Hybrid Circuits of Interacting Computer Model and Biological Neurons

Sylvie Renaud-Le Masson, Gwendal Le Masson, Eve Marder, L. F. AbbottNeurIPS 1992

Researchers have developed a real-time "hybrid" system that connects a computer-simulated neuron to a living biological neural network. By using a high-speed digital signal processor to simulate realistic synaptic conductances, scientists can precisely manipulate individual "virtual" ion channels to observe how they alter the behavior of a real rhythmic biological circuit.

The Challenge of Network Dynamics

In neuroscience, understanding how a single neuron’s electrical properties influence an entire network is notoriously difficult. Experimentalists traditionally rely on pharmacology to block specific ion channels or synapses. However, drugs are often "messy"—they can be non-specific, slow to act, or impossible to wash out completely.

Conversely, pure computer models offer total control but often lack the essential "ground truth" of biological reality. This research introduces a powerful middle ground: the hybrid circuit. By replacing or supplementing a biological neuron with a highly controllable digital model, researchers can perform "surgical" interventions on a network's dynamics that are impossible with traditional biology alone.

Methods: The Real-Time Feedback Loop

The authors built a real-time feedback loop involving three main components to interface with the stomatogastric ganglion (STG) of a crab or lobster—a small circuit that generates rhythmic patterns for digestion.

  1. The Hardware: A Macintosh IIfx equipped with an AT&T DSP32C Digital Signal Processor (DSP) board handles the heavy computation, integrating 10–15 differential equations every 0.7 milliseconds.
  2. The Dynamic Interface: An "Axoclamp" amplifier running in discontinuous current clamp (DCC) mode.
  3. The Biological Preparation: The living STG tissue.
Experimental setup of the hybrid circuit.

The Loop: The system measures the membrane potential ($V_m$) of the biological neurons and feeds that data into the DSP model. The DSP calculates the appropriate "synaptic current" based on a mathematical model of a synapse, and then injects that exact amount of current back into the biological neuron. Crucially, this simulates a realistic conductance change rather than just a simple current injection.

Key Insights and Results

The system achieved an integration time step of 0.7 ms, fast enough to interact with biological neurons without perceptible lag. The model Lateral Pyloric (LP) neuron was successfully "tuned" to fire in the correct phase of the pyloric rhythm, effectively tricking the biological PD and PY neurons into treating the computer model as a peer.

  • Conductance vs. Current: The system doesn't just "shock" the neuron; it simulates a conductance. Because the effect of a synapse depends on the neuron's current voltage, simulating conductances ensures the "artificial" synapse behaves exactly like a biological one. The same artificial synaptic release produced different sized responses depending on the state of the biological post-synaptic neuron, proving the hybrid system captures complex, non-linear biological interactions.
  • The Power of $i_h$: The researchers demonstrated that by increasing the maximal conductance of a specific hyperpolarization-activated current ($i_h$) in the model neuron, they could delay the model's firing and subsequently increase the entire network's cycle period.
Hybrid circuit diagram and traces showing the effect of increased $i_h$ conductance.

Limitations and Future Work

The initial system was limited to two biological inputs and one model neuron due to the DSP board having only two input/output channels. Additionally, the sampling rate of the Axoclamp must be carefully synchronized with the DSP clock to avoid artifacts, presenting a technical hurdle for scaling up.

However, the authors suggest that later generation systems with more channels will allow for the creation of larger hybrid networks. This paves the way for new forms of "biologically-based computing" where silicon and carbon process information together seamlessly.

Summary updated

BibTeX
@inproceedings{masson1992hybrid,
  author = {Sylvie Renaud-Le Masson and Gwendal Le Masson and Eve Marder and L. F. Abbott},
  title = {{Hybrid Circuits of Interacting Computer Model and Biological Neurons}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1992},
}
Figure