The HNN Roadmap

Project Vision

HNN was created as a resource for the MEG/EEG community to develop and test hypotheses on the neural origin of their human data. The foundation of HNN is a detailed cortical column model containing generalizable features of cortical circuitry, including layer specific synaptic drive from exogenous thalamic and cortical sources, that simulates a primary current dipole from a single localized brain area. In addition to calculating the primary current source in units that are directly comparable to source localized data (Ampere-meters, Am), the details in HNN enable interpretation of multi-scale activity including layer specific and individual cell activity. HNN was designed based on workflows to simulate the most commonly measured signals, including ERPs and low frequency brain rhythms based on prior studies.

A main goal of HNN is to create a user-friendly interactive interface and tutorials to teach to the MEG/EEG community how to interact with the model to study the neural origin of these signals, without needing to access the underlying complex neural modeling code. To this end, HNN was constructed with a graphical user interface (GUI) and corresponding tutorials of use for commonly measured signals, which are distributed on the HNN website (https://hnn.brown.edu). Our philosophy is that the interactive GUI is essential for all new users of HNN to develop an intuition on how to interact with the large-scale computational model to study the multi-scale neural dynamics underlying their MEG/EEG data. Once this intuition is gained, users who chose to can dive into the computational neural modeling code, where further command line utily can be developed. As such, an equal goal is to enable the neural modeling and coding community to participate in HNN development. We are prioritizing best practices in open-source software design and the development of a documented API for interoperability and to facilitate integration with other relevant open-source platforms (e.g. MNE-Python, NetPyNE). Our vision is to create a unique transformational software specific to interpreting the neural origin of MEG/EEG.

Timeline Overview

This roadmap timeline outlines the major short-term and longer-term goals for HNNs. The short term goals will entail a substantial reorganization of the HNN code and creation of an API to facilitate HNN expansions, community contribution, and integration with other relevant open-source platforms (e.g. MNE-Python, NetPyNE). To this end, in March 2021, we released the first version of the HNN-core repository. HNN-core contains improved versions of HNN’s non-GUI components following best practices in open-source software design, with unit testing and continuous integration, along with initial API and documentation for command-line coding. We will adopt similar best practices to develop a new HNN-GUI and several new HNN features, including the ability to simulate and visualize LFP/CSD and to use improved parameter estimation procedures. Our process will be to develop all new features in HNN-core, with API and examples of use followed, when applicable, by integration into the HNN-GUI with correspoding GUI-based tutorials on our website. Longer-term goals include integration with the related modeling software MNE-Python and NetPyNe, the development of a web-based interface with ability for simultaneous GUI and Command Line Interface (CLI), and extension to multi-area simulations.

Short-Term Goals

Modularize HNN code to simplify installation, development and maintenance

We are working on cleaning up and re-organizing the underlying code that defines the current distribution of HNN to facilitate expansion and community engagement in its use and development. To minimize the dependencies that are required to install before contributing to HNN development and maintenance, all of the non-GUI components of HNN’s code are being organized into a new repository HNN-core (initial release March 2021). This reorganization will entail continued improvements within the HNN-core repository, along with API development and examples of use, in the following steps:

  • Following best practices in open-source software design, including continuous integration testing, to develop HNN-core. HNN-core will contain clean and reorganized code, and separate all components that interact directly with the NEURON simulator (e.g. cell and network intantiation, external drives, etc..), from those that pertain to post-processing data analysis and plotting functions (e.g. spectra lanalysis). COMPLETED FEB 2021

  • Convert installation procedures to PIP. COMPLETED FEB 2021

  • Parallelization of the simulations in HNN-core via MPI or Joblib. COMPLETED SEP 2020

  • Reorganization of the Network class within HNN-core module to separate cortical column model from exogenous drive, and optimization routines. See gh-104, gh-124, and gh-129 for related discussions. COMPLETED FEB 2021

  • Develop initial HNN-core documentation and example simulations following those detailed in the HNN-GUI tutorials https://jonescompneurolab.github.io/hnn-core/stable/index.html. COMPLETED MARCH 2021

  • First release of HNN-Core 0.1 to the community COMPLETED MARCH 2021

  • Make HNN-Core compatible for windows including installation, testing and continuous integration.

  • Reorganization of Param.py file within HNN-core to multiple files that contain smaller dictionaries of parameters related to different modules of the code. See gh-104 for related discussions.

  • Expand details in HNN-core examples to follow HNN-GUI based tutorials.

Develop a New HNN GUI

A new HNN-GUI will be developed following similar best-practices in open source software design, as employed in HNN-core. The first step will be to ensure all of the functionality of the current GUI distribution is developed in HNN-Core, followed by integration into a new HNN-GUI, with corresponding GUI-based tutorials on the HNN website. Once complete, the current HNN-GUI repository will be deprecated.

  • Development of optimization routines in HNN-core that have the current functionality in HNN-GUI.

  • Develop a new HNN-GUI using ipywidgets in HNN-core that has all of the functionality of the current HNN-GUI.

  • Rename HNN to HNN-GUI and release updated version to the community and deprecate original HNN repository.

LFP/CSD Simulation, Visualization and Data Comparison

Essential to testing circuit-level predictions developed in HNN is the ability to test the predictions with invasive recordings in animals or humans. The most fundamental domain over which the predictions will be tested is local field potential (LFP) recordings across the cortical layers and the associated current source density (CSD) profiles. We will develop a method to simulate and visualize LFP/CSD across the cortical layers and to statistically compare model simulations to recorded data. These components will be developed in HNN-core, with correponding API and examples of use, followed by integration into the HNN-GUI, with corresponding GUI based tutorials on the HNN website, in the following steps:

  • Develop code in HNN-core to simulate and visualize LFP/CSD from cellular membrane potentials.

