|
To obtain the best correspondence between a model of a neuron and the data
recorded from the neuron, the model must include a realistic morphology along
with estimates of the neuron's biophysical parameters. To accomplish the best
estimates of the internal axial resistivity Ri, membrane resistivity Rm, and the
membrane capacitance, as well as several other parameters, you can run the
"modelfit" program which runs the Levenberg-Marquardt algorithm for
least-squares fitting. The program evaluates the model containing the realistic
morphology and compares the output to a set of data points, which are often a
voltage or current trace recorded from the real cell. The modelfit program then
computes the difference between the real trace and the model trace for all the
data points, and computes the sum of the squares of the differences, which
emphasizes the biggest differences. The program then runs the model with
different sets of parameters, computing the sum of the squared differences for
each set, and choosing the parameter set that has the smallest sum of squared
differences.
The modelfit program can work with a parameter set containing up to 12
parameters. Each parameter can be limited to a specified range and can be given
a starting value for the least-squares fit. In practice it is usually best to
give a maximum of 5-8 free parameters, depending on their influence in the
model. The modelfit program passes the parameters to the retsim model which runs
the model and returns with the output data to be compared with the real data.
The parameters are typically Rm (membrane resistivity), Ri (axial restivity), Cm
(membrane capacitance per unit area), and others such as electrode resistance
and capacitance, and injected current. The Rm and Ri parameters can be split
into different sub-parameters for the different regions of the cell.
The modelfit program is given a "charging curve" data set that shows how much
current drawn by an electrode that is voltage clamping the cell with a pulse
voltage. Typically the first few msec represent the charging of the cell soma
where the electrode is located. As time advances, the current recorded from the
electrode represents segments of the cell more distal from the soma, for
example, the axon and dendrites. When the values of Ri, Rm, Cm and the other
parameters are optimal, the model produces the best fit to the original data.
To accelerate the process of running models to compare to the original real
data, several copies of the the modelfit program can be run in parallel with
different sets of parameter values. These parallel "jobs" can be run on a
parallel job system such as "Mosix" which automatically runs computational jobs
on a set of CPUs. The system attempts to load each CPU with an identical load
of jobs so all the Jobs are run at the same overall speed.
Once these basic biophysical parameters have been fitted, the model can be run
using these parameter values with the realistic morphology, for the purpose of
testing the function of the circuit using other parameters.
For more on this fitting process, see Curve Fitting with Neuron-C
Recent studies that used modelfit include:
Percival KA, Venkataramani S, Smith RG, Taylor WR (2017) Directional Excitatory Input to Direction-Selective Ganglion Cells in the Rabbit Retina. J. Comp. Neurol. Accepted Jan 31, 2017. doi: 10.1002/cne.24207
Stincic T, Smith RG, Taylor WR (2016) Time-course of EPSCs in On-type starburst amacrine cells in mouse retina is independent of dendritic location. J. Physiol. 594:5685-5694, doi: 10.1113/jp272384.
|