Root Mean Square Deviation (RMSD)

RMSD calculates the square root of the average squared distance between corresponding atoms after optimal rigid-body superposition. For N atom pairs with positions r_i (model) and r_i^ref (reference), the formula is: RMSD = sqrt(1/N * Σ ||r_i - r_i^ref||²). The Kabsch algorithm is universally used to find the optimal rotation and translation that minimizes this value before calculation. The result is expressed in Ångströms (Å).

RMSD values are context-dependent but follow general thresholds for protein structures:

• < 1.0 Å: Near-identical structures; essentially within experimental error of X-ray crystallography.
• 1.0–2.0 Å: Excellent agreement; typical for high-accuracy predictions like AlphaFold in well-folded regions.
• 2.0–3.0 Å: Good structural similarity; correct global fold with local deviations in flexible loops.
• 3.0–5.0 Å: Moderate similarity; may indicate domain movement or a partially correct fold.
• > 5.0 Å: Significant structural divergence; likely different folds or a failed prediction.

RMSD is computed over different atom subsets depending on the analysis goal:

• Cα RMSD: Calculated using only backbone alpha-carbon atoms. The most common variant for assessing overall fold similarity, as it reduces noise from flexible side chains.
• All-Atom RMSD: Includes all non-hydrogen atoms. More stringent and sensitive to side-chain packing accuracy, used in molecular dynamics refinement and docking validation.
• Backbone RMSD: Uses N, Cα, and C atoms. Balances detail and noise, often used in loop modeling assessment.
• Ligand RMSD: Computed for small-molecule atoms after receptor superposition, critical in drug-target interaction prediction.

RMSD is not a direct pairwise distance metric; it is calculated after an optimal structural alignment. The choice of which residues define the superposition profoundly affects the result. Common strategies include:

• Global superposition: Aligning on all residues, appropriate for comparing highly similar structures.
• Local or domain-based superposition: Aligning on a conserved core or single domain, then calculating RMSD for the entire structure. This reveals domain movements or hinge motions that global alignment would obscure.
• Iterative fitting: Repeatedly aligning and pruning outlier residues to find the most structurally conserved subset.

While RMSD is the classic metric, the Global Distance Test (GDT_TS) is the primary scoring metric in CASP because it is less sensitive to local outlier regions that can dominate RMSD. GDT_TS measures the percentage of residues that can be superimposed under progressively tighter distance thresholds (1, 2, 4, and 8 Å). A model with a single highly disordered loop may have a poor RMSD but an excellent GDT_TS, reflecting its correct global topology. Both metrics are reported together for a complete accuracy picture.

RMSD has known weaknesses that must be considered in structural analysis:

• Length dependence: Larger proteins naturally accumulate higher RMSD values; normalization by chain length is imperfect.
• Outlier sensitivity: A single highly deviating residue (e.g., a flexible terminus) can disproportionately inflate the squared-error sum.
• Symmetry blindness: For symmetric multimers, RMSD may penalize correct but permuted subunit assignments. Specialized symmetry-aware RMSD algorithms are required for quaternary structure prediction.
• No local vs. global distinction: A low global RMSD can mask a completely incorrect local motif. Always pair RMSD with per-residue metrics like pLDDT or local distance difference tests.

A per-residue confidence metric output by AlphaFold2, estimating local prediction accuracy on a 0–100 scale. pLDDT is derived from a neural network trained to predict the local distance difference test (lDDT-Cα) score.

• >90: Very high confidence (backbone often experimental-quality)
• 70–90: Confident backbone prediction
• 50–70: Low confidence (flexible or disordered regions)
• <50: Very low confidence (likely intrinsically disordered)

pLDDT is stored in the B-factor column of AlphaFold PDB files for visualization.

A pairwise confidence metric estimating the expected positional error between any two residues in a predicted structure. PAE is critical for assessing domain packing quality and inter-domain orientation.

• Low PAE (<5 Å): High confidence in relative position
• High PAE (>15 Å): Uncertain relative orientation
• PAE plots: 2D heatmaps revealing domain boundaries and packing confidence

Unlike pLDDT, which is local, PAE captures global structural uncertainty and is essential for evaluating multi-domain proteins.

A stereochemical quality check plotting backbone dihedral angles φ (phi) and ψ (psi) for each residue. Predicted structures should show residues clustering in allowed regions corresponding to α-helices and β-sheets.

• Outliers: Residues in disallowed regions indicate local geometry errors
• Core regions: >90% of residues in most favored regions is expected
• Glycine: Occupies unique regions due to flexibility
• Proline: Restricted to narrow φ angle range

Tools like MolProbity use Ramachandran analysis alongside clashscores and rotamer analysis for comprehensive validation.

A length-independent metric for assessing structural similarity between two protein structures. TM-score ranges from 0 to 1, where scores >0.5 indicate the same fold and <0.17 indicate random structural similarity.

• Scale: Normalized by target protein length
• Weighting: Uses a sigmoid distance function to reduce outlier impact
• Comparison to RMSD: TM-score is less dominated by local deviations

TM-score is widely used alongside RMSD in CASP assessments and protein structure prediction benchmarks.

A physics-based post-processing step that refines predicted structures by simulating atomic motions under force fields like AMBER or CHARMM. MD refinement resolves:

• Atomic clashes: Steric overlaps from imperfect predictions
• Bond geometry: Deviations from ideal bond lengths and angles
• Side-chain rotamers: Suboptimal side-chain conformations

While computationally expensive, MD refinement can improve local RMSD and MolProbity scores, especially for active sites and binding pockets where chemical accuracy is critical.