RMSD (Root Mean Square Deviation)

RMSD is calculated as the square root of the mean squared distances between corresponding atoms after optimal rigid-body superposition. For N atom pairs with coordinates r_i (model) and r_i^ref (reference):

• Formula: RMSD = √(1/N Σᵢ ||r_i - r_i^ref||²)
• Units: Expressed in Ångströms (Å), typically 0–10 Å
• Lower is better: 0 Å indicates perfect identity; <2 Å suggests correct fold
• Weighted variants: Per-residue weighting can emphasize active site accuracy

The choice of atoms included in RMSD calculation dramatically affects the result. Standard conventions include:

• Cα RMSD: Uses only backbone alpha-carbon atoms — the most common convention, insensitive to side-chain rotamer errors
• Backbone RMSD: Includes N, Cα, C, and O atoms — captures main-chain geometry more completely
• All-heavy-atom RMSD: Includes all non-hydrogen atoms — most stringent, penalizes side-chain packing errors
• Ligand RMSD: Symmetry-corrected RMSD for small molecules, accounting for topological symmetries that would otherwise inflate the metric

RMSD has well-known shortcomings that must be understood for proper interpretation:

• Domain motion blindness: A single flexible domain shift can inflate global RMSD while local folds remain perfect — use per-domain RMSD instead
• Length dependence: RMSD scales with √N, making comparisons across proteins of different sizes misleading
• Outlier sensitivity: The squared term heavily penalizes a single poorly modeled loop, obscuring otherwise excellent agreement
• Symmetry ambiguity: Homomeric complexes with rotational symmetry can yield high RMSD despite correct structure if subunit ordering differs

In deep learning for structure prediction, RMSD serves as both a loss function and an evaluation metric:

• Frame Aligned Point Error (FAPE): AlphaFold2's core loss, a generalization of RMSD computed per-residue in local reference frames
• Distogram prediction: Many models predict pairwise distances rather than coordinates, using RMSD only for final evaluation
• Recycling convergence: RMSD between successive recycling iterations indicates whether the model has converged
• TM-score complement: RMSD is often reported alongside TM-score, which is length-normalized and less sensitive to outliers

RMSD is part of a broader ecosystem of structural comparison metrics, each with distinct properties:

• TM-score: Normalized to [0,1], independent of protein size; values >0.5 indicate the same fold
• GDT-TS (Global Distance Test): CASP's primary metric, measuring the fraction of residues within distance thresholds after superposition
• lDDT (local Distance Difference Test): Per-residue metric evaluating local environment preservation without global superposition
• DockQ: Extends RMSD concepts to protein-protein interfaces, combining interface RMSD with native contact recovery

A scale-invariant metric for assessing topological similarity between two protein structures. Unlike RMSD, TM-score is independent of protein size and provides an intuitive measure of structural fold agreement.

• Ranges from 0 to 1, with scores >0.5 indicating the same fold
• Weights closer residue pairs more heavily than distant ones
• Normalized by protein length to enable cross-protein comparisons

A per-residue confidence metric output by AlphaFold2 that estimates local accuracy on a scale from 0 to 100. Higher scores indicate higher predicted reliability.

• pLDDT > 90: High accuracy, suitable for characterizing binding sites
• pLDDT 70–90: Good backbone prediction
• pLDDT 50–70: Low confidence, caution required
• pLDDT < 50: Likely disordered or unstructured region

A pairwise confidence metric that estimates the expected positional error between any two residues in a predicted structure. Critical for assessing domain packing and relative domain orientation accuracy.

• Lower PAE values indicate higher confidence in relative positions
• Visualized as a 2D heatmap with residue indices on both axes
• Used to identify well-predicted domain interfaces versus flexible linkers

A continuous quality score for evaluating protein-protein docking predictions that combines interface RMSD, fraction of native contacts, and ligand RMSD into a single metric.

• Ranges from 0 to 1, with 1 representing a perfect prediction
• DockQ > 0.8: High-quality docking model
• DockQ > 0.5: Medium quality, acceptable for screening
• Integrates multiple evaluation criteria into one interpretable score

A widely used structure-validation software that analyzes all-atom contacts, hydrogen bonding, and backbone geometry to assess the physical realism of protein models.

• Generates a clashscore measuring steric overlaps per 1000 atoms
• Produces Ramachandran statistics for backbone dihedral angle validation
• Identifies rotamer outliers and Cβ deviations
• Essential for validating both experimental and predicted structures

A collection of structurally distinct states representing the intrinsic dynamic flexibility of a protein. Moving beyond a single static prediction captures functional motions relevant to binding and catalysis.

• Generated by methods like Denoising Diffusion Probabilistic Models (DDPMs)
• Essential for understanding allostery and cryptic pocket formation
• Evaluated against NMR ensembles or molecular dynamics trajectories