Biologically informed cortical models predict optogenetic perturbations.

Area of Science:

• Computational neuroscience
• Systems neuroscience
• Machine learning in neuroscience

Background:

• Understanding cortical information processing relies on accurate models of neural circuits.
• Predicting responses to perturbations, like optogenetic stimulation, is crucial for testing model validity.
• Standard recurrent neural networks (RNNs) often fail to generalize to such perturbations.

Purpose of the Study:

• To evaluate the predictive power of recurrent neural networks (RNNs) on optogenetic perturbation data.
• To develop an improved RNN model incorporating biological inductive biases for better generalization.
• To explore the use of RNN gradients for targeted circuit manipulation.

Main Methods:

• Fitting generic RNNs to electrophysiological datasets.
• Developing an alternative RNN model with biologically informed inductive biases (structured connectivity, spiking dynamics).
• Testing model performance on simulated and in vivo mouse datasets with optogenetic perturbations.

Main Results:

• Generic RNNs showed poor generalization to unseen optogenetic perturbations.
• The biologically informed RNN model demonstrated improved prediction accuracy on perturbed trials.
• Theoretical analysis and simulations confirmed the utility of RNN gradients for micro-perturbation targeting.

Conclusions:

• Biologically informed RNNs offer a more accurate approach to modeling cortical information processing.
• These models can predict responses to interventions and potentially guide targeted circuit manipulation.
• This work advances the use of machine learning for understanding and interacting with neural circuits.