardware
Mike Colvin, Lawrence Livermore National Laboratory
3.2. Information Technologies and Database Management
Terry Gaasterland, Argonne National Laboratory
Charles Musick, Lawrence Livermore National Laboratory
3.3. Ensuring Scalability on Parallel Architectures.
Silvia Crivelli, Lawrence Berkeley National Laboratory
Teresa Head-Gordon, Lawrence Berkeley National Laboratory
3.4. Meta Problem Solving Environments
Glenn Martyna, Indiana University
3.5. Usability Evaluation and Benchmarking
Horst Simon, Lawrence Berkeley National Laboratory
References
Huang, S. R., Li, B. M.; Gray, M. D.; Oshima, J.; et al. (1998). The Premature Ageing
Syndrome Protein, Wrn, Is A 3'->5' Exonuclease. Nature Genetics 20, 114-116.
Fischer, D.; Rice, D.; Bowie, J. U.; Eisenberg, D. (1996). Assigning Amino Acid Sequences
To 3-Dimensional Protein Folds. Faseb Journal 10, 126-136.
Sanchez, R.; Sali, A. (1998). Large-Scale Protein Structure Modeling Of The Saccharomyces
Cerevisiae Genome. Proceedings Of The National Academy Of Sciences USA 95, 1359713602.
Tobias, D. J.; Tu, K. C.; Klein, M. L. (1997). Atomic-Scale Molecular Dynamics Simulations
Of Lipid Membranes. Current Opinion In Colloid & Interface Science 2, 15-26.
Arkin, A; Shen, P. D.; Ross, J. (1997). A Test Case Of Correlation Metric Construction Of A
Reaction Pathway From Measurements. Science 277, 1275-1279
Gaasterland, T. (1998). Structural Genomics: Bioinformatics In The Driver's Seat. Nature
Biotechnology 16, 625-627.
Appendix: 1. Glossary
Aggregation information: The information obtained from applying (a set of) operators to the
initial data. For example, monthly sales or expenses, average length of a phone call, or
maximum number of calls during a given time.
Amino acid: Any of a class of 20 molecules that are combined to form proteins in living
organisms. The sequence of amino acids in a protein and hence protein function are
determined by the genetic code.
API (application program interface): An API is a collection of methods that are used by
external programs to access and interact with a class or library. Their main function is to
provide a consistent interface to a library, isolating programs from changes in the library
implementation or functionality.
Automatic Schema Integration: Determination of identical objects represented in the schemata
of different databases by programs using a set of built-in rules, without user intervention.
Base pair: A unit of information carried by the DNA molecule. Chemically these are purine
and pyrimidine complementary bases connected by weak bonds. Two strands of DNA are
held together in a shape of a double helix by the bonds between paired bases.
Bioinformatics: Field of study dealing with management of data in biological sciences.
Change detection: The process of identifying when a data source has changed and how. The
types of changes detected include: new data, modifications to existing data, and schema
changes.
Chromosome: A self-replicating genetic structure in the cells, containing the cellular DNA
that bears in its nucleotide sequence the linear array of genes.
Conflicts: When the same concept is represented in different databases, its representation may
be semantically and syntactically different. For example, a length may be defined in different
units, or contain a different set of attributes. These differences are called conflicts, and must
be resolved in order to integrate the data from the different sources.
Data cleaning: Removal and/or correction of erroneous data introduced by data entry errors,
expired validity of data, or by some other means.
Data ingest: Loading of data into the warehouse.
Data/schema integration: Merging of data/schema from different sources into a single
repository, after resolving incompatibilities among the sources.
Data mining: Analysis of raw data to find interesting facts about the data and for purposes of
knowledge discovery.
Data warehouse: An integrated repository of data from multiple, possibly heterogeneous data
sources, presented with consistent and coherent semantics. Warehouses usually contain
summary information represented on a centralized storage facility.
Distributed database: A set of geographically distributed data connected by a computer
network and controlled by a single DBMS running at all the sites.
DNA (deoxyribonucleic acid): The molecule that encodes genetic information, a double
stranded heteropolymer composed of four types of nucleotides, each containing a different
base. See base pair.
Federated databases: An integrated repository data from of multiple, possibly heterogeneous,
data sources presented with consistent and coherent semantics. They do not usually contain
any summary data, and all of the data resides only at the data source (i.e. no local storage).
Gene: A fundamental physical and functional unit of heredity. A gene is an ordered sequence
of nucleotides located in a particular position on a particular chromosome that encodes a
specific functional product (i.e., a protein or RNA molecule). See gene expression.
Gene expression: The process by which a gene's coded information is converted into the
structures present and operating in the cell. Expressed genes include those that are transcribed
into mRNA and then translated into protein and those that are transcribed into RNA
molecules only.
Genetic code: The sequence of nucleotides, coded in triplets along the mRNA, that determines
the sequence of amino acids in protein synthesis. The DNA sequence of a gene can be used to
specify the mRNA sequence, and the genetic code can in turn be used to specify the amino
acid sequence.
Genetic map (linkage map): A map of relative positions of genes or other chromosome
markers, determined on the basis of how often they are inherited together. See physical map.
Genome: All the genetic material in the chromosomes of a particular organism; its size is
generally given as its total number of base pairs.
Genetic sequencing: The process of identifying the ordered list of amino or nucleic acids that
form a particular gene or protein.
Global schema: A schema, or a map of the data content of a data warehouse that integrates the
schemata from several source repositories. It is "global", because it is presented to warehouse
users as the schema that they can query against to find and relate information from any of the
sources, or from the aggregate information in the warehouse.
