Network-neuron interactions underlying sensory responses of layer 5 pyramidal tract neurons in barrel cortex
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.