Network Modeling

TR&D3: Network modeling

The goal of TR&D3 Network-based approaches have yielded many novel insights, in particular, the identification of common network motifs such as feedback and feedforward loops that give rise to specific classes of signal processing characteristics – like switching, oscillation, and adaptation.

Vision

Cellular systems are profoundly difficult to understand because of the interplay between spatial, biochemical and molecular complexity that occurs on multiple levels of organization, from macromolecular assemblies to synapse architecture to neural circuits. Current cell simulation tools are just beginning to address the complexity that spans these levels in an integrated fashion. Major advances in simulation tools are needed to enable the development of models that capture the required level of detail and at the same time remain computationally tractable. The MCell/CellBlender platform for cell modeling developed in TR&D2 is expressly designed to fulfill these needs, providing insight and understanding of complex cellular systems.

Background and Motivation

The integration of the network modeling component within MMBioS offers numerous opportunities for further developing RBM capabilities and to demonstrating their effectiveness for providing mechanistic insight into complex regulatory networks. A number of specific aims will be pursued to achieve these goals: Accelerated methods for stochastic simulation of rule-based models. Most of the parameters governing the behavior of network models cannot be measured independently but must instead be determined by solving the inverse problem to obtain the parameters from experimental data, which demands many orders of magnitude more computational effort that computing a single trajectory. The structure of rule-based models opens the door to development of more efficient simulation methods that have not been previously explored.

Novel coarse-grained spatial modeling capabilities MCell simulations are prohibitively expensive in terms of computing time and memory for many applications, particularly those that require simulation of protein copy numbers on the scale of a whole cell. Simulations at this scale require more coarse-grained spatial resolution for computational efficiency, which we will develop in Aims 1.2 and 1.3. These coarse-grained network-free simulators will enable parameter estimation and other intensive analysis tasks to be performed on spatially-resolved models, such as those developed in DBP 4 and 8 in collaboration with TR&D4 and TR&D1.

Novel interfaces for model construction, visualization, and comparative analysis Large-scale modeling efforts require scalable and zoomable approaches to visualization. It is currently difficult to compare models in the literature or even those downloaded from databases with each other. Rule-based encoding can facilitates management and comparative analysis of models using tools we developed and we will apply in Aim 2.

Increased accessibility of model analysis and parameter estimation tools for complex models The computational demands of many parameter estimation tasks require HPC resources that many end users lack expertise to utilize. We will develop novel parameter estimation interfaces in Aim 2 that wil provide users with transparent access to HPC resources available through the XSEDE platform in Aim 3.

Unique infrastructure for rule-based modeling capabilities In Aim 3 we will harden the software libraries and interfaces that are critical to the continued availability of RBM capabilities for developers and end users. 

Related work/software Other RBM software tools have been developed, but none appears to have the strengths of the BioNetGen platform. Indeed, some major packages use BioNetGen as an underlying RBM engine, including VCell and Smoldyn.

There is little overlap between current VCell capabilities and the aims proposed here, and in fact the proposed simulator development in Aim 1, the visualization capabilities in Aim 2, and the software infrastructure in Aim 3 may all be used to expand VCell functionality in the future (see Blinov letter of support). Besides BioNetGen, the other major RBM language based on citations, use by non-affiliated groups, and integration into other software frameworks is Kappa, which is open source and written in the OCaml language. Kappa is well-known in the computer science community, but has comparatively few biological applications. Another notable effort is Simmune,  which combines graphical, RBM specification and PDE and ODE simulations. Currently, Simmune does not support stochastic simulations and source code is not available, which limits its use for outside development. A number of other software tools have been developed for RBM, especially in recent years, but most of these were developed for relatively specialized applications and/or are not being actively developed.

Specific Aims

  1. Advance rule-based modeling technology for cell-scale simulations

    Network-free simulation is a key technology for the largest scale simulations because it avoids explicit generation of the full reaction network without sacrificing accuracy. The computational cost of these simulations remains a significant limitation to their application, particularly for computationally intensive tasks such as parameter estimation, which motivates Subaim 1.1: Adapt and further develop methods to accelerate stochastic simulations of rule-based models. The integration of rule-based modeling capabilities into MCell in TR&D2 enables spatially-resolved simulations of large networks exhibiting combinatorial complexity, but simulating cell-sized molecular populations remains prohibitive, which motivates Subaim 1.2: Develop a compartmental network-free simulator; and Subaim 1.3: Develop a network-free subvolume-based simulator to complement TR&D2 and enable simulation-intensive parameter estimation tasks required by multiple DBPs. Both coarse-grained simulators will initially be developed to handle static geometries and subsequently extended to handle dynamic geometries..

