Input-output computation underlying sensory responses of L5PTs

We developed our reduction approach by using one of the multi-compartmental models, whose morphology is shown as example 1 in Fig 1B. For each of its eighty-one network embeddings, we simulated its responses to passive deflections of the PW and the eight SWs, respectively. Within this simulation data, we searched for features in the synaptic input that are most predictive for the generation of an AP at a given millisecond—the ‘prediction time point’ (Fig 2A, methods). For this purpose, we grouped synapses according to their activation time points (1ms bins), by their pathlength distance to the soma (50μm bins), and depending on whether they are excitatory or inhibitory. We found that a weighted count of active excitatory versus inhibitory synapses, where the contribution of each synapse is weighted depending on its soma distance and activation time point, can predict AP output. We therefore determined the spatial and temporal weights (the “spatiotemporal filter”) that maximize the AP output prediction accuracy (methods). According to these weights (Fig 2B), the contributions of both excitatory and inhibitory inputs to AP output decay gradually with soma distance, reaching approximately zero at 500μm distance. Hence, primarily inputs to the proximal (i.e., basal and apical oblique) dendrites of L5PTs are predicted to contribute to their AP output upon sensory stimulation under this specific in vivo condition. In the temporal dimension, the contribution of excitatory and inhibitory synapses has a time course resembling excitatory and inhibitory postsynaptic potentials (EPSPs and IPSPs). Excitatory and inhibitory synapses contribute the most to AP output if they were active around 4ms and 9ms before the prediction time point, respectively. These observations were robust, independent of the method used for estimating the weights (methods, S3 Fig). The probability to observe a sensory-evoked AP in the multi-scale simulations increased with this weighted count of active synapses (Fig 2C), which we in the following refer to as ‘weighted net input’ (WNI). The WNI predicted AP output with high accuracy, as we quantified by the area under the receiver operating characteristic curve (AUROC 0.949).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Input-output computation of L5PTs upon single whisker deflections.

A: Exemplary responses of the multi-compartmental model with respect to the prediction time point (the time point for which the occurrence of an AP is to be predicted) for the three relevant response categories ‘AP’, ‘no AP’, and ‘recent AP’ (AP was elicited shortly before the prediction time point). B: Spatiotemporal input filter that best separates AP and no AP trials assigns strong weight to proximal synapses (top) active in a short time window before the prediction time point (bottom). C: Nonlinear relationship between WNI and AP probability. WNI represents the ‘drive’ a neuron receives; the higher the WNI the higher the probability an AP will be generated. D: Weighted net input–the input filtered by the spatiotemporal filter–separates AP and no AP trials, but not ‘recent AP’ trials, which can be distinguished based on a second measure, ‘time to previous AP’.E: Reduced model structure. APs are generated stochastically based on the AP probability (output of the nonlinearity). If an AP is generated, subsequent APs become less likely due to the post AP penalty, which is subtracted from the WNI. This reduced model directly relates AP output to synaptic input and previously generated APs in the simulated in vivo condition. F: The reduced model’s responses match the biophysically detailed model across many trials (close PSTH match) and on the single trial level (high AUROC score across all time points). Without the post AP penalty, the AUROC score drops during the sensory-evoked response.

https://doi.org/10.1371/journal.pcbi.1011468.g002

We revisited trials which were misclassified by the WNI and found that misclassifications mostly occur if there was a recent AP a few milliseconds before the prediction time point. We found that the multi-compartmental model is less excitable shortly after an AP, reflecting time constants of the involved ion channels, and the chance of eliciting an AP is hence low, even if the WNI is high. We therefore distinguished three categories: no AP, AP and recent AP (Fig 2A). We show that these three categories can be separated based on the WNI and time to previous AP (Fig 2D). To include this AP history-dependent relationship, we determined the separating line of response trials from the others (Fig 2D black line, methods) and incorporated it as a ‘post AP penalty’, i.e., we first computed the WNI as before and then subtracted a penalty value depending on the time since the previous AP. This penalized WNI predicted AP output with higher accuracy (AUROC 0.990) than WNI alone. Thus, three features, the spatial and temporal component of the WNI and the post AP penalty are sufficient to predict AP output from synaptic input in the multi-scale models.

