Embedding Living Neurons into Simulated Neural Networks

M.P. Nawrot, T.Pistohl, S. Schrader, U. Hehl, V. Rodriguez, A. Aertsen

Introduction

The in vitro preparation of acute and cultured brain tissue is widely and successfully used to study neuronal processing at the level of single neurons and synapses. Due to their isolated condition, however, the investigation of their functional interplay with an active network is strongly limited. At the same time, in virtu neural network simulations have reached a state where it is now possible to investigate moderately realistic network models of high complexity. Such model studies lead to predictions which, in turn, can be tested experimentally.

We combine both experimental approaches in real-time in vitro - in virtu experiments. Technically, we iplemented a biderectional interface between the network simulation and the electrophysiological setup with soft real-time capabilities. It uses standard low cost equipment and runs under Linux, which is the best suited platform for the network simulations. A more detailed technical account of this method is given below. For an account of exemplary applications of this method please refer to the following references.

Martin Nawrot

Nawrot MP, Pistohl T, Schrader S, Hehl U, Rodriguez A, Aertsen A (2003)
Embedding Living Neurons into Simulated Neural Networks
Proceedings of the 1st IEEE-EMBS Conference in Neural Engineering
[abstract], [poster],[PDF] below along with further references to recent results.

Technical Implementation

Hardware

As an interface between th cortex neuron and the simulation PC we used a PCI 1200 card (National Instruments). Simulations were performed on a dual P2 500Mhz, running LINUX. Communication between the I/O-card and our programm was provided by a driver from COMEDI.

Real-Time and Simulation Algorithm

A bidirectional communication between a cortex neuron and a network simulation was achieved by two processes: A timing and a simulation process. The first, which controls both the update step (setting the current command for the injection) and  the reading step (measuring the real neuron's voltage) was implemented as a so-called "soft realtime" algorithm meaning that a process uses the system clock as a reference, in contrast to "hard realtime", where the clock directly triggers any procedure. Since the first does not provide precise timing properties, an instance has to be introduced, being able to correct the possible time lags (Figure 1).

Figure 1. Priciple of the realtime-algorithm. Every timestep h the current is updated by a value calculated by the simulation. If a current commant fails to be within the timegrid (grey), as indicated by the red current step, the next step (first green step) is shortened by the duration of this delay, thus being on the time grid again. If a current command lasts for more than two timesteps (purple) the next possible result (second green step) is presented and the number of failed steps N is stored. Every 100 timesteps the algorithm tries to simulate N steps without any current command, keeping the simulation "up to date" while maintaining the current (third green step). The temporal "quality" of this algorithm is measured by the quotient of the average delay time and the simulated time, e.g. 95% accuracy means that the delay time was, in average, 5% of the simulated time. The calibration in our work (Figure 2 ) was determined by 50 test-runs (5 network sizes times 10 frequencies), any of them lasting one Minute.

Figure 2. Calibration of the soft real-time system. For the parameter range indicated by the light gray region a soft real-time accuracy of 95% is achieved. Black line: linear fit to the boundary values.

Simulation is triggered by the timing process (via interprocess communication): after each trigger the simulation starts and returns its result as soon as possible. If this fails, the above correction mechanism tries to maintain the time grid.

Sven Schrader