Template Modeling Score (TM-score)

TM-score solves a critical flaw of RMSD by normalizing the error by the length of the target structure. This ensures a score of 0.5 has the same statistical significance for a 50-nucleotide hairpin as it does for a 500-nucleotide ribozyme. The normalization factor, d₀, scales with the structure size, making scores directly comparable across different RNA families and sizes without bias toward smaller structures.

Unlike RMSD, which treats all atomic distances equally, TM-score uses a Levitt-Gerstein weight function:

• Small errors (< d₀) contribute near-linearly to the score
• Large errors (> d₀) are down-weighted, preventing local outliers from dominating This weighting makes TM-score more sensitive to the global fold topology and less sensitive to flexible loop regions or terminal tails, aligning better with human expert assessment of structural similarity.

TM-score values map to well-defined statistical confidence levels:

• TM-score > 0.5: Indicates the two structures share the same global fold with high probability
• TM-score < 0.17: Corresponds to random structural similarity
• TM-score > 0.6: Considered a high-quality prediction in CASP-RNA benchmarks This statistical grounding allows researchers to set objective quality thresholds without subjective visual inspection.

TM-score calculation involves finding the optimal rotation and translation that maximizes the score between two structures. This is performed using a heuristic iterative alignment algorithm rather than a single least-squares fit. The result is a superposition that prioritizes the core structural motif over peripheral elements, making it particularly effective for comparing RNA structures where conserved helical cores are more important than variable loop conformations.

TM-score is the primary evaluation metric in RNA-Puzzles and CASP-RNA community-wide assessments. It is used alongside:

• INF (Interaction Network Fidelity): Measures base-pairing and stacking accuracy
• GDT-TS (Global Distance Test): Evaluates residue-level superposition
• lDDT (local Distance Difference Test): Assesses per-residue accuracy TM-score's role is specifically to quantify whether the global tertiary architecture has been correctly captured, distinguishing topology-level success from local refinement quality.

TM-score complements rather than replaces other metrics:

• RMSD: Sensitive to local precision but length-dependent; a 5 Å RMSD is excellent for a ribosome but poor for a tetraloop
• GDT-TS: Uses multiple distance thresholds but requires predefined cutoff values
• TM-score: Provides a single, normalized score that balances global topology and local fit In practice, CASP-RNA evaluations report all three metrics, with TM-score serving as the primary indicator of fold-level correctness.

The standard metric for quantifying global structural similarity between a predicted 3D model and an experimental reference. RMSD calculates the average atomic distance after optimal rigid-body superposition.

• Length-dependent: RMSD values scale with system size, making cross-target comparisons difficult
• Outlier-sensitive: A single poorly predicted loop can dominate the score
• Units in Ångströms: Lower values indicate better agreement; values < 2 Å are typically considered high-resolution predictions

TM-score was specifically developed to address RMSD's length-dependence and sensitivity to local outliers.

A per-residue confidence metric output by AlphaFold 3 and related models that estimates the local accuracy of the predicted structure without requiring a reference.

• Scale of 0–100: Higher values indicate higher predicted accuracy
• Self-assessment: Derived from the model's internal uncertainty estimates during structure generation
• Critical filter: Residues with pLDDT < 50 should be treated as low-confidence and often correspond to disordered regions

Unlike TM-score, pLDDT does not require an experimental structure, making it invaluable during de novo prediction.

A length-independent metric originally developed for CASP protein structure assessment that evaluates the percentage of residues that can be superimposed within defined distance cutoffs.

• Multi-threshold: Uses 1, 2, 4, and 8 Å cutoffs to capture both local and global accuracy
• Robust to outliers: Less sensitive to extreme local errors than RMSD
• CASP-RNA adoption: Increasingly used alongside TM-score for RNA structure assessment

GDT-TS and TM-score are complementary; GDT-TS emphasizes coverage across multiple distance scales while TM-score focuses on global topology.

A superposition-free metric that evaluates local structural accuracy by comparing all pairwise inter-atom distances within a defined radius, without requiring global alignment.

• Invariant to domain orientation: Scores individual local environments independently
• Radius-based: Typically uses a 15 Å inclusion radius around each residue
• Stereochemistry-aware: Penalizes violations of local geometry that global metrics may miss

lDDT is particularly useful for multi-domain RNA structures where relative domain orientations may be incorrectly predicted even when individual domains are accurate.

A community-wide blind assessment experiment that evaluates the state-of-the-art in RNA tertiary structure prediction by challenging participants to predict unpublished crystallographic or cryo-EM structures.

• TM-score is a primary metric: Used alongside RMSD and GDT-TS for ranking submissions
• Diverse targets: Includes riboswitches, ribozymes, and viral RNA elements
• Blind assessment: Predictors have no prior knowledge of the experimental structure

RNA-Puzzles has driven methodological advances by revealing that global topology prediction (measured by TM-score) remains challenging for complex RNA folds.

The RNA-specific track of the Critical Assessment of Structure Prediction experiment, providing a standardized, biennial benchmark for comparing computational RNA structure prediction methods.

• Standardized evaluation: Uses TM-score, GDT-TS, and lDDT for consistent ranking
• Regular cadence: Held every two years alongside the protein CASP experiment
• Method-agnostic: Evaluates physics-based, template-based, and deep learning approaches on equal footing

The CASP-RNA track has documented the rapid improvement of deep learning methods like AlphaFold 3, with TM-score values showing significant gains over traditional approaches.