Generating diverse sets of neuron models that capture intrinsic physiology

The purpose of this workflow is to generate neuron models that capture the intrinsic physiology of their cell type. The empirical data this workflow requires as an input are (1) a morphology reconstruction, (2) electrophysiological recordings that capture the cell type’s intrinsic physiology, and (3) bounds on the biophysical parameters that describe the neuron’s passive and active properties (Fig 2A). All parameters listed in ISF’s neuron parameter file [26], including spatial distributions of ion channels and their kinetics, can be used for this workflow (see Methods for more details).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Generating diverse sets of biologically detailed neuron models that capture intrinsic cellular physiology.

(A) A morphology and empirical measurements on the neuron’s response to current injection stimuli and biophysical parameters are required input data. (B) ISF generates neuron models from the input data. (C) ISF takes a random neuron model N1 whose parameters lie within empirical bounds. (D) ISF optimizes the biophysical parameters of N1 until they can reproduce the empirically observed response to the same input stimuli, creating a new neuron model N2. (E) ISF uses neuron model N2 as seed point to explore the space of biophysical parameters and find diverse neuron models (e.g., N3 and N5). (F) ISF finds models that are consistent with the empirical physiology, while being biophysically diverse. (G) The output of this workflow is a large database of empirically constrained, yet biologically diverse neuron models for each morphology.

https://doi.org/10.1371/journal.pcbi.1014617.g002

To generate neuron models from this input (Fig 2B), ISF converts the morphology into a multi-compartmental model, inserts ionic mechanisms such as ion channels and a calcium buffer, (Fig 2C), performs in silico current injection experiments, and optimizes the biophysical parameters until the in silico response lies within the empirically observed response distribution (Fig 2D). This produces a single neuron model that is constrained by empirical data in terms of morphology, biophysical parameters, and intrinsic physiology. This single neuron model is then used as a seed point for exploring the biophysically diverse parameter space (Fig 2E) by which the model produces responses consistent with the empirical observations for each current injection stimulus (Fig 2F). This procedure is then repeated for a set of neurons that capture the morphological diversity of the cell type.

As an output, ISF produces a database of morpho-electrically diverse neuron models whose biophysical parameters lie within the empirically observed ranges, and whose intrinsic electrophysiology is consistent with empirical observations (Fig 2G).

Generating diverse sets of network-embedded neuron models that capture in vivo conditions

The purpose of this workflow is to generate synapse locations and activation patterns onto neuron models that mimic the specific conditions of an in vivo experiment. For clarity, we refer to the neuron models as postsynaptic, while neurons providing input to neuron models are presynaptic. ISF generates synaptic input to neuron models in two stages: (1) it embeds a neuron model into a network model, and (2) assigns in vivo like activity patterns to each synapse.

First, the purpose of embedding a neuron model in a network model is to provide constraints on the number of synapses, the location of these synapses onto the postsynaptic neuron (i.e., including the soma, dendrites, and axon initial segment), and on their neurons of origin in the network. Our strategy is to generate a model of the dense neuropil structure, place the neuron model into the neuropil model, and provide wiring rules on how neurons in the dense neuropil model form synapses with the type of the model neuron. To capture variability in network connectivity, one could change the embedding location of the neuron model, the wiring rules, or generate multiple samples per location if the wiring rules have stochasticity. As a result, ISF provides a set of network-embedded neuron models that are consistent with empirical connectivity data, but which differ in the number and placement of synapses, as well as the neurons from which these synapses originate. This set of network-embedded neuron models provides a way to test whether and how variability in connectivity impacts the simulation results for in vivo-like input patterns.

We reported previously how to generate neuropil models and wiring rules that produce connectivity patterns that are consistent with those observed empirically in dense EM data [10,24,27], or which are derived directly from dense EM data [24,28]. Moreover, we provide examples on how to embed neuron models into a dense neuropil model of the barrel cortex [29,30]. This embedding procedure is set as the default option in ISF. Alternatively, users can explicitly define connectivity themselves by generating the respective connectivity files [31,32].

Second, to constrain the activation patterns of these synapses, ISF requires empirical data on synaptic strength in the form of unitary post-synaptic potential (uPSP) measurements, and activity rates for each cell type under the in vivo condition being investigated (Fig 3D). ISF then scales synaptic weights such that the resulting cell type-specific uPSP distributions match those observed empirically, and activates presynaptic neurons according to the empirically observed cell type-specific activity rates during the in vivo condition of interest (Fig 3E). ISF thereby generates spatiotemporal synapse activity patterns that are constrained by anatomical and functional data to mimic the synaptic input received in vivo. As a result, ISF produces a database of network-embedded neuron models that cover a broad diversity in terms of morphology and biophysics of the neuron, strength and spatial distribution of synapses, and presynaptic origin and timing of synapse activations (Fig 3F).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Generating diverse sets of biologically detailed network-embedded neuron models that capture in vivo activity.

