G-DYNAMICS: Decoding and yielding neural AI models for integrative computational simulation of brain activity

The emergence of high-throughput recording modalities has granted access to large-scale neural population activity, yet interpreting these high-dimensional, non-stationary signals remains a foundational challenge in computational neuroscience. Traditional analytical frameworks have predominantly relied on descriptive or discriminative methods, mapping recorded activity directly to external stimuli or overt behavioural readouts. While valuable, these approaches often treat neural circuits as static input-output systems, failing to capture the autonomous internal dynamics, feedback loops, and latent continuous manifolds that drive biological computation.To bridge this gap, the G-DYNAMICS initiative champions a paradigm shift: leveraging modern generative artificial intelligence as an integrative, biophysically grounded simulation engine for brain activity. Generative architectures, including diffusion probabilistic models, variational autoencoders, generative adversarial networks, energy-based models, and continuous-time neural ordinary differential equations, provide powerful frameworks for learning the underlying statistical and dynamical distributions of neural firing patterns. Rather than merely fitting static observations, these models allow researchers to synthesise realistic neural dynamics, reconstruct unobserved circuit variables, and simulate multiscale brain activity across diverse recording platforms.Within this simulation-first framework, decoding neural representations and generating predictive forecasts serve as rigorous benchmarks of model fidelity. By acting as in silico experimental testbeds, G-DYNAMICS models enable investigators to apply virtual perturbations to test mechanistic hypotheses that remain difficult to investigate in vivo. Crucially, this collection emphasises integrative modelling: harmonising heterogeneous data streams, such as cellular spiking, local field potentials, and whole-brain neuroimaging, within coherent and self-consistent computational simulations.This Research Topic aims to advance the theory, implementation, and biological validation of generative neural models. We welcome all articles addressing themes including, but not limited to:1. Generative architectures for neural population dynamics: Development and application of diffusion models, variational autoencoders, generative adversarial networks, and continuous-time dynamical systems, including Neural ODEs and state-space models, to capture latent neural manifolds.2. Integrative multimodal simulation: Generative frameworks capable of bridging temporal and spatial scales and integrating microscale cellular recordings with macroscale signals such as LFP, EEG, MEG, and fMRI.3. Virtual perturbation and in silico experimentation: Generative neural simulations used to predict circuit responses to interventions such as microstimulation, targeted optogenetics, pharmacological modulation, and neurostimulation.4. Generative decoding and prospective forecasting: Algorithms that decode latent cognitive, sensory, or motor representations and generate prospective predictions of non-stationary brain-state transitions.5. Biological grounding and dynamic fidelity: Benchmarks, mathematical frameworks, and validation metrics assessing whether generative models reflect biological computation rather than statistical overfitting.6. Data augmentation and synthetic neurophysiology: Generative approaches designed to address data scarcity, denoise experimental recordings, and produce biologically constrained benchmark datasets for the computational neuroscience community. Topic editors: Juan Leon, Qiang Luo Source: https://www.frontiersin.org/research-topics/87009/g-dynamics-decoding-and-yielding-neural-ai-models-for-integrative-computational-simulation-of-brain-activity