Based on these three features, we could hence describe the input-output computation of the biophysically-detailed multi-compartmental model by an analytically tractable model. For this purpose, we assembled post AP penalty and the nonlinear relationship between WNI and AP probability into a generalized linear model (GLM, Fig 2E). We applied the GLM to predict AP output at different time points ranging from 25ms before to 25ms after the onset of sensory stimulation. Notably, even though the GLM was developed to predict the peak of AP responses, it also maintained a high AUROC score before and after this time point (Fig 2F). On a single-trial level, the model was able to predict if an AP was elicited in the 25ms after the sensory stimulus with an accuracy of 96.7% (S3 Table), and the timing of APs within this window was accurate to within 1.1±1.6ms (S4 Fig). Thus, the GLM accurately predicts APs throughout the entire 50ms time interval, and the post-stimulus time histogram (PSTH) predicted by the GLM hence matched with the PSTH predicted by the multi-compartmental model (Fig 2F).

We investigated how the performance of the GLM depends on the three features it is based on. Without incorporating the AP history (i.e., post AP penalty) into the GLM, the AUROC score drops during the peak response (Fig 2F, dashed line). Furthermore, when we simplified the GLM to neglect the spatial dimension, such that the weighted count only considers the time point of activation, but not the soma distance, the AUROC score decreased (S3A and S3E Fig). It was also insufficient to incorporate the spatial dimension categorically by only distinguishing a proximal and a distal compartment (S3B and S3E Fig). Therefore, soma distance and synapse activation time point need to be included at high resolution (50μm spatial bins, 1ms temporal bins) to accurately capture the input-output computation. To investigate the relative importance of the temporal and spatial filters on the prediction accuracy of the reduced model, we replaced one of these filters at a time with a fixed filter that only provides a cutoff value but does not weight synapses: replacing the spatial filter with weights of 1 for all inputs with less than 500μm pathlength distance from the soma and 0 otherwise reduced the AUROC score to 0.955. Similarly, modifying the temporal filter by assigning weights of 1 and -1 to excitatory and inhibitory synapses that were active in the 50ms preceding the AP, respectively, and setting the weights to 0 for all other active synapses reduced the AUROC score to 0.959. The modifications indicate that the spatial and temporal components of the WNI contribute equally to the prediction accuracy of the reduced model. In turn, considering the specific cell types of the neurons from which the synaptic inputs originate did not increase the prediction accuracy (S3C–S3E Fig). Taken together, the AP history, soma-distance dependent spatial distribution of synapses and their temporal activation pattern are necessary and sufficient features to accurately predict APs in the investigated L5PT multi-scale model. Thus, reduced models that are based on these three features, such as the GLM described here, provide an analytically tractable description of the input-output computation that a L5PT with this morphology and these biophysical properties performs upon single whisker deflections.

Input-output computation is robust to morphological and biophysical diversity

How does this input-output computation depend on the morphological and biophysical properties of L5PTs? To address this question, we applied the reduction strategy to all multi-scale model configurations of L5PTs with morphologically and biophysically diverse dendrites. The reduction revealed that the shape of the spatiotemporal filters, the shape of the penalty and nonlinearity are qualitatively similar across all L5PTs despite their morphological and biophysical diversity (Fig 3A): all L5PTs count active synapses, strongly weighting proximal input that occurs within the characteristic EPSP/IPSP-like time windows, with a similar AP-history dependent penalty. Each reduced model achieved a high accuracy before and after the sensory stimulus (Fig 3B), and highest accuracies during the peak response (AUROC median/min/max: 0.97/0.93/0.998). At single trial level, the models were able to predict if an AP was elicited in the 25ms after the sensory stimulus with a mean accuracy of 90.9% ± 3.9% (min 85.0%, max 96.7%, confusion matrices for all models are provided in S3 Table) and the timing of AP times was predicted accurately to within a mean of 2.0ms (min 1.1ms, max 3.1ms, S4 Fig). Thus, this input-output computation is robust across morphologically and biophysically diverse L5PTs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Input-output computation is robust to morphological and biophysical diversity.