Required empirical input data consists of (A) a dense connectome model, (B) neuron models, and (D) activity recordings of cell from all populations during an in vivo stimulus. (C) ISF creates multiple network embeddings per neuron model, and (E) generates multiple synapse activation patterns based on the empirically observed network activity per network-embedded neuron model. (F) The output is a database of empirically constrained, yet biologically diverse network-embedded neuron models.

https://doi.org/10.1371/journal.pcbi.1014617.g003

Predicting the origins of in vivo activity from simulations of diverse network-embedded neuron models

To simulate this diverse set of network-embedded neuron models, ISF numerically integrates the dynamics of synapse activations, ion channels and intracellular Ca²⁺-buffering, and the electrical properties of the membrane. These simulated results are then compared to measurements of in vivo observed activity, such as ongoing and sensory-evoked firing rates, receptive field shapes, occurrences of bursts of APs, dendritic spikes, or any other feature of in vivo activity. Importantly, model parameters are not tuned to achieve this match. Instead, all model parameters were already constrained by empirical data that is independent of the in vivo observation in previous steps (see “Input” of Figs 2 and 3). In other words, the empirical data used to constrain these network-embedded neuron models are categorically different from the in vivo observation to be explained, and the in vivo observation is at no point used as a constraint for parameter adjustments.

Can these models produce the in vivo observation under investigation? Each network-embedded neuron model that can do so provides one hypothesis of how the underlying mechanisms could generate the in vivo observed activity, and is selected for further analysis. If, on the other hand, not a single model can reproduce the in vivo observed activity, then the model generation requires further refinement. Likely, a biological mechanism that is important for the phenomenon is not present in the models, the empirical constraints are insufficient, or the space of possible network-embedded neuron models is insufficiently explored (see Discussion).

Those network-embedded neuron models that are consistent with in vivo observed activity allow us to further explore the possible origins of this activity, and to examine whether these models agree about their origins. When the models converge on common mechanisms underlying the in vivo observed activity, across a wide diversity of parameters, it indicates that these mechanisms are robust under the modeled conditions and well supported by the data used to constrain the model, even when this data exhibits intrinsic biological variability. In other words, the models have reached a consensus on how an in vivo function is implemented, despite their diversity. Such “model consensus” represents a strong prediction that these mechanisms may be used in vivo. Mechanisms that are predicted this way are ideally suited for “closing the loop,” i.e., performing an in vivo experiment that directly tests the proposed mechanism. Such an experiment can first be performed in silico until an experimental strategy has been found that is capable of revealing the underlying mechanism, and is achievable with the available experimental techniques.

For the purpose of investigating the underlying mechanisms and designing empirically testable experiments, ISF provides two additional workflows. The first strategy is to manipulate specific properties of network-embedded neuron models and investigate their effect on their simulated neuronal activity (Fig 4B). The second strategy is to reduce these models to a minimal set of parameters that are sufficient to account for an in vivo observation (Fig 4C). Both strategies can be combined as well: it is possible to apply manipulations to reduced models, or apply model reduction to manipulated simulations. The goal of both strategies is to provide empirically testable predictions on how neuron and network mechanisms underlie in vivo activity (Fig 4D).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Predicting the mechanistic origins of in vivo activity from simulated network-embedded neuron models via model reduction and in silico manipulations.

(A) A set of biologically diverse network-embedded neuron models and their simulated neuronal responses to sensory-evoked synaptic activity are a required input for further analysis. Both somatic AP times and the somatic voltage trace are shown. (B) Four examples of manipulations to simulated network-embedded neuron models, clockwise from the top left: scaling the strengths of synapses, changing the activation times of specific synapses, changing the biophysical properties of the neuron model, and changing the connectivity between the presynaptic population and the neuron model. Manipulations that target synapse strengths, activations, and connectivity can be manipulated based on their presynaptic cell type, location on the dendrite, at random, individually, or any other property. (C) network-embedded neuron models are reduced to analytically tractable models that capture their input-output computations. (D) Both the manipulations and reduced models aim to provide concrete in silico predictions that can be tested experimentally.

