Transcriptional response with wide dynamical range. Is it enough to only know the TFs? No, need to know the answer times etc. Cannot ask the genes, yield only expression at one point. - Roland Krause

How to come to the dynamical response? Fitting to a model. Early models by Chechik et al. JCB, 2009 had 7 parameters (peak height, onset and offset rates etc), can be learned by conjugate gradient descent. - Roland Krause

Too few time points lead to overfitting. - Roland Krause

We can cluster time course data as not every gene has its own dynamical prototype. Can work on the clusters for parameter learning. Traditional distance measures don't work well. - Roland Krause

Use the clusters to extract dynamical prototypes to get model parameters, then use the model for cluster. Start with naive k-means, iterate until satisfaction. - Roland Krause

Fitting to priors (prototypes), generate model and paramters. - Roland Krause

Validation by synthetic data, add noise to expression data. Compare to k-means using euclidian and correlation distance. Use mutual information for assessment, works comparable. - Roland Krause

Model avoids overfitting, median fit error compared to Chechik et al. - Roland Krause

Case study: response to LPS and PolyIC in mice, Amit et al , Science 2009, 24h study. Chose 1000 genes repressed or induced in the study, used both stimuli. - Roland Krause

Several clusters, corresponding to early response, stimuli, motifs (every clusters show at least one motif overrepresented) - Roland Krause

Sets are characterized by parameters. - Roland Krause

A general approach for analyzing time course data. # nice method and talk - Roland Krause

Q: Do you think number of clusters change? A: No, is fixed from beginning.Can use iterative approach if clusters are very similar. - Roland Krause

Q/A: Simulated data, number of clusters was fixed. - Roland Krause

Q/A: Dependence on starting conditions, use different starting positions. - Roland Krause