Fig. 1: Overview of the whole brain model, variational autoencoder and perturbation pipeline.

A biophysical model simulates the local dynamics of a 90 region parcellation, with structural connectivity provided by DTI data of healthy subjects. Model parameters are adjusted to fit FC matrices derived from fMRI data of each condition. Pharmacological perturbations are simulated by changing the synaptic scaling of each region weighted by PET receptor maps, which provide a measure of local receptor density. Subsequently, a variational autoencoder (VAE) is trained to reconstruct the input data, resulting in a low-dimensional latent space which facilitates the visualisation of the perturbations, while also providing a geometric interpretation of their effect.