https://doi.org/10.1371/journal.pcbi.1014617.g004

Example application: Explaining sensory responses in barrel cortex

To illustrate all workflows introduced above, we show how we have applied ISF to explain the APs that the pyramidal tract neurons in layer 5 (L5PT) of the barrel cortex (reviewed in [33]) elicit in response to sensory input from different facial whiskers (Fig 5A). L5PTs integrate synaptic input from virtually all excitatory and inhibitory cell types within the barrel cortex, with large and morphologically diverse dendrites spanning all cortical layers (Fig 5A). They are the major output cell type of the cortex by sending axonal projections to various subcortical areas [34]. L5PTs elicit APs in response to whisker stimuli with onset latencies that rival those in the main input layer of the cortex: layer 4. Moreover, these fast responses occur upon stimulation of multiple different individual whiskers (Fig 5A). For decades, the origins of these fast and broadly tuned whisker responses in L5PTs were unclear, as in vivo experiments yielded different and conflicting conclusions [35,36]. Therefore, we decided to revisit which mechanisms account for whisker-evoked responses in L5PTs via our iterative in silico-in vivo approach [19,20].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Example application of ISF to explain whisker-evoked responses in the barrel cortex.

(A) In vivo measurements of L5PTs in barrel cortex during single whisker deflections of anesthetized rats reveal broad and heterogeneous receptive fields in neurons with morphologically diverse dendrites. Diverse L5PT morphologies are reconstructed from in vivo biocytin-filled neurons. Panel adapted from Bast, Fruengel et al. (2024) [20] (B) ISF generates multiple neuron models per L5PT morphology. These models differ in their biophysical parameters, but all reproduce the empirically observed neuronal activity. Shown on the left are the responses of two example neuron models to two current injection stimuli: a somatic current injection and a paired somatic and dendritic current injection. The right side shows the membrane voltage of two example neuron models during the paired somatic and dendritic current injection stimulus. (C) ISF embeds these neuron models into a dense model of the rat barrel cortex, and simulates the responses of these neuron models to synaptic input that mimics the in vivo condition of single whisker deflections in anaesthetized rats. (D) network-embedded neuron models are reduced to analytically tractable models, which weigh the contribution of individual synapses to the response probability based on location and activation time. This is the weighted net input (WNI) to the neuron, captured by spatial and temporal filters. Multiple filters are shown for different neuron models with the same morphology and network embedding. Panel adapted from Bast, Fruengel et al. (2024) [20] (E) Manipulations remove input from the presynaptic L6CCs. Response probability of L5PTs to 2nd surround whisker stimuli is abolished. (F) The in silico receptive fields obtained by simulating network-embedded neuron models reproduce the broad receptive fields as observed in vivo, while accounting for variability in neuron morphology, biophysical properties, and network input. (G) Change in receptive field (pearson R²) when changing either the biophysical properties, morphology, or the network embedding location. Changing network properties has a larger impact compared to biophysical or morphological properties. (H) Reduced models provide a mechanistic explanation by revealing how much weighted net input each presynaptic population provides to the postsynaptic neuron. Input from the thalamus (i.e., VPM) is the largest contribution for PW, but L6CC is the largest contribution for responses. Panel adapted from Bast, Fruengel et al. (2024) [20] (I) This mechanistic insight provides a testable prediction: silencing input from L6CCs in one barrel column removes the whisker that is somatotopically aligned with the manipulated column column from the L5PT’s receptive field (manipulation, red), while leaving the responses to other whiskers unaffected (control, black). Panel adapted from Egger et al. (2020) [19] (J) This testable prediction was verified in vivo by abolishing L6CC input from surround columns using muscimol injections [19]. Panel adapted from Egger et al. (2020) [19].

https://doi.org/10.1371/journal.pcbi.1014617.g005

Here, we review how we have used ISF to investigate the origins of the fast and broadly tuned whisker-evoked responses in L5PTs, including their variability across cells and trials. First, we generated neuron models of L5PTs that account for their morphology, biophysical properties and intrinsic physiology [25] (Fig 5B). Next, we embedded these neuron models into dense network models of the rat barrel cortex [24] and generated synaptic activity patterns that mimic the in vivo activity following passive single whisker deflection in anesthetized rats (Fig 5C). We used model reduction (Fig 5D) and targeted in silico manipulations (Fig 5E) to reveal that the fast onset and broad tuning of L5PTs in the rat barrel cortex are largely due to fast whisker-evoked responses in layer 6 cortico-cortical neurons (L6CCs) (Fig 5G and 5H) that are characterized by elaborate axons that project horizontally to L5PTs across the entire barrel cortex, a finding which we then tested experimentally in vivo [19] (Fig 5I and 5J).

