The analysis of gene expression time-courses poses many methodological challenges due to the complexity of the regulatory interactions producing the observables, their inherent variability and the levels of noise present even under optimal experimental conditions. We propose to use a mixture of Hidden Markov Models (HMMs) for finding groups in gene expression time course datasets. Mixtures have a number of preferable properties. In contrast to clustering, genes do not have to be partitioned into groups, rather the frequent ambiguity of function or regulation can be accounted for. They also exhibit a higher degree of robustness with respect to noise and afford a simple, entropy-based diagnostic for the level of ambiguity in assignments to groups given the mixture. The HMMs used prove effective in modeling the qualitative behavior of time-course data, as they reflect the time-dependencies of measurements and allow to capture subtle signals. Lags and the varying speeds at which the regulatory programs are executed can be handled; simple variants can be used to deal with further questions of biological interest. We will present the mixture framework and the estimation method used. Partially supervised learning makes use of prior information with respect to membership of genes in the same or distinct groups. A heuristic for proposing an initial collection of HMMs and modifying HMMs during the training is discussed, as well as an algorithm to partition an inferred group into synchronous subgroups. The performance of our approach is evaluated on artificial and biological expression data. We conclude with computational aspects of our implementation of the method. This is joint work with Alexander Schoenhuth and Christine Steinhoff