The study published in Nature Communications in 2026 used whole-exome sequencing data from a cohort comprising 6,135 cases and 6,245 controls to predict schizophrenia risk. The researchers evaluated end-to-end genome interpretation neural network models for genetic risk prediction. The findings showed that, when sufficient sample sizes were available, nonlinear neural networks performed significantly better than traditional additive models.
The study indicated that higher-order genetic interactions between schizophrenia-associated alleles and variants should be taken into account in clinical and quantitative genetic models. Explainable artificial intelligence techniques were combined with biological prior knowledge to examine the models’ decision-making processes. While reidentifying previously known schizophrenia risk genes, this approach highlighted BASP1, which plays a role in neuronal development, as a possible and insufficiently studied schizophrenia gene.
Some sections of the paper were adapted from N. Verplaetse’s publicly available doctoral dissertation. The study was conducted by researchers at KU Leuven and Université de Montpellier-CNRS UMR 5535 and was supported by FWO, CNRS Chaire de Professeur Junior, Research Council KU Leuven, CELSA, Flemish Government and AI Research Program. The authors reported no conflicts of interest. The paper was published open access under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 license.
Why it matters
The findings show that approaches based solely on the sum of individual variants may be insufficient to capture biological interactions in large datasets when assessing genetic risk for schizophrenia. This raises questions about the extent to which models used in clinical research account for nonlinear relationships and whether their decisions are biologically interpretable. Combining explainable AI with biological information provides a methodological framework to ensure that model outputs do not remain merely statistical results. The prominence of BASP1 as a candidate gene suggests that gene networks associated with schizophrenia could be expanded; however, the significance of this finding for risk prediction and neuronal development remains to be clarified by subsequent research.