Fig. 5
From: Inferring and validating mechanistic models of neural microcircuits based on spike-train data
Estimation results for synaptic coupling strengths using synthetic data. a Example coupling matrix using 5 min long spike trains from N = 10 neurons and estimation method 1a. b Density of estimated vs. true coupling strengths from 25 networks of N = 10 neurons for uncorrelated (c = 0, left, as in a) and correlated (c = 0.1, right) external input fluctuations. Average Pearson correlation coefficient 〈ϱ〉 and mean absolute error 〈MAE〉 between true and estimated coupling strengths are indicated. Insets: histogram of estimated delays across all networks, arrow marks the true (global) delay. c True and estimated coupling strengths of N = 20 neurons from a randomly coupled network of Ntot = 1000 (800 excitatory, 200 inhibitory) neurons with connection probabilities of 0.1 for excitatory and 0.4 for inhibitory synapses, heterogeneous synaptic strengths and statistics of external inputs; c = 0, delay 1 ms. d–f Inference results from the I&F method compared to GLM- and CCG-based methods for the network in (c). d Estimated vs. true coupling strengths for the I&F method (left, as used in c) and the GLM method (right). Dashed gray lines mark identity. A discrimination threshold for the presence of connections was applied (indicated by green and orange dashed lines; for details see Methods section “Network model and inference details”). e Left: spike train CCGs for an excitatory connection (red), an inhibitory connection (blue) and averaged over all uncoupled pairs (black); right: CCG extremum z-score for positive lags vs. true coupling strengths, with discrimination threshold indicated. f Receiver operating characteristic (ROC) curves for the detection of synapses, i.e., true-positive rate vs. false-positive rate as the discrimination threshold varies, for the three methods. Indicated maximal balanced accuracy (BACC) values correspond to the thresholds visualized in (d) and (e), right. Dashed line represents random guessing. g Pearson correlation coefficient ϱ (top, cf. d and e, right) and max. BACC (cf. f) for the three methods using five different networks as in (c). Results from the same network are connected by gray lines. h–k Effects of connection probability and delay for subsampled networks of 500 excitatory and 500 inhibitory neurons. Left column: spike train CCGs averaged over subsets of excitatory (red) and inhibitory (blue) connections as indicated, and over all uncoupled pairs (black). Center and right columns: ϱ and max. BACC for the three methods (as in g). Connection probability and delay values are indicated in (h–j); k setting as in (j), but spike trains were perturbed by a temporal jitter (random values in [−1, 1] ms). Other parameter values as in (c). Results in (h–k) are from five networks each. In (c–k), 10 min long recordings were used
