Network-level coincidence detection in a computer simulation of primary motor cortex

Coincidence detection of stimuli has been reported for single cells. We evaluated a network model of primary motor cortex (M1) that detects coincidences based on arrival of signals at different network layers, rather than one cell's afferents. Our model was made up of 775 spiking neurons of 10 cell types (4 excitatory and 6 inhibitory) with details of connectivity based on interlaminar connectivity determined using brain activity mapping in mouse M1 slice. Individual neurons were modeled as event-driven, rule-based dynamical units. Simulations corresponded to an in vitro slice preparation with no ongoing input activity driving the cells. We determined thresholds for activation to generate sustained oscillations (> 100 s) across the network by % of cells stimulated in each layer: 13% (L2/3), 54% (L5A), 25% (L5B), and 17% (L6). Frequency and relative phase of the sustained oscillations in L5 depended on the layer of stimulation: those generated by stimulating deeper layers (L5A and L5B) were delayed in phase compared to those generated by stimulation of L2/3, whereas oscillations generated by L6 stimulus showed twice the frequency compared to others. Dual activation was performed with varying strengths and time lags between L5A and L2/3 stimulation. Activation depended on dual simultaneous stimulation, and was all-or-none. Sustained responses occurred at a threshold of 8% L2/3 and 35% L5A, well below the single-stimulation thresholds, demonstrating that a minimal superficial stimulation could trigger a sustained amplification of deep layer signals. This activation required that the two stimuli occur within a time window of 2 ms. The width of this time window increased with increasing L5A stimulus. It has been hypothesized that L2/3 stimulation represents the attentional signal from higher cortical centers. This signal could then be involved in gating feedforward signals generated in L5A. Our results suggest that coincidence detection could play a role in such attentional enhancement of responses. In this context, coincident activation would increase response gain in the stream attended to. This is also consistent with previous work suggesting that models with distinct laminar organization facilitate motor learning and voluntary actions. Modulation of responses with synergistic activation of deep and superficial layers suggests that the network can support predictive coding schemes.