Global Distance Test (GDT_TS)

GDT_TS evaluates structural similarity using four distinct distance cutoffs: 1, 2, 4, and 8 Ångströms. For each threshold, the algorithm performs an optimal superposition of the predicted model onto the experimental structure, then calculates the percentage of Cα atoms within that distance. The final score is the arithmetic mean of these four percentages, yielding a value between 0 and 100.

• 1 Å threshold: Captures near-atomic accuracy
• 2 Å threshold: Assesses high-resolution similarity
• 4 Å threshold: Evaluates secondary structure element placement
• 8 Å threshold: Measures global fold correctness

This multi-scale approach ensures the metric is not dominated by highly variable loop regions while rewarding accurate core topology.

GDT_TS has been the official primary ranking metric of the Critical Assessment of Structure Prediction (CASP) experiment since CASP3 in 1998. It was developed to address the limitations of Root Mean Square Deviation (RMSD), which is disproportionately penalized by large local errors in flexible regions.

In CASP, GDT_TS is calculated for each target domain, and the sum of per-target scores determines the overall ranking of prediction groups. A score above 80 is generally considered to represent a correctly predicted global fold, while AlphaFold2 routinely achieves scores exceeding 90 for well-folded domains.

Unlike RMSD, which squares the deviation and is dominated by the worst-fitting regions, GDT_TS is inherently robust to local errors. By using a percentage-based cutoff rather than a sum of squared distances, a single misplaced loop or flexible terminus cannot catastrophically degrade the score.

This property makes GDT_TS particularly suitable for evaluating comparative models where core regions are well-predicted but surface loops may deviate. The 8 Å threshold effectively ignores large deviations that would otherwise mask the quality of the conserved structural core.

GDT_HA (High Accuracy) is a stricter variant designed to differentiate between near-atomic resolution predictions. It uses the same four-threshold framework but applies a weighted scoring scheme that heavily favors the tighter cutoffs.

• 1 Å: Weighted 4×
• 2 Å: Weighted 3×
• 4 Å: Weighted 2×
• 8 Å: Weighted 1×

This weighting makes GDT_HA particularly sensitive to improvements in side-chain and backbone precision, making it the preferred metric for assessing de novo design accuracy and high-confidence AlphaFold predictions where global topology is already correct.

GDT_TS is computed using the Local-Global Alignment (LGA) program developed by Adam Zemla. LGA performs a sequence-dependent structural superposition that identifies the longest continuous segments of residues that can be aligned within each distance threshold.

The algorithm iteratively:

1. Selects the largest set of residues fitting within the current threshold
2. Performs optimal superposition on that subset
3. Counts all residues within the threshold after superposition
4. Repeats for each of the four distance cutoffs

This iterative selection prevents the superposition from being skewed by divergent regions, ensuring the score reflects the best possible global alignment.

GDT_TS is closely related to the Template Modeling Score (TM-Score), another widely used metric that also addresses RMSD limitations. While both are length-independent and robust to local errors, they differ in their mathematical formulation.

• GDT_TS: Uses discrete distance cutoffs and percentage-based counting
• TM-Score: Uses a continuous distance-dependent weighting function with a length-dependent normalization factor

In practice, the two metrics are highly correlated (Pearson correlation > 0.9), but GDT_TS remains the CASP standard due to its intuitive interpretability and historical precedent. TM-Score is often preferred for automated model quality assessment programs.

The most fundamental metric for structural comparison, RMSD calculates the average distance between the backbone atoms of superimposed protein structures. It is computed by taking the square root of the mean squared distances between aligned atom pairs.

• Sensitivity: Highly sensitive to small, localized outliers; a single poorly predicted loop can dominate the score.
• Calculation: Requires an optimal rigid-body superposition before measurement.
• Limitation: Unlike GDT_TS, RMSD does not reward partial correctness and is not robust to flexible domain movements.

A length-independent metric designed to solve the protein size dependency problem of RMSD. TM-score scales the maximum achievable score by the length of the target protein.

• Scale: Returns a value between 0 and 1, where scores > 0.5 generally indicate the same global fold.
• Weighting: Uses a sigmoidal distance weighting scheme that naturally suppresses the impact of large, noisy errors compared to RMSD.
• Relationship: TM-score is mathematically related to GDT_TS but provides a single, continuous value rather than a set of distance thresholds.

A per-residue confidence metric output by AlphaFold2 that serves as a local proxy for the global GDT_TS concept. It predicts the local LDDT score without knowing the true structure.

• Scale: Ranges from 0 to 100, with higher values indicating higher predicted local accuracy.
• Mechanism: The model predicts the fraction of distances within a 15 Å inclusion radius that are correctly predicted.
• Usage: Researchers use pLDDT to mask disordered regions and identify well-predicted domains, effectively applying the GDT_TS philosophy at the residue level.

The per-residue scoring function that inspired the global GDT_TS metric. LDDT evaluates the preservation of local interatomic distances in the model compared to the reference structure.

• Inclusion Radius: Considers all atom pairs within a default radius of 15 Å.
• Thresholds: Checks if distances are preserved within tolerance thresholds (0.5, 1, 2, and 4 Å).
• Advantage: Unlike GDT_TS which focuses on global superposition, LDDT is stereochemistry-aware and does not require domain parsing, making it ideal for assessing local side-chain environments.

A 2D plot output by AlphaFold that estimates the expected positional error between every pair of residues. It directly visualizes the global topology confidence that GDT_TS measures numerically.

• Interpretation: Low PAE between two domains indicates high confidence in their relative orientation.
• Domain Packing: A key tool for assessing quaternary structure and inter-domain accuracy, complementing the single-score summary of GDT_TS.
• Units: Measured in Ångströms, providing an intuitive physical scale for expected coordinate error.