Nicholas Tolley

@ntolley.bsky.social

163 Followers 269 Following

Postdoc in comp neuro @BrownU, interested in biophysical modeling, deep learning, and open-source software development

The default HNN model is used as a starting point to test hypotheses by either manually altering the values of the chosen model parameters, or using automated optimization and inference algorithms. Differences in parameter values pre-to post-treatment correspond to model-based predictions.
Biophysical modeling allows for testing mechanistic hypotheses that explain how EEG biomarkers emerge and change with drugs. Hypotheses about which drug mechanisms lead to distinct brain activity patterns must be constructed, and corresponding model parameters identified.
Modeling EEG biomarkers begins with picking a specific brain signal that is reliably different between patient populations. An example of a hypothetical EEG biomarker is an auditory event related potential (ERP) that is suppressed in post-treatment (red) relative to pre-treatment (blue).
Biophysical modeling to develop and test mechanistic hypotheses underlying pharmacological EEG biomarkers. Top: Modeling EEG biomarkers begins with picking a specific brain signal that is reliably different between patient populations. An example of a hypothetical EEG biomarker is an auditory event related potential (ERP) that is suppressed in post-treatment (red) relative to pre-treatment (blue). Middle: Biophysical modeling allows for testing mechanistic hypotheses that explain how EEG biomarkers emerge and change with drugs. Hypotheses about which drug mechanisms lead to distinct brain activity patterns must be constructed, and corresponding model parameters identified. Bottom: The default HNN model is used as a starting point to test hypotheses by either manually altering the values of the chosen model parameters, or using automated optimization and inference algorithms. Differences in parameter values pre-to post-treatment correspond to model-based predictions.
Figure showing neural trajectories to modified versions of the NMDA receptor arriving at the dendrite where either 1) the receptor just exhibits slow time constants, or 2) the receptor just exhibits a magnesium block. Neither variant was successfully trained to solve the task, suggesting that the unique properties of NMDA receptors are necessary for the emergence of fixed point attractors.
Figure of loss curves during training for different network variants where extrinsic inputs were fixed. Limiting extrinsic inputs to signal through NMDA receptors produced the best trained networks. AMPA inputs at the dendrite failed to solve the task.
Schematic figure illustrating how different biophysical properties are evaluated for their importance to generate fixed point attractors, 1) fix extrinsic inputs (i.e. NMDA or AMPA inputs to the dendrite or soma), 2) train cell and network parameter (i.e. all other parameters of the model are freely modified), 3) evaluate task performance (the emergence of distinct fixed point attractors for each input is considered a successfully trained network)
Visualization of the different scales of the biophysical model used in the paper. 1) Subcellular where voltage traces of a dendritic action potential are shown alongside the morphology of a multicompartment neuron, 2) cell-level where a spike raster shows how a single neuron exhibits preferential spiking to different inputs to the network, 3) network-level where a spike raster and neural trajectory plot both show that the network produces sustained changes in activity to inputs.
Multipanel figure illustrating the key components of the paper: 1) biophysical reservoir computing where a network of biophysically detailed excitatory and inhibitory neurons are randomly connected, receive a brief input, and produce a sustained spiking pattern in response, 2) an illustration of the task the biophysical reservoir computer is trained on: a simplified working memory task where the network must produce distinct fixed point attractors in response to different inputs.