This subproject is one of many research subprojects utilizing theresources provided by a Center grant funded by NIH/NCRR. The subproject andinvestigator (PI) may have received primary funding from another NIH source,and thus could be represented in other CRISP entries. The institution listed isfor the Center, which is not necessarily the institution for the investigator.1. Introduction. Microphysiological processes involve simultaneous diffusionand reaction of molecules, and incorporation of realistic ultrastructure is amajor challenge for computational studies. It is increasingly important thatsimulations delineate not just a system's average behavior, but also itsvariability and propensity to switch between different operating modes and/orto fail [McAdams and Arkin, 1998]. One of our major interests is synaptictransmission, especially synaptic variability and plasticity, and theirfunctional impact at central and peripheral sites (e.g., learning and memory,neuromuscular pathology) [Stiles et al., 2001]. Several large-scale projects(See 4 below) are particularly in need of fast computing resources withextremely large shared memory. This memory requirement arises from interactivemodel design, and simulation optimizations designed to reduce computation totolerable levels.2. Computational Background. Simulations at atomic resolution are not feasibleon microphysiological scales of space (nm to m) and time ( s to ms), and so weare developing an alternative, MCell ('Monte Carlo cell' or 'MicroCellularphysiology') (www.mcell.cnl.salk.edu and www.mcell.psc.edu). With MCellsimulations, large-scale 3-D reconstruction data from serial electronmicrographs or electron tomography can be used to create tessellated cell andorganelle surfaces, and the defined spaces and structures then can be populatedwith diffusing and static molecules that interact probabilistically. Forspace-independent unimolecular transitions (e.g., conformational changes),events are decided using methods somewhat similar to those introduced byGillespie [1977]. For space-dependent bimolecular associations (e.g., ligandbinding), however, Brownian Dynamics diffusion methods are combined withprobabilistic testing for second order transitions [Stiles and Bartol, 2001]. Details about individual molecular structure are ignored, but the functionaleffects of individual molecular positions within realistic subcellulartopologies are included explicitly.MCell's computational requirements vary widely according to the scale of eachsimulation and the number of simulations to be run [Stiles and Bartol, 2001]. Early versions of our code required Gbyte-level memory on shared memorymainframes (e.g., IBM ES9000 at the Cornell Supercomputer Center) even formodels that are very simple compared to current projects [Bartol et al., 1991;Anglister et al., 1994]. At present, memory use is much more efficient butincreases with the scale and resolution of reconstructions, largely due to acritical optimization that keeps compute speed nearly constant despiteincreasing spatial complexity. Given sufficient memory, run times now scalemostly with the numbers of time-step iterations and diffusing molecules, anddisk use approaches a terabyte for large projects. Typical project scenariosinclude [Casanova et al., 2001]:A. Look-and-see. A single simulation fits in available memory (e.g., 1-4Gbytes) and runs in less than 1 day on a typical workstation. A small set ofruns is used to predict the behavior of a simplified model system. Mostprojects begin in this category and evolve into one of the following (B-D).B. Parameter-fitting. Similar to above, but tens to thousands of runs arerequired to identify a set of input parameter values that produce model outputwhich matches some set of target criteria [Stiles et al., 1996, 1998, 1999,2000].C. Parameter-sweep. As above, but thousands of runs are required to mapthe parameter space encompassed by the model [Bartol et al., 2000]. Thisscenario, like parameter-fitting, is well suited to distributed computing(e.g., using the computational Grid [Casanova et al., 2001]).D. Complex spatial models (as proposed here). Presently the mostchallenging, these can be included within a parameter-fitting or -sweepproject, are presently limited by a lack of large shared memory, and areincreasingly possible due to the availability of high-resolution 3-Dreconstructions. The number of diffusing molecules often is not large enoughto require massively parallel computation, but for high throughput there must be enough shared memory to decouple compute speed from model size. Theseprojects are also critically dependent on interactive visualization, whichrequires large shared memory (see below).3. Interactive Model Visualization. With millions of molecule locations andpolygons in surface meshes, MCell project development is absolutely dependenton optimized data structures and interactive visualization tools. We now use arendering and analysis environment designed for IBM DataExplorer (OpenDX,www.opendx.org). DataExplorer was originally written to exploit shared memoryparallel machines (IBM Power Visualization System). Custom data manipulationand rendering programs can be created within the OpenDX environment, and wehave recently written DReAMM (Design, Render, and Animate MCell Models) at PSCand introduced it at an MCell workshop(www.psc.edu/biomed/training/workshops/2001/micro/index.html; June 2001;see figure below and also movies atwww.compneuro.org/CDROM/docs/chapter4.html).By far the most important bottleneck for interactive visualization is limitedshared memory. Even for 'small' MCell models, we routinely