escience meeting, edinburgh, november 2006. slide 1 carmen code analysis, repository and modelling...
TRANSCRIPT
Slide 1eScience Meeting, Edinburgh, November 2006.
CARMEN
Code Analysis, Repository and Modelling for e-Neuroscience
Jim Austin , Colin Ingram, Leslie Smith, Paul Watson, Stuart Baker, Roman Borisyuk, Stephen Eglen, Jianfeng Feng, Kevin Gurney, Tom Jackson, Marcus Kaiser, Phillip Lord, Stefano Panzeri, Rodrigo Quian Quiroga, Simon
Schultz, Evelyne Sernagor, V. Anne Smith, Tom Smulders, Miles Whittington.
Slide 2eScience Meeting, Edinburgh, November 2006.
CARMEN is a new e-Science Pilot Project, (UK research council funded) in Neuroinformatics.
Objectives:• To create a grid-enabled, real time ‘virtual laboratory’ environment for neurophysiological data
• To develop an extensible ‘toolkit’ for data extraction, analysis and modelling
• To provide a repository for archiving, sharing, integration and discovery of data
To achieve wide community and commercial engagement in developing and using CARMEN
CARMEN is a 4 year project: if it is to last longer, it must become financially self-sufficient.
See http://www.carmen.org.uk
The CARMEN Project
neurone 1
neurone 2
neurone 3
Slide 3eScience Meeting, Edinburgh, November 2006.
CARMEN Consortium
Leadership & Infrastructure
Colin Ingram
Paul Watson
Leslie Smith Jim Austin
Slide 4eScience Meeting, Edinburgh, November 2006.
CARMEN Consortium
Work Packages
University ofSt Andrews
TheUniversity OfSheffield
Slide 5eScience Meeting, Edinburgh, November 2006.
Background: What is Neuroinformatics?
Informatics applied to Neuroscience (of all sorts)
Experimental Neuroscience:Data recording, data analysis have used computers for a long time.But a great deal more can be achieved by pooling data and analysis services
Cognitive and Computational NeuroscienceModelling,
Matching models to more experimental dataMatching models to known appropriate behaviourDefining and running more sophisticated modelsRunning models in real time
Clinical NeuroscienceData-based understanding of neuropathologyNeuropharmaceutical assays and assessment
Slide 6eScience Meeting, Edinburgh, November 2006.
What is Neuroinformatics bringing to Experimental Neuroscience?
Getting leverage from e-Science capabilities to allow better use of data.
Example: Dataset re-use:Experimenter does experiment, records data, analyses data, writes the paper, perhaps makes the data available to a small number of colleagues.
…and then?
The dataset languishes,first on a spinning disk, then later on some DVD’s, then later still, is lost to view, as the experimenter changes
lab,…
Yet the data could be of use to other researchers…
Slide 7eScience Meeting, Edinburgh, November 2006.
What are the basic problems holding back dataset re-use? (1)
Two major technical problems: Data format, and Metadata
Data Format
There are different systems for Neuroscience data collection.
The data format is a particular structureThe structure may be
Proprietary: defined by a particular piece of software, and not made public
Locally generated: defined by a locally written piece of software, but not necessarily well documented
Public, but no suitable converter exists for the intending user
Slide 8eScience Meeting, Edinburgh, November 2006.
What are the basic problems holding back dataset re-use? (2)
Metadata problemsThe data itself is useless unless the re-user knows exactly what the data represents.(Presumably the experimenter knew)
But did they record this information in an accessible way?
Metadata is data about the datasetHow was it generated?
What were the experimental conditions? What was the culture, or what preparation, or what animal,…? What was the temperature of the recording? Etc. etc.
If the data is to be readily re-used these metadata problems need to be solved in a directly usable way
Simply describing the protocol in English is not enoughCan’t automate reading J. Neurosci yet!
There needs to be an automatically processable way of describing the experimental protocol.
Particularly true is datasets are used for a large-scale survey of data
e.g. for data-mining.
Slide 9eScience Meeting, Edinburgh, November 2006.
