AI and Inverse Methods for Building Digital Twins in Neuroscience

Deep learning is revolutionizing Neuroscience and Healthcare. One driver for this change is the exponential growth in the volume of biomedical data generated by modern medical imaging technology and automated data acquisition systems, such as automated patch clamps. Another driver is the increasing reliance of clinical diagnosis on deep learning algorithms to detect abnormalities in MRI scans. By automating the analysis of MRI images, algorithms free clinical staff from repetitive time and consuming tasks while removing subjectivity in the diagnosis and classification of brain tumors. Deep learning algorithms routinely assist neurosurgeons during critical operations, for example by stimulating surrounding brain tissue during tumor resection. Deep learning algorithms are also increasingly capable of forecasting epileptic seizures and assisting researchers in understanding how language is coded in the brain. Digital twins of brain activity trained on electroencephalographic time series have predicted epileptic seizures and connectivity changes in the brain during language comprehension. Machine learning is also progressing neuroscience by modelling biocircuits at the single neuron level. Recursive neural networks trained on electrophysiological data are making accurate predictions of the voltage oscillations of central pattern generators. The dynamics of individual ionic currents is also inferred to a good degree of accuracy when additional information in the form of surrogate model is provided. This importantly suggests that the dynamics of the membrane voltage which is observed and the ionic current waveforms which cannot be directly measured may be reconstructed from the analysis of electrophysiological time series.