Principal Investigator: Mr Bernard Stopak
Ada Health, Berlin, GermanyTags: 34802, artificial intelligence, complex disease, modelling, prediction, risk factors
There is a currently a large focus in patient-facing genomics to accurately identify individuals who are at risk for heritable or partially heritable diseases. To-date, there have been advances in accounting for this heritable risk, but gaps in performance remain as there remains a portion of unexplained inheritance. Our goal is to produce genetic risk prediction models that can accurately assess whether an individual is at an increased risk for disease. We believe that complex phenotypes have differing underlying genetic architecture. Therefore, it may be that different modelling methods will work better for different conditions. For example, for a condition with several high-effect variants, such as age-related macular degeneration, a simple genetic model may perform well with few genetic variants. However, for a condition like Crohn’s disease where hundreds of genetic loci have been identified, more complex models with more genetic variants are likely to be more powerful. We will create and test our genetic predisposition models for their validity to predict individuals at-risk of genetically correlated diseases using polygenic risk scores, machine learning, and hand-curated genetic models. These models will then be compared to see which kind of modelling best reflects the genetic architecture of each condition, then see if there are any patterns amongst different conditions.
Do different modeling types work best for different diseases?
It is unclear whether there is a single genetic-modeling framework that works best across many complex conditions. We believe it is unlikely that there is a single modeling method that works best because of differing genetic architecture amongst traits. We will compare different genetic modeling methods, such as machine learning, polygenic risk scores, and manually curated models of known pathogenic variants to see which work better for predicting the SNP variance for different conditions. The type of best performing modeling method may also give insights into the genetic architecture underlying these traits. Since we are interested in investigating the ubiquity of potential genetic models, we want to sample conditions from across the condition-space. We will include common conditions that have a widespread health impact, such as coronary artery disease, type 2 diabetes, asthma, allergic rhinitis, and depression, while also sampling conditions with both high and low heritability, including bipolar disease, schizophrenia, Crohn’s disease, age-related macular degeneration, and a number of cancers. The final production and refinement of models will depend upon how many cases are available for the respective conditions.
Cardiovascular disease is still the number one cause of death worldwide according to the WHO. Currently, genetic risks and heart arrhythmias are analyzed separately. Combining them both for cardiovascular disease detection could lead to a more holistic and sophisticated patient risk profile. Identifying risk factors early on enables a shift towards proactive healthcare with meaningful widespread health impact. With the rise of artificial intelligence and high-performance computing, we aim to explore how combining genetic cardiovascular risk factors with other factors helps to predict more serious conditions in order to prevent them. We want to benchmark several different approaches against each other for cardiovascular risk profiling.
Last updated Jul 22, 2019