2026-09-01

タンパク質間相互作用の界面を検索・分類する Foldseek-Interface

タンパク質間相互作用は、細胞内のきわめて多様な機能を媒介しており、そのため結合様式も多岐にわたる。近年、タンパク質構造の全体像を網羅的に把握し分類しようとする大規模な取り組みが進んでいる一方で、相互作用界面の構造多様性をインタラクトーム規模で評価し、クラスタリングするための同等の手法はこれまで存在しなかった。本研究では、三次元の相互作用界面構造を検索可能な配列へと変換することで、タンパク質相互作用界面の高速なアラインメントとクラスタリングを可能にする手法、Foldseek-Interfaceを開発した。本手法は、最先端の既存ツールと同等の精度を示しながら、最大230倍高速に動作する。Foldseek-InterfaceをPDBに登録されているすべての生物学的アセンブリに適用した結果、310万の二量体を77,167個の相互作用界面クラスターに分類した。さらに、このリソースを用いて、相互作用界面の多様性、進化、および病原体による分子擬態の特徴を解析した。また、予測されたタンパク質複合体構造のデータベースにFoldseek-Interfaceを適用することで、これまで知られていない可能性のある新規な相互作用界面タイプを迅速に見いだすことができ、今後の実験的検証に値する候補を抽出できることを示した。Foldseek-Interfaceおよび相互作用界面クラスターのリソースは、検索用Webサーバー（https://search.foldseek.com/interface）および探索用Webサーバー（https://interface.foldseek.com）として無償で公開されている。

インストール

WSLで AVX2版 (GPU build)をダウンロードしてテストした｡

Github

# Linux AVX2 build (check using: cat /proc/cpuinfo | grep avx2)
wget https://mmseqs.com/foldseek/foldseek-linux-avx2.tar.gz; tar xvzf foldseek-linux-avx2.tar.gz; export PATH=$(pwd)/foldseek/bin/:$PATH

# Linux AVX2 & GPU build (req. glibc >= 2.17 and nvidia driver >=525.60.13)
wget https://mmseqs.com/foldseek/foldseek-linux-gpu.tar.gz; tar xvfz foldseek-linux-gpu.tar.gz; export PATH=$(pwd)/foldseek/bin/:$PATH

# Linux ARM64 build
wget https://mmseqs.com/foldseek/foldseek-linux-arm64.tar.gz; tar xvzf foldseek-linux-arm64.tar.gz; export PATH=$(pwd)/foldseek/bin/:$PATH

#mamba (注;2026年9月1日現在のバージョンでは Foldseek-Interfaceには対応していない)
mamba create -n foldseek -y
conda activate foldseek
mamba install -c conda-forge -c bioconda foldseek -y

> foldseek

Foldseek enables fast and sensitive comparisons of large structure sets. It reaches sensitivities similar to state-of-the-art structural aligners while being at least 20,000 times faster.

Please cite:

van Kempen, M., Kim, S.S., Tumescheit, C., Mirdita, M., Lee, J., Gilchrist, C.L.M., Söding, J., and Steinegger, M. Fast and accurate protein structure search with Foldseek. Nature Biotechnology, doi:10.1038/s41587-023-01773-0 (2023)

foldseek Version: e28c8d2045b2603f192f20ffdc81b6084b53cc5d

usage: foldseek <command> [<args>]

Easy workflows for plain text input/output

easy-search Structual search

easy-cluster Slower, sensitive clustering

easy-rbh Find reciprocal best hit

easy-multimercluster Multimer level cluster

easy-interfacecluster Interface level cluster

easy-interfacesearch Interface level search

easy-multimersearch Multimer level search

Main workflows for database input/output

createdb Convert PDB/mmCIF/tar[.gz]/DB files or directory/TSV to a structure DB

search Sensitive homology search

rbh Reciprocal best hit search

cluster Slower, sensitive clustering

structureclusterupdate Update structure clustering with new sequences using structural alignment

multimercluster Multimer level cluster

interfacecluster Interface level cluster

interfacesearch Interface level search

multimersearch Multimer level search

Input database creation

databases List and download databases

createindex Store precomputed index on disk to reduce search overhead

createclusearchdb Build a searchable cluster database allowing for faster searches

