A schizophrenia genes AI study published in Nature Genetics has identified 766 genes associated with the condition by modeling both nearby and more distant genetic influences on brain gene expression. The result broadens the field’s list of associations. It does not establish that any one of those genes causes schizophrenia, diagnose an individual, or deliver a treatment.
The study addresses a stubborn genetics problem: many variants linked to complex conditions sit outside protein-coding parts of the genome and are thought to affect risk through changes in gene expression. Standard transcriptome-wide association studies have largely modeled local, or cis, effects. In the paper’s framework, those are effects within roughly 1 megabase of a gene.
How did the schizophrenia genes AI study work?
The researchers built two computational models, called INGENE and MODULE, using RNA-sequencing data from six post-mortem human brain regions. The models included candidate trans-acting variants within gene co-expression networks, meaning they accounted for patterns in which genes show coordinated expression rather than relying only on local genetic predictors.
WIRED described the approach as AI-based computational modeling of coordinated activity among thousands of genes. The primary paper describes elastic-net models and network analysis, not a generative AI system. That distinction is not cosmetic: this was statistical genomics, not a chatbot finding a cure.
When the team combined the new trans-informed models with conventional cis-based predictors, gene-expression imputation improved for 18,744 genes across the brain regions, according to the paper. They then applied the combined framework to Psychiatric Genomics Consortium wave 3 genotypes, testing associations between genetically imputed expression and schizophrenia.
At a false-discovery-rate threshold below 0.01, the analysis identified 766 schizophrenia-associated genes. Of those, 641 had not been reported in previous transcriptome-wide analyses. The authors said the findings point to contributions from distal regulatory mechanisms and interactions among gene networks in schizophrenia risk.
What does the 766-gene finding mean for patients?
For now, it is an association-focused research result. Schizophrenia has substantial heritability in twin studies, the paper notes, but the proportion of heritability explained by known findings remains limited, and translating statistical links into molecular mechanisms is still difficult.
The study also has an obvious evidence boundary: its expression data came from post-mortem brain tissue. The reported material does not include clinical validation, patient-level risk-prediction results, or therapeutic testing. A larger gene list can guide further scientific scrutiny, but it is not a clinical test or a personalized-treatment menu.
Separate work illustrates another use of machine learning in psychiatric genomics. A Stanford-led study published in Cell in 2024 used an AI-based system called ARC-SV to detect complex structural variants in whole-genome sequencing data. Stanford Medicine reported that those variants tended to be near or overlap genome-wide association locations for schizophrenia or bipolar disorder and affected expression of nearby genes, suggesting they could contribute to disease. That was a different study, using a different kind of genetic variation, and it likewise stopped short of proving causation.
This story draws on original reporting from WIRED.
