CNS*07 Electrophysiology Databases and Analysis Software Post-Workshop Summary

Following the introduction by workshop organizer Cengiz Gunay, Padraig Gleeson's talked about the use of the neuroConstruct tool to build neuronal network models for multiple simulators. Dr. Gleeson's work aimed at unifying different approaches to neural simulation under a flexible neural model description language, NeuroML.   The Neural Open Markup Language project, NeuroML [1, 2] [http://www.neuroml.org], is an international, collaborative initiative to create standards for the description and interchange of models of neuronal systems. MorphML and ChannelML are standards under the NeuroML framework, and they are designed to model morphology and ion channels, respectively. These languages take their strength from the XML language which already changed many other fields, such as e-commerce, by standardizing the way computers and software communicate. Dr. Gleeson's neuroConstruct software is a Java application that can read model descriptions and visualize neuron and network topologies [3]. It can delegate simulations to Neuron and Genesis neural simulators, collect and visualize their outputs. To do this, it creates Neuron and Genesis input files from the model's description in NeuroML. It can load morphology files from a number of formats and convert between them, as well. NeuroConstruct's major advantage is its flexibility to attach its models to new simulators, and understand new input formats. This was made possible by employing modern computer science concepts such as XML style-sheets (XSLT). These recipes allow transforming, for example a channel description in ChannelML, to a web page, a kinetics plot, or transforming a neural model to a Genesis input script. This makes adding new target simulators as easy as defining a new XSLT transformation. In fairness, neuroConstruct cannot yet support all Genesis and Neuron model descriptions. Given the limitless programming options available in a full simulator, NeuroML cannot convert an arbitrary Genesis or Neuron script into NeuroML. Dr. Gleeson recommended that researchers starting new models would choose an existing simulator, and transfer the model to NeuroML only after its maturation. The CNS audience appreciated Dr. Gleeson's software and several people showed interest in using it. The consensus was that neuroConstruct could provide a useful medium for collaboration between modelers and for testing and validations of available models.

Tom Morse talked about the use of computational intellligence for electrophysiological databases (EPDBs). This subject was changed from his proposed title on data sharing methods with NeuronDB and ModelDB, because these has already been discussed in a spontenously-formed workshop the previous day. Dr. Morse made a comprehensive effort to justify the need for making EPDBs widely available and feasible. He suggested that software utilities, such as spike sorting methods, should be collected in a central repository, similar to the SimToolDB repository [http://senselab.med.yale.edu/simtooldb/]. He identified several reasons that require EPDBs. One was parameter extraction approaches which required a seed of single cell data to verify the models found. The major obstacle to creating public EPDBs was that each researcher keeps their own specialized EPDBs. Many people agreed that this is most simple for many research projects, and maintaining a large common database requires a lot of time that experimentalists in the field cannot afford. A solution was offered to prepare EPDB support into existing data acquisition systems such that the experimentalist did not spend extra time for entering data. Another question that was raised was how the experimenter can be credited if his/her data was used. Unfortunately, the alternative to EPDBs is the use of "data thief" software to get data from published papers. Many people agreed from their own experience that this is tedious and inadequate solution, but may be the last resort in certain cases. Dr. Morse's conclusion was that the much-hyped semantic web failed to bring its promise so far, and there is still a dire need for data sharing among modelers and electrophysiologists.

Workshop organizer Tomasz Smolinski introduced the next session on special data analysis methods. As first speaker, Bill Lytton focused on data-mining algorithms in spike-wave detection and seizure classification. He reviewed the need for data-mining in biological projects. He pointed to the Structured Query Language (SQL) as one of the widely adopted and easy-to-use database software for data-mining. His Neural Query System (NQS) is a software package that allows making similar queries from within the Neuron simulator, as well as connecting to an SQL engine. He applied this method to seizure prediction from recorded traces by analyzing "bumps" in the data. In this process, he introduced a graph that can convey information in five-dimensions using various parameters of circles to represent the different aspects of the data. His method involved using K-means clustering of bump intervals.

