Semi-Supervised Pre-trained Neural Networks for Schizophrenia Subtyping

Dates: Jan 2026 – Dec 2027

Funder: MSCA-CZ (No. 250507)

Role: Lead researcher

Status: Active

---

Objectives

This MSCA Postdoctoral Fellowship project develops a pre-trained neural network model using healthy brain data (Human Connectome Project, N=1200) to learn a normative template of brain function. The model uses self-supervised pre-training tasks (generative, predictive, contrastive) on resting-state fMRI functional connectivity graphs. When applied to schizophrenia patients, the embeddings quantify deviations from the healthy template, enabling discovery of neurobiologically meaningful subtypes.

Methodology

• Graph neural networks and transformers on ROI-level fMRI data (Craddock atlas)
• Self-supervised pre-training tasks:
  • Generative (masked autoencoders)
  • Predictive (forecasting fMRI signals)
  • Contrastive (brain “fingerprinting”)
• Meta-learning to combine multiple pre-training tasks for a more general model
• Subtype validation via cross-site generalization and consensus clustering
• Interpretation through associations with symptoms, clinical variables, and structural MRI

Team

• Supervisor: Dr Jaroslav Hlinka (Complex Networks & Brain Dynamics Group, ICS CAS)
• Collaborators: National Institute of Mental Health (NIMH) Klecany

Outputs

• Open-source models (GitHub/OSF)
• 2+ first-author manuscripts
• Dissemination via ENIGMA consortium
• Conference presentations (OHBM, EMBC)