Combining machine-learning approaches with mechanistic models to integrate expertise

Partners involved: Theis*, Hammer, Mukherjee

To close the gap between black box models based on ML and mechanistic modeling approaches, we will

(i) develop regression and classification models regularized by mechanistic models and

(ii) statistically embed small-scale mechanistic submodels into large-scale models based on ML.

Both strategies allow us to integrate prior knowledge, whereby intelligent regularizations can be achieved even for pure prediction problems in the case of limited sample sizes - an important challenge in the field of high-dimensional models of biological data. In many cases, the limitation of sample sizes is a matter of principle. We will concentrate on mechanistic models described by dynamic systems, especially on systems of parameterized ordinary differential equations. These differential equations are special cases of the graph structures from AP2 and have already been successfully used in systems biology to model the behavior of systems from protein signals to gene dynamics and cell movements. We will focus on transcriptomics and the related questions of gene regulation.