Unite and intersect databases

createsubdb Create a subset of a DB from list of DB keys

createinterfacedb Create an interface DB of a dimer DB

createstructinterfacedb Create an interface DB of a dimer DB

createdimerdb Create a dimer DB of a DB

Format conversion for downstream processing

structty Launch StrucTTY viewer for foldseek results

convertalis Convert alignment DB to BLAST-tab, SAM or custom format

compressca Create a new C-alpha DB with chosen compression encoding from a sequence DB

convert2pdb Convert a foldseek structure db to a single multi model PDB/mmCIF file or a directory of PDB/mmCIF files

createmultimerreport Convert complexDB to tsv format

Prefiltering

expandmultimer Re-prefilter to ensure complete alignment between multimers

Alignment

tmalign Compute tm-score

lolalign LoLalign. creating structure alignmetns by optimizing the Local distance log odds (LoL) score

structurealign Compute structural alignment using 3Di alphabet, amino acids and neighborhood information

structurerescorediagonal Compute sequence identity for diagonal

aln2tmscore Compute tmscore of an alignment database

scoremultimer Get multimer level alignments from alignmentDB

Clustering

clust Cluster result by Set-Cover/Connected-Component/Greedy-Incremental

Profile databases

result2profile Compute profile DB from a result DB for both amino acid and 3di

>foldseek easy-interfacecluster -h

usage: foldseek easy-interfacecluster <i:PDB|mmCIF[.gz]> ... <i:PDB|mmCIF[.gz]> <o:clusterPrefix> <tmpDir> [options]

By Sooyoung Cha <ellen2g77@gmail.com>

options: prefilter:

--comp-bias-corr INT Correct for locally biased amino acid composition (range 0-1) [1]

--comp-bias-corr-scale FLOAT Correct for locally biased amino acid composition (range 0-1) [1.000]

--seed-sub-mat TWIN Substitution matrix file for k-mer generation [aa:3di.out,nucl:3di.out]

-s FLOAT Sensitivity: 1.0 faster; 4.0 fast; 7.5 sensitive [4.000]

-k INT k-mer length (0: automatically set to optimum) [0]

--target-search-mode INT target search mode (0: regular k-mer, 1: similar k-mer) [0]

--k-score TWIN k-mer threshold for generating similar k-mer lists [seq:2147483647,prof:2147483647]

--max-seqs INT Maximum results per query sequence allowed to pass the prefilter (affects sensitivity) [300]

--split INT Split input into N equally distributed chunks. 0: set the best split automatically [0]

--split-mode INT 0: split target db; 1: split query db; 2: auto, depending on main memory [2]

--split-memory-limit BYTE Set max memory per split. E.g. 800B, 5K, 10M, 1G. Default (0) to all available system memory [0]

--diag-score BOOL Use ungapped diagonal scoring during prefilter [1]

--exact-kmer-matching INT Extract only exact k-mers for matching (range 0-1) [0]

--mask INT Mask sequences in prefilter stage with tantan: 0: w/o low complexity masking, 1: with low complexity masking [0]

--mask-prob FLOAT Mask sequences is probablity is above threshold [1.000]

--mask-lower-case INT Lowercase letters will be excluded from k-mer search 0: include region, 1: exclude region [1]

--mask-n-repeat INT Repeat letters that occur > threshold in a rwo [6]

--min-ungapped-score INT Accept only matches with ungapped alignment score above threshold [30]

--spaced-kmer-mode INT 0: use consecutive positions in k-mers; 1: use spaced k-mers [1]

--spaced-kmer-pattern STR User-specified spaced k-mer pattern

--local-tmp STR Path where some of the temporary files will be created

--exhaustive-search BOOL Turns on an exhaustive all vs all search by by passing the prefilter step [1]

align:

--sort-by-structure-bits INT sort by bits*sqrt(alnlddt*alntmscore) [1]

-a BOOL Add backtrace string (convert to alignments with mmseqs convertalis module) [0]

--alignment-mode INT How to compute the alignment:

0: automatic

1: only score and end_pos

2: also start_pos and cov

3: also seq.id [0]

--alignment-output-mode INT How to compute the alignment:

0: automatic

1: only score and end_pos

2: also start_pos and cov

3: also seq.id

4: only ungapped alignment

5: score only (output) cluster format [0]

-e DOUBLE List matches below this E-value (range 0.0-inf) [1.000E+07]

--min-seq-id FLOAT List matches above this sequence identity (for clustering) (range 0.0-1.0) [0.000]

--min-aln-len INT Minimum alignment length (range 0-INT_MAX) [0]

--seq-id-mode INT 0: alignment length 1: shorter, 2: longer sequence [0]

--alt-ali INT Show up to this many alternative alignments [0]

-c FLOAT List matches above this fraction of aligned (covered) residues (see --cov-mode) [0.000]

--cov-mode INT 0: coverage of query and target

1: coverage of target

2: coverage of query

3: target seq. length has to be at least x% of query length

4: query seq. length has to be at least x% of target length

5: short seq. needs to be at least x% of the other seq. length [0]

--max-rejected INT Maximum rejected alignments before alignment calculation for a query is stopped [2147483647]

--max-accept INT Maximum accepted alignments before alignment calculation for a query is stopped [2147483647]

--gap-open TWIN Gap open cost [aa:10,nucl:10]

--gap-extend TWIN Gap extension cost [aa:1,nucl:1]

--min-assigned-chains-ratio FLOAT Minimum ratio of assigned chains out of all query chains > thr [0.0,1.0] [0.000]

