Modeling Complex RNA Tertiary Folds with Rosetta

Clarence Yu Cheng*; Fang-Chieh Chou*; Rhiju Das*,†,1    * Department of Biochemistry, Stanford University, Stanford, California, USA
† Department of Physics, Stanford University, Stanford, California, USA
1 Corresponding author: email address: rhiju@stanford.edu

1 Introduction

Computational modeling of RNA structures is advancing rapidly, with recent developments improving prediction and design of both secondary and tertiary structures of RNA. Continuing improvements to secondary structure prediction algorithms (Tinoco et al., 1973), classification of RNA structural motifs (Petrov, Zirbel, & Leontis, 2013), molecular dynamics and quantum mechanical techniques (Ditzler, Otyepka, Sponer, & Walter, 2010), conformational sampling with energy scoring (Das, Karanicolas, & Baker, 2010), atomic-scale loop and motif modeling (Sripakdeevong, Kladwang, & Das, 2011), integration with conventional crystallographic (Chou, Sripakdeevong, Dibrov, Hermann, & Das, 2013) and NMR approaches (Sripakdeevong et al., 2014), and connections with recent single-molecule (Chou, Lipfert, & Das, 2014) and internet-scale videogame (Lee et al., 2014) technologies hold promise for eventually attaining confident 3D modeling and design of RNAs with high spatial resolution. An important driver of recent innovation has been the establishment of blind prediction trials, proposed during a community-wide collation of 3D RNA modeling methods in 2010 (Sripakdeevong, Beauchamp, & Das, 2012) and begun soon thereafter. The RNA-Puzzles trials (Cruz et al., 2012), modeled after the 20-year-old CASP trials in protein structure prediction, challenge participating groups to create accurate 3D models of RNAs from sequence alone; the submitted models are compared to unreleased crystallographic structures of the targets to assess the methods’ predictive power. These trials provide a rigorous testing ground for current computational as well as hybrid experimental/computational structure prediction methods on RNA domains that are of strong biological interest.

This chapter describes methods from our laboratory of medium computational and experimental expense that achieve subhelix-resolution accuracy for 3D models of 100- to 300-nucleotide RNAs, a typical size range for many riboswitch and ribozyme domains and representative of RNA-Puzzles target sizes. Subhelical resolution, while not the ultimate achievable, has still been useful in guiding mutational experiments in vitro and in vivo, detecting partial structure in riboswitches without their ligands, and in revealing or illustrating evolutionary connections that are not obvious from sequence comparisons alone. The primary tools for this approach are constraints from chemical mapping experiments, which we discuss briefly here and will be described in more detail elsewhere, and computational modeling to integrate chemical mapping data into 3D portraits.

Our laboratory is developing several tools that seek to advance 3D macromolecule modeling at multiple length scales. For small RNA motifs, we leverage algorithms based on a “stepwise ansatz,” which enable modeling of RNA loops and motifs with near-atomic accuracy (better than 2 Å RMSD), particularly if limited NMR or crystallographic data are available (Chou et al., 2013; Sripakdeevong et al., 2014, 2011). Unfortunately, the computational expense of those high-resolution tools is currently prohibitive for de novo modeling of large RNAs. Instead, our practical tools for large RNAs have largely been built on Fragment Assembly of RNA with Full-Atom Refinement (FARFAR) in the Rosetta framework, which was first introduced to model small motifs of RNAs in 2007 (Das & Baker, 2007) and was initially based on Rosetta protein structure prediction methods that we had helped in advance. Since that time, FARFAR has been progressively developed to allow for nucleotide-resolution building of not just individual RNA motifs but also more complex RNA folds involving dozens of helices. This chapter is intended to offer a practical guide to getting started with Rosetta using an up-to-date workflow from our laboratory, laid out in Fig. 1. We will illustrate this workflow below using the ligand-binding region of a tandem glycine-binding riboswitch from F. nucleatum, which forms a complex pseudosymmetric fold stabilized by A-minor interactions between two glycine-binding subdomains. A homolog of this domain was posed as an RNA-Puzzles challenge (Cruz et al., 2012), and crystallographic and biochemical work on this system by several RNA laboratories (Butler, Xiong, Wang, & Strobel, 2011; Cordero, Kladwang, VanLang, & Das, 2012; Erion & Strobel, 2011; Kladwang, VanLang, Cordero, & Das, 2011) have made this RNA a useful model system for calibrating and illustrating experimental and computational methodologies.

2 Setting the Stage for 3D Modeling Using Experimental Data

Several pieces of information can provide powerful constraints to help construct accurate 3D models of RNA. The most fundamental of these is the RNA's secondary structure. If phylogenetic inference of secondary structure is precluded by the lack of sequence homologs, difficulties in sequence alignment, or targeting of “alternative” states of the RNA (e.g., without ligands or in misfolded conformations), chemical mapping techniques provide useful guides to computational secondary structure prediction (Cordero, Kladwang, VanLang, & Das, 2014; Hajdin et al., 2013; Kladwang et al., 2011). In traditional “one-dimensional” (1D) chemical mapping experiments, solution-state RNAs are exposed to chemical modifiers which form adducts to the backbone or nucleobases depending on backbone flexibility or base-pairing status (Fig. 2A). These modifications are traditionally detected by reverse transcription, which stops at the modified location, followed by gel or capillary electrophoresis or, more recently, deep sequencing to identify the sequence position of each modification. The reactivity of each nucleotide position to the chemical modifier can be quantified using several publically available software suites, with HiTRACE (Kim, Cordero, Das, & Yoon, 2013; Yoon et al., 2011) (https://github.com/hitrace/hitrace) and MAPseeker (Seetin et al., 2014) (https://github.com/DasLab/map_seeker) particularly optimized for high-throughput analysis of capillary electrophoresis and deep-sequencing data, respectively. Secondary structure prediction servers such as RNAstructure (Reuter & Mathews, 2010) (http://rna.urmc.rochester.edu/RNAstructureWeb) and the RNA mapping database structure server (Cordero, Lucks, & Das, 2012)...