Section: Research Program

RNA and protein structures

RNA

Participants : Julie Bernauer, Alain Denise, Rasmus Fonseca, Feng Lou, Yann Ponty, Mireille Régnier, Philippe Rinaudo, Jean-Marc Steyaert.

Common activity with P. Clote (Boston College and Digiteo).

From RNA structure to function

We are currently developing a combinatorial approach, based on random generation, to design small and structured RNAs. An application of such a methodology to the Gag-Pol HIV-1 frameshifting site will be carried out with our collaborators at Igm . We hope that, upon capturing the hybridization energy at the design stage, one will be able to gain control over the rate of frameshift and consequently fine-tune the expression of Gag/Pol. Our goal is to build these RNA sequences such that their hybridization with existing mRNAs will be favorable to independent folding, and will therefore affect the stability of some secondary structures involved in recoding events. Moreover it has been observed, mainly on bacteria, that some mRNA sequences may adopt alternate folds. Such events are called a conformational switch, or riboswitch. A common feature of recoding events and riboswitches is that some structural element on the mRNA initiates and unusual action of the ribosome, or allows for an alternate fold under some environmental conditions. One challenge is to predict genes that might be subject to riboswitches.

Beyond secondary structure

One of our major challenges is to go beyond secondary structure. Over the past decade, few attempts have been made to predict the 3D structure of RNA from sequence only. So far, few groups have taken this leap. Despite the promises shown by their preliminary results, these approaches currently suffer to a limiting scale due to either their high algorithmic complexity or their difficult automation. Using our expertise in algorithmics and modeling, we plan to design original methods, notably within the AMIS-ARN project (ANR BLANC 2008-2012) in collaboration with PRISM at Versailles University and E.Westhof's group at Strasbourg.

1. Ab initio modeling: Starting from the predicted RNA secondary structure, we aim to detect local structural motifs in it, giving local 3D conformations. We use the resulting partial structure as a flexible scaffold for a multi-scale reconstruction, notably using game theory. We believe the latter paradigm offers a more realistic view of biological processes than global optimization, used by our competitors, and constitutes a real originality of our project.

2. Comparative modeling: we investigate new algorithms for predicting 3D structures by a comparative approach. This involves comparing multiple RNA sequences and structures at a large scale, that is not possible with current algorithms. Successful methods must rely both on new graph algorithms and on biological expertise on sequence-structure relations in RNA molecules.

RNA 3D structure evaluation

The biological function of macromolecules such as proteins and nucleic acids relies on their dynamic structural nature and their ability to interact with many different partners. Their function is mainly determined by the structure those molecules adopt as protein and nucleic acids differ from polypeptides and polynucleotides by their spatial organization. This is specially challenging for RNA where structure flexibility is key.

To address those issues, one has to explore the biologically possible spatial configurations of a macromolecule. The two most common techniques currently used in computational structural biology are Molecular Dynamics (MD) and Monte Carlo techniques (MC). Those techniques require the evaluation of a potential or force-field, which for computational biology are often empirical. They mainly consist of a summation of bonded forces associated with chemical bonds, bond angles, and bond dihedrals, and non-bonded forces associated with van der Waals forces and electrostatic charges. Even if there exists implicit solvent models, they are yet not very well performing and still require a lot of computation time.

Our goal, in collaboration with the Levitt lab at Stanford University and H. van den Bedem at the Stanford Synchrotron Radiation Laboratory (Associate Team Itsnap http://pages.saclay.inria.fr/julie.bernauer/EA_ITSNAP/ ) is to develop knowledge-based (KB) potentials, based on measurements on known RNA 3D structures and provide sampling for experimental structure fitting and docking conformation generation. KB potential are quick to evaluate during a simulation and can be used without having to explicitly address the solvent problem. They can be developed at various levels of representation: -atom, base, nucleotide, domain- and could allow the modelling of a wide size range: from a hairpin to the whole ribosome. We also intend to combine these knowledge-based potentials with other potentials (hybrid modelling) and template-based techniques, allowing accurate modelling and dynamics study of very large RNA molecules. Such studies are still a challenge. We will also study conformations for experimental data fitting by extension the innovative, robotics-inspired Kino-Geometric Sampler conformational search algorithm for proteins to nucleic acids and to include experimental data. KGS models a protein as a kinematic linkage and additionally considers hydrogen bonds that "close" kinematic cycles. In closed kinematic cycles rotatable bonds can no longer be deformed independently without breaking closure. KGS preserves all kinematic cycles, and thus hydrogen bonds, by sampling in a subspace of conformational space defined by all closure constraints. KGS exhibits a singularly large search radius and optimally reduces the number of free parameters. These unique features enable flexible 'docking' of atomic models in the data while moderating the risk of overfitting at low resolution. The KGS procedure can also accommodate knowledge-based potentials to improve evaluation of putative conformations and their interactions (see below).