Limitations of the study

Our strategy of model consensus comes at a cost. Without tuning parameters, models generated with ISF may fail to predict the desired in vivo observations. One reason for such “model failure” could be that parameter diversity at subcellular, cellular or network scales is not sufficiently captured — i.e., the configurations that could predict the in vivo observations are simply not yet present in the set of models. Users should therefore always ensure that the set of models is sufficiently diverse, to capture both the empirically observed ranges of each parameter at each scale, as well as the diversity of possible parameter combinations across scales.

However, despite best efforts to construct models that are biologically realistic and as diverse as possible, ISF may still fail to predict a desired in vivo observation. For example, early on, we had generated models of a single cortical (barrel) column, which predicted in vivo observed responses in layers 4 [49] and 2 [18], but not of the pyramidal tract neurons in layer 5 (L5PTs) [19]. For this purpose (see also our application example), we needed to systemically quantify the substantial horizontal intracortical connectivity [50] and then incorporate these empirical data to constrain models that comprised virtually the entire barrel cortex [24]. Thus, models that assume (entirely) columnar processing, or which rely (entirely) on connectivity data from brain slices — i.e., where horizontal axons are severely truncated [24]— could, in principle, not reveal the mechanistic origins of the in vivo observed fast and broadly tuned responses of L5PTs. Similarly, while our models predicted the probabilities, timings, and tunings of L5PT responses, they could not predict that responses in some cells and trials occur as bursts of APs [19,20]. For this purpose, we had to incorporate our empirical observation into the models that thalamocortical axons, which provide sensory input form the whiskers to the barrel cortex, target specifically and most densely the dendritic domain of L5PTs that initiates calcium APs [21]. Model failure is hence an integral part of ISF to reveal the limits of bottom-up models and to identify experiments that could improve them. In essence, model failure addresses the questions: are the model assumptions valid (e.g., columnar processing), is the quality of the empirical data sufficient (e.g., connectivity in slices), and are the strategies and algorithms to build these models plausible (e.g., axo-dendritic overlap vs. axons target specific dendritic domains)?

Another cost of our model consensus strategy is that it requires a large set of models. Consequently, the computational resources that ISF requires for generating, simulating and analyzing these models can be enormous, as we estimate in the Supplement for our application example in the barrel cortex (S1 Table). We hence designed ISF for high-performance compute infrastructures. However, the computational cost depends on the complexity of both the models and the in vivo observations under investigation, and may be substantially less compared to our use-case, where neuron models needed to capture the highly complex dendritic physiology of L5PTs, as well as inputs from neurons across the entire barrel cortex.

Our model consensus strategy could also set the stage to investigate the mechanistic origins of in vivo observed network dynamics at unprecedented levels of detail. However, we did not yet implement a workflow within ISF to simulate recurrent dynamics in networks of biophysically detailed neuron models. Even so, there is no principle reason why ISF could not be used for simulations of recurrent network dynamics. This would require to generate sets of network models where both the network configurations and their constituent biophysically detailed neuron models capture the empirically observed ranges of each parameter at each scale, as well as the diversity of possible parameter combinations across scales. Simulating and analyzing such sets of biologically diverse neuron-network models, and achieving model consensus about the mechanistic origins of in vivo observed network dynamics, will however require enormous computational resources. Incorporating advances in simulation technology (e.g., Jaxley [51]) into ISF may provide the necessary means to overcome these computational challenges by scaling to large GPU infrastructures.

In addition, our approaches in ISF to reduce the biophysically detailed neuron models to analytically tractable models, as well as other model reduction approaches [52], may provide the means to simulate recurrent dynamics in networks with similar biological realism, but at substantially lower computational costs. Well-established frameworks for simulating network dynamics such as NEST [53] and The Virtual Brain [54] indeed provide strategies for incorporating more biological details into network models of integrate-and-fire point neuron models, e.g., by constraining connection probabilities based on data from brain slices. Incorporating such biological details generally impacts the predicted dynamics of such simplified network models significantly. For example, by incorporating connectivity measurements from the barrel cortex into such models, we showed how empirically observed diversity in the number incoming synaptic inputs per neuron impacts the dynamical balance between excitation and inhibition in layer 4 [55]. Thus, replacing integrate-and-fire point neuron models with those generated by ISF via model reduction, may be a way to investigate network dynamics at increasing levels of biological realism. By making our code available and well-documented [56]  we hence seek to facilitate collaborative efforts that combine ISF with other frameworks [46,52,53,57] to simulate cortical dynamics at unprecedented levels of detail.

