Ivan Raikov

The Layer-Oriented Approach to Declarative Languages for Biological Modeling

We present a new approach to modeling languages for computational biology, which we call the layer-oriented approach. The approach stems from the observation that many diverse biological phenomena are described using a small set of mathematical formalisms (e.g. differential equations), while at the same time different domains and subdomains of computational biology require that models are structured according to the accepted terminology and classification of that domain. Our approach uses distinct semantic layers to represent the domain-specific biological concepts and the underlying mathematical formalisms. Additional functionality can be transparently added to the language by adding more layers. This approach is specifically concerned with declarative languages, and throughout the paper we note some of the limitations inherent to declarative approaches. The layer-oriented approach is a way to specify explicitly how high-level biological modeling concepts are mapped to a computational representation, while abstracting away details of particular programming languages and simulation environments. To illustrate this process, we define an example language for describing models of ionic currents, and use a general mathematical notation for semantic transformations to show how to generate model simulation code for various simulation environments. We use the example language to describe a Purkinje neuron model and demonstrate how the layer-oriented approach can be used for solving several practical issues of computational neuroscience model development. We discuss the advantages and limitations of the approach in comparison with other modeling language efforts in the domain of computational biology and outline some principles for extensible, flexible modeling language design. We conclude by describing in detail the semantic transformations defined for our language.

NineML – a description language for spiking neuron network modeling: the abstraction layer

With an increasing number of studies related to large-scale neuronal network modeling, the International Neuroinformatics Coordinating Facility (INCF) has identified a need for standards and guidelines to ease model sharing and facilitate the replication of results across different simulators. To create such standards, the INCF has formed a program on Multiscale Modeling to develop a common standardized description language for neuronal network models. The name of the proposed standard is Network Interchange for Neuroscience Modeling Language (NineML) and its first version is aimed at descriptions of large networks of spiking neurons. The design of NineML is divided in two semantic layers: an abstraction layer that provides the core concepts, mathematics and syntax with which model variables and state update rules are explicitly described and a user layer that provides a syntax to specify the instantiation and parameterization of a network model in biological terms. The key concepts of spiking neuron network modeling are 1) spiking neurons 2) synapses 3) populations of neurons and 4) connectivity patterns across populations of neurons. Accordingly, the INCF task force on multiscale modeling has identified a set of mathematical abstractions that are capable of representing these concepts in a computer language. First, we propose a flexible block diagram notation for describing spiking dynamics. The notation represents continuous and discrete variables, their evolution according to a set of rules such as a system of ordinary differential equations, and the conditions that induce a regime change, such as the transition from subthreshold mode to spiking and refractory modes. The notation we have developed is an explicit formalization of event handling and is an important step in ensuring model simulation consistency. In addition, the abstraction layer provides the notation to describe a variety of topographical arrangements of neurons and populations, and to describe random connectivity patterns between neuronal populations, based on structural properties of neuronal networks.

Spatiotemporal network coding of physiological mossy fiber inputs by the cerebellar granular layer

The granular layer, which mainly consists of granule and Golgi cells, is the first stage of the cerebellar cortex and processes spatiotemporal information transmitted by mossy fiber inputs with a wide variety of firing patterns. To study its dynamics at multiple time scales in response to inputs approximating real spatiotemporal patterns, we constructed a large-scale 3D network model of the granular layer. Patterned mossy fiber activity induces rhythmic Golgi cell activity that is synchronized by shared parallel fiber input and by gap junctions. This leads to long distance synchrony of Golgi cells along the transverse axis, powerfully regulating granule cell firing by imposing inhibition during a specific time window. The essential network mechanisms, including tunable Golgi cell oscillations, on-beam inhibition and NMDA receptors causing first winner keeps winning of granule cells, illustrate how fundamental properties of the granule layer operate in tandem to produce (1) well timed and spatially bound output, (2) a wide dynamic range of granule cell firing and (3) transient and coherent gating oscillations. These results substantially enrich our understanding of granule cell layer processing, which seems to promote spatial group selection of granule cell activity as a function of timing of mossy fiber input.

The Promise and Shortcomings of XML as an Interchange Format for Computational Models of Biology

aims for a wider scope of model description and isnot specific to any one field of biology.XML, a data exchange language closely related toHTML, is a widely accepted standard for describingstructured textual data. Its often advertised advantage isthat XML documents with different structure can be readby the same generic reusable parser. However, XML byitself does not enable information interchange. Humanreaders of XML may be able to guess the meaning of astatement such as g ¼ 0:1 but to a computer program, and and areallequallymeaningless.Whataprogramshoulddo with XML data is undefined by the XML standard.Frequently, in practice, the semantics (or meaning) ofXML-based model description languages are described inspecification documents written in natural human language.Softwareprogrammersmustreadthespecificationdocumentsand convert the requirements into programs (in Python, Java,etc.) for reading and writing model descriptions. But as newmodeling approaches emerge, and new simulation code iswritten, the semantics of model description languages mustbe reimplemented. A specification written in a human lan-guage often hides ambiguities, and as the complexity of thelanguages and the number of supported software platformsincreases, it is difficult or impossible to ensure that everylanguage construct is implemented consistently in every soft-ware package.The XML community has developed several schema lan-guages that can specify rules for structuring and validatingXML documents, but they offer weak support for data types,procedures, or complex dependencies between elements.Therefore, the expressive power of XML is greatly affectedby its interpretation.The use of XML for the syntactic structure of a modelinglanguage does indeed eliminate the need for specialized