A: Reduced models inferred on the different multi-compartmental models are qualitatively similar, with similar temporal and spatial filters, nonlinearity and post AP penalty. B: All models have high AUROC scores, specifically during the sensory-evoked (peak) response. C: We quantify latency, spontaneous AP rate (before the stimulus) and response probability for each pair of multi-compartmental and reduced model. D: Comparing response properties between multi-compartmental and corresponding reduced model shows close match.

https://doi.org/10.1371/journal.pcbi.1011468.g003

Notably, while the properties of the reduced models are qualitatively similar, they are not identical. For example, the decay of the spatial filter with soma distance differs slightly between reduced models, which largely reflects diameter differences of the apical trunk dendrite (S5 Fig). To what degree do these small variations in the reduced model capture differences of the input-output computations across L5PTs? We compared the predicted PSTHs between the multi-scale configurations of each multi-compartmental model with the respective reduced model and found that the slight differences between their shapes are captured well by the reduced model (S4 Fig). To quantify this similarity, we compared the time to maximum response (latency), AP rate before the stimulus, and response probability between multi-compartmental models and corresponding reduced models (Fig 3C). Each reduced model generated responses which matched those of the corresponding multi-compartmental models in all of these properties, while preserving the considerable variation across them (Fig 3D). Thus, the small variabilities in the reduced models account for differences in the input-output computation between the multi-compartmental models.

The model structure that we found to accurately predict AP output in simulations of the multi-compartmental models is surprisingly linear. This indicates that synaptic input patterns that mimic passive single whisker deflections during our specific in vivo condition do either not strongly activate nonlinear dendritic mechanisms in L5PTs, or these mechanisms are activated, but do not strongly influence the prediction of AP output. To test these possibilities, we quantified the occurrences of calcium APs and sodium currents in the dendrites, and the contributions of NMDA and AMPA currents in apical and basal dendrites, respectively. For this purpose, we recorded currents through voltage-gated ion channels, and through AMPA and NMDA receptors during simulations of PW deflections in a multi-compartmental model (example morphology 2). We recorded these currents for 16 locations across the basal and apical dendrites, including the primary branchpoint of the apical trunk (S6A Fig). Dendritic nonlinear mechanisms were frequently activated in the simulations. For example, in the apical tuft, we found peaks of sodium influx (S6E Fig). At the primary branchpoint, we found that calcium APs occurred in 18% of the simulation trials (S6B Fig). The calcium APs, however, did generally not result in somatic burst firing (response during 179 trials with calcium APs: 13% no AP, 83% 1 AP, 4% burst with 2 APs). The reduced model predicted the response equally well for trials with and without calcium APs (86% vs 90%, p = 0.07). Moreover, in our simulations, NMDA currents dominate in apical trunk and tuft dendrites (S6C, S6D and S6F Fig). We compared the prediction accuracy between trials with low NMDA contribution (33^rd percentile and below) and high NMDA contribution (66^th percentile and above) as measured in the tuft dendrites. The strength of these distal NMDA contributions did not affect the prediction accuracy of the reduced models (88% vs 91% for trials with low vs high NMDA contributions, p = 0.17). Thus, nonlinear synaptic and dendritic mechanisms are frequently activated in the simulations that mimic synaptic inputs to L5PTs upon passive single whisker deflections, but either their impact on AP output is low in the investigated condition, or the reduced models are able to capture their impact on the effective input-output computation. To investigate this, we re-simulated all whisker-evoked responses for one multi-compartmental model (example morphology 2) and reduced by 50% or removed NMDA conductance in all synapses. Reducing NMDA reduced response probabilities to all whiskers (S7A and S7B Fig), indicating that NMDA strongly contributes to AP output. When we inferred reduced models (S7C Fig) from these modified multi-compartmental models, we found that reducing NMDA changes both the spatial and temporal filters. More specifically, NMDA is predicted to increase the influence of distal synapses (S7D Fig) and to increase the integration time window (S7E Fig). Notably, nonlinear effects of NMDA thereby decrease the prediction accuracy of the linear reduced models (S7F Fig), which hence reach a maximal AUROC score of 0.99999 in the absence of NMDA.