  2. Further develop advanced visual interfaces for building, managing and analyzing rule-based models

    Visual interfaces are essential to the modern collaborative modeling endeavor: they make modeling accessible to non-specialists and enable collaborative development and re-use of models developed by other groups, motivating Subaim 2.1: Extend RuleBender, the graphical front end to BioNetGen, to enable visual model building without detailed knowledge of the BioNetGen language and provide additional model visualization capabilities that we have developed as prototypes in the initial funding period. Facilitating model annotation, comparison, and re-use motivates Subaim 2.2: Develop a database of molecules, rules, and models that can be used for comparative analysis of existing models and development of new models. This aim will take advantage of the Atomizer tool we have developed that enables any reaction network model to be recast as a rule-based one. Finally, facilitating analysis of model dynamics to enable mechanistic discovery from data motivates Subaim 2.3: Develop RuleBender interfaces for parameter sweeps, sensitivity analysis, and parameter estimation.  

  3. Extend and refine software infrastructure for rule-based modeling

    Although the BioNetGen framework already provides rule-based modeling capabilities for a number of major software tools including the Virtual Cell, pySB, rxncon, BioUML, MCell and WESTPA, the core capabilities provided under the BioNetGen umbrella are currently distributed among multiple codebases in different languages, motivating Subaim 3.1: Develop a hardened C/C++ library and application programming interface (API) for BioNetGen that exposes both high- and low-level functionality. The computationally intensive nature of parameter estimation and other model analysis tasks motivates Subaim 3.2: Develop a web portal that will provide end users direct access to rule-based modeling capabilities and HPC resources.

Modeling with BioNetGen Gives MMBioS Team Insight into How Immune System Decides to Attack — or Let Be

T-Cell Receptor SignalingFriend or Foe?

A mix of computer modeling and laboratory experiments has helped reveal how the body differentiates “friend from foe.” Using their BioNetGen computer tool for simulating biochemistry, MMBioS members and colleagues have painted a sharper picture of how T cells, the advance scouts of the immune system, decide when to protect bodily tissues from immune attack—and when to lead the attack. The finding may guide future efforts to control human diseases like diabetes and cancer.

When T cells—a type of white blood cell—encounter other cells in the body, “they have to decide what kind of a response to make,” says James Faeder, project co-leader of MMBioS’s Technology Research and Development Project 2 Team and associate professor of computational & systems biology, University of Pittsburgh. “Is that a threat, or is it something benign? Depending on their assessment of a possible threat they can either become activated, immune-boosting cells that kill pathogens, or they can tamp down those responses.”

Read the entire article

See more about MMBioS research in cell modeling.


Improved Sampling of Cell-Scale Models using the Weighted Ensemble Strategy

The “weighted ensemble” (WE) strategy for orchestrating a large set of parallel simulations has been established as an effective tool for efficiently calculating kinetic and equilibrium observables in molecular systems – and now has been extended to spatially resolved cell-scale systems by MMBioS researchers. In a collaboration among several MMBioS groups, the WESTPA implementation of WE has been used to control MCell simulations, including models built using a BioNetGen-CellOrganizer pipeline for situating complex biochemistry within spatially realistic cell models. As sketched, the WE strategy enables computational effort to be focused on rare, difficult-to-sample events – essentially increasing the precision of measurements in the tails of distributions. In an application to an MCell model of the frog neuro-muscular junction (NMJ), WE simulation enabled calculation of vesicle release probabilities under particularly challenging low-calcium concentrations that could not be well sampled by conventional MCell simulation: see red-circled points in the figure. This in turn, enabled validation of the model by confirmation of an established NMJ empirical power-law relationship between calcium-concentration and release probability. WE yielded estimates of observables in less overall computing time than would be required in ordinary parallelization, thus exhibiting super-linear parallel performance.

