Authors :
Presenting Author: Jacob Hull, PhD – Stanford University
Surya Ganguli, PhD – Associate Professor, Applied Physics, Stanford University; John Huguenard, PhD – Professor, Neurology, Stanford University
Rationale:
Absence seizures (AS) are characterized by synchronized spike wave discharges over multiple cortical areas. Numerous mutations, brain regions, and cell types are identified as generating AS. A focus on single regions or ion channels may however overlook common dynamics which achieve similar effects on inter-regional communication. Due to the brain’s complexity, describing these interactions requires a mathematical framework. However, without prior knowledge of the equations this approach is intractable. Recent advancements in machine learning enable model identification directly from primary data, simultaneously identifying model terms and their combinations which account for many simultaneous observations. Here we use interpretable data-driven model discovery, identifying a system of equations describing AS generation, capturing multiple phase, amplitude, and frequency-dependent interactions. Using silicon probes, we then identify the specific cortical and subcortical processes corresponding to the identified functions.
Methods:
We used 16-site electrocorticogram (Ecog) recordings from Scn8a+/- and Hcn2EA/EA mouse models of absence and the sparse identification of nonlinear dynamics (SINDy) machine learning algorithm, to discover AS governing equations. We then used silicon probes (1152 sites) in Scn8a
+/- mice to identify physiological correlates over thousands of individual neurons within 21 brain structures in awake and behaving mice during AS.
Results:
In the Ecog recordings, SINDy identified equations for nonlinear oscillators with phase and amplitude-dependent coupling.
When simulated, it recapitulates 36 measures of oscillation phase, amplitude, and frequencies across regions simultaneously. Phase plane analysis identifies seizure activity depends on the coupling strength between somatosensory and frontal/secondary motor regions with weaker contributions of autonomous activity. Using silicon probes, we find synchronous neuronal firing is highly correlated with the oscillations described by the model in somatosensory and motor integration related brain regions while limbic, prefrontal, visual, and olfactory related regions are less robustly recruited. We detect spike pattern shifts in the posterior thalamic nucleus (PO) two seconds before seizure onset, identifying a robust pre-seizure state only apparent when observing numerous neurons simultaneously. Using current source density analysis, we identify strong sink source pairs in layer one and layer five of cortex during absence seizures corresponding to the projection zones of PO axons. Importantly, PO is a region known to be involved in enhancing the strength of long range cortico-cortico communication across the somatosensory/motor axis, providing a physiological basis for the enhanced cortico-cortico coupling detected by SINDy. Conclusions:
Our interpretable machine learning approach identifies a quantitative mechanism of AS generation resulting from thalamic, specifically PO, coordination of long range cortico-cortico communication, providing multiple therapeutic targets which are not apparent by studying these brain regions in isolation.
Funding:
Work supported by (T32-NS07288 and F32-NS123009) to JMH and (R01-NS34774) to JRH.