MMKnots: A max-margin model for RNA secondary structure prediction including pseudoknots

Motivation: The ideal algorithm for the prediction of pseudoknotted RNA secondary structures will provide fast and accurate predictions for pseudoknots of arbitrary complexity. However, existing algorithms are typically lacking on one of these three axes. Energy-based methods suffer from the intractability of pseudoknotted structure prediction under realistic energy models, while statistical approaches struggle with inference on the vast space of possible structures. Results: In this paper, we present MMKnots, an algorithm for secondary structure prediction including pseudoknots. MMKnots leverages a max-margin framework to train a model with a simple scoring scheme, which then enables the use of efficient (approximate) algorithms at prediction time. Experiments on datasets not observed in training show that MMKnots outperforms the state-of-the-art pseudoknot prediction algorithm without sacrificing speed. Furthermore, whereas existing algorithms tend to be conservative in predicting pseudoknots, MMKnots can include these complex interactions more easily due to the flexibility of the model representation. MMKnots illustrates the potential of this framework in excelling on the three major criteria of accuracy, generality and speed. Availability: MMKnots will be made available at http://ai. stanford.edu/ ̃csfoo/mmknots upon publication. Contact: [email protected]