We recently released rHVDM (Hidden Variable Dynamic Modelling), an R/Bioconductor package that predicts targets of a known transcription factor using time course microarray data. The key feature behind the algorithm is a simple ODE model of mRNA concentration. In the first stage of rHVDM, transcription factor activity (the hidden variable) is deduced from the expression time profile of a small number of known targets. This information is then used to screen other genes for dependency on that transcription factor. The accuracy of the technique has been demonstrated with Affymetrix microarray time course data and verified experimentally using siRNA knockdown of a targeted transcription factor (p53).

While implementing the rHVDM algorithm and refining it for release we encountered a number of problems. These included parameter identifiability, parameter count reduction, algorithmic speed, parameter domain restriction, confidence interval estimation, and measurement noise. I will discuss each of these issues individually, along with the techniques we used to address them.