In systems biology, we aim at deriving gene-regulatory or signaling models based on multivariate readouts, thereby generating predictions for novel experiments. However any model only approximates reality, leaving out details or other types of regulation. Here I ask why a given model fails to predict a set of observations with acceptable accuracy and how to refine the model using this experimental knowledge. This resembles a question from signal processing, namely the blind identification of hidden (latent) variables in a mixing model. Many, powerful methods have been proposed to answer it. However, they have not been extended to dynamical systems due to the involved strong nonlinearities.I propose to infer additional upstream species in a given model, denoted as latent causes, that improve the prediction and at the same time are subject to the model dynamics. Multiple causes are estimated using statistical assumptions such as minimum mutual information. The model estimation will be performed within a Bayesian framework. This will allow for the efficient but crucial inclusion of prior biological information. The method will be applied to infer a differentiation model describing lineage segregation of embryonic stem (ES) cells to endo- and mesoderm. Here, latent causes are known to be transcription factors and microRNAs, but also small molecules/drugs. Identified off-target effects of these causes will be validated in collaboration with experimental partners.This study will establish links between information-theoretic signal processing and dynamical systems. Its application to a detailed ES cell model will foster our understanding of differentiation and may ultimately contribute to the development of more efficient differentiation protocols for cell replacement therapy.