Beyond traditional GWAS: New methods reveal overlooked genetic associations

Complex diseases such as schizophrenia are difficult to decode because they are influenced by thousands of genetic variants together with environmental factors. Although genome-wide association studies (GWAS) have identified many disease-associated variants over the past two decades, conventional approaches often rely on linear models that may miss more complex biological relationships.

Hidden biases in conventional genetic analyses

Research led by Burak Yelmen suggests that these simplifying assumptions can have important consequences. Using mathematical modelling and simulations based on Estonian Biobank data, the researchers showed that ignoring interactions between genetic variants can distort statistical results and, in some cases, produce false-positive associations.

The findings highlight an important limitation of standard GWAS analyses. While these methods are highly effective at detecting strong and straightforward signals, they may struggle to capture more complex genetic patterns. As genomic datasets continue to grow, accounting for gene-gene interactions may become increasingly important for accurately interpreting genetic associations.

Artificial intelligence reveals previously undetected genetic signals

To address these challenges, the researchers explored whether artificial neural networks could identify genetic patterns that conventional methods fail to detect. Unlike traditional linear models, neural networks can capture nonlinear relationships and interactions among genetic variants.

Using explainable AI approaches together with genome-wide genotype data, the team analysed schizophrenia data from the Estonian Biobank. They identified several potentially associated genomic regions that had not been detected using standard GWAS methods. Most of these regions overlapped with loci previously linked to schizophrenia, bipolar disorder and other brain-related traits, while genes within those regions were enriched for functions related to brain development and neurological processes.

Complementary approaches to the same challenge

Together, the two studies tackle the same problem from different directions. One reveals how conventional GWAS analyses can become unreliable when biologically relevant gene-gene interactions are overlooked. The other demonstrates how explainable AI may help uncover genetic signals that traditional approaches miss.

While schizophrenia served as the test case, the implications extend far beyond a single disorder. Many common diseases are shaped by numerous genetic and environmental factors, and methods capable of identifying subtle patterns could prove useful across a wide range of complex traits.

The findings do not mean that AI has solved the genetics of schizophrenia. The newly identified signals still require extensive validation and many modelling challenges remain for AI to be used more broadly. However, the research demonstrates how explainable AI, combined with the rich resources of the Estonian Biobank, can open new avenues for discovery. “With further improvements, these approaches may help build a more complete picture of the genetic architecture of complex diseases and ultimately support the development of better diagnostics and treatments,” explained Yelmen.

Read more:

NAR Genomics and Bioinformatics: https://academic.oup.com/nargab/article/8/1/lqag019/8494699?searchresult=1

Journal of Computational Biology https://journals.sagepub.com/doi/epub/10.1177/15578666261479749