Schizophrenia Affects 23 Million People. AI Just Found 766 Genes That Could Help Explain It

Schizophrenia genetics stopped being a list and became a network this month, and the shift looks durable rather than a one-off study. A Nature Genetics analysis of genetic data from more than 102,000 people, paired with brain tissue samples from six regions, links the disorder to 766 genes, 641 of which no earlier study of gene activity in the brain had flagged. The count matters, and so does the change in how those numbers are read.

For decades, psychiatric genetics hunted for individual risk variants. Each one explained a sliver of risk, and the catalog grew without ever resolving into a mechanism. The new study, led by researchers at the Lieber Institute for Brain Development and collaborators including the University of Bari, reframes the problem. Risk is not carried by scattered letters in the genome. It is carried by coordinated sets of genes turning on and off in patterns that span long stretches of DNA, sometimes far from the genes they ultimately affect.

To see those patterns, the team trained AI-based computational models on gene expression data from hundreds of brain donors across six regions. The models learned which genes behave as a unit, then used those units to predict which genetic variants would alter expression elsewhere in the network. A variant that looks harmless on its own can, in this view, behave as a remote control for a gene that actually does the work. Schizophrenia risk emerges from the network, not from any single switch.

This is why the 641 new genes matter. They were not invisible because earlier studies lacked statistical power. They were invisible because earlier methods asked a different question. Standard genome-wide scans look for variants that change protein coding or sit next to a single gene. Co-expression models, by contrast, ask which variants shift the activity of gene sets that act in concert, including variants that sit hundreds of thousands of base pairs from the genes they regulate. The new paper is the largest coordinated map of those long-range signals in schizophrenia to date, and the trade press corroborated the headline number within days of publication.

The WHO puts schizophrenia's global prevalence at about 23 million people, roughly one in every 345. That scale is the reason the mechanistic shift matters. A list of 766 candidate genes, even a long one, is not a treatment. A network of 766 genes, with co-expression structure and long-range wiring mapped, is a triage list: it tells mechanistic researchers which gene sets to perturb in the lab, and which drug targets sit on the same regulatory rails.

The caveats are real, and the paper itself makes them. Family history is a risk factor, not a sentence; the new gene list does not change clinical risk for any individual. Specialists still do not fully understand how the variants interact, and the AI co-expression model could carry artifacts of its own training data. Replication in independent cohorts, and orthogonal validation in cell and animal models, will determine which of the 641 survive into the working map. A preprint version of the analysis circulated in February 2026, and the published Nature Genetics paper should be treated as the canonical version for any downstream use.

What the field has now is a working model. The same AI pipeline that produced this map can be re-run on other psychiatric disorders, on larger cohorts, and on finer brain-region atlases. Each pass will narrow the list and test the network structure, either reinforcing or pulling apart the long-range regulatory picture. The model is the durable part. The 766-gene count is the first snapshot it has produced.