Journal of Chemical Information and Modeling-2024 TM-search：An Efficient and Effective Tool for Protein Structure Database Search

2024-01-25 · Journal of Chemical Information and Modeling· Biology · protein, structure search

Paper: TM-search：An Efficient and Effective Tool for Protein Structure Database Search

Code: https://zhanggroup.org/TM-search/

TM-search: fast structural alignment

Abstract

Existing protein structure search algorithms incur substantial computational cost. This work introduces TM-search, a program built on TM-align with a new iterative clustering algorithm. Benchmarks show it runs about 27× faster than TM-score while retaining ~90% hit rate, and 2–10× faster than other current tools such as Foldseek, Dali, and PSI-BLAST.

Introduction

By 2020 the PDB contained roughly 178,000 structures and ~400,000 protein chains—too large for exhaustive TM-align search, which averages ~0.5 s per protein pair.

Two mainstream strategies accelerate structure database retrieval:

Map 3D structures to 1D structural identifiers (e.g., Foldseek). This discards much structurally critical information and can reduce sensitivity to remote homolog detection.
Use clustering to cut redundant computation, performing pairwise alignment only between the query and representative structures. Methods such as MMseq2 or CD-HIT cluster by sequence similarity rather than structure.

TM-search builds a hierarchical database from a structure-similarity matrix.

Materials and Methods

Preparation of the TM-search Database

Build a hierarchical structure database.

Split each structure in the database into domains; use SCOPe domain definitions when available, otherwise Protein Domain Parser.
After excluding structures shorter than 30 amino acids, ~470,000 structures form the initial database (PDBall).
CD-HIT clusters at 70% sequence identity; the largest structure in each cluster serves as the representative, yielding 71,115 non-redundant entries (PDB70).

Next, compute pairwise TM-scores for PDB70.

Choosing cluster representatives is key to accuracy and speed; three strategies are used:

type-α: the representative is the structure with the most neighbors in the cluster with TM-score > 0.5.
type-αβ: same as type-α, but if several proteins tie for the largest neighbor count, pick the longest.
type-β: the longest protein not already assigned to an existing cluster becomes the representative.