Network vs neuron contributions to receptive field variability

How can the large variability in L5PT responses arise from such small variability in input-output computations? We repeated the simulations of all multi-scale model configurations, this time with the reduced models instead of the multi-compartmental models. The probabilities that a network input elicited a response (i.e., ≥ 1 AP 0-25ms post passive whisker deflection) in the reduced and corresponding multi-compartmental models were highly correlated for input from each whisker, all network embeddings, and all L5PT models (Fig 4A, Pearson correlation coefficient = 0.97). Consequently, also the receptive fields were virtually identical between the reduced and multi-compartmental models (Pearson R between receptive fields 0.96±0.01, Fig 4B and 4C). Thus, the reduced models capture the variability of responses for each multi-compartmental model across simulation trials, and the variability of responses across multi-compartmental models with different network embeddings, biophysical properties, and/or dendrite morphologies.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Reduced models predict origins of receptive field variability.

A: Comparison of responses of 7 different multi-compartmental models and their corresponding reduced models to 9 different whisker stimuli (PW and 8 SW) in 81 different network embedding locations. Response probability is the probability that one or more APs are generated 0-25ms after the sensory stimulus. B: Comparison of exemplary receptive field shapes shows close match between biophysically detailed and reduced model. C: Quantification of receptive field similarity for all cell positions and biophysically detailed models. D: Comparison between in vivo and reduced model responses to 9 different whisker stimuli (PW and 8 SW). Green dots represent the mean response probability. E: Exemplary receptive fields of reduced models. F: Influence of biophysics, morphology and cell position on receptive field shape, quantified by computing the correlation coefficient between receptive fields if one of these properties is changed.

https://doi.org/10.1371/journal.pcbi.1011468.g004

How well do the sensory responses predicted by the reduced models match with the in vivo data? We compared response probabilities to each whisker stimulus (PW and 8 SWs) between in vivo data and in silico predictions (Fig 4D). The distribution of response probabilities closely matches for any whisker (p values ranging between 0.08 and 0.77) and the means of these distributions, i.e., the ‘mean receptive field’, is significantly correlated (R value 0.77, p = 0.015). The simulated receptive fields even contain matches to the extreme cases observed in vivo, for example weaker response to the PW than to a SW, selective response to few whiskers, unselective response to virtually all whiskers (Fig 4E). Thus, the reduced models predict means, variance and outliers consistent with those observed in vivo, despite the small variability in input-output computations.

Because they capture trial-to-trial and cell-to-cell variability of sensory responses as observed across L5PTs in vivo, the reduced models can dissect how variations in network input or neuron properties could contribute to these observations. For this purpose, we computed the pairwise correlation between receptive fields across different morphologies, biophysical parameters, and network embeddings (Fig 4F). We found a change in network embedding, and therefore synaptic input patterns, to have the strongest effect (i.e., leads to the largest drop in the Pearson correlation coefficient), followed by the neuron’s morphology and biophysical properties. The models thereby indicate that variations in network input are the primary determinant of receptive field variability from cell to cell, whereas variations in morphology and biophysics play a minor role. Thus, the reduced models predict that L5PTs with diverse morphologies and biophysical properties perform the same input-output computation upon passive single whisker deflections in anesthetized rats, and the variability of their responses across trials and cells is determined by the variability in input that they receive from the network–i.e., from VPM thalamus and barrel cortex.

