Researchers have developed a deep learning algorithm designed to predict the effects of rare genetic variants. This technique enables more accurate identification of individuals at high risk for diseases and aids in pinpointing genes involved in disease development.
Our susceptibility to specific diseases is largely determined by the numerous variants in our genome. However, understanding the impact of rare genetic variants, which appear infrequently in the population, has been challenging. A research team from the German Cancer Research Center (DKFZ), the European Molecular Biology Laboratory (EMBL), and the Technical University of Munich has created a deep learning-focused algorithm that can forecast the implications of these rare genetic variants. The approach enhances the ability to differentiate individuals at high risk for diseases while aiding in the identification of disease-related genes.
Each person’s genome contains millions of unique differences called variants, which can be linked to specific biological characteristics and diseases. These connections are typically explored through genome-wide association studies.
However, rare variants that occur at a frequency of 0.1% or lower are frequently disregarded in these studies. “Rare variants can have a disproportionately large effect on biological traits or diseases,” states Brian Clarke, one of the initial authors of this study. “They may help us uncover the genes responsible for disease development, leading us to new treatment possibilities,” adds co-first author Eva Holtkamp.
To enhance the prediction of rare variant effects, the teams led by Oliver Stegle and Brian Clarke at the DKFZ and EMBL, along with Julien Gagneur at the Technical University of Munich, have devised a risk assessment tool powered by machine learning. Dubbed “DeepRVAT” (rare variant association testing), this method is groundbreaking as it employs artificial intelligence (AI) in examining genetic associations to interpret rare variants.
The model was initially trained using exome sequence data from 161,000 individuals within the UK Biobank. Additionally, it included information about the genetically influenced biological traits of these individuals, as well as the genes linked to those traits. In total, the training comprised approximately 13 million variants, each accompanied by detailed “annotations” that provided quantitative insights into their potential effects on cellular functions or protein structures; these annotations were integral to the training process.
Once trained, DeepRVAT can predict which genes are likely affected by rare variants for each individual. The algorithm evaluates these individual variants alongside their annotations to generate a numerical value that reflects the degree to which a gene’s function is compromised and its possible health consequences.
The researchers validated DeepRVAT using genomic data from the UK Biobank. It identified 352 associations with genes involved in 34 tested traits, which are relevant blood test results linked to diseases. This performance significantly surpassed all previously established models. The results produced by DeepRVAT displayed remarkable robustness and were better replicable with independent datasets compared to other methods.
A notable application of DeepRVAT is in assessing genetic predisposition to various diseases. The researchers merged DeepRVAT with polygenic risk scoring that considers more common genetic variants, greatly enhancing prediction accuracy, particularly for high-risk variants. Moreover, DeepRVAT uncovered genetic correlations for numerous diseases—including several cardiovascular conditions, types of cancer, and metabolic and neurological disorders—that had remained undetected by current testing methods.
“DeepRVAT has the potential to significantly propel personalized medicine forward. Our method is versatile and can be effectively combined with other testing approaches,” explains physicist and data scientist Oliver Stegle. His team aims to further evaluate the risk assessment tool in extensive trials and expedite its implementation. The scientists are currently in discussions with the organizers of INFORM, which aims to identify personalized treatments using genetic data for children with relapsed cancer. DeepRVAT could clarify the genetic factors contributing to specific childhood cancers.
“I’m particularly excited about the potential impact of DeepRVAT on rare disease research. A significant challenge in this field is the scarcity of expansive, systematic data. By harnessing AI and the half a million exomes from the UK Biobank, we’ve identified which genetic variants have the most substantial impact on gene functions,” remarks Julien Gagneur of the Technical University of Munich.
Looking ahead, the integration of DeepRVAT into the German Human Genome Phenome Archive (GHGA) infrastructure is planned to promote its application in diagnostics and fundamental research. Another benefit of DeepRVAT is its relatively low computing power requirement compared to similar models. DeepRVAT is available as a user-friendly software package, allowing researchers to utilize pre-trained models or adapt the tool using their own datasets for specific objectives.