Enabling Neuroinformatics based collaboration
Solving data format problems
Force users to adopt a common format?Alienates users: they won’t do it unless they can see real benefits
Support documented formatsAdopt a common internal format, providing translators to & from this formatRely on proprietary format owners to come aboard because of customer pressure
Slide 10eScience Meeting, Edinburgh, November 2006.
Enabling Neuroinformatics based collaboration: solving metadata problems
Difficult problem: there are a number of attempts at solving it:
BrainML: (Cornell) brainml.org“BrainML is a developing initiative to provide a standard XML metaformat for exchanging neuroscience data. It focuses on layered definitions built over a common core in order to support community-driven extension.”
NeuroML: http://www.neuroml.org/“NeuroML is an XML-based description language for defining and exchanging neuronal cell, network and modeling data including reconstructions of cell anatomy, membrane physiology, electrophysiological data, network connectivity, and model specification”
Relevant not only for Neuroinformatics and experimental neuroscience: Part of a cross-cutting problem for all aspects of neuroscience.
Slide 11eScience Meeting, Edinburgh, November 2006.
As well as metadata systems for neuronal systems, there are related metadata systems which can be used by BrainML and NeuroML,
ChannelML: for defining ion channel modelsMorphML: for defining the morphology of a neuronSBML: Systems Biology markup language: models of biochemical reaction networksCellML:to store and exchange computer-based mathematical models
SBML is particularly well advanced: see http://sbml.org/index.psp.
MathML: for describing mathematical notation and capturing both its structure and content. See http://www.w3.org/Math/
Metadata is a big but soluble problem.
It is a multi-level problem, but the systems above provide a multi-level solution.
Solving Metadata problems continued:
Slide 12eScience Meeting, Edinburgh, November 2006.
Enabling Neuroinformatics based collaboration: Sociological problems
There is a reluctance to permit re-use amongst some experimental neuroscientists.
What do experimental neuroscientists get from allowing others to reuse their data?
If the answer is only better science, then some experimental neuroscientists will not come on board.They need to be convinced sharing that their hard-earned datasets will be of benefit to them
Names on papers?The ability to be involved in the further research?At the very least, some credit!
Some neuroscientists fear that their data will be used without their knowledge
There is therefore some reticence amongst the experimental neuroscience community.
Slide 13eScience Meeting, Edinburgh, November 2006.
Solving sociological problems
There are technical aspects to solutions:
Security aspects on the holding of data:Ensure that datasets can be secured: for example that they can only be re-used with the experimenter’s permission.
Security is critically important for holding of data which is still being analysed prior to publication.
…and non-technical aspects too
Bringing experimental neuroscientists on board
Ensuring that the Neuroinformatics community is properly cross-disciplinary, with good representation from the experimentalists.
Getting journals on-side
Many journals are demanding that raw/processed data be made available in order to check results.
Slide 14eScience Meeting, Edinburgh, November 2006.
Neuroinformatics and clinical neuroscience
QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture. QuickTime™ and aTIFF (Uncompressed) decompressorare needed to see this picture.
Clinical Neuroscience is about treatment of
Mental illnessBrain diseasesTrauma
Neuroinformatics has major application here, ranging from 3d imaging technologies to EEG recordings: much broader than the focus of CARMEN.
CARMEN is primarily concerned with neural recordings.
These can provide data on neurochemical effects on neural function.Overall brain states (mental illness, disease) are believed to originate in the neurochemistry (research in depression and schizophrenia suggests this).
Slide 15eScience Meeting, Edinburgh, November 2006.
Neuropharmaceutical assays
Neural cell culturesDifferent types:
Slice preparationsCultures grown from neural
cell linesCultures from neonate neurons
…have recordings made from them with and without added neuropharmaceuticals.
Interest is on changes in behaviour in these preparations.
Requires instrumentation and analysis techniques
Sharing these results can lead to major advances
Pharmaceutical companies are interestedSecurity implications
Slide 16eScience Meeting, Edinburgh, November 2006.
Work Packages
WP 0Data Storage& Analysis
WP1 Spike Detection& Sorting
WP2 Information TheoreticAnalysis of Derived Signals
WP 3 Data-Driven ParameterDetermination in Conductance-
Based Models
WP4 Measurement and Visualisationof Spike Synchronisation
WP5 Multilevel Analysis andModelling in Networks
WP4 Intelligent Database Querying
Slide 17eScience Meeting, Edinburgh, November 2006.