Outlook

Dissecting the cellular and circuit mechanisms that underlie in vivo recorded neuronal activity remains a major challenge. Current experimental techniques are insufficient to simultaneously measure all relevant variables under in vivo conditions. However, our vision for the rodent whisker system to dissect the mechanistic origins of its functions by generating, simulating and analyzing anatomically and functionally realistic models is shared broadly across the neurosciences. For example, large-scale initiatives, such as the Retinal Functomics consortium or the Allen Institute for Brain Science combine functional imaging of neuronal activity in the mouse retina [58,59] or visual cortex [60,61] with dense EM reconstructions. Similar experimental efforts are made for fruit flies, where an entire brain has already been reconstructed at EM resolution [62–65]. Thus, the empirical data that is necessary to generate bottom-up models similar to those that we developed for the barrel cortex is becoming increasingly available. ISF may hence be readily applicable to a growing number of model systems.

In summary, ISF’s key concepts are (1) generating sets of empirically well-constrained models that capture biological diversity and degeneracy within and across subcellular, cellular and network scales, (2) simulating in vivo conditions without parameter tuning, (3) analyzing the simulations to achieve model consensus about mechanistic origins that could account for in vivo recorded activity and its variability across animals, cells and trials, and (4) using in silico manipulations to identify experimental strategies that would be best suited to test whether the predicted mechanisms are indeed utilized in vivo. While we demonstrate the value of ISF for the barrel cortex with empirical data that is accessible today, we believe that the power of ISF and of its underlying concepts will continue to increase as new experimental techniques become available. For this reason, we make ISF available as a fully documented online resource that will facilitate iterative in silico – in vivo investigations into the mechanistic origins of in vivo functions.

Ethics statement

No animal experiments were carried out in this study. The previously reported animal experiments [40], were carried out after evaluation and approval by the local German authorities, and in accordance with the animal welfare guidelines of the Max Planck Society.

Implementation of in silico framework

In this section, we provide a detailed overview of how ISF implements the aforementioned workflows, as well as a general implementation overview. A high-level dependency graph and an overview of the most important file formats used in ISF are given in S1 and S2 Figs. The online documentation provides a complete API reference [22], in addition to tutorials [23] for the workflows outlined in the Results, and a specification on all file formats used in ISF [66]. We highlight some key features below and refer to the corresponding sections of the API reference.

ISF is written entirely in Python 3, with the exception of the NEURON mechanisms for channels and synapses, which are written in NMODL. It uses the NEURON simulation environment [67] as a backend, and provides extensive parallellization capabilities with Dask [68]: a mature and widely adopted multiprocessing library. Installation and environment management are handled through pixi [69], providing version-controlled environment lock files to ensure reproducibility. ISF is available for Linux, macOS and Windows (Windows support is experimental) and only requires minimal dependencies: pixi and git. Windows additionally requires a user-based installation of NEURON. Installation instructions for all OS’s are available online [70]. ISF is open-source under Apache 2.0, subject to continuous integration testing, and fully documented [22]. In addition, we provide tutorials [23] where we apply the workflows from the Results on L5PTs in the barrel cortex.

The Interface package provides top-level API access to all important workflows and classes, often only requiring a single function call to start a workflow. It is recommended to use ISF through Interface:

Neuron modeling

A neuron model in ISF is defined by a .hoc morphology file, and a single parameter file that contains passive and active properties in the .param format [26]. The single_cell_parser package provides convenient API to build multi-compartmental neuron models for simulation in NEURON.

Generating neuron models in ISF consists of two steps: optimization to find a seed model whose responses to current injection stimuli are consistent with empirical physiology, and then exploration to discover the full parameter space (i.e., degeneracy) of models consistent with empirical physiology. The empirical data this workflow requires as an input are (1) a morphology reconstruction, in .hoc format, (2) electrophysiological recordings that capture the neuron’s intrinsic physiology, and (3) bounds on known biophysical parameters (e.g., ion channel distributions) which describe the neuron’s passive and active properties, including intracellular dynamics if applicable. The electrophysiological recordings are parametrized into separate attributes (such as AP height and width, spike frequency etc.), hereafter called “objectives.” Given the empirically observed diversity of neuronal activity, each objective may be specified as a distribution.

