oNeura: Physics-Native Neural Simulation

Abstract

Built entirely in Rust with native CUDA acceleration, oNeura runs the same molecular brain on a laptop (M4 Pro), datacenter GPUs (A100, H200), or in the browser via WebGPU - with aligned physics across those targets. We demonstrate learning on Pong (inspired by the DishBrain line of work), exploratory autonomous Doom behavior, and a 139,000-neuron Drosophila brain simulation with real connectome topology. Those public experiments are the dONN path; the same local codebase also carries whole-cell and atomistic layers below that slice, and supported pharmacology now bridges explicit molecules, docking/MM-GBSA calibration, and live ligand MD into the neural runtime instead of stopping at a detached side tool. All reported values in this document are tied to the experiment conditions stated in the relevant sections.

Scope of Claims

This page is therefore best read as a technical paper for the validated dONN path plus the strongest currently connected pharmacology pieces, not as a claim that every future whole-cell or ecosystem layer is already public and benchmarked.

Motivation: Why Not ANNs?

Artificial neural networks are mathematical abstractions. They model neurons as scalar weights and biases, discarding the molecular machinery that gives biological brains their most interesting properties: graceful degradation, pharmacological sensitivity, sleep-dependent consolidation, damage recovery, and developmental plasticity.

Architecture Overview

oNeura is built as one simulation stack spanning browser demos, native neural runtimes, whole-cell scheduling, and atomistic molecule modeling. The core engine handles neuron state, synaptic transmission, plasticity, and pharmacology in a unified timestep loop, while the lower layers provide local chemistry, permeation, binding, docking calibration, and ligand-MD perturbation for the supported pharmacology bridge. The point is not a public repo layout; the point is that these scales now operate as one connected system.

The same neuron model runs identically across CPU (Rust), GPU (CUDA), and browser (WebGPU). Hardware-agnostic physics means results are reproducible on any platform.

Current Repo Reality: ~173 State Variables Per Neuron, ~45-Stage dONN Pipeline

The older public copy described this system as 25 molecular subsystems per neuron. The current oNeura repo now presents the flagship Drosophila dONN path as substantially deeper: roughly 173 state variables per neuron flowing through an approximately 45-stage pipeline on each timestep, with many more channel, signaling, structural, organelle, and gene-regulation layers explicitly tracked in the working project.

dV/dt = (1/C_m) * (
    I_ext
    - g_Na * m³ * h * (V - E_Na)
    - g_K  * n⁴     * (V - E_K)
    - g_L            * (V - E_L)
    - g_Ca * s²      * (V - E_Ca)
    - g_AMPA         * (V - E_AMPA)
    - g_NMDA * B(V)  * (V - E_NMDA)
    - g_GABA         * (V - E_GABA)
)

The right claim is therefore not that oNeura has become less mechanistic. It is the opposite: the public site has lagged behind a repo that now describes a materially deeper dONN neuron and a much broader bottom-up digital-biology stack than the earlier 25-subsystem summary.

Three-Factor STDP Learning

Learning in oNeura is biologically grounded. Rather than backpropagation, synaptic weights change through spike-timing-dependent plasticity (STDP) with three factors: pre/post spike timing, NMDA receptor gating, and neuromodulatory eligibility traces.

delta_t = t_post - t_pre

if delta_t > 0:                              // Pre before post → LTP
    dw = A_plus * exp(-delta_t / tau_plus)   // tau_plus = 20ms
    // Insert AMPA receptors into postsynaptic membrane
else:                                        // Post before pre → LTD
    dw = -A_minus * exp(delta_t / tau_minus) // tau_minus = 20ms
    // Remove AMPA receptors from postsynaptic membrane

// Factor 2: NMDA gating (requires postsynaptic depolarization)
dw *= nmda_gate(V_post, Mg_block)

// Factor 3: Dopamine-modulated eligibility trace
eligibility += dw
eligibility *= decay_rate                    // Slow decay
w += learning_rate * eligibility * dopamine  // DA gates consolidation

// BCM metaplasticity: sliding threshold prevents runaway potentiation
theta_m = f(mean_postsynaptic_activity)

Pharmacology: Atomistically Calibrated Molecules into Neural State

Supported compounds in the native dONN path now run through an actual molecule-to-behavior bridge. They start as explicit molecules, build a molecular graph from SMILES, dock into a target-specific surrogate pocket, get rescored with MM-GBSA, permeate into the live run, and then update occupancy with a ligand MD microprobe before they alter neural conductance or transmitter state.

