henhaobio
Sign in

ScienceBioinformaticsPeer-reviewed journal article

Multiple Protein Structure Alignment at Scale with FoldMason

Cameron L. M. GilchristiD, Milot MirditaiD, Martin SteineggeriD

PDFDOIPreprintCodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 3 (27 Aug 2024) · CC BY 4.0

FoldMason uses Foldseek's 3Di+AA alphabet for progressive multiple structure alignment, beating structure aligners' LDDT on AlphaFold DB clusters at near sequence-aligner speed, scaling to 100,000 structures.

Structure prediction has populated resources like AlphaFoldDB and ESMAtlas with hundreds of millions of models, yet multiple structure alignment (MSTA) tools such as US-align, MUSTANG and Matt trade speed for accuracy. FoldMason brings MSTA close to sequence-aligner speed, adds a reference-free local distance difference test (LDDT) quality score, and supports structure-based phylogeny below the sequence twilight zone.

Key results

  1. ShownOn a reference-free benchmark of divergent multidomain AlphaFold2 models, FoldMason gave the highest average LDDT of any tested aligner, structure- or sequence-based. Optional LDDT-maximizing refinement raised it further.

    Across 150 AFDB clusters (20 structures, 2-4 domains), mean MSTA LDDT was 19.7%, 23.9%, 9.19%, 33.8%, 4.15% above US-align, mTM-align, MUSTANG, Matt, Caretta-shape and 2.16%, 2.55%, 1.36%, 7.77% above Clustal Omega, FAMSA2, MAFFT, MUSCLE; 100 refinement iterations added 1.47%. · Introduction; Benchmarks: AFDB cluster database; Tables

  2. ShownFoldMason runs near sequence-aligner speed on small clusters and scales to 100,000 structures, whereas US-align failed to finish far smaller subsets. FAMSA2 still scales better than FoldMason.

    Mean under a second per AFDB cluster: 3261x faster than Matt, 4x slower than FAMSA2. On subsets of cluster A0A258JAY9 (533,657 members), 100,000 structures took 92899.10 s (64 cores) vs 147.49 s for FAMSA2; US-align did not finish 5000 structures within 2 days. · Introduction; Benchmarks; Benchmarks: AFDB cluster database; Benchmarks: Scalability

  3. ShownOn 1,032 Homstrad families with curated reference alignments, structure-aware aligners beat sequence aligners. FoldMason reached structure-aligner accuracy but ranked below three of them.

    Sum-of-pairs: structure-based tools averaged 6.5% higher sensitivity and 8.3% higher specificity than sequence-based tools (7.1% and 9.5% excluding Caretta-shape); FoldMason F1 0.894, below MUSTANG, US-align and mTM-align. · Introduction; Benchmarks: Homstrad; Tables

  4. ShownIQ-Tree phylogenies built directly from FoldMason MSTAs of Flaviviridae E, E1 and E2 glycoproteins recovered every annotated clade of earlier manually assembled structural trees (Mifsud et al.), with similar node support.

    Quartet similarity to published trees 92.71% (E), 90.72% (E1), 81.72% (E2); bootstrap support 87.3-90.2% vs 87.1-93.3% for reference trees; on corresponding nodes (85/245, 109/188, 100/187) support differed by 1.2-1.6%. · Introduction; Benchmarks: Glycoprotein Phylogeny

  5. SuggestedFoldMason scores local structural interactions rather than global superposition. In three families it aligned flexible proteins (hinged lobes, mobile loops, domain swaps) more accurately than US-align.

    Average MSTA LDDT 22.3%, 16.6% and 3.8% above US-align for calmodulin, odorant-binding proteins and globins; Caretta-shape scored 2.7% higher than FoldMason on the globin MSTA. Pairwise cases: CaM 1ncx/3cln, myoglobin 2eb8/5vzq, OBP 1a3y/1g85. · Introduction; Benchmarks: Flexibility; Tables

Methods, data and reproducibility

Structures are encoded as 3Di+AA strings; SIMD-accelerated all-vs-all ungapped alignment yields a minimum-spanning guide tree, then parallel progressive alignment of AA and 3Di profiles (PSSMs) uses Gotoh affine-gap dynamic programming. msa2lddt scores any MSTA by per-column LDDT, and optional refinemsa re-splits and re-aligns to maximize LDDT.

