Publication
How does the brain generate behavioral responses from sensory stimuli? This sensor-to-motor transformation is a fundamental question in systems neuroscience. While recent experiments have highlighted the complexity of neural representations across the cerebral cortex, computational models still struggle to capture these observations effectively. In this thesis, we advance the development of computational models that succinctly describe neural and behavioral data, enabling the formulation of hypotheses about the underlying causal mechanisms. First, based on the data collected by my experimentalist colleagues in three different areas of the mouse cortex during a whisker detection task, I designed and implemented a decoding model to quantify the relationship between neural activity and behavioral signals. My analysis, along with additional analysis conducted by colleagues, revealed that stimulus, decision, and movement information could be decoded across all recorded areas, namely the primary whisker sensory cortex, the medial prefrontal cortex, and the primary tongue-jaw motor cortex. Consistent with existing literature, stimulus information was most prominent in the sensory area, movement information in the motor area, and decision-making signals in the prefrontal area. Second, we introduce a novel method called trial-matching, which enables a data-constrained spiking recurrent neural network to capture trial-to-trial variability. Our model replicates trial-averaged and population-averaged neural activity from a whisker-delayed detection electrophysiology dataset that includes recordings from six cortical brain areas. This method features a one-to-one mapping between recorded and simulated neurons, with each neuron in the model assigned to a specific cortical area and neurotransmitter type (i.e., excitatory or inhibitory population). The model simulates a complete sensor-to-motor transformation by processing sensory stimuli across six simulated cortical areas and generating a behavioral response in the form of a ``jaw'' movement. Third, we evaluate whether the networks generated by our method implement the same causal mechanisms as biological circuits. We conduct a perturbation test replicating experimental optogenetic inactivations in silico by applying a transient current to an inhibitory population of neurons in the fitted models. We then compare whether this inactivation produces the same behavioral effects in our model as observed in the experimental recordings. Our results indicate that incorporating known biological inductive biases enhances the model's robustness against optogenetic perturbations. We believe our modeling approach lays the groundwork for future multi-area data-constrained models validated with causality tests, such as optogenetic perturbations. This methodology can ultimately facilitate the generation of experimentally testable hypotheses about the causal mechanisms underlying behavior.