CARMEN Objectives
• Create ‘virtual laboratory’ for neurophysiological data• Provide repository for:
• archiving, sharing, integration and discovery of data• services that operate on the data
• Develop extensible ‘toolkit’ for data extraction, analysis and modelling
• Achieve wide community and commercial engagement in CARMEN • CARMEN must become financially self-sufficient after
4 years
Slide 18eScience Meeting, Edinburgh, November 2006.
• Bowker’s “Standard Scientific Model”1
1. Collect data
2. Publish papers
3. Gradually loose the original data
1The New Knowledge Economy and Science and Technology Policy, G.C. Bowker, E1-30-03-05
• Problems:• papers often draw conclusions from unpublished data• inability to replicate experiments• data cannot be re-used
Data in Science
Slide 19eScience Meeting, Edinburgh, November 2006.
• Bowker’s Model• Collect data• Publish papers• Gradually loose the original data
• Problems:• papers often draw conclusions
from unpublished data
• inability to replicate experiments• data cannot be re-used
• Solution• Data Repositories
• Computational Science• Write codes• Publish papers• Gradually loose the codes
• Problems:• papers often draw conclusions
from the results of unpublished codes
• inability to replicate experiments• codes cannot be re-used
• Solution• Service Repositories
but… codes can be lost too
Slide 20eScience Meeting, Edinburgh, November 2006.
CARMEN Active Information Repository
Data
Metadata
Compute Cluster on which Services are Dynamically
Deployed
External Client.......
.......
External Client
Security
Workflow Enactment
Engine
RegistryServiceRepos -
itory
Slide 21eScience Meeting, Edinburgh, November 2006.
C WSP
req
res
1
Host Provider
node 1s2, s5
…
node 2
node ns2
Web Service Provider
3
2: service fetch &deploy
SR
Service Repository
Dynamic Service Deployment - Dynasoar
R
Slide 22eScience Meeting, Edinburgh, November 2006.
DAME developed a tool to analyse large volumes of distributed signal data
CARMEN will extend this to: allow search and management of labelled data link the search results to data descriptions to allow better ranking and data analysis
Data Exploration
Slide 23eScience Meeting, Edinburgh, November 2006.
Metadata• Import tools enable users to describe
experimental conditions• Analysis services describe their own functionality• Registry of data and services
• “is there any data captured under conditions x, y & z?”• “what services are available to process this spike train
data?”
• Automatic provenance generation
Slide 24eScience Meeting, Edinburgh, November 2006.
e-Science Stretch
• Tool for locating patterns in time-series data across multiple levels of abstraction
• Dynamic service provisioning over a grid• Extensible, standardised metadata for
neuroscience• Fine-grained access control• Integrating data from multiple repositories
Slide 25eScience Meeting, Edinburgh, November 2006.
Work Packages
WP 0Data Storage
& Analysis
WP1 Spike Detection& Sorting
WP2 Information TheoreticAnalysis of Derived Signals
WP 3 Data-Driven ParameterDetermination in
Conductance-Based Models
WP4 Measurement and Visualisationof Spike Synchronisation
WP5 Multilevel Analysis andModelling in Networks
WP4 Intelligent Database Querying
Slide 26eScience Meeting, Edinburgh, November 2006.
Spike Detection & Sorting (WP1: Stirling & Leicester)
Analogue recording (digitised)
Cluster 1 Cluster 2
Doesn’t fit!
(Clustering using wave_clus)
Slide 27eScience Meeting, Edinburgh, November 2006.
CARMEN and spike detection and sorting
Idea is to provide many services
Several different types of spike detection algorithmsSeveral different types of spike sorting techniques
(including different types of data reduction, as well as different types of clustering)
Allow the user to test with a variety of techniques, and then choose the techniques they prefer
High speed links should allow immediate transfer of some datasets to Grid based systems
Allow experimentalist to choose near-real-time detection and sorting for immediate feedback
To assist during the experimentSlower (and more effective) techniques for later analysis off-line.
Allow comparison of different techniques on a wide variety of dataWhich is best, and for what?
Slide 28eScience Meeting, Edinburgh, November 2006.
