PSICS - the Parallel Stochastic Ion Channel Simulator
previous ←   → next
Under development
Please let us know of errors, unclear text or if you have suggestions form improvements:

If you don't find what you want here, try the forum.

The rate in the PSICS Hodgkin Huxley channels for n and m transitions looks to be ten times larger than the original model. Why is this?

The expression used for the Rallpack HH model is: alpha = A (v-V0) / (exp((v-V0)/B) - 1) where A has units 1/(mV msec). This is slightly different from the rate(A), scale(B) and midpoint(V0) parameters used by PSICS where rate = A x / (1 - exp(-x)) where x = (v - v0) / B.

The PSICS expression converts the potential to a dimensionless quantity, x, using the midpoint and scale, applies the function and then multiply by the rate constant, A. The units of A in PSICS are just per_ms. This differes from the Rallpack/HH form of the expression which just has A (v - V0) as the numerator, not A (v - V0)/B. So A in psics is larger than A in Rallpack by a factor of B, and the units of A in PSICS are just per_mV. It just happens that B=10 in this model which adds to the confusion.

The motivation for the form in PSICS is to make the parameter A a rate, rather than a rate per mV as in HH. It makes it a bit easier to understand and allows the parameterizations to be uniform across the different functional forms.

The documentation includes ranges for most variables. Are these hard limits or just suggestions? What effect do they have?

These are just suggestions. There is a rather loose convention that square brackets imply hard limits (eg gamma being between 0 and 1 for a gating transition would be written "[0, 1]" ) and round brackets for soft limits. At present, the only use is to give the user an idea of what might be plausible. But there are potential future uses - eg, a facility to compare a model's parameters with the suggested ranges and generate a report about any that are way out of range (this might be most useful for models that result from parameter fitting that aren't necessarily biologically plausible). The other place where they could be used is for an eventual user interface with sliders: you need to know what range to start with for the slider, but this is a long way off.

In the example models, quantites are sometimes written with a decimal point as in "-65.mV", rather than just "-65mV". Is this significant?

No, it doesn't make any difference. This is just a programming convention to be clear about whether a quantity is an integer or a floating point number. The purpose is to avoid accidental integer arithmetic, like writing (10 / 3) which evaluates to 3 in most programming languages, whereas (10. / 3) implies a floating point division and evaluates to 3.333... But in PSICS it doesn't make any differnce since the data types are internally defined already.

How are the tips of compartments treated? Is the area included? Can they contain ion channels?

The default behavior is to assume hemispherical ends. The end area goes into the surface area calculation so it affects both the compartment capacitance and the number of channels. You can also make a morphology with a single point and then it is assumed to be a sphere.

You can change the behavior by setting squareCaps="true" in the StructureDiscretization block. The main reason for this is to allow it to express models that exactly match the rallpack specification.

When I run it, a number of files with different extensions are created. What are they all?

For a simple run there will be three files: one ".ppp" file, one ".dat" and one ".txt". The .ppp file is the Psics Pre-Processed form of the model. This is a complete description of the model as it is to be run, with tabulated transitions rates for the ion channels and details of the compartements making up the discretized form. The .ppp file is the only input to the core calculation, so if PSICS is running on a grid, it is this file that gets moved around to the separate nodes.

The .dat file contains the raw results as binary data in the order in which they arrive. The core calculation simply writes to the end of the file as it goes along.

The .txt file is created for plotting convenience. It contains the same data as the .dat file but formatted as text. Also, if the .ppp file model specifies multiple runs with the same parameters, these pile up one after another in the .dat file. In the .txt file they are rearranged into columns with just one set of time points in the first column.

If the model specifies a paramter sweep, then there is one set of the three files for each paramter value.

Why another neural simulator?

PSICS was written to address a class of problems that can't be handled with existing tools: the efficient simulation of the stochastic behavior of kinetic scheme ion channels on complex geometries.

Since it also computes continuous channel behavior, there is some overlap with other systems, but this kind of choice is a good thing. It helps improve software quality in general and introduces new ideas that may or may not take hold depending how users respond to them.

It is also something of a myth to imagine that the community would be better served by concentrating development effort on a few large projects. As was eloquently been explained in The Mythical Man Month back in 1975, the dynamics of the software development process leads to drastically reduced efficiency as group sizes grow. The more effective approach is to focus on cross-system compatibility and modularization. These are core aspects of PSICS with its simple declarative file formats and standalone core computations.

You say it uses Fortran, but I didn't have to install any Fortran. Where is it?

The core that does the calculations is indeed written in Fortran. It is built as a single exectable that does not depend on any non-standard system libraries. The executable is included inside the .jar file. Each time you run PSICS, it unpacks the executable to a temporary directory and runs it from there. This avoids the need for a separate installation and eliminates possible version compatibility issues between the Java preprocessing and the core calculation.

Why Fortran and Java? Isn't c++ the best language for this kind of thing?

For a simultor, c++ does provide an appealing combination of flexibility with its object oriented structure and computational efficiency, but it is not the only answer.

PSICS is designed in such a way that the simulation task breaks down into two very different parts: processing a complex model specification into something ready to be run; then repeating the same small set of core steps over and over again. For the first the focus is on eas of writing and debugging, and Java, with its garbage collection, simple program structure, and cross platform deployment is an easy choice.

For the core calculation, Fortran has two great benefits: first it is a high level language with built in support for sophisticated array based operations, linear algebra, maths, data parallelization and smart IO constructs. This makes the code shorter and easier to write, read and maintain. And second, it is a compiler-aware language with numerous attributes that can be attached to variable declarations to tell the compiler how they will be used and, more importantly, how they will not be used. This includes things like saying you will never use a pointer to refer to an array element which lets the compiler use transient storage or even optimize away the array completely. The result is that, for calculations like the ones done by PSICS, Fortran makes it relatively easy to get a well optimized executable in a fairly predictable way.

Can I save a model from XYZ and open it with PSICS?

Not for a complete model, but you can import some components such as morphologies and channel definitions. This is partly because PSICS covers a different problem domain from other simulators, but also because the interchange formats needed for this kind of comaptibility are still under development (compare, for example, the effort behind SBML which arguably has a less diverse domain to address).

The ones that are supported are morphologies in SWC and MorphML formats, kinetic scheme channels in ChannelML and some of the more standard HH style channels in ChannelML.

previous ←   → next