Mechanistic modelling for mind, brain, and intelligent systems.

We use task and resting-state M/EEG with biophysical generative models (DCM; Variational Laplace with thermodynamic integration) to infer synaptic and circuit mechanisms across disorders including schizophrenia, depression, catamenial epilepsy, and neurodevelopmental CNVs (e.g., 22q11.2DS).

• Thalamo‑cortical conductance modelling for psychosis‑risk CNVs (22q11.2DS)
• Receptor‑informed network models for antidepressant response in depression
• Visual LTP paradigms and circuit modelling in catamenial epilepsy
• Hierarchical/group VL and empirical Bayes for cohort‑level inference

Selected manuscripts:
• Catamenial epilepsy: perimenstrual visual LTP (bioRxiv, 2025)
• 22q11.2DS sleep & thalamo‑cortical pathology (medRxiv, 2025)
• Restoring synaptic balance in schizophrenia (Schizophrenia Bulletin, in press)
• Ketamine response: FPN connectivity & GABA_A (Translational Psychiatry, 2024)
• Microdose LSD & visual LTP captured by TCM (BMC Neurosci, 2024)
• GABAergic modulation of beta & motor adaptation (Alzheimer’s & Dementia, 2025)
• Impaired GABAergic inhibition in schizophrenia (Schizophrenia Bulletin, 2020)

The Free Energy Principle and Active Inference originated in neuroscience as a unifying account of perception, learning, and action. We view them as the foundation for next‑generation AI: truly agentic systems that maintain beliefs about the world and themselves, minimise expected free energy, and act adaptively.

Our work spans proof‑of‑concept agents, from predictive Pong and gridworlds to autonomous drones, culminating in an 11‑node “general generative model” that demonstrates reusable active‑inference components. These agents illustrate planning, action selection, and learning with the same principles we deploy in computational psychiatry.

• Thermodynamic Variational Laplace for online inference in agents
• Seven‑component reusable architecture with precision modulation and message passing
• Pong, gridworld, and drone demos with multi‑step policy rollout and belief visualisation

Preprints:
• Toward a Reusable Architecture for Intelligent Agents (2025)
• A Neuro‑Inspired Computational Framework for AGI: Predictive Coding, Active Inference, and Free Energy (2025)

Code: VariationalLaplace (Julia) · aLogLikeFit (MATLAB) · GitHub

We develop inference methods that make biophysical modelling fast, robust, and scalable. Our thermodynamic Variational Laplace (Thermo-VL) augments classical VL with annealed likelihoods, low-rank curvature, and smarter noise updates for reliable fits to nonlinear dynamical systems (e.g., neural mass models, agents).

• Adaptive low-rank covariances factorise the Hessian into a rank-k subspace (V D V^T) updated each iteration, capturing dominant curvature while keeping inference tractable.
• Heteroscedastic noise update observation variance per feature/epoch to prevent over-confidence and absorb structured residuals (e.g., frequency-dependent M/EEG noise).
• Thermodynamic integration (TI) anneal the likelihood with β∈[0,1] to stabilise optimisation and enable path-sampling estimates of the log evidence.
• Solenoidal mixing optional skew-symmetric term in the natural-gradient preconditioner to align updates with posterior geometry.
• Laplace-domain transfer functions linearise dynamics with delays to compute cross-spectra efficiently for DCM-style fits.
• Mixture-of-Gaussians features represent data/model features with Gaussian bases for flexible spectral fits and denoising.

How fitVariationalLaplaceThermo works (brief)

Core loop: refine mean m, structured covariance V,D, and noise, while annealing β and tracking ELBO.

// Pseudocode (see MATLAB implementation)
init m=m0, S0 → low-rank V,D; set β schedule; set σ²
repeat until convergence:
  // Tempered objective
  Lβ(m) = β·log p(y|m,σ²) + log p(m)
  // Curvature & low-rank update
  H ≈ ∇²Lβ(m); [U,S]=svd(H); V=U(:,1:k)·sqrt(S₁:k)
  D = diag(diag(H) - rownorm(V)^2)
  // Mean update (preconditioned / natural gradient)
  m ← m + step · (V,V,D)⁻¹ ∇Lβ(m)
  // Noise & heteroscedastic variance
  σ² ← update_per_feature(residuals)
  // Anneal β and compute ELBO / g_ELBO for line-search
until |ΔELBO| < tol

MATLAB source: fitVariationalLaplaceThermo.m · Overviews: Comp-Neuro · Neuro-AI

We investigate how sleep disruption in neurodevelopmental CNVs (e.g., 22q11.2DS) impacts cognition and psychosis risk. We combine overnight EEG with mechanistic thalamo-cortical modelling to derive circuit-level biomarkers and candidate intervention targets.

• Spindle and slow-wave phenotyping linked to thalamo-cortical mechanisms
• Cross-species assays and translational pipelines
• Partnerships with clinics, patient groups, and methodologists

Lead researcher: Lio Berndt. Collaborators include Prof Matt Jones, Prof Jeremy Hall, Prof Marianne Van den Bree, Prof Chris Jarrold, and others.

Preprint: Sleep as a window into thalamo-cortical pathology in 22q11.2DS (medRxiv, 2025)

Funding: Wellcome — Sleep Detectives: sleep stratification in young people

Ketamine is a rapid-acting antidepressant, with downstream effects from NMDA receptor modulation that reshape cortical network dynamics within hours. We model fronto-parietal circuits (FPN) as treatment targets, integrating receptor-informed priors with M/EEG connectivity and biophysical models to explain symptom improvement and individual heterogeneity in response.

• Receptor-informed DCM/TCM of ketamine effects on oscillations and connectivity
• Fronto-parietal network mechanisms underlying antidepressant response
• Biomarker discovery for patient stratification and treatment optimisation

We collaborate widely, including with Dr Rachael Sumner (University of Auckland), Prof Celia Morgan (Exeter), and Joy Krecke (Exeter), among others.

Key papers:
• Ketamine for depression: M/EEG connectivity and GABA_A changes (Translational Psychiatry, 2024)
• Generative modelling of thalamo-cortical mechanisms underlying ketamine effects on oscillations (NeuroImage, 2020)

Clinical study: BAM study (Awakn Life Sciences)

We develop EEG/MEG-based in-silico assays of GABAergic function to characterise cortical dynamics in health and disease, and to compare the neuronal and behavioural effects of GABA-modulating compounds and functional drinks.

• Drug studies: tiagabine (GABA uptake inhibitor), alcohol, propofol, various benzodiazepines and more — analysed with receptor-informed spectral/connectivity models.
• Biophysical modelling: DCM/TCM fits of oscillations and connectivity; parameter-to-phenotype mapping for circuit-level interpretation.
• Assay design: task and resting paradigms optimised for GABA_A sensitivity (beta/gamma power, connectivity, and visual LTP interactions).

Selected papers: Tiagabine matches PET-derived GABA_A receptor maps (ENP, 2021) · Propofol vs dexmedetomidine oscillations (NeuroImage, 2021) · GABA-ergic dynamics in frontotemporal networks (J. Neurosci., 2020)

We have teamed up with Prof David Nutt and GABA Labs to investigate Sentia and related GABA-ergic functional drinks, benchmarking their neural and behavioural signatures against pharmacological GABA agents.

Links: GABA Labs team · PhD project: Neurophysiology of GABAergic modulation