Theoretical neuroscientists use neuron models to predict, understand, and explain biological neuron behavior. They often work with reduced neuron models that abstract away biological details but capture essential neuronal dynamics. This choice facilitates mathematical tractability, conceptual analysis, and computational speed. However, the tradeoffs inherent in using such models (instead of biologically detailed ones) are not transparent. It is often unclear if a model is faithful to essential observed dynamics of the neuron, and if so, under what model parameters and stimulus conditions. It is also rare for multiple types of reduced models to be compared in this regard, making it difficult to select the most appropriate one for a scientific question. Lastly, such models, once developed and parameterized, usually are not shared among researchers in a form that facilitates reproducibility and re-use, nor can they be easily discovered. As a result, status quo behavior in the use of reduced models is often simply to choose a favorite model regardless of merit, to optimize it for the scientific question at hand, and then to discard it. A standard practice in professional software development is unit testing. A unit test is a procedure that validates a single component of a computer program against a single correctness criterion. An ongoing effort to develop an analogous unit-testing procedure for neuron models, NeuronUnit, enables the construction of validation tests-executable functions that validate models against a single empirical observation to produce a score indicating agreement between the model and an observation. NeuronUnit facilitates the construction and logical grouping of tests for neuron models, the parameterization of tests using a wide range of empirical data, and the execution of tests against models in a continuous and transparent fashion. Aggregate results provide both theoretical and experimental neuroscientists with an overview of model suitability for targeted research questions. Merits and deficiencies of competing models are clearly visible, benefiting ongoing modeling efforts and informing new theoretical and experimental directions. This proposal aims to expand NeuronUnit to create data-driven, neuron-type-specific validation tests for reduced models. The ability of a range of reduced models to capture the relevant membrane potential and spiking dynamics of specific biological neuron types in response to specific stimuli, using publicly available experimental data from numerous sources, will be quantitatively tested and visualized. In doing so, the merits and deficiencies of each reduced model-as well as tradeoffs in model complexity, speed, and analysis-become transparent, providing critical information for model choice. Project aims are to 1) express a large number of reduced models using NeuroML/LEMS, 2) implement NeuronUnit testing of these models against data from a wide range of neuron- and experiment-types, 3) provide web-based search and visualization for test results and corresponding simulations, and 4) make these models available both as NeuroML documents and as code for every NeuroML-supported simulator. Collaborations with multiple existing initiatives will promote uptake of these tools, which for the first time, will provide a rigorous, transparent process for evaluation and selection of reduced models to address scientific questions about neurons. This project goes beyond model sharing by facilitating the dissemination of information about the performance and applicability of reduced neuron models in the context of specific datasets, complementing the existing dissemination mode of manuscript publication. By making model choice more deliberate and model appropriateness more objective, this work highlights which models should be used to address which scientific questions and why, without the need for a deep literature search (for models and data) or the installation of new tools or re-coding of models for simulation. The project also serves neuroscience educators by providing an interactive platform for visualization of reduced model dynamics accessible to any student, using data from biological neurons. This work broadly transforms theoretical neuroscience: by giving modelers a tool to select models quickly and with clear purpose; by rigorously identifying the models best-suited for further research efforts; and by helping experimentalists enhance the impact of their work.
