TM-score-guided iterative superposition

teaching-model

The Refocusing Projector: match, superpose, then match again

This is a teaching analogy, not the paper's story. Two silhouette slides begin with only a rough match. Each refocusing round superposes them, observes which positions are close, and updates the next match.

1. Chapter 1: begin from three rough matches

   The technician does not know the best starting match, so three initial calibrations are prepared and each enters refinement.

   Decoded paper language: TM-align uses secondary-structure DP, gapless threading, and a secondary-structure/distance hybrid as three initial alignments, retaining the highest TM-score.

   Zhang & Skolnick 2005, pp. 2302–2303.

2. Chapter 2: superpose using the current match

   The technician temporarily trusts the current match and rigidly moves one silhouette so the matched positions come as close as possible.

   Decoded paper language: Each round computes a rigid superposition from the current alignment so later distance scoring occurs in a shared coordinate frame.

   Zhang & Skolnick 2005, p. 2303.

3. Chapter 3: nearby positions speak loudly, distant ones softly

   After superposition, nearby silhouette pairs influence the next round more strongly, while distant pairs retain a rapidly diminishing score.

   Decoded paper language: The DP match score is 1/(1+dᵢⱼ²/d₀²), smoothly downweighting distance instead of using only a hard cutoff.

   Zhang & Skolnick 2005, p. 2303.

4. Chapter 4: trace a new monotone path

   The technician fills a grid with updated closeness scores, finds a high-scoring route that moves forward on both slides with skips allowed, and refocuses again.

   Decoded paper language: TM-align updates the alignment with DP, then repeats superposition and rescoring. The paper reports typical convergence in two to three rounds, with gap opening −0.6 and no gap extension penalty.

   Zhang & Skolnick 2005, p. 2303.

5. Chapter 5: compare different canvas sizes on a common scale

   Short and long silhouettes should not be compared by raw match count alone. The technician normalizes by target length for a more comparable score.

   Decoded paper language: TM-score is normalized by target length and d₀ varies with length. The paper's empirical 0.5 same-fold reference is not a functional or clinical verdict.

   Zhang & Skolnick 2005, pp. 2303–2304, eq. (3).

A five-stop plain-language map of TM-align

Read it as a loop: initialize, superpose, score by distance, update with DP, repeat, and compare TM-score.

1. 1. Do not bet everything on one start

   Three initializations provide different rough alignments, each entering the same refinement process.

   Keep only this: Starts affect local search, so try several.

   Zhang & Skolnick 2005, pp. 2302–2303.

2. 2. Rigidly superpose from current matches

   The current alignment supplies point pairs, and superposition places both structures in a comparable pose.

   Keep only this: Matches determine the fit, and the fit changes the next match scores.

   Zhang & Skolnick 2005, p. 2303.

3. 3. Score with TM-score-style distance weighting

   Close pairs score near 1 and decay smoothly with distance; d₀ controls the scale.

   Keep only this: It is a continuous weight, not one hard cutoff.

   Zhang & Skolnick 2005, p. 2303.

4. 4. Use DP to find a new ordered alignment

   Distance scores fill a match matrix, and DP finds a high-scoring path preserving order on both structures while allowing gaps.

   Keep only this: The new path becomes the point pairs for the next superposition.

   Zhang & Skolnick 2005, p. 2303.

5. 5. Read TM-score together with normalization

   The score is normalized by target length, so direction can affect it. The 0.5 fold-level reference is empirical, not a universal truth.

   Keep only this: The score is a comparison tool, not a biological-function verdict.

   Zhang & Skolnick 2005, pp. 2303–2304.

Check understanding

1. What is TM-align's core loop?

   • Superpose → distance-score → DP-update → repeat
   • Compute RMSD once and stop
   • Compare sequence letters only

   The current alignment produces a fit, the fit produces new distance scores, and DP updates the alignment.

2. What happens to 1/(1+d²/d₀²) as distance grows?

   • It decreases smoothly
   • It always increases
   • It always equals 1

   The weight gives close pairs more influence and distant pairs less.

3. Does TM-score 0.5 directly prove identical function?

   • No; it is an empirical fold-level reference, not a function verdict
   • Yes, and it means identical sequence
   • Yes, without exceptions at any length

   Structural scores require context from coverage, length, direction, and biology.

Completion task: Draw a five-box cycle for initialization, superposition, distance scoring, DP, and repeat; add one note about TM-score normalization.

paper-fact

Three initial alignments

TM-align starts from secondary-structure DP, gapless threading, and a hybrid secondary-structure/distance alignment. Each enters refinement, and the highest TM-score is retained.

Zhang & Skolnick 2005, pp. 2302–2303.

paper-fact

Superpose, score, run DP, and repeat

Under the current superposition, each residue-pair DP score is 1/(1+dᵢⱼ²/d₀²). The paper uses a −0.6 gap-opening penalty, no gap-extension penalty, and typically converges after two to three iterations.

Zhang & Skolnick 2005, p. 2303.

paper-fact

Why TM-score is length-normalized

TM-score is normalized by target length, and d₀ also changes with target length, making random-background scores more comparable across sizes. The paper uses 0.5 as an empirical same-fold reference, not a biological-function or clinical verdict.

Zhang & Skolnick 2005, pp. 2303–2304, eq. (3).

Glossary

TM-score

A structure-similarity score with distance weighting and target-length normalization.

d₀

A target-length-dependent distance scale controlling how quickly distant matches are downweighted.

Gap opening

The score cost paid when a dynamic-programming path starts a gap.

Interactive lab

Interactive lab loads when JavaScript is available.

Sources and limits

• Open-access article

• The current lab does not reproduce the three initializations, formal d₀ edge handling, or TM-score rotation.
• A single score cannot establish biological function, evolution, or clinical conclusions.