Contribution of different input pathways to sensory responses

Which presynaptic neuron populations underlie the variable AP responses? While so far, we only distinguished between excitatory and inhibitory cells, the network model provides information about the location and cell type of the neuron from which each active synapse originates [15], for both thalamocortical (i.e., VPM) and excitatory intracortical cell types: pyramidal neurons in L2/3/4 (L2PY, L3PY, L4PY), spiny neurons in L4 (L4SP), intratelencephalic neurons in L5 (L5IT), L5PT, corticocortical neurons at the L5/6 border (L6CC) and in deep L6 (L6INV), and corticothalamic neurons in L6 (L6CT). To dissect how these presynaptic populations contribute to AP output, we utilized the reduced models and calculated the WNI separately for each cell type across whisker stimuli and network embeddings (i.e., we applied the spatiotemporal filters to synaptic input from each presynaptic cell type one at a time, Fig 5A). First, we analyzed L5PTs embedded at the center of the barrel column that is somatotopically aligned with the PW defined in this study (i.e., C2). The analysis shows that presynaptic neurons contribute in two ways to APs: by their spontaneous activity preceding the stimulus (‘baseline WNI’) and/or by their increase in activity upon whisker deflection (’sensory-evoked WNI’). Predominantly, L5PTs contribute to the baseline WNI (see Fig 5A baseline), due to their high spontaneous firing rates [12,17]. To isolate the effect of the sensory-evoked WNI, we subtracted the baselines (Fig 5B). After the stimulus, VPM provides the first sensory-evoked contribution to the WNI, followed by L6CC, which is the first intracortical cell type that responds reliably to VPM input [14]. For a SW stimulus (Fig 5C), the situation is different: VPM contributes little to AP output, while L6CCs provide the main drive. Thus, the reduced models predict that depending on the stimulated whisker, thalamocortical and intracortical pathways–primarily VPM and L6CCs–contribute differently to the sensory-evoked response.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Reduced models predict contribution of input pathways to sensory responses.

A: Absolute contribution of presynaptic populations to WNI following a PW stimulus to a model located at the center of the C2 column: pyramidal neurons in L2/3 (L2PY, L3PY), spiny neurons in L4 (L4SP), intratelencephalic neurons in L5 (L5IT), L5PT, corticocortical neurons at the L5/6 border (L6CC), and relay cells in the ventral posterior medial nucleus of thalamus (VPM). Despite the lack of sensory-evoked responses by L4PY, L6CT and L6INV [14], their contributions were considered in the overall WNI calculations (Figs 2 and 3). B: Sensory-evoked contribution (i.e. absolute contribution minus baseline for each input pathway) of presynaptic populations to WNI following a PW stimulus. C: Sensory-evoked contribution of presynaptic populations to WNI following a SW stimulus. D: Contribution of the main input pathways–VPM and L6CC–depending on the soma location of the L5PT model in a 9x9 grid across the C2 column for a PW stimulus. The black circle denotes the C2 column border. E: Comparison between model responses to 9 different whisker stimuli (PW and 8 SW) under control conditions and when removing sensory-evoked input from L6CC. Removing evoked L6CC activity attenuates responses, in particular to surround whiskers.

https://doi.org/10.1371/journal.pcbi.1011468.g005

Could variability in input from these two pathways underlie variations in the receptive fields across L5PTs? We computed, which fraction of the WNI is provided by VPM versus L6CC depending on the network embedding. This analysis revealed that L5PTs receive different amounts of input from these two pathways depending on their location (Fig 5D). As a result, L6CCs provide strong input to L5PTs for both PW and SW stimuli, whereas VPM provides strong input for PW stimuli only. Inactivating L6CCs–a manipulation which we performed previously in silico and in vivo [14]–results in narrow receptive fields, as it largely abolishes L5PT responses to SWs, but it also reduces the probability of PW responses. We repeated this manipulation now with the reduced models. Indeed, inactivation of L6CCs resulted in a major reduction of all sensory responses, and a narrowing of receptive fields across all models and network embeddings (Fig 5E). Thus, the reduced models capture the origins of the fast and broadly tuned responses that L5PTs in barrel cortex elicit upon single whisker deflections–i.e., drive to proximal dendrites via the VPM➔L6CC➔L5PT pathway [14].