The optimizer currently used in ISF is BluePyOpt’s [71] implementation of an indicator-based evolutionary algorithm (IBEA) [72]. Each iteration, BluePyOpt generates new sets of biophysical parameters, simulates a set of stimuli to each neuron model and evaluates the resulting responses to current injection stimuli by comparing the simulated result to each empirical objective. Neuron models that cannot be compared to a given objective (e.g., because they have no AP at all) are given a penalty value. The neuron models that had the best evaluation are then used to construct new models in the next “generation,” with some stochasticity. This continues until a model is found which is empirically consistent, meaning all attributes of its responses lie in-distribution of the empirically observed objectives.

ISF’s exploration algorithm [25] relies on the availability of a valid “seed” model that captures the desired empirically observed intrinsic physiology. For our example of L5PTs, BluePyOpt [71] was most successful in identifying seed models that captured the complex dendritic and perisomatic physiology of this neuronal cell type. However, any method that provides one or more seed points is compatible with ISF, including manual tuning, randomly sampling parameters (e.g., as in Roy and Narayanan (2023) [73]), gradient-based optimization (e.g., with Jaxley [51]), and recent extensions and frameworks around BluePyOpt [74]. The exploration algorithm then makes small random variations on these biophysical parameters, simulates the same stimuli and checks if the responses are still in-distribution of the empirically observed objectives. For all newly generated models that pass this check, this process is repeated. The direction of variation in the biophysical parameters can be configured to favor various metrics such as variance, increasing euclidean distance in parameter space, novelty, or others. In our experience, random variations were sufficient to explore a broad diversity of neuron models per morphology. A common stopping criterion for this exploration is to keep track of the variance of all found neuron models, and interrupt the exploration once the marginal increase in variance fails to reach a user-defined limit.

These routines are implemented in the biophysics_fitting package, which allows the user to configure any set of repeatable stimulus and evaluation protocols, and vary the biophysical parameters as a single parameter vector. Any parameter listed in ISF’s neuron parameter specification [26] can be modified, and hence used during optimization and exploration. Specifically, the ISF neuron parameter specification includes (1) ISF-specific parameters for spatial profiles of channel densities, (2) all NEURON parameters defined in the PARAMETER block of a loaded NMODL file, and (3) parameters used for ISF-specific cell modification functions, defined in the single_cell_parser.cell_modify_functions module. NEURON parameters for ion channels typically only include the conductance densities of ion channels for a given compartment of the neuron model. For calcium buffers, these are typically the parameters of the diffusion kinetics, such as the time constant, shell depth, and ratio of unbuffered Ca²⁺. However, any parameter of a NEURON mechanism can in principle be exposed by defining them in the PARAMETER block of the corresponding NMODL file, including channel opening and closing rates, temperature scaling coefficients, or calcium buffering and extrusion rates. In addition, the custom functions under single_cell_parser.cell_modify_functions can be applied to a neuron model by listing the function name, and their keyword argument names and values in the neuron parameter file. These parameters are explained in detail on our documentation

page [66].

Both the optimization and exploration algorithm can be started from Interface:

The tutorials provide explicit examples on how to use these packages for simulation [75], model evaluation [76], and model generation [37].

Network-embedded neuron models

The singlecell_input_mapper [30] package provides all functionality to generate anatomical network embeddings for a given morphology as described in Udvary et. al. (2022) [24] under singlecell_input_mapper.udvary2022. It is provided as a stand-alone package to allow the addition of other network embedding methods in the future. The purpose of any current or future network embedding method is to generate the synapse locations and presynaptic origin for a neuron morphology in the form of .syn [31] and .con [32] files. Alternatively, users can define these files themselves, e.g., based on dense EM reconstructions.

The input to the network embedding workflow is a model or reconstruction of the dense neuropil populated with annotated neuron morphologies, i.e., the cell ID of each neuron and their associated cell type, dendrite and axon reconstructions, the number and locations of pre- and postsynaptic structures (e.g., dendritic spines and axonal boutons), and the morphology of the biophysically detailed neuron model to be embedded (hereafter referred to as postsynaptic). Our dense neuropil model of the rat barrel cortex has been described and validated previously [24], can be downloaded and explored interactively [77], and can now additionally be downloaded via the pixi front-end in ISF:

Briefly, singlecell_input_mapper.udvary2022 discretizes the neuropil into voxels of a user-specified size, and identifies where the postsynaptic morphology intersects with these voxels. Synaptic connections are only possible if pre- and postsynaptic structures are located in the same voxel. For each of these voxels, synapses are created onto the postsynaptic neuron model from presynaptic neurons in the dense neuropil model/reconstruction based on their respective pre- and postsynaptic structures in the voxel and user-defined synaptic wiring rules [24,28]. The anatomical network embeddings are saved in the form of two files. A synapse distribution file (.syn [31]) describes where each synapse connects onto the postsynaptic morphology. A synapse connectivity file (.con [32]) describes to which presynaptic cell each synapse belongs, and which cell type this is. Any dense EM reconstruction is compatible with ISF as long as its description of connectivity onto the postsynaptic cell can be translated to these .syn and .con files. More information on this workflow can be found in the associated tutorial [29].

