Biologically informed cortical models predict optogenetic perturbations

Ideally a recurrent spiking neural network fitted to large electrophysiological data sets should illustrate the chain of cortical information transmission. In particular, successful network model reconstruction methods should enable a model to predict the response to optogenetic perturbations. We propose a method that extracts a recurrent spiking neural network from electrophysiological data sets and test it on unseen optogenetic interventions. Our method uses multiple objectives to reproduce various aspects of neuronal activity patterns as well as biology-based inductive biases like structured connectivity and separation of neurons into dominant excitatory and inhibitory cell types. We show that theses biological inductive biases are crucial to predict responses to optogenetic perturbations. Furthermore, once a model is fitted to experimental data, gradient evaluations in the extracted model network can, in principle, be used to design a minimal low-power optogenetic stimulus that causes a maximal effect by stimulating a limited subset of neurons. Our computational methods provide a systematic path towards translating observed spike data into a matching recurrent neural network that is able to predict the results of a variety of experimental interventions.