Software Systems for Neuroinformatics
Nigel Goddard
Institute for Adaptive & Neural Computation
Division of Informatics
University of Edinburgh
IST2001 #2
Overview
• Neuroinformatics: what and why?
• Methodological challenges
• Software solutions: simulation
• Software solutions: collaboration
IST2001 #3
This talk
Neuroinformatics: What is it?
Computational models and analytical techniques to help understand information-processing in the nervous system
Information processing methods in the nervous system inspire new IT tools
IT techniques to help collect, analyze, archive, share, simulate and visualize knowledge of information processing in the nervous system
Computational Models
Neural Engineering
Software Systems
IST2001 #5
Overview
• Neuroinformatics: what and why?
• Methodological challenges
• Software solutions: simulation
• Software solutions: collaboration
IST2001 #6
Methodological Challenges• Scale and heterogeneity
– data, models, computations– multiscale, multiformalism methods needed– Parallel/GRID resources required
• Collaboration is essential– Data and its understanding is distributed – Computional models are valuable expressions of
understanding– we need tools to support exchange, discussion and
comparison of models and data
IST2001 #7
Overview
• Neuroinformatics: what and why?
• Methodological challenges
• Software solutions: simulation
• Software solutions: collaboration
IST2001 #9
Large Scale Network Modeling
• Compartmental cell modeling understood…
• .. but network modeling support needs study
• Parallel computing needed for networks of spiking cells…
• … and amenable to effective parallel simulation using Discrete Event Simulation
Tdata
IST2001 #10
NEOSIM Simulation Approach
• Optimize simulation kernel for network activity
• Plug in single-cell models and other components from other packages
• Design for parallel computers
• NEural Open SIMulation
IST2001 #11
Multilevel Simulation
Purkinje
granuleDiffusion
Modeling Component
Connectionist NetworkModeling Component
Rastorgram component
Voltage trace component
http://www.neosim.org
IST2001 #12
Overview
• Neuroinformatics: what and why?
• Methodological challenges
• Software solutions: simulation
• Software solutions: collaboration
IST2001 #13
Need Improved Methodologies• Progress will accelerate when neuroscientists
can share model components to build more complex simulations
• Key technical requirements:– Need a common model exchange format
• The model description language must be translatable into forms suitable for simulation
– Need software tools that support simulation of models, development, visualisation, exchange and storage of computational models and components of models
IST2001 #14
Architecture
NeuroML is the language all components use to communicate data and models
Some components can implement other interfaces
IST2001 #15
Neural Open Markup Language• We propose NeuroML/Model as a candidate
model representation & exchange language– Uses a simple, well-supported, textual substrate (XML)– Adds components that reflect the natural conceptual
constructs used by modelers in the domain• Data structures – a simulator independent model
description– neuroml.model.cell, .synapse, .network…
• Extensible – tools can add tool-specific annotations
• Code is hidden behind the NeuroML declarative interface
IST2001 #18
From specifications to working tools
• A standard data structure is only useful if tools can read & write it• We have released a development kit for providing easy access to
NeuroML for tool developers• Features:
– XMLIn / XMLOut : to read & write NeuroML files– A module loader : to download code modules on the fly– A generic model editor
IST2001 #19
NeuroML & Simulation Tools
Catacomb channel simulator
NEOSIM network simulator
Kinetic scheme model of a sodium channel
NeuroML Models
Multicompartment network model
Simplified integrate/firenetwork model with learning rules
Genesis compartmental solver
Neuron compartmental solverBall/stick style cell model
Integrate/fire point neuron model
Monte-carlo synaptic transmission model
Monte-carlo synaptic transmission model
Cell generated from L-system growth rules
Hodgkin-huxley style channel models
Visualisation tools
Reconstructed 3D cellwith channels distributedover structure
IST2001 #20
Order in the chaos• Step 1 : models declared with the languages of
neuroml.model.*– side benefit: parallelisation is easier
• Step 2 : where simulators need to interoperate during a simulation run, they can implement interfaces in: neuroml.sim.run.* (for execution) and neuroml.sim.state.* (for access to instantiated model state variables)
IST2001 #22
Futures
• Make more simulators NeuroML-aware
• Starts to become possible to construct multi-scale models, with different simulators cooperating
• Napster-like service for modellers : models and software components are valuable – it makes sense to share + reuse as much as possible
IST2001 #23
NEOSIM/NeuroML CollaboratorsInformatics
Nigel Goddard, Fred Howell, Paul Rogister - EdinburghGreg Hood - Pittsburgh, Michael Hines - YaleOliver Gewaltig - Honda R&D, Robert Cannon - BostonMichael Hucka - Caltech, Hugo Cornelis - AntwerpPaul Verschurre - Zurich, Ronan Reilly – DublinSimon Thorpe - CNRS
NeuroscienceErik De Schutter - AntwerpTerrence Sejnowski - SalkWilliam Levy - VirginiaDavid Willshaw, Andrew Gillies – EdinburghAngus Silver - UCL
Funded by the Human Brain Project, National Institutes of Health and National Science Foundation, USA