The Book of GENESIS

The last several years have seen a dramatic increase in the number of neurobiologists building or using computer-based models as a regular part of their efforts to understand how different neural systems function (Eeckman and Bower 1993, Bower 1992). As experimental data continue to be amassed, it is increasingly clear that detailed physiological and anatomical data alone are not enough to infer how neural circuits work. Experimentalists appear to be recognizing the need for the quantitative approach to exploring the functional consequences of particular neuronal features that is provided by modeling. This combination of modeling and experimental work has led to the creation of the new discipline of computational neuroscience (Eeckman and Bower 1993).

Before beginning to explore the tutorials, it is important to understand something about the assumptions and the mathematical models that underlie these simulations. Thus, although the first section of the book describes what are essentially “point and click” tutorials, it is important not to use these tutorials blindly. Their effective use requires some understanding of the basics of neural modeling, as well as the concepts in neuroscience that are introduced along with the tutorials. Entire books have been written on this subject, so obviously we can only highlight the issues here. However, throughout the text we have referenced other sources of information. If you are seriously considering building models yourself, we would strongly recommend that you consult these references.

Now that we have briefly described the numerical basis for the tutorials included in the first part of this book, we are ready to get started with running the tutorials. The tutorials included in this manual are all constructed using GENESIS, the General NEural SImulation System that has been under development in our laboratory at Caltech since 1985. This chapter is intended to introduce each of the tutorials, as well as provide a demonstration as to how to use the GENESIS graphical interface. First, however, we provide some basic information about the GENESIS system on which the tutorials are based.

In the previous chapter, we used a single compartment model to study the mechanisms for the activation of voltage-activated channels, which produce neuron firing. Next, we need to understand how inputs to the neuron affect the potential in the soma and other regions that contain these channels. The following chapter deals with the response of the neuron to synaptic inputs to produce postsynaptic potentials (PSPs). In this chapter, we concentrate on modeling the spread of the PSP through the dendritic tree.

The simple neuron model that we have built up over the past three chapters is what Llinás (1988) has referred to as the “Platonic Neuron.” In this idealized model, postsynaptic potentials in the dendrites (Chapter 6) propagate passively through the dendritic “cable” (Chapter 5) to the soma. Here, near the axon hillock, the summed and attenuated PSPs may activate voltage-dependent sodium and potassium channels that are very much like those found in the squid giant axon (Chapter 4). Although it was once assumed that this simplified description applied to most neurons, we now know that the situation can be much more complex.

Many organisms exhibit repetitive or oscillatory patterns of muscle activity that produce rhythmic movements such as locomotion, breathing, chewing and scratching. Examples include the escape swimming of the mollusc Tritonia diomedia, the digestive rhythms of the lobster, the undulatory swimming movements of the fish or the lamprey, the stepping movements of the cockroach, the rapid wing motion of the locust during flight, and the more complicated locomotion of a quadruped mammal such as the domestic cat. The neuronal circuits that give rise to the patterns of muscle contractions which produce these movements are referred to as central pattern generators, or CPGs. Various experimental preparations in which the CPG is isolated from external influence demonstrate that these circuits require no external control for the generation of temporal sequences of rhythmic activity. However, these animals move through the world in an adaptive manner where the same motorneurons are involved in the production of a variety of rhythmic behaviors. Thus, many CPGs are capable of producing multiple patterns of activity in the intact behaving animal (Getting 1989). The ability to switch between different motor behaviors and blend different rhythms relies on feedback from proprioceptors and influence from higher centers of the nervous system; therefore, it is most appropriate to view every CPG as one piece of a distributed control system (Cohen 1992).

Previous chapters in this volume have considered detailed models of single cells and small networks of cells. In this chapter we consider a large scale multicellular model of the mammalian olfactory cortex. The simulation consists of three distinct neuronal populations of 135 cells each for a total of 405 interconnected neurons. With this simulation, we will explore the possible physiological basis for experimentally recorded electroencephalographic patterns in this cortex.

In the preceding chapters, we have taken the building blocks for neuronal models in the form of ion channels, membrane compartments and synapses (Chapters 4, 5 and 6) and put them together to form neurons, small neural circuits, and then networks (Chapters 7, 8 and 9). We conclude this section of the book by opening the lid on the building blocks, and seeing what happens inside them. There used to be a view of the atom as a tiny solar system, which led to exotic visions of an infinite series of ever-decreasing worlds as one examined matter at finer and finer scales. Neurobiology now has to face a similar unsettling concept: that hidden within “atomic” compartments and under all the ion channels, there exists a whole new universe of subcellular networks. These networks, of course, are the multitudes of interacting biochemical signaling pathways. They profoundly affect everything from the properties of single channels to the morphology of neurons and the wiring of the brain itself. Recent work on biochemical signaling reactions has emphasized the sheer complexity of these networks. There are dozens of known major signaling pathways, each having at least five to ten enzyme isoforms, each of which communicates with a different subset of other messengers and pathways. As with neuronal networks themselves, these biochemical networks cannot be analyzed in isolation. Neuronal signaling events from above, and nuclear and metabolic events from below, provide a barrage of signals that this network must process.