HGP (human genome project): The U.S. Human Genome Project is a 15-year effort
coordinated by the U.S. Department of Energy and the National Institutes of Health to
identify all the estimated 100,000 genes in human DNA, determine the sequence of the 3
billion chemical bases that make up human DNA, store this information in databases, and
develop tools for data analysis. To help achieve these goals, a study of the genetic makeup of
several non-human organisms, including the common human gut bacterium Escherichia coli,
the fruit fly, and the laboratory mouse is also underway.
Homology: Here, a relationship of having evolved from the same ancestral gene. Two nucleic
acids are said to be homologous if their nucleotide sequences are identical or closely related.
Similarly, two proteins are homologous if their amino acid sequences are related. Homology
can also be inferred from structural similarity.
Intelligent search: Procedures which use additional knowledge to eliminate a majority of
space to be searched in order to retrieve a set of data items which satisfy a given set of
properties.
JGI (joint genome institute): A massive DOE sponsored project that will integrate human
genome research based in three of its national laboratories in California and New Mexico. The
Joint Genome Institute is a "virtual laboratory" that will sequence approximately 40 percent of
the total human DNA by 2005 and share information through public databases.
Knowledge discovery: The process of identifying and understanding hidden information and
patterns in raw data which are of interest to the organization and can be used to improve the
business procedures. Data mining is a single, but very important, step in this process.
Metadata: Description of the data. Metadata is used to explicitly associate information with
(and knowledge about) data in a repository. This information may range from the simple (it is
always an integer) to arbitrarily complex (it is a floating point value representing the
temperature of an object in degrees Celsius, and is accurate to 5 significant digits).
Mediated data warehouse architecture: A data warehouse architecture providing access to data
from various sources, where not all the data is stored is cached at the warehouse. Mediators or
software agents determine the best source of data to satisfy user requests and fetch data from
the selected sites in real-time.
Micro-theories: A micro-theory is an ontology about a specific domain, that fits within, and
for the most part is consistent with, an ontology with a broader scope. For example, structural
biology fits within the larger context of biology. Structural biology will have its own
terminology and specific algorithms that apply within the specific domain, but may not be
useful or identical to, for example, the genome community.
mRNA (messenger RNA): RNA that serves as a template for protein synthesis. See genetic
code.
multi-databases: A repository of data from several, possibly heterogeneous data sources. They
do not provide a consistent view of the data contained within the data sources, do not provide
summary information, and rarely have a local data store.
nucleic acid: A large molecule composed of nucleotide subunits.
Nucleotide: A subunit of DNA or RNA chemically consisting of a purine or pyrimidine base,
a phosphodiester, and a sugar.
object identification: The process of determining data items representing the same real-world
concepts in data sources, and merging the information to a single consistent format. This will
allow users to obtain accurate and unique answers to information requests.
object model: A data model for representation real-world entities using the concept of objects
which can belong to one of several classes. The classes have a specialization-generalization
relationship and contain both data and methods. Objects themselves can be comprised of other
objects.
OLAP (on-line analytical processing): Analysis of data along various dimensions (such as
time periods, sales areas, and distributors) so as to obtain summary information as well as
very specific targeted information to predict business performance based on various
parameters.
ontology framework: A framework for data management and manipulation based on a
descriptions of inter-relationships of the data items in a specific domain. This allows a more
targeted search of the data items using a semantic understanding in addition to using the
syntactic values.
PDB (protein data bank): The Protein Data Bank (PDB) is an archive of experimentally
determined three-dimensional structures of biological macromolecules, serving a global
community of researchers, educators, and students. The archives contain atomic coordinates,
bibliographic citations, primary and secondary structure information, as well as
crystallographic structure factors and NMR experimental data.
partially materialized views: In a mediated warehouse, a portion of the data represented in the
global schema may not be represented in the warehouse, but is only accessible through the
mediator. The portion of the data represented locally is said to be materialized. Thus, if there
is both local and non-local data represented in the schema, it is said to be partially
materialized.
physical map: A map of the locations of identifiable markers on DNA, regardless of
inheritance. For the human genome, the lowest-resolution physical map is the banding
patterns on the 24 different chromosomes; the highest-resolution map would be the complete
nucleotide sequence of the chromosomes. See genetic (linkage) map.
Protein: A large molecule composed of one or more chains of amino acids in a specific order;
the order is determined by the sequence of nucleotides in the gene coding for the protein.
Proteins are required for the structure, function, and regulation of the body's cells, tissues, and
organs, and each protein has unique functions.
protein sequence: The ordered list of amino acids that make up the protein.
protein taxonomy: Proteins can be grouped according to several criteria including their
sequence or structural similarity with other proteins. These groupings are often referred to as
taxonomies.
RNA (ribonucleic acid): A molecule chemically similar to DNA, involved in the readout of
information stored in DNA.
schema (pl. schemata): A description of the data represented within a database. The format of
the description varies but includes a table layout for a relational database or an entityrelationship diagram.
SCOP: A database of distinct protein structures and their structural classifications, along with
detailed information about the close relatives of any particular protein.
Sequencing: Determination of the order of nucleotides in a DNA or RNA molecule, or the
order of amino acids in a protein molecule.
SWISS-PROT: is a protein sequence database with high level of annotations (such as the
description of the function of a protein, its domains structure, post-translational modifications,
variants, etc.), and a minimal level of redundancy.
terabyte: A measurement of the amount of computer storage. Equivalent to 1,000 gigabytes or
1,000,000 megabytes of information.
unstructured data: Refers to data whose structure is not well defined or strictly enforced. For
example, web pages and flat files are unstructured data while a relational database is structure.
http://cbcg.lbl.gov/ssi-csb/Meso.html#anchor584580