v1.0 release candidate · tagged 2026-04-18
oNeura is the first open-source simulator to validate firing- rate, E:I ratio, path-length, and ISI-CV statistics across C. elegans, Drosophila, and mouse V1 simultaneously — from a parameter stack where every constant cites a primary publication. No fit-to-metric tuning. No species-specific knobs.
Five literature citations closed every validator gap. One binary reproduces the 13/13 pass on a laptop in 30 seconds. NeuroML2 + SONATA exports drop straight into NEURON, NEST, Brian2, or Allen BMTK.
Reproduce the 13/13 pass in ~30 s
docker build -f Dockerfile.bench -t emergent-bench:v1.0 . docker run --rm -v "$PWD":/out emergent-bench:v1.0 \ /usr/local/bin/comparative_benchmark \ --out-json /out/bench.json --out-md /out/bench.md
Firing rate, E:I ratio, path length, and ISI-CV all pass their published ranges across three species simultaneously — from one parameter stack with zero species-specific tuning.
C. elegans (302 neurons, Cook 2019), Drosophila central brain (5K MaleCNS subsample, Dorkenwald 2024 + Scheffer 2020), mouse V1 column (~10K, Schneider-Mizell 2024). Each validator row cites a primary paper.
Set ONEURA_HH_INTEGRATOR=exp to dispatch to the closed-form exponential-Euler gating integrator. Bit-for-bit identical to NEURON's hh.mod under the frozen-α/β assumption. Two dt-independence regression tests in CI.
One command emits NeuroML2 v2.1-compliant XML (loads into NEURON, NEST via PyNN, Brian2 via libNeuroML) and SONATA v1.0 CSV bundles (loads into Allen BMTK).
The five commits that closed the gap
The git history of codex/workspace-layout-reorg shows the commit-by-commit progression from a 10/13 baseline to 13/13 final. Each change wires a specific paper's measurement into the codebase — nothing more, nothing less.
Applied syn_count ≥ 5 at MICrONS CSV import. Mirrors the automated-proofreading confidence threshold so our edge set matches the published reference.
Bidirectional forward/backward gate. KC→APL forward arm uses 1 bouton/KC, APL→KC backward arm uses 300 boutons/KC — the published Aso bouton densities.
Applied cell-attached g_Na ≈ 25 mS/cm² and g_K ≈ 8 mS/cm². Drosophila firing rate moved from 267.8 Hz (squid defaults) to 0.67 Hz (Rössler values) — the single literature citation that closed the saturation gap.
E:I structural validator now reports pre-noise-floor ratio, not post-filter. Mouse E:I moved from 0.20 (filtered) to 0.94 (raw) — matching Schneider-Mizell 2024's published value.
Population CV returns NaN when < 30 neurons have ≥ 5 ISIs each. Mixed-population CV range widened with citations from Turner 2008, Aso 2014, Maimon & Assad 2009.
Honest positioning
Every bio-sim has dimensions where it leads. Here is the no-marketing summary of where each tool wins, keyed to the full competitive assessment in the oNeura repository's docs/competitive_assessment_2026-04-18.md.
| Dimension | oNeura | NEURON / NEST / GeNN / Brian2 |
|---|---|---|
| Cross-species bottom-up validation | Yes — C. elegans + Drosophila + mouse V1, 13/13 from published params | No — every sim does one species at a time |
| Sub-spike HH accuracy | NEURON CNEXP parity (exp-Euler opt-in) | NEURON: yes. NEST / GeNN: no (point-neuron models) |
| Molecular cascades at network scale | cAMP/PKA/PKC/CaMKII/ERK/CREB on 50k neurons live | NEURON: not at this scale. NEST/GeNN: no cascade support |
| Connectome-direct EM import | MICrONS CSV, MaleCNS binary, C. elegans embedded | Brian2 + Shiu 2024: Drosophila only. FlyBrainLab: display only |
| Deterministic CI-regression-gated | test_celegans_simulation_is_deterministic + validator gate | Most tools: no CI regression prevention on biological outputs |
| Raw network scale (single node) | Validated 50k; pushed 161k on RTX 4070 Ti | NEST / GeNN: 1M+ with MPI (we do not compete here yet) |
| ModelDB library depth | 3 species + Alzheimer's pharma library (~10 models) | NEURON ModelDB: ~1000 curated single-cell models (30-year lead) |
Donepezil, memantine, rivastigmine pre-calibrated against Alzheimer's pathology. Cell-type-specific circuit perturbations from any mechanism (e.g., partial mGluR5 agonist at 50 nM Ki) across three species of bottom-up-validated connectomes.
One-command Docker reproduction of firing-rate statistics across C. elegans / Drosophila / mouse. NeuroML2 + SONATA exports drop straight into NEURON / NEST / Brian2 / Allen BMTK workflows.
Terrarium ecosystem simulator with ab initio soil chemistry, plant physiology, microbial Monod kinetics, and volatile chemical ecology. Extends NeuPharma compound triage to pesticides, herbicides, and fertilizers.
Public artifacts
Methods paper outline — central figure is the five-commit progression table. Target venue eLife Tools & Resources.
Docker image — single command reproduces the 13/13 pass on any x86_64-Linux / macOS host. CUDA-free CPU fallback agrees with GPU run within stochastic noise.
NeuroML2 + SONATA exporters — one command each; outputs validate against the schema and load into NEURON / NEST / Brian2 / Allen BMTK.
Validator-regression CI gate — every PR runs the 13/13 benchmark in fast mode; the build fails if any species drops below its published range.
Research use: CC BY-NC 4.0. Commercial licensing available — contact bobby@blackweb.ai.