Donovan RM, Tapia JJ, Sullivan DP, Faeder JR, Murphy RF, Dittrich M, Zuckerman DM (2016). Unbiased Rare Event Sampling in Spatial Stochastic Systems Biology Models Using a Weighted Ensemble of Trajectories PLoS Comput Biol. 12(2):e1004611

See more about MMBioS research in molecular modeling


 

Development and improvements to MCell

The MCell modeling and simulation platform provides the core capabilities for spatial simulation reaction-diffusion dynamics at complex biological interfaces. Highlights of progress include:

libMCell API We have developed a high-level set of routines, which enable the creation and simulation of MCell models via API calls alone, i.e., without invoking any parsed MDL description of the model. Current functionality includes creation of species, defining reactions, defining molecule releases, creating model geometry, and simulation and runtime control.

MCell testing framework This framework, called nutmeg (https://github.com/haskelladdict/nutmeg), has completely replaced our previous, python based unit test framework. During the year we made nutmeg much more robust and added many additional test cases.

Implementation of new simulation capabilities that extend the range of simulations that can be performed in order to address important biological questions, such as the how membrane dynamics and signal transduction interact or how multisite phosphorylation, binding, and cooperativity affect signaling. Highlights include:

  • Dynamic geometries Users can now create dynamical meshes using Blender and use these to drive an MCell simulation of reaction and diffusion with moving boundaries.
  • Modeling pipeline Users can build complex biochemical models using BioNetGen and sample complex cellular geometries from imaging data using CellOrganizer and combine those into a single model that can be simulated in CellBlender (in collaboration with TRD3).
  • Recovery of protein-protein interactions and other aspect of biochemical mechanisms from reaction network models using Atomizer. The automated conversion of these models into a rule-based format allows for comparative analysis of models and reuse of existing model components.
  • Visualization of regulatory structurein rule-based models New visualization methods have been developed that enable global visualization of models exhibiting combinatorial complexity.

D. P. Sullivan, J. J. Tapia, R. Arepally, J. Czech, R. F. Murphy, M. Dittrich, and J. R. Faeder, “Design Automation for Biological Models : A Pipeline that Incorporates Spatial and Molecular Complexity,” in 25th edition of the Great Lakes Symposium on VLSI, 2015, pp. 321–323.

J. A. P. Sekar, J.-J. Tapia, and J. R. Faeder, “Visualizing Regulation in Rule-based Models.” Submitted to Bioinformatics . arXiv:1509.00896 [q-bio.QM].

See more about MMBioS research in cell modeling.


 

Development of CellBlender

CellBlender is a graphical interface for model construction, simulation, and analysis of complex spatial models of reaction-diffusion systems  Highlights of progress  include

  • A major redesign of the interface to consolidate functionality based on a single toolbar.
  • Implementation of a robust parameter handling system
  • Functionality for running simulations from within the interface taking advantage of parallel computing resources allowing users to manage on-going simulation streams.
  • Direct access to run-time error logs within the Blender interface.
  • Model and geometry importers that allow import of compartmental reaction network models in either SBML or compartmental BNGL format.
  • Implementation of internal data model provides backward compatibility with future CellBlender versions.
  • Release of CellBlender version 1.0 and publication of an article featuring MCell and CellBlender in the Encyclopedia of Computational Neuroscience

T. Bartol, M. Dittrich, and J. Faeder, “MCell,” in Encyclopedia of Computational Neuroscience, D. Jaeger and R. Jung, Eds. Springer New York, 2014, pp. 1–5.

See more about MMBioS research in cell modeling.


 

Stochastic Simulation 

fig1 weightedensemble biggreySystems biology, the quantitative study of complex interacting biological systems, is becoming increasingly demanding of computational resources. As models become able to capture truly physiological behavior, some of the most difficult computations may also be the most important, such as a rare transition from normal to pathological behavior.

A novel approach to meet growing computational demands is described in a paper recently accepted to the Journal of Chemical Physics ("Efficient Stochastic Simulation of Chemical Kinetics Networks Using A Weighted Ensemble Of Trajectories"). First-author Donovan, in collaboration with MMBioS investigators Faeder and Zuckerman, applied a sophisticated parallel simulation algorithm originally developed for small-scale molecular systems to systems biology models.

The results were eye-catching: the new approach was more computationally efficient at modeling rare events than standard methods by orders of magnitude, even for a very complex model with thousands of reacting species. 

Donovan RM, Sedgewick AJ, Faeder JR, Zuckerman DM (2013) Efficient Stochastic Simulation of Chemical Kinetics Networks Using A Weighted Ensemble Of Trajectories J. Chem. Phys. 139:115105.

 

See more about MMBioS research in molecular modeling.

Copyright © 2020 National Center for Multiscale Modeling of Biological Systems. All Rights Reserved.