PROTEINS

Participants : Jérôme Azé, Julie Bernauer, Adrien Guilhot-Gaudeffroy, Jean-Marc Steyaert.

Docking and evolutionary algorithms

As mentioned above, the function of many proteins depends on their interaction with one or many partners. Docking is the study of how molecules interact. Despite the improvements due to structural genomics initiatives, the experimental solving of complex structures remains a difficult problem. The prediction of complexes, docking, proceeds in two steps: a configuration generation phase or exploration and an evaluation phase or scoring. As the verification of a predicted conformation is time consuming and very expensive, it is a real challenge to reduce the time dedicated to the analysis of complexes by the biologists. Various algorithms and techniques have been used to perform exploration and scoring [49] . The recent rounds of the Capri challenge show that real progress has been made using new techniques [46] , [3] . Our group has strong experience in cutting edge geometric modelling and scoring techniques using machine learning strategies for protein-protein complexes. In a collaboration with A. Poupon, Inra -Tours, a method that sorts the various potential conformations by decreasing probability of being real complexes has been developed. It relies on a ranking function that is learnt by an evolutionary algorithm. The learning data are given by a geometric modelling of each conformation obtained by the docking algorithm proposed by the biologists. Objective tests are needed for such predictive approaches. The Critical Assessment of Predicted Interaction, Capri , a community wide experiment modelled after Casp was set up in 2001 to achieve this goal (http://www.ebi.ac.uk/msd-srv/capri/ ). First results achieved for Capri'02 suggested that it is possible to find good conformations by using geometric information for complexes. This approach has been followed (see section New results). As this new algorithm will produce a huge amount of conformations, an adaptation of the ranking function learning step is needed to handle them. In the near future, we intend to extend our approach to protein-RNA complexes.

Such as in the protein case, the function of RNA molecules also depends on their interaction with one or many partners. Upon interaction, RNA molecules often undergo large conformation changes. Understanding how these molecules interact with proteins would allow better targeting for therapeutic studies. The CAPRI (Critical Assessment of PRediction of Interactions) challenge1 has shown that classical docking procedures largely fail when large conformation change occurs and when RNA is involved. This is especially true for RNA molecules, whose large-scale dynamics remain often unknown. Modeling RNA conformational changes is made hard by the inherent flexible nature of their structure but also by the electrostatics involved. These are hard to model and often lead to computationally expensive simulations. Even if for small RNA molecules, molecular dynamics can be used, such simulations are hard to extend to larger molecules and protein-RNA complexes.

For many diseases, such as cancer and HIV, microRNA molecules play a very important role regulating gene expression by guiding the RISC. Some miRNAs have been shown to suppress tumors and are thus ideal candidates for the development of therapeutic agents. Even if various computational techniques have been developed to predict miRNA targets, none of these consider the structural aspects of the interactions between components of the RISC and miRNA. We aim to target these problems, in collaboration with the Huang lab at Hkust (PHC Procore)

The combination of Voronoi models at a coarse-grained level and powerful machine learning techniques allows the accurate scoring of protein-protein complexes [12] . Our actual machine learning approach for proteins is a combination of several machine learning approaches (evolutionary algorithm, decision trees, decision rules,...). By adapting these approaches to protein-RNA, we would have a fast and efficient technique for scoring large protein-RNA complexes where conformational changes are involved.

Working with RNA instead of protein introduce many major differences in the machine learning approaches. RNA conformations are often smaller than protein conformations, which has an impact on the values of the descriptors used to describe objects. Due to the size differences between RNA and protein, it is often more difficult to generate (during the modeling stage) conformations closed to the biological solution (near native solution). The machine learning algorithms therefore need to take into account all theses specificities to be able to learn good predictive models from data that are not very close to the real solution.

The acquired knowledge on RNA flexibility, dynamics and the importance of the sequence will be a strong advance in the modeling of protein-RNA interactions we are working on. Il will help the development of scoring functions based on Voronoi models for RNA and provide us with the level of flexibility needed in complex conformational search. We also intend to develop hybrid KB potentials for complexes from hybrid RNA KB data. These could be incorporated in leading-edge flexible docking modeling software such as Rosetta.