To stratify risk, uncover the causal biology, and predict consequences of low-dose radiation, we need models designed for forward prediction and reverse engineering of molecular and cellular phenotypes. Physiological consequences of low-dose radiation emerge from a dynamic interplay between regulatory events across the biological hierarchy.  Rich -omic datasets offer an opportunity to integrate information across length-scales from epigenome to proteome in data-driven predictive models. However, such models don’t readily offer the interpretability or extrapolation power of knowledge-based mechanistic models. Our ability to define mechanistic models, on the other hand, is limited by gaps in biological knowledge, parameterization challenges, and the computational demands of numerical simulation. We propose to combine the two approaches: informing knowledge-based models via AI/ML predictions while constraining predictive models with known “rules” of biological signaling and relationships between molecular and cellular phenomena.

Leveraging the coherent set of epigenetic, transcriptomic, proteomic data acquired by our multi-institutional team, we hypothesize that the integration of AI/ML models with stochastic, knowledge-based models will allow us to identify and predict the probabilistic, longitudinal occurrence of multiscale signaling patterns that confer risk of – or protection against – the negative physiological consequences of low-dose radiation.