Design the loop before the tissue exists.
A simulated neural culture on a virtual electrode array: it bursts on its own, is read through the same acquisition chain a lab would use, and answers to stimulation. Every burst, spike and pulse is known, so detectors and closed-loop controllers are built against ground truth first, and carried unchanged to living platforms later.
Why a synthetic bench is necessary
Living neural platforms are the most interesting substrate for closed-loop neuromodulation research, and the hardest place to learn on. Three problems make an in-silico stage the right first step.
Tissue time is scarce and expensive
Cultures are shared, booked by the hour, and replaced every few weeks. Protocol bugs found on tissue cost days. Here the same experiment runs thousands of times, at any hour, for the price of a CPU.
Biology has no ground truth
On a real array nobody knows which bursts a detector missed or which pulses did nothing. The sandbox keeps every event beside the recording, so latency, false triggers and suppression are measured, not guessed.
Each culture is different
Batch-to-batch variability makes results hard to reproduce and hard to review. Seeded simulations reproduce exactly, and heterogeneity is a dial instead of a surprise.
What the sandbox guides
It is an engineering instrument for the questions that come before biology: does the loop close fast enough, does the detector trip on the right events, and what does a manipulation do to a network we fully observe?
Closed-loop controller design
Responsive stimulation from detector to pulse, with refractory rules and rate limits, tuned on known bursts. Latency, false triggers per hour and stimulation load come out as numbers.
Detector validation
The burst detector is scored against true onsets: recall, precision, latency distributions, and how each threshold trades suppression against stimulation load, in a sweep.
Hypotheses about the medium and the field
Oxygen, temperature, adenosine, dopamine, toxin load and a weak DC-like field are dials on the model. They generate predictions to test, not results to publish.
A hardware on-ramp
Experiment code talks to one contract and an SDK-shaped façade. When access to a living platform arrives, the backend swaps and the protocol runs as written.
How it works
One loop, simulated end to end. Ground truth is recorded at every stage but never exposed to the parts of the system that would not have it on hardware.
Two culture models, one contract
Phenomenological
No neurons. Bursts, spikes, noise and stimulation responses are generated from culture statistics. Thirty times faster than real time; ideal for detector and loop development.
- Poisson bursts with a chirped 10–40 Hz oscillation
- Evoked and truncated bursts with tunable probabilities
- Deterministic by seed
Spiking network
Excitatory and inhibitory neurons in a dome on the array, distance-dependent wiring, short-term depression. Bursts are not scripted; they emerge, and ground truth is a population-rate definition.
- About real-time speed on a workstation CPU
- Quiet baseline, ~9 bursts per minute per organoid, 0.4 s each
- Medium and field dials act on synapses and drive
Principles
Every page, file and result carries the word. Nothing here is a biological finding.
Bursts, spikes, pulses and their effects are stored with the recording and used only for scoring.
Protocol code imports one façade. The simulator and a living platform are interchangeable backends.
Seed, configuration hash, code version and detector settings are written with the results.
Start with a culture that is already bursting.
The live page shows a simulated array in real time; experiments run in the background and report against ground truth.
Sign inNo accounts exist yet. The first administrator is created from the server environment (NPSIM_BOOTSTRAP_EMAIL) or with npsim users add you@example.com --role admin