require 2-4 Gbytesof memory. For large models, this will increase by 50-100 fold (presentlyunavailable anywhere). Thus, in addition to running simulations, we will usethe entire proposed machine heavily as a high-end interactive visualizationserver.4. MCell Simulations of Realistic Synaptic Models.4.1. Pre- and postsynaptic factors that contribute to synaptic variability atthe vertebrate neuromuscular junction. Miniature endplate currents (mEPCs) atthe vertebrate neuromuscular junction (NMJ) show a large degree of variabilityin amplitude and time course, assumed to arise from variability in the numberof acetylcholine (ACh) molecules released from synaptic vesicles. However,preliminary simulations suggest that a variety of pre- and postsynaptic factorsmust contribute, including molecular organization and topology of the endplatemembrane [Stiles et al., 1996, 1999, 2001].With the proposed machine, we will undertake systematic investigation of thesepossibilities for the first time, using large-scale 3-D reconstructions ofmouse sternomastoid NMJ ultrastructure. We have (see figure below) a firstreconstruction from 60 serial sections (out of ~400 available), ~9x5x4 m,~350 m2 postsynaptic surface area, containing ~106 polygons and at least twicethat many ACh receptors, ACh esterase sites, choline reuptake sites, anddiffusing ACh molecules. We anticipate ~20 Gbytes of memory and from minutesto ~1 hour for each simulation (processor), depending on conditions such asposition and amount of released ACh, and presence or absence of active AChesterase sites. Different choices of input parameters will predict different relative effectson the distributions of and correlations between mEPC amplitude, rise time, andfall time, and available mEPC data will be used to test modeling predictionsdirectly. Varied input parameters will include: (1) position of ACh release;(2) density and distribution of ACh receptors; (3) density and distribution ofACh esterase sites; (4) inhibition of ACh esterase; (5) receptor block; (6)density and distribution of choline reuptake sites. Each of these conditionswill require thousands of runs using different parameter values and randomnumber seeds, and the project will require weeks to months of machine time.4.2. Extracellular calcium dynamics in a cerebellar glomerulus. A class ofsynaptic learning models, in which a sum of postsynaptic activity from manyneurons drives plasticity, has generally been considered biologicallyinfeasible. Postsynaptic cell bodies may be far apart, and there are nobackward signals known to sum activity in a terminal-specific manner. However,some specialized synapses, known as glomeruli, are tightly ensheathed by glial cells. The ensheathment may force enclosed, neighboring dendrites to share alimited resource of extracellular calcium (ECa), and thus may allow the synapseto detect and respond to summed postsynaptic activity.We hypothesize that the ECa concentration in glomeruli may encode the level ofspike activity in postsynaptic cells. We plan to investigate this hypothesis inMCell simulations of cerebellar glomeruli, where dendrites from granule cellsentwine around a mossy fiber terminal, and the ensemble is tightly ensheathedby an astrocyte. Computer simulations of glomeruli will indicate the range ofconditions under which ECa will be proportional to the sum of granule cellactivity.Realistic synthetic or reconstructed glomerular structures will span a cuboidvolume of ~6x6x6 m, requiring ~3 x 107 ECa atoms. We anticipate thatindividual runs spanning ~1 sec of simulated time will require ~40 Gbytes ofmemory and will take up to several days on the proposed machine, even if usingall 10 processors. Thus, this project will require months of machine time.We will investigate whether ECa changes suggest a novel learning rule forcontrol of plasticity at the mossy fiber/granule cell synapse. This learningrule approaches a sparsely distributed and statistically independent coding inthe parallel fibers. Although traditional neural models emphasize onlyneurotransmitters and connectivity, these simulations will highlight the 3-Dcontext of axons, terminals, and dendrites.4.3. Glutamatergic synaptic transmission and calcium dynamics in hippocampus.In some areas of brain cortex, the density of synapses can be ~3 x 109 per mm3and 80% of these are excitatory glutamatergic synapses. We have beguninvestigating the biophysical properties of glutamatergic synapses using MCellsimulations and simplified neuropil ultrastructure [Franks et al., 2001a, b, c,d], focusing on neurotransmitter-receptor interactions and subsequentintracellular calcium dynamics in the postsynaptic density of individualsynapses. However, given the complexity of the neuropil environment and theissues discussed in the two preceding projects, a deeper understanding ofsynaptic function will only come through increasingly realistic computersimulations.We will perform MCell simulations of glutamatergic synaptic transmission andcalcium dynamics in a large-scale 3-D reconstruction of hippocampal neuropil. The reconstruction will span ~10x10x10 m and thus will contain about 3000synapses. Each individual simulation will require ~200 Gbytes of memory.Through reconstructions and simulations of this system we intend to study suchissues as variability of synaptic efficacy, degree of neurotransmitter spillover and synaptic crosstalk, frequency dependence of the ambient glutamateand calcium concentration in the extracellular space, reuptake of glutamateinto glial cells, and trafficking of glutamine back to neurons.
Showing the most recent 10 out of 292 publications