  • Homstrad families with manually curated reference structure alignments (May 2024 release) · n = 1,032 families
  • AFDB cluster benchmark: AlphaFold2 multidomain clusters with TED/CATH annotations, same CATH topology but distinct homologous superfamilies · n = 150 clusters of 20 structures (50 each with 2, 3 or 4 domains)
  • Largest AFDB cluster with mean length >300 aa, subsampled at sizes from 10 to 100,000 structures for the scaling test · n = 533,657 members · UniProt A0A258JAY9 (cluster representative)
  • Calmodulin structures with a flexible linker between N- and C-lobes (RCSB PDB) · PDB 1jfj, 1ncx, 2sas, 3cln
  • Globin structures including apo and nitric oxide-bound myoglobins (RCSB PDB) · sperm whale (myoglobins 2eb8, 5vzq); others not stated · PDB 1h97, 1hho, 1mbn, 1oj6, 1urv, 2eb8, 2hhb, 2oif, 5vzq
  • Odorant-binding proteins, including monomeric pig OBP and domain-swapped dimeric cow OBP (RCSB PDB) · pig (1a3y), cow (1g85); others not stated · PDB 1a3y, 1g85, 3fiq, 3zq3, 5ngh
  • Flaviviridae E, E1 and E2 glycoprotein structures, alignments and reference trees from Mifsud et al. · Flaviviridae (viruses) · Zenodo 10.5281/zenodo.10616318
Methods & tools
FoldMason (easy-msa, msa2lddt, refinemsa modules), Foldseek 3Di+AA structural alphabet (used as internal library), TM-align, US-align v20240730, mTM-align v20220104, MUSTANG v3.2.4, Matt v1.0, Caretta-shape (commit 998125b), MAFFT v7.520, MUSCLE v5.1, FAMSA v2.2.2-7eb7612, Clustal Omega v1.2.4, T-Coffee aln_compare (sum-of-pairs, TC and CS scores), trimAl v1.4.rev15, IQ-Tree v2.3.0 with ModelFinder, Quartet R library v1.2.6 (tqDist), UCSF ChimeraX 1.6.1; ggtree v3.12
Software & environment
FoldMason commit 92b1690 (GPL-3.0); comparators at fixed versions, e.g. US-align v20240730, mTM-align v20220104, MAFFT v7.520, MUSCLE v5.1, Caretta-shape commit 998125b; all run with default parameters; IQ-Tree v2.3.0, trimAl v1.4.rev15.
Assay conditions & compute
One AMD EPYC 7702P server (64 cores, 1,024 GB RAM); Homstrad and AFDB cluster benchmarks single-threaded; scaling test used 64 cores for FoldMason and FAMSA, single thread for US-align.
Reproducibility gaps
Some glycoprotein phylogeny inputs came via private communication rather than the public Zenodo record; webserver uploads currently limited to PDB format (mmCIF support pending).
Limitations (6)
  • Homstrad reference alignments are high quality but cover mostly small families of short, single-domain proteins, so they poorly represent typical protein families; this motivated the reference-free AFDB benchmark. · stated by the authors
  • LDDT-maximizing iterative refinement is off by default because of its computational cost; with 100 iterations it ran 38x slower than FAMSA2 for a 1.47% average LDDT gain. · stated by the authors
  • Inputs must be 3D structures; the authors note structure prediction is slow, especially for long proteins, and propose but have not yet implemented ProstT5-based 3Di prediction from sequence. · stated by the authors
  • Protein complexes are not yet supported; extension via Foldseek-Multimer and tighter Foldseek webserver integration are described as future work. · stated by the authors
  • The AFDB benchmark ranks tools by average LDDT, the same score FoldMason uses to pick first-round pairwise alignments and to drive refinement, which may favor FoldMason relative to reference-based evaluation. · noted in this summary
  • US-align's scaling failure is reported inconsistently: the main text says it did not complete 5000 structures within 2 days, while the methods say it was stopped after 214 hours on the 500-structure subset. · noted in this summary

Read this if you need to align thousands of predicted protein structures, assess MSTA quality without a reference, or build phylogenies for proteins too divergent for sequence alignment.

Citations
117
FWCI
114.92
Field percentile
top 1%

Abstract

Protein structure is conserved beyond sequence, making multiple structural alignment (MSTA) essential for analyzing distantly related proteins. Computational prediction methods have vastly extended our repository of available proteins structures, requiring fast and accurate MSTA methods. Here, we introduce FoldMason, a progressive MSTA method that leverages the structural alphabet from Foldseek, a pairwise structural aligner, for multiple alignment of hundreds of thousands of protein structures, exceeding alignment quality of state-of-the-art methods, while two orders of magnitudes faster than other MSTA methods. FoldMason computes confidence scores, offers interactive visualizations, and provides essential speed and accuracy for large-scale protein structure analysis in the era of accurate structure prediction. Using Flaviviridae glycoproteins, we demonstrate how FoldMasons MSTAs support phylogenetic analysis below the twilight zone. FoldMason is free open-source software: foldmason.foldseek.com and webserver: search.foldseek.com/foldmason.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISEdgar '24Hong '21Trinquier '25Wang '26Malik '26Gudyś '26
73% semantic matchBioinformatics
R. C. Edgar, Igor Tolstoy·bioRxiv·Oct 26, 2024· 5 citations