Structure Evaluation Metrics

Question 1

Which score(s) should I use to assess whether my model’s OVERALL FOLD is correct (model vs native)?

Answer 1

Reference needed: Yes
Most relevant: TM-score; GDT_TS
Also check: lDDT; RMSD (Cα/backbone)
Why: TM/GDT are robust to outlier loops and more comparable across lengths than RMSD.

Question 2

Which score(s) should I use if I want a physical distance error in Å between two structures?

Answer 2

Reference needed: Yes
Most relevant: RMSD (define atom set: Cα/backbone/all-atom)
Also check: TM-score; GDT; lDDT
Why: RMSD gives an Å scale but is sensitive to domain motions/outliers and atom selection.

Question 3

Which score(s) best capture LOCAL DETAIL accuracy without being affected by global superposition?

Answer 3

Reference needed: Yes
Most relevant: lDDT (global or per-residue)
Also check: per-residue RMSD; MolProbity geometry
Why: lDDT is superposition-free and focuses on preservation of local distance patterns.

Question 4

How should I score DOMAIN ARRANGEMENT correctness for multi-domain proteins?

Answer 4

Reference needed: Yes (for direct comparison)
Most relevant: Domain-wise RMSD/TM (per domain) + relative placement analysis
Also check: lDDT per-domain
Why: Global RMSD can look bad even if each domain is right; evaluate domains separately and their relative placement.

Question 5

Which AlphaFold scores tell me which parts of a MONOMER prediction I can trust (no reference structure)?

Answer 5

Reference needed: No
Most relevant: pLDDT (local confidence)
Also check: PAE (relative placement uncertainty); pTM (global fold confidence)
Why: pLDDT is local; PAE/pTM report global packing/relative-domain certainty.

Question 6

Which AlphaFold scores are best for judging whether a MULTIMER INTERFACE is believable (no reference complex)?

Answer 6

Reference needed: No
Most relevant: ipTM
Also check: Inter-chain PAE blocks; ipSAE (if available)
Why: ipTM targets inter-chain placement; PAE shows which interfaces are confident; ipSAE can be more interface-focused/robust to extra regions.

Question 7

Which score should I use to evaluate a DOCKING POSE against a known reference complex?

Answer 7

Reference needed: Yes
Most relevant: DockQ
Also check: iRMSD; LRMSD; Fnat
Why: DockQ combines contact recovery + interface geometry + global placement into a single 0–1 score.

Question 8

Which scores matter most if I want to reason about an ACTIVE SITE or BINDING SITE?

Answer 8

Reference needed: Depends
Most relevant: Local confidence—pLDDT (or lDDT if reference exists)
Also check: side-chain rotamers/clashes; local RMSD
Why: Binding sites require high local accuracy AND reasonable side-chain geometry.

Question 9

How do I choose the best model among several predictions when I DON’T have a reference structure?

Answer 9

Reference needed: No
Most relevant: pTM (monomer); ranking_confidence = 0.8·ipTM + 0.2·pTM (multimer); pLDDT
Also check: PAE patterns; geometry checks
Why: pTM/ipTM capture global placement; pLDDT highlights unreliable regions; PAE shows domain/interface uncertainty.

Question 10

TM-score

Answer 10

Description: Global fold similarity between two structures; higher means more similar and is relatively length-normalized.
Range: 0–1.
Formula: TM = max{ (1/L_norm) · Σ_k=1..Lali 1 / (1 + (d_k/d₀(L_norm))²) ) }, d₀(L) = 1.24·(L−15)^1/3 − 1.8.

Question 11

pTM

Answer 11

Description: AlphaFold’s predicted global TM-score-like confidence (expected fold correctness vs the unknown true structure).
Range: 0–1.
Formula: pTM = max_i{ (1/L) · Σ_j=1..L E[ 1 / (1 + (e_ij/d₀(L))²) ] }.

Question 12

ipTM

Answer 12

Description: AlphaFold-Multimer’s predicted interface confidence, focusing on how well different chains are positioned relative to each other.
Range: 0–1.
Formula: ipTM = max_i{ (1/|D_−chain(i)|) · Σ_{j∈D−chain(i)} E[ 1 / (1 + (e_ij/d₀(|D_−chain(i)|))²) ] }.

Question 13

PAE

Answer 13

Description: Predicted Aligned Error; expected positional error (Å) at residue x when the structure is aligned using residue y as the reference frame.
Range: ~0–31.75 Å (typically capped).
Formula: PAE(x,y) = E[e_yx] ≈ Σ_b q_yx,b · c_b.

Question 14

pLDDT

Answer 14

Description: AlphaFold’s per-residue local confidence (expected lDDT-Cα) indicating how reliable the local geometry is around each residue.
Range: 0–100.
Formula: pLDDT_i = Σ_b p_i,b · v_b.

Question 15

ipSAE

Answer 15

Description: Interface score derived from PAE that emphasizes confidently placed inter-chain regions and is more robust to extra flexible/disordered parts.
Range: 0–1.
Formula: ipSAE(X→Y) = max_i∈X{ (1/n₀(i)) · Σ_{j∈Y, PAEij<c} 1 / (1 + (PAE_ij/d₀(n₀(i)))²) ) }, n₀(i)=|{j∈Y : PAE_ij<c}|; often report ipSAE(X,Y)=max(ipSAE(X→Y), ipSAE(Y→X)).

Question 16

RMSD

Answer 16

Description: Root-mean-square deviation (Å) between paired atoms after optimal rigid-body superposition; lower means more similar.
Range: 0–∞ Å.
Formula: RMSD = √( (1/N) · Σ_i=1..N ‖x_i − y_i‖² ).

Question 17

GDT

Answer 17

Description: Global Distance Test; percent-like measure of how many residues can be superposed within several distance cutoffs (GDT_TS commonly uses 1, 2, 4, 8 Å).
Range: 0–100.
Formula: (No single standard formula required here; computed from fractions within multiple distance thresholds.)

Question 18

lDDT

Answer 18

Description: Local Distance Difference Test; superposition-free accuracy based on how well local inter-atomic distances are preserved vs a reference.
Range: 0–1 (often reported as 0–100).
Formula: lDDT = (1/4)·(C(0.5)+C(1)+C(2)+C(4)), C(t)=(1/|P|)·Σ_(a,b)∈P 𝟙(|d_ab^model−d_ab^ref|<t), P={(a,b): d_ab^ref<R₀} (often R₀=15 Å).

Question 19

DockQ

Answer 19

Description: Docking-quality score (0–1) combining contact recovery and RMSD-based interface/global placement terms vs a reference complex.
Range: 0–1.
Formula: DockQ = (1/3)·(Fnat + 1/(1+(LRMSD/8.5)²) + 1/(1+(iRMSD/1.5)²)).