Root Mean Square Deviation (RMSD)

RMSD is calculated as the square root of the average squared distance between corresponding atoms after optimal rigid-body superposition. For N atoms with coordinates r_i (predicted) and r_i^ref (reference):

• Formula: RMSD = √(1/N Σ ||r_i - r_i^ref||²)
• Units: Typically reported in Ångströms (Å)
• Superposition: Uses the Kabsch algorithm to find the optimal rotation and translation that minimizes RMSD
• Atom Selection: Commonly calculated on C1' backbone atoms for RNA to capture overall fold while ignoring flexible base orientations

RMSD values must be interpreted in the context of RNA size and flexibility. General guidelines for RNA structure prediction:

• < 2.0 Å: Near-native accuracy, comparable to experimental resolution limits
• 2.0–5.0 Å: Correct global fold with local deviations in flexible loops or junctions
• 5.0–10.0 Å: Partially correct topology; significant domain-level deviations
• > 10.0 Å: Incorrect fold; model fails to capture the overall architecture
• Length Dependence: Larger RNAs naturally accumulate higher RMSD values; a 100-nucleotide ribozyme at 5 Å RMSD may be more accurate than a 20-mer at 3 Å

RMSD has well-documented shortcomings that must be considered when evaluating RNA structure predictions:

• Domain Dominance: A small, poorly predicted domain can be masked by a large, well-predicted domain in the global average
• Outlier Sensitivity: Squared distances heavily penalize individual large deviations, making RMSD sensitive to a single misplaced helix
• Alignment Ambiguity: Symmetric or repetitive RNA structures can produce artificially low RMSD values through alternative atom correspondences
• No Local Information: RMSD provides a single global number with no per-residue or per-motif accuracy breakdown
• Complement with TM-score: The Template Modeling Score is length-independent and more sensitive to global topology, making it a preferred complement in RNA-Puzzles assessments

The Kabsch algorithm is the standard method for computing the optimal rotation matrix that minimizes RMSD between two sets of corresponding points:

• Step 1: Center both coordinate sets at their centroids to remove translation
• Step 2: Compute the 3×3 covariance matrix between the centered coordinate sets
• Step 3: Perform Singular Value Decomposition (SVD) on the covariance matrix
• Step 4: The optimal rotation matrix is derived from the SVD components, with a determinant check to prevent reflections
• Computational Cost: O(N) for N atoms, making it efficient even for large ribosomal RNA structures

RMSD serves as a primary evaluation metric in community-wide blind assessment experiments:

• RNA-Puzzles: Reports RMSD alongside TM-score, GDT-TS, and INF (Interaction Network Fidelity) to provide a multi-faceted accuracy assessment
• CASP-RNA: Uses RMSD calculated on C1' atoms as a standard metric, but emphasizes that pLDDT and local distance difference tests are more informative for per-residue confidence
• Multi-reference RMSD: When multiple experimental structures exist (e.g., different crystal forms or NMR ensembles), the best RMSD against any reference is often reported
• Ensemble RMSD: For methods that predict structural ensembles, RMSD is calculated against the closest ensemble member to assess conformational sampling accuracy

Advanced RMSD variants provide spatially resolved accuracy information critical for RNA structure refinement:

• Per-Residue RMSD: Calculated on a sliding window of 3–5 nucleotides to identify locally inaccurate regions such as bulges or internal loops
• Domain-Level RMSD: Computed independently for each structural domain after domain decomposition, revealing which motifs are correctly predicted
• Contact RMSD: Evaluates only atom pairs within a specified distance cutoff (e.g., 8 Å), focusing on tertiary interaction accuracy rather than global fold
• L-RMSD vs. G-RMSD: Local RMSD ignores long-range deviations, while Global RMSD captures overall fold; both are reported in RNA-Puzzles to distinguish local precision from global accuracy