Ethics statement

No animal experiments were carried out in this study. The previously reported animal experiments [12], 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.

Extracellular recordings

We reported the data for receptive fields (RFs) and morphologies used in this study previously [12]. Briefly, Wistar rats (P25-30, both sexes) were anesthetized with urethane and responses to passive single whisker deflections (applied with a piezo manipulator) of the PW and SWs were recorded via a cell-attached pipette. The in vivo recorded neurons were filled with biocytin and morphologically reconstructed to identify their cell types. We determined the PW by identifying the barrel column in which the soma was located–i.e., the PW is not necessarily the one that evoked the strongest response. The in vivo data for L5PTs used in this study comprised exclusively L5PTs located in the D2 column.

Morphological reconstructions

Neuronal structures were extracted from image stacks using a previously reported automated tracing software [30]. For reconstruction of biocytin labeled neurons, images were acquired using a confocal laser scanning system (Leica Application Suite Advanced Fluorescence SP5; Leica Microsystems). 3D image stacks of up to 2.5mm × 2.5mm × 0.05 mm were acquired at 0.092 × 0.092 × 0.5μm per voxel (63x magnification, NA 1.3). Image stacks were acquired for each of 45–48 consecutive 50μm thick tangential brain slices that range from the pial surface to the white matter. Manual proof-editing of individual sections, and automated alignment across sections were performed using custom-designed software [31]. Pia, barrel and white matter outlines were manually drawn on low-resolution images (4x magnification dry objective). Using these anatomical reference structures, all reconstructed morphologies were registered to a standardized 3D reference frame of rat vS1 [32].

Multi-compartmental models

We selected 5 L5PT reconstructions that are representative of the morphological variability of this cell type. Multi-compartmental models were generated for these morphologies as described previously [14,16]. Briefly, a simplified axon morphology was attached to the soma of the reconstructed L5PT dendrite morphology [33]. The axon consisted of an axon hillock with a diameter tapering from 3μm to 1.75μm over a length of 20μm, an axon initial segment of 30μm length and diameter tapering from 1.75μm to 1μm diameter, and 1 mm of myelinated axon (diameter of 1μm). Next, a multi-objective evolutionary algorithm was used to find parameters for the passive leak conductance and the density of Hodgkin-Huxley type ion channels on soma, basal dendrite, apical dendrite and axon initial segment, such that the neuron model is able to reproduce characteristic electrophysiological responses to somatic and dendritic current injections of L5PTs within the experimentally observed variability, including back-propagating APs, calcium APs, and AP responses to prolonged somatic current injections [16]. We augmented the original biophysical model of L5PTs [14,16] with two ion channel parameters as previously described [34]: in accordance with a previous report [35], the density of the fast non-inactivating potassium channels (Kv3.1) was allowed to linearly decrease with soma distance until it reaches a minimum density (i.e., the slope and minimum density are two additional parameters, see [34]). The diameter of the apical dendrites was optimized by a scaling factor between 0.3 and 3. We incorporated the IBEA algorithm [36] for optimization. The optimization was terminated if there was no progress or when acceptable models had been found. We repeated the optimization process several times. From each independent run, we selected one model for which the maximal deviation from the experimental mean in units of standard deviation across all objectives was minimal (0.9–1.9 mean STDs across objectives).

Network embeddings

Cell type-specific thalamocortical (from the ventral posteromedial thalamic nucleus (VPM) which is the primary thalamic nucleus of the whisker) and intracortical (from excitatory and inhibitory cells in vS1) connections are derived from an anatomically realistic circuit model of rat vS1 [13], a procedure, which has been described in detail previously [14]. We embedded the dendrite morphologies selected for multi-compartmental modelling in the network model at 81 locations within the cortical barrel column representing the C2 whisker, which is located approximately in the center of vS1. For all in silico data presented in this study, the PW is hence the C2 whisker. Thus, the morphologies of the in vivo recorded L5PTs were registered to the C2 column, while preserving their in vivo soma depths and laminar dendrite distributions that was observed empirically in the D2 column [32]. The locations were the column center, and equally spaced grid with a distance of 50μm between adjacent somata. For each of the 81 locations, we estimated the location of presynaptic neurons in VPM and vS1 that provide input to the respective L5PT (Fig 1D) and where along the L5PT dendrite they synapse (Fig 1E).

