Neural Emulation | VectorScaleDB

Three-Layer Architecture

A unified stack for neural data: connectome structure at the bottom, dynamic emulation in the middle, and cross-model comparison at the top.

Layer 1

Connectome Storage

Ingest neuron-level connectome data with synapse weights, neuropil assignments, and neurotransmitter types. Neurons are stored as temporal vectors with structural features. Synaptic connections form a queryable weighted graph.

Layer 2

Emulation Engine

Run neural dynamics on stored connectomes. Five emulation modes from simple rate-coding to biophysically detailed spiking models. Emulation state is captured as temporal vectors, enabling behavioral compression and time-travel queries.

Layer 3

Cross-Model Comparison

Compare outputs from different models running on the same connectome. Quantify agreement between neural network approaches, detect where models diverge, and correlate neural activity with behavioral outcomes.

Multi-Species Support

A species registry manages connectome definitions, neuron models, and gap junction configurations. Each species has its own neural parameters while sharing the same query and emulation infrastructure.

Nematode

C. elegans — 302 neurons

The most completely mapped nervous system in biology. Full connectome with chemical synapses and gap junctions. Ideal for whole-organism neural emulation and validation of emulation accuracy against decades of experimental data.

302 neurons, ~7,000 synaptic connections
Complete sensory-to-motor pathway tracing
Gap junction support for electrical coupling
Behavioral correlation with locomotion data

Fruit Fly

Drosophila — 139,000+ neurons

The FlyWire connectome: 139,000+ neurons with millions of synaptic connections across 78 neuropils. Ingest directly from FlyWire CSV exports. Query neural pathways, compute neuropil connectivity matrices, and analyze circuit activation patterns.

139,000+ neurons from FlyWire dataset
78 neuropil regions with typed connections
Excitatory/inhibitory balance analysis
Sensorimotor latency estimation

Connectome Query Functions

Nine built-in query functions for structural and functional connectome analysis.

Structural

Circuit topology

POST /v1/connectome/trace-pathway — trace neural pathways with signed weights
POST /v1/connectome/neuropil-connectivity — connectivity matrix between brain regions
POST /v1/connectome/ei-balance — excitatory/inhibitory ratio analysis

Functional

Dynamic analysis

POST /v1/connectome/circuit-activation — activation spread through circuits
POST /v1/connectome/neuropil-activity — activity summary per brain region
POST /v1/connectome/sensorimotor-latency — estimated latency from sensor to motor

Comparative

Cross-model analysis

POST /v1/connectome/cross-model-agreement — quantify model convergence
POST /v1/connectome/ablation-analysis — predict effects of neuron removal
POST /v1/connectome/behavioral-correlation — correlate neural activity with behavior

Five Emulation Modes

Choose the right fidelity-performance tradeoff for your research question. All modes produce temporal vectors that flow through the same compression, indexing, and query pipeline.

Mode 1

Rate Coding

Fastest mode. Neurons represented by firing rates. Suitable for large-scale population dynamics and rapid parameter sweeps across entire connectomes.

Mode 2

Leaky Integrate-and-Fire

Classic spiking neuron model with membrane potential dynamics. Captures spike timing and basic temporal coding. Good balance of fidelity and computational cost.

Mode 3

Conductance-Based

Ion channel dynamics with excitatory and inhibitory conductances. Captures synaptic integration, shunting inhibition, and dendritic processing effects.

Mode 4

GNN-Accelerated

Graph neural network acceleration for connectome-scale dynamics. Trained on biophysical simulations, runs orders of magnitude faster while preserving key dynamical properties.

Mode 5

Hybrid

Mixed-fidelity emulation: high-fidelity models for neurons of interest, simplified models for the surrounding network. Automatically routes computation based on region of focus.

All Modes

Unified output format

Every emulation mode produces temporal vectors in the same format. Switch between modes without changing downstream queries, dashboards, or analysis pipelines. Compare mode outputs side-by-side.

Cross-Model Comparison

Run multiple emulation approaches on the same connectome and quantify where they agree and diverge. Essential for validating computational models against each other and experimental data.

Agreement Metrics

Quantified model convergence

Compare outputs from different emulation modes or entirely different modeling frameworks. VectorScaleDB computes agreement scores per neuron, per neuropil, and across the whole connectome. Identify which circuits are model-sensitive and which are robust across approaches.

Behavioral Correlation

Neural activity meets behavior

Correlate emulated neural activity with observed behavioral data. Import locomotion traces, stimulus responses, or decision outcomes and find which neural circuits predict which behaviors. Temporal alignment handles different sampling rates automatically.

Neural Data Compression

Neural emulation generates enormous data volumes. Drift-based compression exploits the structure of neural activity — long quiescent periods punctuated by bursts — to achieve extreme compression ratios.

Quiescent Neurons

1,000x+ compression for silent circuits

Most neurons in a large connectome are silent at any given time. Intelligent write elimination suppresses storage for quiescent neurons entirely. A 139,000-neuron emulation where 95% of neurons are silent stores only the active 5% at full resolution.

Spike Trains

Regime-aligned segmentation

Active neurons produce spike trains with natural regime structure: bursting periods, tonic firing, and silence. Compression segments align with these neural regimes, preserving firing rate statistics, burst boundaries, and inter-spike interval distributions.