Information Theoretic Analysis of Electrically- and Optically-Derived Signals (WP2: Imperial College, Manchester, and UCL)
Action potentials will be detected both electrically and optically.
Action potentials (spikes) are the primary electrical communication mechanism between neurons.
How can one interpret neuronal action potentials?
Information Theory, can be used to establish the ‘neuronal code’
quantifying how much information is carried by different potential neuronal coding mechanisms.
Issues:
Sampling problems, Spike correlation, Multimodal recording
By using Grid technology, we can assemble large quantities of optical and multi-electrode recordings and apply existing and novel techniques to its analysis.
We can make these techniques available as services.
Slide 29eScience Meeting, Edinburgh, November 2006.
Data-Driven Parameter Determination in Conductance-Based Models (WP3: Sheffield: Gurney et al)
Neuron modelling uses conductance based models.
[Many ionic species cross the neural membrane.Ion channels embedded in the membrane accomplish this transportThere are many different ion channel typesSetting the parameters for each type would enable better understanding of neuron operation]
Determining parameters for neural models is difficult and requires a great deal of data.
The parameters are not constant, and vary with (e.g.)Cell typePresence and concentration of neuromodulatorsTemperature
CARMEN aims to provide this volume of data, and hence to enable many of these parameters to be determined.
Slide 30eScience Meeting, Edinburgh, November 2006.
‘Compartmental’ modelling of morphology
Real neural morphology
approximation
(Passive) electrical equivalent
Slide 31eScience Meeting, Edinburgh, November 2006.
Fitting the model to current clamp data
100 ms
50 mV
100ms
50mV
Data (Wilson)
Model
Wood et al., Neurocomputing, 2003
Slide 32eScience Meeting, Edinburgh, November 2006.
Measurement and Visualisation of Spike Synchronisation (WP5: Newcastle, Plymouth)
More advanced analytical techniques are required to handle large scale, simultaneous recordings arising from MEAs.
Visualisation is critical to understanding what is happening
WP5 aims to: •develop reliable and robust analysis techniques to address these issues, particularly sweeping statistical methods to test if measures show significant changes•develop novel visualisation methods for displaying the results from these techniques, particularly those working in high-dimensional space•conduct real-time analysis of spike coding through a distributed Grid-enabled virtual laboratory.
Advanced (but fast) visualisation techniques are important to the whole community using CARMEN
Slide 33eScience Meeting, Edinburgh, November 2006.
Gravitational Clustering
Particle aggregation in gravitational clustering, (Gerstein and Lindsey(2006)). Each particle represents a cell; charges on each cells are incremented with each spike, and a force occurs between particles dependent on the charges.
Slide 34eScience Meeting, Edinburgh, November 2006.
Multilevel Analysis and Modelling in Networks (WP6: Newcastle, St. Andrews, Cambridge)
This WP aims to integrate the work of WP1-4, using the technology of WP0.
Understanding activity dynamics within neuronal networks is a major challenge in neuroscience
requires simultaneous recording from large numbers of neurons.
This WP will provide•integration of existing and novel network analysis techniques into CARMEN in order to build comprehensive models of network dynamics•data of exceptional quality and detailed provenance for the CARMEN repository for analysis of network properties•development of new dynamic Bayesian network algorithms to trace paths of neural information flow in networks.
For example: waves of activity in early turtle retina, recorded using Ca++ sensitive dye. (Thanks to Evelyne Sernagor, ION, Newcastle University)
Slide 35eScience Meeting, Edinburgh, November 2006.
QuickTime™ and aCinepak decompressor
are needed to see this picture.
Slide 36eScience Meeting, Edinburgh, November 2006.
Slide 37eScience Meeting, Edinburgh, November 2006.
Concluding remarks
CARMEN is a recent project (funding started October 2006).The baseline support technology is still being assembled.
It’s not the first attempt at making neurophysiological recordings re-usable
But:
CARMEN will contain more than recordingsServices, workflows, capability of using multiple data formats
CARMEN builds on earlier e_Science projectsRe-use not re-invention
We have experimental neuroscientists, informaticians, and computational neuroscientists all on board
Tackling the broad range of issues from multiple perspectives.