Boundary representation of neural architecture and connectivity

The general objective of this work is to develop a description language for constructive 3D boundary representation [5] of neuroanatomical structures and connectivity at various levels of granularity (from coarse-resolution solids to fine meshes). This approach is motivated by a desire to capture regularities in neural circuitry as revealed by neuro-architectonic studies [4], while at the same time to explore hypotheses about anatomic variability of various origins, such as experimental uncertainties, cross species scaling factors, individual differences, and assess them for their impact on connectivity, and ultimately on network dynamics. Boundary representation is a general approach to describe 3D objects solely by their surface. Boundary representations consist of topological objects and their concrete geometrical representations in terms of enclosing boundaries. Topological elements include vertices, edges and faces, with corresponding geometrical elements being points, curves, and surfaces. The relationships between topological elements in a structure are expressed by means of a graph of topological connections. The description language is being implemented on top of the GNU Triangulated Surface library [3], and provides the ability to: • Specify compound topological objects with parametric geometry; • Specify geometric parameters for the instantiation of topological objects, such as coordinates for placement, or probability distributions for random placement of a group of identical objects; • Specify individual coordinate systems for different cell populations; • Define categories of topological objects, such as stellate, basket and Golgi cells, which may be part of a morphological continuum; • Define rules for connectivity between different categories of objects. We present the anisotropic cerebellar circuitry as a case study, and define boundary representations of the arrangement of Purkinje and Golgi cells in the cerebellar cortex. We use the hexagonal grid pattern suggested by Palkovits et al. in [1,2], while allowing for small variability in the placement of cells within a hexagon. The implementation of the language instantiates the topological objects, computes the intersections of the resulting surfaces, and given connectivity rules for the different categories of objects, computes the potential synaptic connectivity (producing graph theoretical measures, as well as connectivity histograms), and ultimately aims to generate connectivity descriptions for the NEURON simulator.

A NineML-based domain-specific language for computational exploration of connectivity in the cerebellar granular layer

The patterns of connectivity within a neuronal network can strongly influence its function. In neuroanatomical models of the cerebellum, the dimensions and topology of the neuronal arbors are crucial components as the cerebellum contains some of the most spatially extended cells in the brain (Golgi and Purkinje), which are connected by very long axonal projections of the granule cells (the parallel fibers). However, it is not known how these spatial features influence the overall cerebellar dynamics. We begin to address this question by means of a prototype domain-specific language for constructing large-scale models of the cerebellar granular layer. The model currently implemented encompasses a patch of 1500 × 700 × 200 microns, and consists of 800 000 granule cells and 2000 Golgi cells. Input is provided by spatially embedded mossy fibers that function as non-homogeneous Poisson processes. The language framework is based on the emerging NineML network description language [1] and can interface to the NEURON simulator [2]. We show how the language can be used to investigate diverse neuroanatomical architectures. Variations of the basic granular layer architecture include random perturbations of the parallel fibers, exploring various shapes of synthetic Golgi cell morphologies and using experimentally obtained cell reconstructions. We characterize the effect of architecture changes on the network dynamics by comparing the spatial and temporal correlations of granule cell spiking activity. The NineML language is intended to allow different methods for describing connectivity implemented as separate modules. Current working proposals include incorporating the Connection-Set Algebra [3], a general purpose graph library [1], and an equation-based format similar to the one implemented in the Brian2 simulator [4]. The present work demonstrates that an approach to describing connectivity based on geometric shape is compatible with the NineML object model.

Exploring the functional implications of brain architecture and connectivity: a multi-simulator framework for biophysical neuronal models

