Skip to content

NeuroML 2 Import Guide

Import neuron models from NeuroML 2 XML files into SC-NeuroCore.

Supported Cell Types

NeuroML 2 Tag SC-NeuroCore Model Notes
<iafCell> StochasticLIFNeuron Conductance-based, converted to tau-based
<iafRefCell> StochasticLIFNeuron With refractory period
<iafTauCell> StochasticLIFNeuron Direct tau mapping
<iafTauRefCell> StochasticLIFNeuron tau + refractory
<izhikevichCell> SCIzhikevichNeuron 2003 dimensionless model
<izhikevich2007Cell> Izhikevich2007Neuron Exact biophysical NeuroML units
<adExIaFCell> AdExNeuron Brette & Gerstner 2005

Usage

Python
from sc_neurocore.adapters.neuroml import import_neuroml, create_neuron

# Import all cells from a NeuroML file of point-cell definitions
cells = import_neuroml("my_cells.nml")

for cell in cells:
    print(f"{cell.cell_id}: {cell.cell_type}")
    for note in cell.notes:  # what the mapping assumed or could not carry
        print(f"  {note}")
    neuron = create_neuron(cell)
    spike = neuron.step(10.0)

What Is Refused

The importer never invents a value. It raises ValueError, naming the cell and attribute, when:

  • an attribute the NeuroML schema requires is missing, including the cell id;
  • a dimensional attribute is a bare number, or its unit has the wrong dimension (for example C="100mV"), or the value is not finite;
  • a dimensionless attribute (the 2003 Izhikevich a, b, c, d) has a unit;
  • the document holds an element it does not model -- a network, population, projection, pulseGenerator, ion channel or morphological cell -- because keeping only the cells would silently change what the document describes.

Documentation elements (notes, annotation, property) are ignored.

Mapping Notes

Each ImportedCell.notes lists what its mapping assumed or could not carry: the timestep the model runs at (NeuroML cells carry none), voltages taken relative to the leak reversal, the normalised input resistance of the LIF mapping, a refractory period rounded to whole timesteps, the 2003 Izhikevich v0 and thresh (the model has no place for them), and an AdEx refract other than zero (AdExNeuron has no refractory period).

Parameter Conversion

LIF (iafCell -> StochasticLIFNeuron)

NeuroML specifies LIF with conductance and capacitance: - C (pF), leakConductance (nS), leakReversal (mV)

SC-NeuroCore uses time constant: - tau_mem = C / leakConductance (ms) - Voltages normalised relative to leak reversal

Izhikevich 2007

NeuroML <izhikevich2007Cell> uses physical units and is imported as Izhikevich2007Neuron without conversion to the dimensionless 2003 model:

  • C in pF, k in nS/mV, voltages in mV
  • a in 1/ms, b in nS, d and runtime input current in pA
  • RK4 integration is selected by default for imported cells

AdEx

Direct parameter mapping. Units match (pF, nS, mV, ms).

Limitations

  • HH neurons (<cell> with <biophysicalProperties>) not yet supported. These require parsing ion channel definitions separately.
  • Network topology (populations, projections) is not imported; a document that contains it is refused.
  • Multi-compartment morphologies not supported.
  • NeuroML is import-only: SC-NeuroCore has no NeuroML exporter.

Validating Documents

The optional neuroml extra (pip install "sc-neurocore[neuroml]") installs libNeuroML, the reference library for writing NeuroML 2 documents and checking them against the schema (neuroml.utils.validate_neuroml2). The importer's tests import documents that libNeuroML wrote and validated.

Reference

NeuroML 2 specification: https://docs.neuroml.org/Userdocs/Schemas/Cells.html