Transcript Covert2012_overview

A tremendous modeling feat
• Whole-cell, molecular-level computational
model of the life-cycle of a procaryote:
Mycoplasma genitalium
• Integrates 28 different sub-modules, each
employing its own modeling technique
(ODEs, CBM, Stochastic processes, etc.)
Some food for thought
Model
Simulation
Research
Engineering
Understanding
“biological reality”
.
Understanding
“biology as it could be”
• What does it mean “to validate” a
construct of such complexity?
• How could such a tool be used?
The modeling approach
• Each of the 28 submodels was built,
parametrized and
tested independently
• The processes were
assumed to be
independent on sub-1s
time scales
• Synchronous simulation was performed at 1s intervals,
with 16 cellular variable types passed between the submodels at each step
#genes associated
with the process
RNA weight, length,
composition, etc.
Translation
progress
Status (active/
stalled), RNA
species, codon
position
Fluxes vector
Molecule counts for
3 compartments:
cytosol, membrane,
extracellular space
Boolean functions
Poisson process
ODE system
FBA + MOMA
Model reconstruction
• Chromosome reconstruction:
–
–
–
–
Genes
Transcription units
Promoters
Protein binding sites
• Functional annotation (new
annotations for 237 out of 525
genes!)
Model reconstruction (cont.)
• Structural reconstruction of each gene product:
– Protein sequence
– Post-transcriptional and post-translational processing
and modification of RNA and protein
– Signal sequence and localization
– DNA footprint of each DNA-binding protein
– Chaperones and prosthetic groups required to fold each protein
– Subunit composition of each protein and ribonucleoprotein
complex
– Disulfide bonds of each protein and complex
• Curation and complementation of chemical reactions for
all processes
– The source metabolic model of M. genitalium by Maranas (iPS189 Suthers et al., PLoS CompBio 2009) had 262 reactions and 274 metabolites
– The current model has 645 metabolic reactions (out of 1857
reactions overall) and 722 metabolites!
•••
Parameter reconciliation
• Constraints between parameters in
multiple modules were identified
• Model parameters were then tuned to
– Satisfy these constraints
– Deviate minimally from empirical observations
Determining initial conditions
• Cell proerties right after division were
assumed to be statistically identical in
consequtive generations
• Initial conditions for all cell state variables
were iteratively modified, until the postdivision variable distributions converged to
a steady state
Example: metabolic modeling
• Derivation of optimal flux distribution (maximizing
biomass) through “modified FBA”:
– Network expansion to include the metabolic requirements of the
27 other processes
– Internal exchange reactions added to recycle the metabolic
byproducts of the other processes
– Optimization objective expanded to include the recycling and
export of the metabolic byproducts of the other processes
– Flux bounds taking into account enzyme copy numbers and
catalytic rates
• MOMA to then fit the growth rate to the observation
(apparently with no distance limitation)
• Updating metabolite counts due to transfer reactions
between the three compartments
Using the simulation
• “Wildtype” simulations:
– Initializing the cell state from random initial
conditions
• “Knocked-out strains”:
– Set the half-life of the RNA and protein
products of the deleted gene to zero
– Deleted all RNA and protein products of the
deleted gene
Movie time…
Verification on training data
•
Simulated 128 wild-type cells in a typical Mycoplasma culture
environment, to predict
–
–
•
Cellular properties: cell mass; growth rate
Molecular properties: count; localization; activity
Consistency with observations on:
(A)
(B)
(C)
(D)
(E)
C
Doubling time
Cellular chemical composition
Major cell mass fractions B
Gene expression
Gene essentiality
D
E
Validation on ‘external’ data sets
• The flux through
glycolysis is >100-fold
more than that through
the pentose phosphate
and lipid biosynthesis
pathways (E)
• Predicted metabolite
concentrations are
within an order of
magnitude of those
measures in E. coli for
all the metabolites in
one study and for 70%
of them in a more recent
study (F)
• ‘‘Burst-like’’ protein synthesis due to the local effect of
intermittent messenger RNA (mRNA) expression and the
global effect of stochastic protein degradation on the availability of free
amino acids for translation, which is comparable to recent reports (G)
• The mRNA and protein level distributions are consistent with recently
reported single-cell measurements (H)
Novel predictions (1)
Interactions of DNA-binding proteins
(A)
Predicting chromosomal protein occupancy
(averages over 128 wildtype simulations):
•
•
•
•
(B)
(C)
Overall average
DNA polymerase
RNA polymerase
Replication Initiator DnaA
Temporal dynamics of chromosome
exploration
Temporal dynamics of gene expression
Interactions of DNA-binding proteins (cont.)
(D) Spatio-temporal dynamics of DNA and
RNA polymerases
(E) Collision and displacement frequencies
for pairs of DNA-binding proteins
(F) Correlation between proteing density
and frequency of collision across the
chromosome
Novel predictions (2): Metabolism
as an emergant cell cycle regulator
• More cell-to-cell variation in the durations of the
replication initiation (64.3%) and replication
(38.5%) stages than in cytokinesis (4.4%) or the
overall cell cycle (9.4%)
• DNA replication proceeds at two distinct rates:
– Initially, replication proceeds quickly due to
availability of free dNTP in the cell
– When the dNTP pool is exhausted the rate
limiting factor becomes dNTP synthesis
More predictions
• Global energy distribution
• Molecular pathologies of single-gene
disruptions
• Model driven discovery: putative alternative
NAD electron transfer pathway for
• Who knows what more…
Great results, but…
“Are they real?” ™
™ Seinfeld: season 4, episode 19 – “The implant”
Here are some of the problems
• A lot of the model was based on empirical results from other
organisms, some of which pretty distant from M. genitalium
• The MOMA stage of the
metabolic model wasn’t
controlled for distance
• Due to the complexity of the
system, the “External validation”
is prone to multiple-hypothesis
bias
• Validation results don’t
necessarily require whole-cell
model
• No rigorous robustness analysis
of the results to perturbations in
model parameters (1900 of
them…)
M. capricolum
M. arthritidis