Jean-Marc Fellous talked about a method for discovering spatio-temporal spike patterns in multi-unit recordings. Timing and reliability of timing from multiple trials or animals has been an interesting question [5]. He introduced a method that involved sorting spike rasters for finding order among them. This method used the similarity matrix obtained by comparing spike rasterograms after convolving with a Gaussian kernel. Then, fuzzy-clustering was used to organize the matrix into distinct regions, which was used to sort the raster plots. In the discussion, a method based on random shuffling was proposed to replace the fuzzy clustering.

Workshop organizer Bill Lytton introduced the next session on community software projects. Cengiz Gunay presented his PANDORA Matlab toolbox for analyzing simulated or recorded intracellular traces. He demonstrated databases can be created from recorded or simulated data alike, and complex analysis can be performed to result in descriptive plots. The toolbox's features produced substantial interest but there was some concern that its dependency on a commercial package (Matlab)  rather than a free software variant (such as GNU Octave or Python) could limit its adaptation.  A request was made to have more database templates suitable for researchers using different experimental and model setups. During the discussion, the need for common data to test new algorithms was voiced again. Dr. Gunay's toolbox can be downloaded for free [http://userwww.service.emory.edu/~cgunay/pandora].

Horatiu Voicu demonstrated a very low-maintenance method to feed parameter values into custom simulation software as an alternative to creating sophisticated graphical user interfaces. He demonstrated this method using a free text editor software, GWD [http://www.gwdsoft.com/], and the message-passing capabilities of the Windows operating system to drive a hippocampal simulation system. He was able to change arbitrary parameter values of the simulator on-the-fly. His software can be downloaded from [http://www.voicu.us/software.zip].

Workshop organizer Cengiz Gunay introduced the final session on parameter search and other analysis methods. Adam Taylor talked about mapping from model neuron parameters to functional output. He demonstrated methods for correlating variability of channels with other measurements. He aimed to model the results of mRNA measurements predicting channel densities [4]. He created a model database with 80k models and found 100 models matching target data within 1 STD. He used a scatter plot matrix to explain these matches. From this database he concluded that real cells are more constrained in changing their conductance densities than the models he found. To find such constraints in real neuron data he used linear and quadratic fits to channel dependencies from recorded data.

Gloster Aaron presented a method for finding repeating synaptic inputs on a single neuron. He used Matlab programs for finding repeating synaptic inputs in intracellular voltage data from recordings [6]. However, he showed that his method did not work with new data. There was a discussion on the non-stationarity of the recorded data affecting method results. The resolution was to adjust the window of comparison to have sufficiently good estimate of the mean and variance of the data.

The workshop closed with the audience's wishes for its repeat in the next year's CNS meeting.

References

1. Goddard, NH, Hucka, M, Howell, F, Cornelis, H, Shankar, K, David Beeman: Towards NeuroML: model description methods for collaborative modelling in neuroscience. Philos Trans R Soc Lond B Biol Sci, 2001. 356(1412): p. 1209-28.

2. S Crook, P Gleeson, F Howell, J Svitak, RA Silver: MorphML: Level 1 of the NeuroML standards for neuronal morphology data and model specification. Neuroinf 2007, in press

3. Gleeson P, Steuber V, Silver RA: neuroConstruct: a tool for modeling networks of neurons in 3D space. Neuron 54 (2): 219-235 APR 19 2007.

4. Schulz DJ, Goaillard JM, Marder E: Variable channel expression in identified single and electrically coupled neurons in different animals. NATURE NEUROSCIENCE 9 (3): 356-362 MAR 2006.

5. Mainen ZF, Sejnowski TJ (1995) Reliability of spike timing in neocortical neurons. Science 268:1503-1506.

6. Ikegaya YI, Aaron GB, Cossart R, Aronov D, Lampl I, Ferster D, Yuste R. Synfire Chains and Cortical Songs: Elastic Temporal Modules of Cortical Activity. Science 304: 559-564, 2004.