Can an artificial neural network do more than predict brain activity—and actually interact with living neurons in real time?

Reading time: about 6 minutes.

Most spiking neural networks (SNNs) are used either to study neural activity in simulations or to build efficient AI systems. BiœmuS asks a different question: what would it take to build an artificial neural circuit that can exchange signals with a living brain or brain organoid while the experiment is happening?

In a 2024 Nature Communications paper, Beaubois and colleagues presented a small FPGA-based platform that runs detailed biophysical neuron models, reproduces selected neural activity patterns, and connects to real biological recording and stimulation devices. The important achievement is not a new learning algorithm. It is turning a simulated network into a real-time component of a biohybrid experiment.

1. Why is “real time” harder than “fast simulation”?

Imagine recording a burst of spikes from a brain organoid. You want an artificial network to respond, then electrically stimulate that organoid based on the response.

It is not enough for a simulation to finish one second of neural activity in less than one second on average. In a closed-loop experiment, a late response may arrive after the biological event you intended to influence. The system also needs predictable timing and working connections to lab equipment.

There is another design choice: bio-inspired versus biomimetic neurons. A simple leaky integrate-and-fire (LIF) neuron makes a convenient model for efficient computation. A Hodgkin–Huxley (HH) family neuron instead tracks how membrane voltage changes through ion-channel currents. That richer description costs more computation, but its parameters have a more direct relationship to neural electrophysiology.

BiœmuS uses a Pospischil-style HH model with six conductance-based currents and configurable biophysical synapses, including AMPA, NMDA and GABA receptor dynamics. Its goal is to emulate relevant neural dynamics at biological time scales, not just generate spikes.

2. How does BiœmuS run a complex SNN on a small board?

The trick is to divide responsibilities between two parts of an AMD Xilinx Kria KR260 system:

  • FPGA programmable logic (PL): carries out the neuron and synapse updates on a fixed, hard-real-time schedule.
  • CPU processing system (PS): runs Ubuntu and a C++ controller to configure the network, monitor activity and communicate with equipment through familiar interfaces.

Users create network configurations with Python. They can test them in software before loading them onto the FPGA. This keeps experiment design flexible without sacrificing deterministic computation.

Figure 1 — Overview of the BiœmuS platform

Figure 1 | Original paper Figure 1 (unmodified). The crucial boundary is between the FPGA that performs time-critical HH calculations and the Ubuntu/CPU side that handles configuration and communication. Source: Beaubois et al., Nature Communications (2024), CC BY 4.0.

The reported capacity is 1,024 HH neurons and 2²⁰ (1,048,576) synapses, updated every 31.25 μs—32,000 model steps per second. Total board power consumption is reported as approximately 6.50 W.

How can HH calculations fit into that schedule? One practical optimization is to precompute expensive voltage-dependent ion-channel functions in lookup tables. At runtime, the FPGA mostly performs table reads and multiply-add operations instead of repeatedly evaluating expensive functions such as exponentials. This is an example of hardware–model co-design, not a simplification to LIF neurons.

3. Does the artificial network behave like a biological one?

A configurable SNN is useful only if it can reproduce some meaningful population behavior.

One demonstration models two cortical organoids. The authors compared a disconnected configuration, where the two networks have no links between them, with a connectoid-like configuration, where neurons in the two groups can connect across the boundary.

Figure 3 — Artificial disconnected and connected organoids

Figure 2 of this blog | Original paper Figure 3 (unmodified). Compare the raster plots: in the connected configuration, bursts across the two artificial organoid populations become more synchronized. Source: Beaubois et al., Nature Communications (2024), CC BY 4.0.

This teaches a useful distinction. A spike is an individual firing event; a network burst is a brief episode in which many neurons fire together. If the two populations become connected, their bursts can influence one another and align in time. BiœmuS reproduces this qualitative change in network dynamics.

In a second test, the authors built a clustered SNN to emulate activity in the rat rostral forelimb area (RFA). They tested 80 configurations and compared their mean firing rate, mean bursting rate and burst duration against recordings from six rats. The selected artificial networks approached the biological recordings on these summary measures.

But similar spike statistics do not prove that an artificial network replicates an entire brain circuit. The authors demonstrated matching on particular activity features, not functional equivalence.

4. The key experiment: connecting living and artificial neurons

The strongest part of this study comes after the simulations. In the organoid experiment, living connected organoids sat on a high-density microelectrode array (HD-MEA), which records extracellular spikes and delivers electrical stimulation. A gateway computer ran Python software connecting the HD-MEA to BiœmuS.

The loop worked in both directions:

  1. A burst in the living organoid is detected, and software stimulates the artificial SNN.
  2. A burst in the artificial SNN is detected, and software stimulates the living organoid through the HD-MEA.

Figure 5 — Biohybrid experimental prototypes

Figure 3 of this blog | Original paper Figure 5 (unmodified). Panel A shows the bidirectional organoid–FPGA experiment and the activity traces; panel B shows the separate rat-brain stimulation prototype. Source: Beaubois et al., Nature Communications (2024), CC BY 4.0.

That is a genuine biohybrid closed-loop prototype: living activity changes the artificial network’s input, and artificial activity can trigger a new intervention in living tissue.

The authors also showed a different, open-loop animal experiment. An artificial RFA model generated spikes that triggered stimulation in the rat’s somatosensory cortex (S1) via an INTAN interface. The interval from SNN spike detection to delivered stimulation was reported as less than 1 ms. This is an interface demonstration; it does not establish a new therapy’s clinical efficacy.

5. What has not been solved?

There is an important difference between the FPGA being real-time and the whole biohybrid loop being precisely synchronized.

  • The 31.25 μs step applies to computation inside the FPGA, not to every external device and software process.
  • The organoid loop still had around ±100 ms timing uncertainty when timestamps were aligned across systems. Better clock synchronization and interface characterization remain necessary.
  • Stimulation can create large electrical artifacts that resemble spikes. If those artifacts are mistakenly detected as neural events, they could trigger unwanted repeated stimulation.
  • 1,024 neurons is tiny compared with a real brain region. Memory for synaptic weights and precomputed lookup tables is also an important hardware bottleneck.

Takeaway

BiœmuS is best understood as an experimental bridge between computational neuroscience and living electrophysiology. It shows that a detailed, configurable neural model can be packaged into embedded hardware and plugged into real recording and stimulation setups.

The next challenge is not simply to simulate more neurons. It is to make the entire read–compute–stimulate loop reliable, precisely timed and biologically meaningful. This paper demonstrates that the bridge can be built—not that it can already cure neurological disease.

References

  • Beaubois, R. et al. (2024). BiœmuS: A new tool for neurological disorders studies through real-time emulation and hybridization using biomimetic Spiking Neural Network. Nature Communications, 15, 5142. DOI: 10.1038/s41467-024-48905-x.
  • Open-source BiœmuS repository.
  • Original figures: Figures 1, 3 and 5, reproduced without modification from Beaubois et al. (2024) under CC BY 4.0.