Learning to design RNA polymers with graph kernels

RNA polymers have many functions in cells and are relatively easy to create. Finding ways to construct RNA sequences that correspond to a desired function is thus desirable. Conventionally inverse folding is used to tackle this problem. Inverse folding however is dependent on user provided very strict constraints. In this work I present an approach that will learn those constraints before generating new RNA molecules. [1] describes a general purpose approach to sample distributions of graphs. Metropolis Hastings sampling is used with a graph grammar to propose graphs and a graph kernel to aid the evaluation of proposed graphs. Since the described grammar is locally context-sensitive, long range constraints can not be represented in the grammar. If only local constraints are considered when proposing a graph, ignoring long range dependencies will increase the rejection rate of the MH algorithm. This is especially true for RNA secondary structure graphs. The structures themselves impose constraints on the graphs that cant be represented in a locally context-sensitive grammar. In this work I present a method which is using graph minors (a minor is obtained by contracting adjacent nodes in a graph) to enhance the grammars expressiveness.