The whole-cell drug-interaction module still handles permeation, target binding, and rate modifiers, while the atomistic molecule layer supplies SMILES-defined molecules, docking, MM-GBSA, reactions, and ADMET. The important change is that these layers are now bridged for the supported neural pharmacology set rather than living as parallel tracks with no runtime contact.

SMILES
  -> MoleculeGraph
  -> surrogate-pocket flex_dock()
  -> compute_mmgbsa()
  -> calibrated baseline Kd
  -> Bateman PK -> plasma concentration
  -> membrane permeation
  -> live ligand MD microprobe
  -> effective Kd / occupancy
  -> conductance or transmitter modulation
  -> neural dynamics -> behavior

The remaining limit is scale, not missing plumbing. oNeura still does not run explicit receptor-pocket MD for every receptor instance in every neuron at every timestep, but the neural pharmacology path is now atomistically calibrated rather than hand-waved.

Experiment 1: DishBrain Pong

Replicating and extending the Cortical Labs DishBrain experiment. A 25,250-neuron spiking brain learns to play Pong using free-energy-style learning - the brain receives unpredictable stimulation on misses, and the absence of stimulation on hits acts as a reward signal. No explicit reward function. No backpropagation. Learning emerges from STDP.

Sensory layer:    250 neurons (ball position encoding)
Cortical layer: 5,000 neurons (STDP learning circuits)
  ├── Excitatory:  4,000 (80%)
  └── Inhibitory:  1,000 (20%)
Motor layer:      4× populations of 5,000 neurons each
  ├── UP    (5,000 neurons)
  ├── DOWN  (5,000 neurons)
  ├── STAY  (5,000 neurons)
  └── NULL  (5,000 neurons)

Total: 25,250 neurons, ~7M synapses
Connectivity: Sparse random (p=0.1-0.3 within populations)
dt = 0.1 ms, 12 simulation frames per rally

All runs: seed 42, 80 rallies, dt=0.1ms. The brain consistently outperforms random baseline (30%) on every hardware platform. Results are deterministic given the same seed.

Experiment 2: Doom Combat

A 5,050-neuron spiking brain plays Doom (VizDoom, "Defend the Center" scenario). The brain sees Doom through a molecular retina (50 retinal ganglion cells), processes visual input through STDP-learning cortical circuits, and outputs motor commands to move, turn, and shoot. After 12 teacher-guided episodes, the brain plays fully autonomously - zero decoder biases, zero teacher shaping.

Drug Effects on Combat Performance

These combat runs are not gameplay hacks. They sample the same neural target classes the native pharmacology stack modulates, with supported compounds now entering through explicit molecules, docking/MM-GBSA calibration, permeation, and occupancy before circuit behavior changes. Three drugs were tested across 3 runs each, measuring average score in Defend the Center.

Teacher pretraining builds synaptic connections via STDP by stimulating correct motor populations alongside visual input. During autonomous play, the brain chooses actions solely from L5 cortical spike counts - whichever motor population fires most wins.

Hardware Benchmarks

oNeura is hardware-agnostic. The same brain architecture runs identically on a laptop, datacenter GPU, or in the browser. All benchmark data is from actual experiment runs (seed 42, 80 rallies, dt=0.1ms, 25,250 neurons, ~7M synapses).

The H200 achieves 2.55x wall-time speedup over M4 Pro and 1.59x over A100. Hit rate variations between platforms are due to floating-point order-of-operations differences in parallel reductions, not physics divergence - the equations are identical.

Prior Art: What Makes oNeura Different

Limitations

Future Directions

Selected References

Future revisions should convert these selected references into inline citations and a fully formatted bibliography/appendix.