Research

What we work on.

Biological subtypes of psychiatric illness, personalized treatment, machine learning for neural and biomedical data, and the neurotechnology that produces those data.

Biological subtyping & biomarkers

Psychiatric diagnoses group together patients whose brains and treatment responses differ enormously, which is part of why finding an effective therapy is so often trial and error. We develop statistical and machine-learning methods that define data-driven biological subtypes from neuroimaging, multi-omics, and clinical data, and that predict which patients will respond to which treatment, from ketamine and TMS in depression to molecularly grounded subtypes of autism.

Personalized treatment & generative digital twins

Choosing a treatment means reasoning about a counterfactual: what would happen to this patient under that therapy. We build generative digital twins and counterfactual models (including diffusion world models and reinforcement-learning methods for dynamic treatment regimes) that estimate the treatment most likely to help an individual, and that make those predictions interpretable enough to trust in the clinic.

Machine learning for neural & biomedical data

Brain and biomedical data are noisy, high-dimensional, and irregularly structured. We develop new machine-learning methods built for them: graph neural networks that learn over the structure of brain networks, robust and reproducible estimators of connectivity, and representation-learning tools for comparing neural dynamics across animals and patients. Methods are released as open-source software.

Neuroengineering & neurostimulation

Better models need better measurements. We build and apply neurotechnology to read and perturb brain activity: TMS combined with high-density EEG in patients undergoing accelerated stimulation for treatment-resistant depression, and closed-loop approaches that turn those signals into individualized, adaptive interventions.