In order to generate synaptic activity onto the postsynaptic neuron, singlecell_input_mapper [78] converts all presynaptic cells to point neurons that emit spikes according to cell type-specific activity recorded during the in vivo condition to be explained. ISF provides functionality to define activity with different distributions, such as Poisson (default), normal, uniform, log-normal, or exact activations. ISF generates activation times at each individual synapse on the postsynaptic neuron based on its release probability, which can be defined for each presynaptic cell type individually. Each presynaptic cell type’s synapses are parameterized by delay, weight and dynamical parameters such as rise and decay time constants, and could include additional empirically constrained parameters, e.g., to capture different plasticity mechanisms [19,21]. Explicit code examples on how to perform this workflow is shown in the associated tutorial [41].

A single network model is captured by one parameter file [79], containing information on the anatomy (references to .syn [31] and .con [32] files), cell type-specific activation probabilities for an in vivo stimulus, synapse dynamics, and synaptic weights. It is important to note that two of these parameters are stochastic: the spike probability of presynaptic cells, based on their respective distributions, and the synapse activation probability upon a presynaptic spike. Therefore, each parameter configuration can be used to generate a set of synapse activation patterns, capturing trial-to-trial variability.

Simulating network-embedded neuron models

The simrun package simulates network-embedded neuron models from neuron and network parameter files using NEURON as a simulation engine. A tutorial for this process is available online [80]. It expects both parameter files, and can run multiple trials per neuron-network parameter configuration to capture trial-to-trial variability in cell and synapse activation patterns. These trials are parallelized across processes by default. A full workflow can be started with a single function call:

The simulation results can be efficiently parsed and stored by the data_base package. By default, these parsed simulation results include the membrane voltage of the neuron model at the soma and other specified recording sites, the exact cell and synapse activations that were used for that trial, and a copy of the parameter files that were used to generate these results for each simulated trial [81]. Converting the raw simulation output to high-performance dask dataframes in the msgpack or parquet format can be done by:

Manipulations

Re-running existing simulations with various manipulations to the neuron, network and synapses is achieved as follows:

Manipulations are defined by functions that accept a neuron parameter file, network parameter file, or synapse activation dataframe, and return it changed. This allows to re-run existing simulation trials (or a subset thereof) while making targeted manipulations, such as (but not limited to): scaling synaptic weights by distance, receptor type, presynaptic origin, or individually; changing synaptic activation patterns by the same categories; removing connectivity to presynaptic cells or populations; changing biophysical parameters; any combination of the above. A specific example that silences synaptic activations of a specific presynaptic population while keeping all other synapses exactly the same is provided in the online tutorials [42], which replicates the same manipulations as described in Egger et al. (2020) [19] on a small sample size.

Reduced models

The goal of model reduction is to capture the relationship between synaptic input and neuronal output in simulated network-embedded neuron models with a minimally sufficient description. A reduced model is a function f that predicts the neuronal output o from synaptic input , based on tunable parameters :

(1)

To build reduced models, ISF provides three modular components under simrun.modular_reduced_model_inference: a DataExtractor, Strategy, and Solver. These components define and o, f, and respectively. Each component can be configured by the user to best address their research question, depending on the in vivo condition, neuron models and network models under investigation. First, the DataExtractor allows the user to define data processing workflows that extract relevant features from the input and output of fully detailed simulated network-embedded neuron models. This defines and o respectively. For example, the output commonly takes the form of somatic spike times, but any aspect of the neuronal output can in principle be used, including somatic voltage traces, burst occurences, or dendritic events. Second, the Strategy object allows the user to define the exact mathematical form of the inference function f, including tunable parameters . ISF provides predefined Strategy objects, such as Generalized Linear Models (GLM) and Linear Discriminant Analysis (LDA) classifiers. Third, the Solver object defines a cost function and an optimization algorithm. The cost function g captures how well the output o produced by the inference function matches the actual neuronal output, e.g., a simple mean squared error (MSE) or the area under the receiver-operator curve (AUROC). The Solver object then adapts according to the user-defined optimization algorithm, e.g., COBYLA [82].