  • Develop code in HNN-core to statistically compare and visualize model LFP/CSD to invasive animal data.

  • Develop functions in HNN-GUI to enable simulation, visualization and data comparison in the GUI.

Parameter Estimation Expansion

Parameter estimation is an inherent difficulty in neural model simulation. HNN currently enables some parameter estimation, focussing on parameters relevant to an ERP. New methods have been recently developed that apply a machine learning approach to parameter estimation, namely Sequential Neural Parameter Estimation (SNPE) (Gonçalves et al Elife 2020: DOI: 10.7554/eLife.56261). We will adapt this method for parameter estimation to work with HNN-core, enabling estimation of a distribution of parameters that could account for empirical data, and then integrate it into the HNN-GUI, with GUI-based tutorials, in the following steps:

  • Extending HNN-core to run batch simulations that enable parameter sweeps.

  • Development of functions in HNN-GUI to enable parameter sweeps via the GUI.

  • Develop code for SNPE parameter estimation and visualization in HNN-core.

  • Develop functions in HNN-GUI to enable SNPE estimation in the GUI.

Different Cortical Model Template Choices

HNN is distributed with a cortical column model template that represents generalizable features of cortical circuitry based on prior studies. Updates to this model are being made by the HNN team, including a model with alternate pyramidal neuron calcium dynamics, and an updated inhibitory connectivity architecture. We will expand HNN-core to enable a choice of template models, beginning with those developed by the HNN team and ultimately expanding to model development in other platforms (e.g. NetPyNE), see Longer-Term goals. These models will first be developed in HNN-core, with corresponding API and examples of use, followed by integration into HNN-GUI, with GUI-based tutorials.

  • Develop new cortical column template models with pyramidal neuron calcium dynamics, in HNN-core.

  • Create flexibility to change local connectivity and to visualize connectivity in HNN-core.

  • Create flexibility to change exogenous connectivity and to visualize connectivity in HHN-core.

  • Develop functionality in HNN-GUI to chose amng different template models.

  • Develop function in HNN-GUI to choose among different template models in the GUI.

See gh-111 for more discussions.

API and Tutorial development

The ability to interpret the neural origin of macroscale MEG/EEG signals in a complex high-dimensional non-linear computational neural model is challenging. A primary goal of HNN is to facilitate this interpretation with a clear API and examples of use in HNN-core, and interative GUI-based tutorals for all HNN-GUI functionality on our HNN website. Following the process for creating new featuers in HNN, the process for documenting new features will be to first develop them with API and examples of use in HNN-core, followed by integration into the HNN-GUI, with corresponding GUI-based tutorials on the HNN-website. Developmental goals are only complete once the corresponding documentation is available.

Longer-Term Goals

Develop a framework to import cortical column models developed in NetPyNE or other modeling platforms into HNN: The core of HNN is a cortical column model that simulates macroscale current dipoles. Currently, HNN is distributed with a template cortical column model based on generalizable features of cortical circuitry and as applied in prior studies. Essential to future expansion of HNN is the ability to use other cortical column models that include different cell types and or different network features. We have begun creation of a framework where models built in NetPyNE can be adapted to the HNN workflows of use. As a test bed, this currently entails integration of the HNN cortical column model and exogenous drives into the full NetPyNE platform (https://github.com/jonescompneurolab/hnn/tree/netpyne). See also update from MARCH 2021 https://github.com/jonescompneurolab/hnn/tree/hnn2 .

To limit the scope of this effort to HNN-specific goals, i.e. neural modeling designed for interpretation of human EEG/MEG signals, we will work with NetPyNE team to develop clean modularized framework for integrating NetPyNe developed cortical models that have laminar structure and multicompartment pyramidal neurons into HNN design and workflows of use to simulate ERPs and low frequency brain rhythms work.

Integrate HNN and MNE-Python tools: We will work to create a framework where source localization using MNE-Python is seamlessly integrated with HNN for circuit-level interpretation of the signal. We will develop workflows that enable users starting with sensor level signals to perform both source localization using MNE-Python and circuit interpretation using HNN-core. We begin with use open-source median nerve datasets and develop examples using three different inverse methods (Dipole, MNE, Beamformer).

Convert HNN to web-based platform with dual GUI and Command Line Interface (CLI): We have begun working with MetaCell (metacell.org) to convert HNN to a web-based interactive GUI with updated graphics (https://github.com/MetaCell/HNN-UI). This conversion will eliminate the installation process and enhance computational efficiency. Additionally, MetaCell is facilitating the transformation to a dual GUI and CLI interface enabled through Jupyter notebooks. There are advantages to both GUI and CLI in adapting HNN to user goals. GUIs provide a framework for teaching the community the workflow to use such models to study the biophysical origin of MEG/EEG signals, like ERPs and brain rhythms. Once a meaningful parameter set is identified to account for the data of one subject, CLI scripts can be useful to investigate how well this parameter set accounts for the data from multiple subjects or how parameter changes impact the signal. CLIs can be used to generate sequences of processing steps that can then be applied to multiple data sets, ensuring rigor and reproducibility. Further, simultaneous viewing of GUI and CLI can help advanced users quickly adapt the code with scripting, and ultimately help create a community of HNN software developers. This framework will also facilitate the integration with other open-source platforms, including MNE-Python and NetPyNE.

Expand HNN to include study of multi-area interactions: HNN is designed for detailed multi-scale interpretation of the neural origin of macroscale current dipoles signals from a single brain area. A long term vision is to create a framework where multi-area interactions can be studied. We will begin with simulations of the interactions between sensory and motor cortices during median nerve stimulation.

Footnotes

1

We do not claim all the neural mechanisms of these signals are completely understood, rather that there is a baseline of knowledge to build from and that HNN provides a framework for further investigation.