Predicting early intervention outcomes in autism via individual participant data mega-analysis.

Understanding the brain requires linking theoretical mechanisms of neurons to empirical observations in vivo. Biophysical neural models in neuroscience [–] help bridge fundamental neuronal properties, neurophysiology, cognition, and behavior. Models differ in their level of abstraction: conductance-based models can expose channel- and receptor-level mechanisms, population models can preserve efficient cell-level dynamics, and population or neural-mass models can summarize large-scale dynamics. Hodgkin-Huxley-style models [,] are useful when the scientific question concerns ionic currents, conductances, or their voltage and time dependencies. These modeling choices make it possible to ask how changes in ion-channel properties [–], cell types [–], cortical layers [–], neurotransmitters [–], neuronal metabolism [–], oscillations [–], and synaptic integration [–] affect perception, cognition, and behavior. Models of neuronal oscillations are particularly important for predictive processing theories [–], which propose that top-down and bottom-up processing are associated with distinct oscillatory channels. Top-down processing reflects internal or cognitive state variables such as attention or prediction [,], whereas bottom-up processing reflects how external sensory states are signaled [–]. Predictive routing (PR) [] is one theory that links predictive processing to empirically observed oscillatory dynamics. PR proposes that gamma-band activity and increased spiking carries bottom-up signals from lower-order sensory cortex toward higher-order areas, whereas lower-frequency alpha/beta activity reflects top-down predictions that modulate sensory processing [,,,]. In the present manuscript, PR is used as a motivating neuroscience case study because its proposed building blocks include spiking dynamics and beta/gamma state changes. The goal is not to prove PR or to implement a full predictive task; instead, we test whether automated optimization can fit spiking neural models to spectral and empirical objectives relevant to this framework. A mechanistic account of how these oscillatory dynamics emerge from neuronal circuits, and how they might support predictive routing, remains incomplete. Numerous studies have modeled oscillatory interactions with conductance-based, and population-level approaches [,,,–]. These models can generate oscillatory dynamics, including stimulus-evoked gamma oscillations. Many use Pyramidal-Interneuron networks (PIN), consisting of input drive to pyramidal neurons and feedback inhibition from interneurons. The oscillation frequency depends on inhibitory synaptic time constants, recurrent connectivity, and input structure, such that a network can be manually tuned to express PIN-gamma (PING) or PIN-beta (PINB). Networks of interconnected inhibitory neurons can also generate gamma oscillations without pyramidal participation, an alternative mechanism referred to as Interneuron Network Gamma (ING) []. In both PING and ING, synaptic connectivity, internal noise, and bottom-up inputs can generate rhythms that are either highly synchronous (strong PING, [,,,]) or sparse and irregular across neurons (weak PING, [–]). Biologically observed dynamics in healthy brains are often more consistent with the latter sparse and irregular regime.