We introduce a framework for implementing networks of neuronal models with conductance-based mechanisms and morphology (where applicable) across multiple simulators. The framework extends the existing NINEML language [1] by adding two independent modules, NINEML-Conductance and NINEML-BREP [2], which allow the specification of conductance-based mechanisms and geometrically derived connectivity respectively. The PyNN API [3] is utilised to reproduce connectivity across multiple simulators, with adapters added where necessary to accommodate the proposed extensions to NINEML. PyNN was chosen to handle the multi-simulator connectivity because it offers translations to a wide range of neural simulators and provides a standardised Python interface for simulation control. It is also straightforward to load predefined connectivity into the PyNN-Connector API from a sparse-matrix-like format, allowing a general interface to NINEML-BREP. Neuronal mechanisms are precompiled into simulator-dependent formats from the NINEML-Conductance declaration, and are then integrated into PyNN via a novel “conductance standard model” class. Depending on whether the selected simulator supports multi-compartment neuronal models, cell morphology is optionally loaded from the NINEML-BREP description and incorporated into the conductance standard model, with flags set in the declarative model description to handle the required adjustments to mechanism parameters. By the meeting we aim to have completed the extensions to the NINEML language and the required interface between the extended NINEML language and PyNN for the NEURON [4] and NEST [5] simulators, and have a working network model of the cerebellar cortex within this framework. This will enable us to test the effect of varying the biophysical detail of neuronal models and different simulators on the proposed cerebellar cortex model. Figure 1

Exploring the limitations of simulator independence via an implementation of a biophysically detailed cerebellar cortex model in NEURON and NEST

The ability to develop models of complex neural networks in a simulator independent manner has been a longstanding goal of the computational neuroscience community [1,2]. One of several important reasons behind this is because the effect of subtle differences in simulator implementations, the timing of spike propagation for example [3], on the qualitative behavior of networks with complex neuronal models is unclear a priori. In addition, the relevance of constraints placed on model design by fundamental assumptions in the simulator architecture and features that are not available in all simulators, such as gap junctions and active dendritic compartments, has not been extensively studied. To begin to address such issues, we investigate differences between NEURON [4] and NEST [5] simulations of a biophysically detailed model of the cerebellar cortex. Following the approach outlined in [6], the cerebellar cortex model is defined using a custom declarative architecture, which is based on NineML [1] and NeuroML 2.0 [7] where possible, and otherwise extended to meet the requirements of the model. Neuronal dynamics are described using a custom extension to the NineML language for conductance-based dynamics, which is compiled directly into simulator-native model formats [8]. Connections between neuronal populations within the model are generated from a combination of morphologically based [9] and soma-to-soma geometric connectivity rules. These rules are integrated into PyNN framework [10], which handles the appropriate simulator-dependent connection routines. There are a number of factors that make the cerebellar cortex a good test case to study the effect of simulator disparities. The cerebellar cortex is strongly hypothesized to be involved in the fine-tuning of movement [11], and is therefore likely to be sensitive to spike timing. Also, in previous modeling, the behavior of the granular layer sub-network has been shown to be strongly affected by the Golgi-to-Golgi gap junctions [12], making the cerebellar cortex an interesting system in which to study the implications of different implementations of gap junction connections. Therefore, the differences between the NEURON and NEST implementations of the cerebellar cortex model should lend considerable insight into the practical issues that could limit the development of truly simulator independent models.

Challenges of declarative modeling of conductance-based neurons in diverse simulation environments

In the study, we attempt to quantify some of the important differences that exist between simulators, and we present a code generation approach that can solve the challenges caused by these differences. Furthermore, we highlight practical issues encountered while developing a convenient Python wrapper class for model code generated from the prototype language. This work is a step towards establishing a significant body of declarative models of neurons and identifies some of the issues related to interoperability of diverse neuroscience software.

Exploring the functional implications of brain architecture and connectivity: a declarative language framework

We present a prototype framework for exploring hypotheses about the neuroanatomical structures and connectivity in the cerebellar cortex at various levels of granularity, based on experimental data and hypotheses from the scientific literature [1]. As illustrated in Figure ​Figure1,1, the framework consists of declarative and algorithmic components. The declarative components include languages for describing connectivity and neuronal and synaptic mechanisms, built as an extension to the NineML description language [2]. The algorithmic components are Python scripts and the PyNN program, which are used for interfacing to specific simulator platforms and for simulation control. Figure 1 The core assumptions of the framework are: 1) connectivity rules are specified as probability distributions for overlapping volumes of objects of different categories; 2) synapse locations are randomly generated from the distribution associated with an overlapping volume 3) the volumes that represent dendritic trees have regions of uniform synaptic density. We use the NineML language for declarative descriptions of integrate-and-fire neuronal dynamics, and we have built two extensions to NineML to describe conductance-based neuronal spiking mechanisms and geometric connectivity. The NineML Conductance language is an extension of NineML for describing Ohmic and GHK currents based on the Hodgkin-Huxley formalisms or Markov chains. The NineML BREP language is an extension of NineML for constructive 3D boundary representation [3] of neuroanatomical structures and connectivity at various levels of granularity (from coarse-resolution solids to fine meshes). NineML BREP is implemented on top of the GNU Triangulated Surface library [4], and provides the ability to specify geometric parameters for the instantiation of topological objects, such as coordinates for placement, or probability distributions for random placement of a group of identical objects; define categories of topological objects, such as stellate, basket and Golgi cells; define rules for connectivity between different categories of objects.