Section: New Results

RNA structures

RNA secondary structures: folding, design and evolution

In a collaboration with J. Waldispuhl (McGill, Canada) (Presented at the Recomb'11 conference [32] ), we used weighted grammatical models, introduced by members of the group [2] , to perform an efficient exploration of the mutational landscape of RNA. We proposed an adaptive sampling algorithm, where weights were used to compensate an identified bias toward regions of higher GC-content within sampled sequences, thereby allowing for the exploration of more relevant portions of the evolutionary landscape. These adaptive sampling principles can be adapted into a method for the RNA design following similar principles. This constitutes a competitive alternative to local search strategies used by all existing tools for this problem. This work is ongoing as a collaboration with B. Berger group (Mit ) and J. Waldispuhl (McGill, Canada), and a manuscript was recently submitted.

RNA knowledge-based potentials and 3D studies

We used the curated database of biologically interesting structures we have set up to perform a statistical analysis and developed knowledge-based potentials. The database server is available at http://csb.stanford.edu/rna .

We obtained RNA knowledge-based potentials that now performs well at different representation levels. They can be used in three well-known Molecular Dynamics (MD) and modeling software suites Encad [44] , Gromacs (v3 and 4) [45] and Mosaics [41] and are available for the community. The study we performed on a large number of new decoys showed that our potential outperforms Rosetta RNA scoring function [37] which is the gold standard. We show that not having correction terms for base-stacking and pairing can be of advantage when modelling loops at high resolution. The study was welcomed by the RNA community and published in RNA Journal (IF 6.5) [8] .

We also refined the mixture model strategy we developed for building knowledge-based potentials. In collaboration with O.Schwander at Lix , we compared different mixture models: Dirichlet Process Mixture models (DPM), Kernel Density Estimation (KDE) models, Expectation Maximization models (MM) with different number of components (including a simplified version based on a post-processing step using K-Means). We showed that the Dirichlet Process Mixtures (DPM) is a good tradeoff despite its longer precomputation time as it provides a smooth potential having relatively few components. This study was presented at the Mcmmb'11 conference and was submitted as a journal paper.

This work was done in collaboration with A. Sim, X. Huang an M. Levitt (Stanford University - Gnapi Associate team).