Synapse models

Synapse models and synaptic parameters (rise and decay times, release probabilities, reversal potentials) were reported previously [14]. Briefly, conductance-based synapses were modeled with a double-exponential time course. Excitatory synapses contained both AMPA receptors (AMPARs) and NMDARs in 1:1 ratio. Inhibitory synapses contained GABA ARs. The peak conductance of excitatory synapses from different presynaptic cell types was determined by assigning the same peak conductance to all synapses of the same cell type, activating all connections of the same cell type (i.e., all synapses originating from the same presynaptic neurons) one at a time, and comparing parameters of the resulting unitary postsynaptic potential (uPSP) amplitude distribution (mean, median and maximum) for a fixed peak conductance with experimental measurements in vitro (IC input [37]) or in vivo (TC input [18]). The peak conductance for synaptic inputs from each cell type was systematically varied until the squared differences between parameters of the in silico and in vitro/in vivo uPSP amplitude distributions were minimized (i.e., the mean, median and maximum of the distributions were used, and mean and median were weighted twice relative to the maximum). This procedure was repeated for each multi-compartmental model using the connectivity model for the location in the center of the C2 column. The peak conductance at inhibitory synapses was fixed at 1nS [17].

Synaptic input patterns

Synaptic input patterns to the L5PT model were estimated as described previously [14]. Briefly, we first chose the whisker stimulus. Then presynaptic neurons were grouped by their cell type and the column in which they are located. For each column, the relative position to the stimulated whisker was determined. Neurons in this column were activated based on the electrophysiologically recorded responses for this cell type and relative position. Each AP in a presynaptic neuron is registered at all synapses between the presynaptic neuron and the L5PT model without delay and may cause a conductance change, depending on the release probability of the synapse. Depending on the network embedding, neurons receive different ratios of excitatory and inhibitory inputs, which is not guaranteed to maintain functional E/I balance [38]. We therefore introduced a scaling factor for evoked inhibitory input strength, which we constrained based on the empirically observed PW response probabilities. The scaling factors ranged between 0.79 and 1.56 depending on the multi-compartmental model.

Multi-compartmental model simulation data

Simulations were performed using Python 2.7, dask [39] and NEURON 7.4 [40]. We simulated 1000 trials for each of the 9 multi-compartmental models and each of the 81 locations and 9 whisker stimuli (PW, all 8 SWs). Additionally, the database contained 81000 trials of one 2^nd SW (E2), which was however not incorporated in the analysis. This results in 810000 simulation trials per multi-compartmental model, which reflects the effect of biophysics, morphology, network embedding and stimulus on sensory-evoked responses.

Reduced model inference

We performed the reduced model inference for each multi-compartmental model separately. We split the multi-compartmental model dataset for the respective model into a training and test dataset (split ratio: 70% to 30%, respectively). Synapses are binned based on their time point of activation (1ms bins) and soma distance (50 micron bins). We excluded trials of the recent AP category (in which there was an AP in the last 50ms). Spatial and temporal filters were constructed as a weighted sum of basis functions f_i and g_j. The soma distance dependent weight w_z and the time dependent weight w_τ of a synapse are given by where a_i and b_j are free parameters.

We used raised cosine functions [41] as basis functions. The temporal basis functions f_i were: for τ such that k·log(t + c) ϵ [ϕ_i – π, ϕ_i + π] and 0 elsewhere. Values used were k = 3, c = 5, and φ ϵ [3, 12]. Analogously, the spatial basis functions g_i were: for z such that k·log(z + c) ϵ [ϕ_i – π, ϕ_i + π] and 0 elsewhere. Values used were k = 2, c = 1, and φ ϵ [1, 11]. The basis functions are visualized in S9 Fig.

