Combining Hidden Markov Models and Stochastic-Context Free Grammars

Currently, gene finding and RNA secondary structure prediction are performed separately. This is only valid if, for any DNA sequence the location of genes and corresponding RNA secondary structure can assumed to be independent. I develop methods that take a hidden Markov model for gene finding and a stochastic context-free grammar for RNA secondary structure prediction and from these, construct a combined model which can make joint predictions. I show that in general, the combined model is a stochastic context-free grammar and compare how an example performs relative to the separate models when data is simulated from the combined model. An investigation of the combined model reveals that apparent dependence can be an artefact of the model rather than dependence present in the data. My results are inconclusive and further work is required to determine whether future implementation of a the joint model is appropriate and feasible on DNA sequence data.