Efficient support for Sonata in NEST

Project summary

One of the central goals of the Human Brain Project is to advance brain simulations. Achieving this goal depends crucially on the efficient and convenient simulation packages. Such packages exist—for example, NEST, NEURON, Arbor, and others—and have demonstrated computational efficiency and applicability to a broad range of modeling tasks. However, the utility of even the most advanced simulation software is limited without a broadly applicable, flexible, and high performance modeling data format. The current trends towards collaborative projects demand standardized formats for model sharing and reproducibility, as well as for interoperability between tools. Meanwhile, high computational performance of such formats becomes increasingly important to enable efficient representation of the growing biological complexity of models.

We recently developed the SONATA (Scalable Open Network Architecture TemplAte) data format [Dai et al., PLOS Comp. Bio., 2020], as a result of a collaboration involving the Blue Brain Project and developers of NeuroML, PyNN, NetPyNE, and the NWB format for neurophysiology data. SONATA describes all portions of the modeling workflow: it represents neuronal circuits, simulation configurations, and simulation outputs. It utilizes computationally efficient binary formats for storing large datasets while also offering text-based formats for less data-rich model components. And, it provides much flexibility in describing models at different levels of resolution, including hybrid models.

Due to these features, SONATA is quickly growing in popularity among researchers in Europe, the US, and around the world. It is especially useful for large-scale, biologically realistic, data-driven simulations, where the size and complexity of the models require efficient solutions for reducing the file storage footprint and streamlining disk I/O operations during massively parallel simulations. As a universal modeling format, SONATA contributes crucially to improving model sharing and reproducibility in the whole field.

Here we propose to implement native support for SONATA in NEST, which is a highly efficient and widely used software package for simulations of spiking neural networks and a core network simulation tool in EBRAINS. Currently, it is possible to instantiate NEST simulations from SONATA models using the Allen Institute software suite called the Brain Modeling ToolKit. However, native support of SONATA in NEST will provide convenient ways to exchange and reuse models for a much broader user base. That includes not only many thousands of NEST users, but also users of other software tools that may exchange models in SONATA format between those tools and NEST.

Furthermore, there are known issues with the existing support of instantiating NEST simulations from SONATA files in BMTK. The major problem is a long time it takes for constructing networks in NEST after reading their specification from the SONATA files. As an example, instantiation of a 230,000-neuron Allen Institute model of the mouse cortical area V1 [Billeh et al., Neuron, 2020] may take tens of minutes, whereas the subsequent simulation runs at approximately 3 minutes per simulated second (on a single CPU). This results in a substantial bottleneck in terms of setup time. While not prohibitive for extensive simulations, this reduces the efficiency of such simulations. This problem is due to the network instantiation algorithms in NEST being focused on generating network connectivity from relatively simple rules, whereas in SONATA the connectivity is defined as a list of all the connections, which have been previously generated at the model building stage.

The proposed work will focus on resolving this problem, which will be beneficial to both NEST and BMTK. Most importantly, it will also focus on developing native support for SONATA in NEST, so that users do not need to use BMTK or other intermediate tools. As a result, the new capability of NEST (and, by extension, BMTK) to interface efficiently with SONATA will give users the ability to generate large network simulations from SONATA specifications in a much more efficient and general way than what is possible now.

Thus, the proposed activity will contribute strongly to EBRAINS Service Category 3 (Brain Simulation). Using SONATA, scientists can share models much more efficiently and precisely than what has been possible so far, thus supporting FAIRness (Findability, Accessibility, Interoperability, and Reuse of digital assets) in neuroscience modeling. Importantly, efficient support for SONATA network specifications in NEST will make it much easier for scientists to migrate models to NEST and EBRAINS.

With respect to the gender, age, and other characteristics of scientists participating in research activities, SONATA and NEST (as well as BMTK) are all open-source packages freely accessible to everyone. Therefore, by their very nature they promote free, undiscriminated, and unrestricted access of all researchers to the important tools enabling scientific progress and excellence.

Partnering organisations

Key facts

Time Frame: July 2021 – March 2023

Origin: EBRAINS Research Infrastructure Voucher Programme 2020

Funding: HBP SGA3