hybrid-biophysical-neuron-neural-ode

name	hybrid-biophysical-neuron-neural-ode
description	Hybrid biophysical neuron modeling methodology combining conductance-based models with neural ODEs. Captures unknown ion channel kinetics while preserving mechanistic interpretability. Enables single-compartment reduction of multi-compartment models.

Learning Hybrid Biophysical Neuron Models with Neural ODEs

Hybrid modeling framework that embeds neural ODEs into conductance-based biophysical models to capture unknown currents or mis-specified channel kinetics while preserving mechanistic interpretability.

核心问题

传统困境:

• Ion channel kinetics poorly characterized
• Practical simplifications introduce systematic gaps
• Model vs. biology mismatch

解决方案: Hybrid approach that discovers unmodeled dynamics while preserving mechanistic structure

技术架构

Neural ODE 参数化

• Voltage-dependent steady-state functions
• Time-constant functions
• Recover interpretable gating dynamics
• No functional form assumption

混合模型设计

Conductance-based model + Neural ODE component → Hybrid model

关键特性:

• Plug-and-play replacement of unknown components
• Mechanistic interpretability preserved
• Data-driven discovery of unmodeled dynamics

实验验证

Ion Channel 模型拟合

• 数据集: 2400 ion channel models
• 结果: Fits gating kinetics accurately
• 泛化: Out-of-distribution stimulus regimes

多室模型降维

• 原始: Multi-compartment cortical neuron model
• 降维: Single-compartment hybrid model
• 增益: Learned axial current
• 效率: Up to 10x computational cost reduction

关键优势

可解释性

• Voltage-dependent functions recovered
• Gating dynamics interpretable
• Mechanistic structure retained

数据驱动发现

• Unknown gating dynamics from single current-clamp recordings
• Generalizes to realistic inputs
• Handles parameter misspecification

模型降维

• Multi-compartment → Single-compartment
• Learned axial current surrogate
• Computational efficiency

实现要点

神经 ODE 设计

• Parameterize by steady-state & time-constant
• Voltage-dependent functions
• Interpretable gating dynamics

混合策略

• Selectively replace unknown components
• Preserve known mechanistic structure
• Plug-and-play framework

验证机制

• Current-clamp recordings
• Out-of-distribution test
• Parameter misspecification scenarios

应用场景

离子通道建模

• Discover unknown channel kinetics
• Correct mis-specified models
• Fit large-scale channel datasets

神经元模型降维

• Multi-compartment simplification
• Axial current learning
• Computational cost optimization

计算神经科学

• Bridge model-biology gaps
• Mechanistic + data-driven hybrid
• Interpretable neural dynamics

技术指标

• 拟合规模: 2400 ion channel models
• 泛化能力: Out-of-distribution stimuli
• 降维效率: Up to 10x computational reduction
• 数据需求: Single current-clamp recording
• 可解释性: Voltage-dependent gating functions

理论贡献

混合建模范式

• 传统: Fully mechanistic OR fully data-driven
• 新范式: Hybrid mechanistic + data-driven
• 价值: Best of both worlds

参数化策略

• Neural ODE in interpretable form
• Steady-state + time-constant functions
• No black-box dynamics

降维方法

• Learned axial current as surrogate
• Multi-compartment → Single-compartment
• Mechanism-preserving reduction

论文信息

标题: Learning Hybrid Biophysical Neuron Models with Neural ODEs

作者: Jonas Beck, Michael Deistler, Dóra Viktória Molnár, Jakob H. Macke, Philipp Berens

arXiv: 2606.16693 (Submitted 2026-06-15)

领域: q-bio.NC (Neurons and Cognition), cs.LG (Machine Learning)

引用

@article{beck2026hybrid,
  title={Learning Hybrid Biophysical Neuron Models with Neural ODEs},
  author={Beck, Jonas and Deistler, Michael and Molnár, Dóra Viktória and Macke, Jakob H. and Berens, Philipp},
  journal={arXiv preprint arXiv:2606.16693},
  year={2026}
}