You have now explored several tutorials intended to both provide some insights into fundamental concepts in neurobiology and to introduce you to the use of neural simulations. The remainder of this book is intended to guide you in the creation of your own simulations. In general, we have found that the most efficient way to develop new simulations is to modify one that already exists and that you understand well. We would encourage you to consider starting with a tutorial most similar to the simulation you would like to build, after having learned a few basics of GENESIS programming from the following chapters.

Part I of this book has used several existing GENESIS simulations to introduce some of the theory underlying neural modeling. In Part II, we will build upon this background, using the GENESIS script language to create our own simulations. We will begin by simulating a simple neural compartment like that described in the first two sections of Chapter 2. Before proceeding with this tutorial, you may find it useful to review that material.

In the previous chapter, we modeled the simple passive compartment shown in Fig. 12.1. The membrane resistance R _m is in series with a leakage battery E _rest , and is in parallel with a membrane capacitance C _m and a current source I _inject . Assembled into a script, the commands used would look something like the listing shown in Fig. 13.1. Note the use of “//” for comments. Multi-line comments may be entered by bracketing the commented lines with “/*” and “*/”, as in C. This script also illustrates the use of the backslash character, “\”, in order to continue a statement (the second addmsg command) to another line. This will be useful when entering lengthy setfield commands.

In the previous chapter, we created a simple multi-compartmental neuron with a dendrite compartment, a soma, and an axon. The dendrite contained a synaptically activated channel and the soma contained voltage-activated Hodgkin—Huxley sodium and potassium channels. The script tutorial4.g, listed in Appendix B, contains the function definitions and commands used to construct this neuron, provide synaptic input to the cell, and plot the results of the simulation. In this tutorial, we will modify the script in order to construct the same neuron from a data file, using the GENESIS cell reader. The cell reader, which is implemented in the readcell command, allows one to build multi-compartmental neurons by reading cell parameters from a cell descriptor file. This file contains the names and dimensions of the compartments that will be used in the cell, along with the names of the channels and other elements that are contained within each compartment. The cell reader then uses this information to put together a cell and to establish the necessary messages between the various elements. This can significantly reduce the effort needed to construct a complex cell with many compartments and channels.

In the first four GENESIS programming tutorials, we covered the features of the GENESIS/XODUS script language needed to construct a simple model neuron. This neuron contains a dendrite compartment with a synaptically activated excitatory channel and a soma with Hodgkin—Huxley sodium and potassium channels. A source of randomly distributed spikes is used to excite the synapse. Action potentials produced in the soma trigger a spike generator that may be used to provide input to a synapse on another cell. In our model, we used a feedback connection to the cell’s own synapse that can be toggled on and off. Enough details of XODUS were introduced to create graphs for the membrane potential and channel conductance, along with buttons, toggles and dialog boxes for controlling the simulation.

As your GENESIS simulations grow in complexity, either through increased level of detail in multi-compartmental cell models, or through increased size of network models, you will eventually begin to look for ways to increase the speed of your simulations. In the following section, we present some general hints, methods and “tricks” that you can use to make your simulations run faster. For models that contain many compartments, you can obtain large speedups by using a more efficient numerical method for the integration of the compartmental equations. In fact, as we will see in Sec. 20.3.4, the default integration method that is used in GENESIS is much better suited to simple cell models with few compartments. The rest of this chapter describes how you may implement these improved methods, with emphasis on the fast implicit methods that are associated with the GENESIS hsolve object.

PGENESIS is a parallel form of GENESIS that enables simulation of very large models. Simulation models are critical for integration of behavioral data with anatomical and physiological data. Although explanations of behavioral data are possible without resort to neural simulation models (Chomsky 1957, e.g.), those integrative accounts that make contact with the anatomical and physiological data require large-scale simulation models at the neural level. The scale of the models required can be seen in theories about the function of the hippocampus in learning and memory (McClelland and Goddard 1996, Levy 1996). These theories assert that statistical properties of firing rates, synaptic transmission efficiencies, and connection structures are crucial in explaining information processing in the hippocampus. The validity of these statistical properties is conditioned on sufficient sample sizes that cannot hold if the model scales down the real system by more than one or two orders of magnitude. Even scaling down by two orders of magnitude leaves us with very large models that, as we shall see, go beyond the capabilities of existing simulation environments.

One of the biggest hazards in developing a simulation is the pressure to make it user-friendly. If you ever make the mistake of making a simulation easy for people to play with and understand, they will suddenly discover gaping holes in your model design, and start to think up all sorts of “improvements” for you to make An even more unpleasant situation can arise when you have sent off the final page proofs of your simulation paper, and decide that now is a good time to provide it with a colorful display so that people reading your paper can run the simulation themselves. Inevitably, the display will reveal a fundamental bug in the simulation that no one (least of all yourself) would ever have noticed in all the hundreds of lines of simulation code. The prudent builder of simulations will avoid any compromises when it comes to obfuscation. This chapter, then, is for the reckless, since its stated goal is to reveal all, to shine the bright light of day on the hidden crannies of simulations where bugs lurk, and to display the gory details using the rainbow colorscale in an animated three-dimensional draw widget.