One size does not fit all: Calibrating microstimulation to individual subjects using spiking network models

Microstimulation is an effective tool for manipulating brain activity, but its drawbacks are sobering: electrode locations are not known exactly, electrode efficacy is variable, the cells being stimulated are rarely those being recorded from, and the number of independent electrodes is much smaller than the dimensionality of the systems being stimulated. One approach is to use optimal control algorithms to design microstimulation protocols through trial and error. However, the effectively infinite space of stimulation protocols, the slowness of training optimal control models, and the limited lifespan of experimental subjects make it unlikely that such approaches could find the globally optimal solution. Computer simulations, by contrast, are omniscient and immortal; their sole limitation is their potential dissimilarity to real brains. In this pilot study, we show how a spiking network model can be used with optimal control algorithms to expedite the development of microstimulation protocols. Specifically, we tuned and validated large-scale spiking network models against data from individual rats, then used these calibrated models to design microstimulation protocols specific to each subject. Electrophysiological data were recorded from the somatosensory cortex and thalamus of five rats during both natural touch and microstimulation. The network models consisted of 24,000 spiking Izhikevich neurons with cell types and connectivities drawn from empirical data. Global model parameters (including overall balance of excitation versus inhibition, connection probability, average axon length, and relative strength of internal versus external input) were calibrated to experimental data (including firing rates, local field potential [LFP] spectra, peristimulus time histograms, stimulus fields, and inter-electrode mutual information) using a nonlinear optimization algorithm called Bayesian adaptive locally linear stochastic descent. Once tuned, the spiking network models were used with a model predictive control (MPC) algorithm to design a microstimulation protocol that minimized mismatch with an output LFP. Compared to applying MPC to a subset of simulation data comparable to that actually available from the subjects, applying it to the full set of data available in the simulation significantly reduced prediction error. Furthermore, the microstimulation protocols tuned to individual brains showed less error than protocols designed using pooled data only. These results demonstrate the potential of using spiking network models calibrated to individual brains for the development of microstimulation protocols.