We are developing a multivariate survival analysis model for age dependent disease risk prediction based on a patient’s proteome and genome. Our method takes both the influence of a patient’s proteome and other diseases that a patient might already have into account and is able to model comorbidities.
We introduce a multivariate accelerated failure time model. Diseases that are caused by some slow, progressive, degradative process that ultimately leads to either catastrophic failure or to gradual manifestation can be well described by this model. Most age related diseases, such as heart attack, stroke or neurodegenerative diseases are of this type.
Our model describes the influence of proteome levels as fixed and the influence of diseases on each other as random effects. We use variational inference to train the model.
The way we model the influence of the proteome level on the disease risk of a patient allows us to decompose it into the influence of a patient’s protein levels on ageing processes and the influence of these ageing processes on the development on certain diseases, i.e. we can find different factors of ageing. We will use a group lasso penalty to identify a small set of predictive proteins for ageing.
Our main goals of this project are the discovery of these factors of ageing and their influence on various age-related diseases. We are planning to investigate possible disease mechanisms linking proteins most predictive for the factors of aging to the disease risks and to analyze why they might be causal or correlated with disease risk. Furthermore, we want to use the accelerated failure time model to develop an easily understandable and interpretable risk score. We hope to demonstrate that our risk score outperforms competing methods such as the proteomic ageing clock developed by Argentieri et al. (Nature Medicine 2024). Lastly, our model allows us to analyze the comorbidity structure of various diseases.