Using the spatiotemporal filters, we compute a weighted sum of active synapses at a time point t, in the following referred to as weighted net input (WNI). The WNI is computed as follows: where, τ ϵ [0, 80ms] is the time before t, z ϵ [0, 1300μm] is the distance of the synapse from the soma, and n_z,t–τ is the number of active synapses at a given 1ms time and 50 micron soma distance bin. w_τ and w_z are the temporal and spatial filter, respectively. n, w_τ and w_z are split by synapse type, as indicated by the subscripts E (excitatory) and I (inhibitory).

To adjust the free parameters (a_i and b_j for excitatory and inhibitory input) which determine the shape of the spatial and temporal filters, we use a gradient-free optimization method (COBYLA, implemented in SciPy: [42]) to maximize the area under the receiver operating characteristic curve (AUROC, see section ‘Analysis’) between the WNI and AP output for a selected 1ms time bin. The 1ms time bin selected for the optimization is in the following referred to as the inference time point t_inference. For each multi-compartmental model, we performed this optimization for t_inference ϵ [0,25]. We then manually selected one inference time point that resulted in reduced models which generalized well to other time points. The selected inference time point of each selected model is visualized in S4 Fig. Finally, we normalized the spatial and temporal filters such that the peak of the temporal filter is 1 for excitatory synapses and -1 for inhibitory synapses, and the value of the spatial filter for excitatory synapses at spatial bin 0 is 1.

In order to calculate the spiking probability for a given WNI value, we used these filters to calculate the WNI for all multi-compartmental model simulation trials from the training dataset at ti_nference, and recorded whether a AP occurred at this time bin or not. The trials were then binned by WNI. The AP probability for each bin corresponds to the proportion of trials that produced an output AP in the multi-compartmental model simulation (Fig 2C). Bins with few data points were combined to ensure a minimum of ten data points per bin. Linear interpolation was used to find the spiking probability corresponding to any WNI value, and WNI values greater/smaller than values seen in the biophysical model simulation were assigned the highest/lowest spiking probability respectively. We hereafter refer to this as the “nonlinearity” function. In order to estimate the effect of recent APs on spiking probability, WNI values were plotted against the time since the previous AP (Fig 2D), and a boundary was drawn to best separate spiking from non-spiking trials as follows. We dropped points in the lowest 5% of all WNI values (to remove the effect of outliers) and then drew a boundary along the minimum remaining spiking WNI values. This boundary was normalized such that it is zero for time to previous AP → ∞ by subtracting the respective offset. We found this two-step procedure (first estimate spatiotemporal filters based on a dataset in which recent APs are filtered out, second determine the post AP penalty based on these filters) to be more robust than a joint inference of both. WNI with penalty applied maintains a high AUROC over all stimulus periods (Fig 2F), while the uncorrected AUROC drops after the sensory stimulus.

Reduced model simulations

After the inference of the reduced models, we computed the WNI for all time points for which we wish to predict the occurrence of APs (in the manuscript referred to as a prediction time point). AP output was generated iteratively for one millisecond time bin at a time. For each time point, we computed the WNI by applying the temporal and spatial filter to the synaptic input of the preceding 80ms. The WNI was transformed to an AP probability based on the nonlinearity derived above. APs were randomly sampled based on this probability. If an AP occurred, the WNI of consecutive time points was updated based on the post AP penalty.

Analysis

‘AP probability’ is used to denote the probability that an AP is evoked within a 1ms time bin. ‘Response probability’ denotes the probability that one or more APs are generated in the 25ms following a whisker stimulus. ‘Accuracy’ is the percentage of trials in which the multi-scale and reduced model agree on whether or not an AP is elicited in a 25ms window following the whisker stimulus. We used the SciPy function ‘metrics.roc_auc_score’ to compute the AUROC score, which by default uses the full range of values (i.e. thresholds are automatically chosen such that the false positive rate varies between 0% and 100%). Distributions of response probabilities were compared with the two-sided Kolmogorov-Smirnov test using the SciPy function ‘scipy.stats.ks_2samp’.