Biomedical studies increasingly incorporate longitudinal data, enabling us to track individual disease processes over time at the molecular level, and to discover associations of the molecular profiles with the outcome of interest, such as the onset of a disease. Despite the potential of statistical methods that jointly model longitudinal and time-to-event data, they have not yet been widely adopted in high-throughput omics studies. Therefore, we evaluated multiple approaches for joint modeling of longitudinal and time-to-event data, and we introduce a joint modeling strategy for longitudinal proteomics studies. The focus is on assessing the utility of the methods in predicting the dynamic disease risk of an individual from longitudinal proteome profiles. To benchmark the methods, we used a range of simulated datasets that reflected real proteome profiles with varying complexities. Our results clearly demonstrated the advantages of the longitudinal methods over conventional Cox proportional hazards models with single time point studies. This was further supported by re-analysis of data from a proteomics study of early type 1 diabetes prediction, where we discovered new early candidate proteins associated with the disease onset that were not detected in the original study.
Lindén et al. (Mon,) studied this question.