As a concrete example application, we provide an online tutorial [43] reproducing the results from Bast, Fruengel et al. (2024) [20]. Here, the DataExtractor extracted the soma distance and activation time of each synapse as input data , and the AP time points as output o. Next, the Strategy object defined f as a Generalized Linear Model (GLM), whose linear component were filter functions that weigh the contribution of each synaptic activation to the neuronal output o, depending on their soma distance and activation time. The filter functions were linear combinations of raised cosine basis functions, and the tunable parameters were defined as the weights of this linear combination. The nonlinear component of the GLM assigned a penalty to recent APs, to capture refractoriness. Finally, the Solver object defined the cost function as the AUROC of the AP response probability, and defined COBYLA [82] as the optimization algorithm to adapt the weights . Each reduced model was generated for one particular neuron model, embedded at multiple locations to capture network variability. The resulting reduced models were able to predict the AP response probability based on the soma distance and activation times of each synapse. The tuned filter functions of these reduced models allowed an analytically tractable comparison for how different neuron models weigh synapse activations, across a wide diversity of input patterns. Inference with these reduced models sidestep simulating compartmental neuron models, but the synapse locations (or at least their distance to the soma) and activation times need to be known, which still requires the network embedding workflow to be run. Each component used in the creation of this particular reduced model — i.e., the spatiotemporal synapse data extractor, raised cosine basis functions, wrapper to the COBYLA optimizer, and cost function — is available in ISF as predefined components.

Data management

ISF implements an intuitive data_base package with modern and performant data formats, such as parquet [83] and zarr [84], as well as an updated pandas interface to the msgpack format using fast and efficient compression such as blosc [85,86]. Extra care is taken into account to ensure all data is: (1) reproducible, (2) can be written or converted to standard file formats, and (3) saved efficiently, both in terms of IO speed and file size. To ensure reproducibility, all data saved with ISF’s data_base package contains rich metadata in human-readable JSON format. This metadata includes the commit hash of the code used to generate the data, creation date, file format, and which reader to use in order to load in the data. In case the data was generated in an interactive IPython session, such as Jupyter notebooks, data_base also saves a history of the IPython code as metadata using the %history magic command.

For high performance computing (HPC) contexts, ISF provides compatibility with multiple file locking servers: Apache Zookeeper, Redis, and simple file-based locking with fasteners. In addition, data_base is easy to use thanks to the Python dictionary-like syntax:

These considerations ensure high and stable performance, as well as data portability without the need for an additional interface or specialized data format.

Visualization

ISF provides extensive visualization capabilities for neuron models and network-embedded neuron models. Individual frames or videos of neuron models and network-embedded neuron models can visualize their membrane voltage, ion currents, and synapse activation patterns. The three outputs of the example code below is shown in S3 Fig.

These visualizations can be exported to VTK for postprocessing in VTK-compatible software suites like ParaView [87]. ISF additionally includes BlenderSpike [88] as a dependency for postprocessing in Blender. The combination of the two can create high-quality figures and animations. An example of both is shown in S4 Fig.

Stability and extensibility

ISF is designed such that it can integrate future work, without breaking existing functionality. ISF takes several measures to facilitate this goal.

First, ISF explicitly defines a stable API boundary in a dedicated interface, aptly called Interface. By convention, any function or class exposed in the Interface module may not change its API or functionality. Extensions can be made by adding keyword arguments to functions, whose default value should maintain the previous functionality. Classes can be extended with new methods and attributes. In case of incompatibilities it is preferred to create a new class or function with a new name, while keeping the previous function accessible and unchanged.

Second, the data_base module separates how data is saved (i.e., the data format) from the syntax on how to load it. This design makes it is possible to change the data format post hoc, should a better format become available. The data remains accessible via unchanged syntax. Thereby, previous code, including data analysis notebooks, do not need to be modified when updating data formats. New data formats can be easily added by adding a module to the data_base.IO.LoaderDumper package, which only needs to implement a set and get method. Thanks to this convention, simulation data that we saved more than a decade ago (in Python 2) is still readable with ISF today.

Third, we provide a single python environment to run all functionality within ISF. This minimizes the need to set up project specific environments and facilitates the exchange of code and data across projects and lab members. This python environment is version-tracked thanks to pixi’s lockfile capabilities. As an additional precaution, the git commit is saved in the metadata of each data object, providing a link to the exact Python environment that was used to save the data.