rsomics-tm-align

Pairwise structural alignment of two protein structures from PDB — reports the TM-score (normalized by each chain length), RMSD over the aligned core, aligned length, the rotation + translation that superposes chain 1 onto chain 2, and the residue alignment. A clean-room Rust reimplementation of TMalign.

rsomics-tm-align a.pdb b.pdb            # TM-scores, RMSD, rotation, alignment
rsomics-tm-align a.pdb b.pdb --json     # machine-readable scores
rsomics-tm-align a.pdb b.pdb -o out.txt

The first chain of the first model is read from each file (Cα trace, blank/A altloc), matching TMalign's default monomer behaviour.

Method

TM-score of an aligned residue set is

TM-score = (1/L) · Σ 1 / (1 + (d_i / d0)²),   d0 = 1.24·(L − 15)^(1/3) − 1.8

with L the normalization chain length and d_i the distance between an aligned pair after optimal superposition. The alignment is found by Zhang & Skolnick's heuristic: several initial correspondences (gapless threading at every offset, a secondary-structure Needleman-Wunsch, and a fragment seed) are cheaply ranked by a single Kabsch + TM-score; the strongest few are refined by iterating {TM-score-maximizing superposition ↔ dynamic-programming re-alignment} to a fixed point. The reported aligned length and RMSD use TMalign's d8 cutoff.

Compatibility

Validated against TMalign (Version 20220412) on same-fold pairs, where the TM-score is a well-defined, reproducible quantity:

pair	fold relation	TMalign TM (1/2)	ours TM (1/2)	Δ
1UBQ vs 1UBQ	self	1.0000 / 1.0000	1.0000 / 1.0000	0
1UBQ vs 1UBI	identical (ubiquitin)	0.9991 / 0.9991	0.9991 / 0.9991	0.0000
1MBN vs 1HHO	globin fold	0.8397 / 0.9048	0.8396 / 0.9046	0.0002

Max deviation: ΔTM = 0.0002, ΔRMSD = 0.00 Å, Δaligned = 0. tests/compat.rs runs this differential when a TMalign binary is present (loud-skip otherwise).

For unrelated folds (TM ≲ 0.4) the superposition landscape is flat and full of near-degenerate optima; independent heuristics legitimately disagree there by ~0.02, and the TM-align paper itself treats TM < 0.5 as "different fold". That regime is exercised only as a bounded smoke test, not an equality gate.

Performance

Single-threaded, against TMalign 20220412 on an Intel Xeon Gold 6238R (hyperfine, ≥12 runs):

pair	size	ours -t1	TMalign	speedup
1DPE vs 1GAI	507 × 472	58.7 ms	627 ms	10.7×
1J6Z vs 3HBT	368 × 359	31.2 ms	123 ms	3.9×

With -t8, 1DPE vs 1GAI reaches 24× (the initial-seed refinement parallelizes over rayon). The win comes from ranking gapless-threading offsets cheaply and paying the full O(n·m) DP only on the few strongest seeds, rather than refining every offset.

Origin

This crate is an independent Rust reimplementation of TMalign based on:

• The published method: Y. Zhang and J. Skolnick, "TM-align: a protein structure alignment algorithm based on the TM-score", Nucleic Acids Research 33.7 (2005) 2302–2309, doi:10.1093/nar/gki524.
• The TM-score definition: Y. Zhang and J. Skolnick, "Scoring function for automated assessment of protein structure template quality", Proteins 57 (2004) 702–710, doi:10.1002/prot.20264.
• The public PDB file-format spec.
• Black-box behaviour testing against the TMalign binary.

No source code from TMalign was copied into this implementation. Test fixtures are real public PDB entries (1UBQ, 1UBI, 1MBN, 1HHO, 1CRN).

License: MIT OR Apache-2.0. Upstream credit: TMalign (Zhang lab) https://zhanggroup.org/TM-align/.