Using explainable AI to formulate candidate gene risk scores for AD

In complex diseases like Alzheimer’s disease (AD), the disease risk is conveyed by genetic variants each one with a small effect on disease susceptibility, which may add up to a relevant impact in combination. Such complex synergistic effects can be detected by means of machine learning (ML) models such as decision tree ensembles, which are excellent candidates for detecting genetic variant interactions even beyond only pairwise interactions in candidate gene approaches. Explainable artificial intelligence can then explain the models and deliver information about effects and importance of the single variants.
In a first application of this technique, we have formulated an aquaporin-4 (AQP4) single nucleotide polymorphisms (SNPs) based risk score for brain amyloid burden using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). We plan now to replicate and extend our method with a larger dataset and include further candidate genes, e.g. circadian, sleep related or blood-brain-barrier related genes.
The data from the NIAGADS data repository will expand the previously used database significantly. Of special interest are the data from the Alzheimer’s Disease Sequencing Project (ADSP). Besides a general increase in data, the racial/ethnic diversity of ADSP is also of high interest. Genetic variants might occur at different frequencies in different populations, so that rare combinations of variants in one population might be more frequent in another population. Access to these data will not only allow to replicate our previous findings, but also to extend our methodology to other candidate genes.
Information on SNP status, together with further information like age, gender, APOE4 status, diagnosis and endophenotypes of AD, as available, will be used to train and explain a ML model. Information gathered from the explanations will be used to assign importance as well as protecting or deleterious effects to the SNPs, eventually allowing to formulate a risk score for (endo)phenotypes of AD or diagnosis.

In diseases like Alzheimer’s, genetic variations each with a small effect may together affect the disease risk. These variations can add up to have a significant impact. We can use machine learning, like decision tree ensembles, and explainable AI to find these combined effects.
We've already used this method to create a risk score for brain amyloid burden in Alzheimer's. Now, we want to expand this work using a larger dataset and looking at more genes, e.g. genes related to sleep. We want to use a new dataset that includes more data and diverse genetic information. Different populations may have different genetic variations, so having more diverse data will help us to understand these differences.
We'll use this data, along with other information like age and gender, to train a machine learning model. By understanding how the model makes its decisions, we can figure out which variants of a gene are most important and how their combination affect Alzheimer's risk. This will help us to create gene-specific risk scores for Alzheimer's and related conditions.