Rosetta FARFAR2

FARFAR2 operates on the principle of fragment assembly, constructing 3D models by stitching together short, overlapping RNA fragments (typically 3 nucleotides long) from a library of experimentally determined structures. The algorithm initiates with an extended chain conformation and performs a Monte Carlo simulated annealing search. Each move replaces a segment with a torsionally compatible fragment, and the new conformation is accepted or rejected based on the Metropolis criterion using the Rosetta all-atom energy function.

To efficiently explore the vast conformational landscape, FARFAR2 initially represents RNA in a coarse-grained centroid mode. In this phase, each nucleotide side chain is reduced to a single pseudo-atom (centroid) located at the geometric center of the base. This simplification dramatically reduces the degrees of freedom, allowing the Monte Carlo search to rapidly identify globally correct helix packing arrangements and tertiary folds before computationally expensive all-atom refinement.

Following the low-resolution stage, FARFAR2 transitions to a high-resolution all-atom refinement phase. The centroid models are recovered to full atomic detail, and the energy landscape is sampled using the Rosetta all-atom energy function (ref2015). This function integrates:

• Lennard-Jones van der Waals packing terms
• Lazaridis-Karplus implicit solvation model
• Orientation-dependent hydrogen bonding potentials
• Knowledge-based backbone torsion and base-stacking potentials

FARFAR2 is designed to seamlessly incorporate sparse experimental constraints to guide folding and improve accuracy. The energy function can be augmented with pseudo-energy terms derived from:

• SHAPE chemical probing reactivity profiles
• DMS methylation protection data
• Cryo-EM density maps (via map-to-model fitting)
• NMR residual dipolar couplings (RDCs) These restraints penalize conformations that violate experimental observations, effectively biasing the Monte Carlo trajectory toward the native state.

Due to the stochastic nature of Monte Carlo simulation, FARFAR2 requires extensive independent trajectories to converge on the global energy minimum. The algorithm is massively parallelized to run thousands of independent folding simulations simultaneously across high-performance computing clusters. Results are clustered by root-mean-square deviation (RMSD), and the lowest-energy models from the largest clusters are selected as the final predicted structures, providing a consensus-based confidence metric.

The predictive power of FARFAR2 depends critically on its fragment library, a curated database of RNA backbone torsion angles and base-pair geometries extracted from high-resolution X-ray crystallographic structures. The library is non-redundant and filtered by resolution. For a given target sequence, fragments are selected based on sequence similarity and predicted secondary structure propensity, ensuring that local structural building blocks are physically realistic and evolutionarily plausible.

The core algorithmic strategy of FARFAR2, where the target RNA sequence is divided into short, overlapping fragment windows (typically 3 nucleotides). Candidate conformations for each window are drawn from a library of experimentally determined RNA structures. The full-length model is built by Monte Carlo sampling, which randomly selects and inserts these fragments to explore the conformational space, accepting or rejecting moves based on the Rosetta energy function.

A high-resolution refinement method used within FARFAR2 that builds RNA models one nucleotide at a time. Unlike fragment insertion, SWM samples the discrete torsion angle space of the sugar-phosphate backbone for each residue. It employs a branch-and-bound search strategy to systematically enumerate and evaluate possible conformations, enabling the precise modeling of non-canonical motifs and ligand binding sites that are poorly represented in fragment libraries.

A hybrid scoring function that combines physics-based and knowledge-based terms to evaluate the free energy of an RNA conformation. Key components include:

• fa_atr/fa_rep: Lennard-Jones van der Waals interactions
• hbond_sr_bb/sc: Short-range backbone and sidechain hydrogen bonding
• rna_torsion: Statistical torsion angle potentials derived from known structures
• stack_elec: Base stacking electrostatics The total score guides the Monte Carlo acceptance criterion, driving the simulation toward the native fold.

A geometric ontology used by FARFAR2 to annotate and score RNA interactions. It classifies base pairs by three interacting edges (Watson-Crick, Hoogsteen, Sugar) and the relative glycosidic bond orientation (cis or trans). This produces 12 geometric families, enabling the algorithm to precisely recognize and enforce specific tertiary contacts like the A-minor motif and ribose zipper during fragment assembly and scoring.

A statistical energy term derived by inverting the Boltzmann distribution of observed geometries in a non-redundant set of high-resolution RNA crystal structures. FARFAR2 uses these potentials to score the likelihood of observed torsion angles, base pair geometries, and inter-residue distances. A low score indicates a conformation that is statistically similar to those found in nature, effectively encoding the 'rules' of RNA folding without explicitly solving the physics.

A ubiquitous RNA tertiary interaction where an unpaired adenine inserts its minor groove edge into the minor groove of a neighboring Watson-Crick helix. FARFAR2 explicitly models this motif through geometric constraints and base pair classification. Type I and Type II A-minor interactions are critical for stabilizing the compact folds of large RNAs like the ribosome and group I introns, and their accurate prediction is a hallmark of successful de novo modeling.