--monomer-include-mode INT Monomer Complex Inclusion 0: include monomers, 1: NOT include monomers [0]

--expand-multimer-evalue DOUBLE E-value threshold for multimer chain expansion (range 0.0-inf) [1.000E+04]

clust:

--cluster-mode INT 0: Set-Cover (greedy)

1: Connected component (BLASTclust)

2,3: Greedy clustering by sequence length (CDHIT) [0]

--max-iterations INT Maximum depth of breadth first search in connected component clustering [1000]

--similarity-type INT Type of score used for clustering. 1: alignment score 2: sequence identity [2]

--set-mode BOOL 0: Cluster by each entry

1: Cluster by set [0]

kmermatcher:

--weights STR Weights used for cluster priorization

--cluster-weight-threshold FLOAT Weight threshold used for cluster priorization [0.900]

profile:

--mask-profile INT Mask query sequence of profile using tantan [0,1] [1]

--e-profile DOUBLE Include sequences matches with < E-value thr. into the profile (>=0.0) [1.000E-01]

--wg BOOL Use global sequence weighting for profile calculation [0]

--filter-msa INT Filter msa: 0: do not filter, 1: filter [1]

--filter-min-enable INT Only filter MSAs with more than N sequences, 0 always filters [0]

--max-seq-id FLOAT Reduce redundancy of output MSA using max. pairwise sequence identity [0.0,1.0] [0.900]

--qid STR Reduce diversity of output MSAs using min.seq. identity with query sequences [0.0,1.0]

Alternatively, can be a list of multiple thresholds:

E.g.: 0.15,0.30,0.50 to defines filter buckets of ]0.15-0.30] and ]0.30-0.50] [0.0]

--qsc FLOAT Reduce diversity of output MSAs using min. score per aligned residue with query sequences [-50.0,100.0] [-20.000]

--cov FLOAT Filter output MSAs using min. fraction of query residues covered by matched sequences [0.0,1.0] [0.000]

--diff INT Filter MSAs by selecting most diverse set of sequences, keeping at least this many seqs in each MSA block of length 50 [1000]

--pseudo-cnt-mode INT use 0: substitution-matrix or 1: context-specific pseudocounts [0]

--profile-output-mode INT Profile output mode: 0: binary log-odds 1: human-readable frequencies [0]

--num-iterations INT Number of iterative profile search iterations (0: auto (select optimal), 1: default, 1-n), N≥2 exactly N) [1]

misc:

--mask-bfactor-threshold FLOAT mask residues for seeding if b-factor < thr [0,100] [0.000]

--save-res-index BOOL Saves PDB residue indices of all residues in the input structures [0]

--input-format INT Format of input structures:

0: Auto-detect by extension

1: PDB

2: mmCIF

3: mmJSON

4: ChemComp

5: Foldcomp [0]

--input-compression-format INT Compression format of input structures:

0: Auto/File-ending based

1: gzip

2: zstd [0]

--file-include STR Include file names based on this regex [.*]

--file-exclude STR Exclude file names based on this regex [^$]

--tmscore-threshold FLOAT accept alignments with a tmsore > thr [0.0,1.0] [0.000]

--tmscore-threshold-mode INT 0: alignment, 1: query 2: target length [0]

--lddt-threshold FLOAT accept alignments with a lddt > thr [0.0,1.0] [0.200]

--alignment-type INT How to compute the alignment:

0: 3di alignment

1: TM alignment

2: 3Di+AA

3: LoL alignmnet [2]

--exact-tmscore INT turn on fast exact TMscore (slow), default is approximate [0]

--prefilter-mode INT prefilter mode: 0: kmer/ungapped 1: ungapped, 2: nofilter, 3: ungapped&gapped [0]

--tmalign-hit-order INT order hits by 0: (qTM+tTM)/2, 1: qTM, 2: tTM, 3: min(qTM,tTM) 4: max(qTM,tTM) [0]

--tmalign-fast INT turn on fast search in TM-align [1]

--lolalign-multidomain INT MultiDomain Mode LoLalign [1]

--candidate-seeds INT Number of candidate seeds to consider for expansion [10]

--refine-seeds INT Whether to refine seeds by re-aligning top candidate seeds and picking the best one for expansion [3]

--allow-deletion BOOL Allow deletions in a MSA [0]

--cluster-search INT first find representative then align all cluster members [0]

--interface-lddt-threshold FLOAT accept alignments with a lddt > thr [0.0,1.0] [0.000]

--chain-tm-threshold FLOAT accept alignments with a tmsore > thr [0.0,1.0] [0.000]

--multimer-tm-threshold FLOAT accept alignments with a multimer tmsore > thr [0.0,1.0] [0.400]

--min-aligned-chains INT save alignments with at least n chain aligned between query and target [2]

--distance-threshold FLOAT Residues with C-alpha below this threshold will be part of interface [10.000]

--min-interface-residues-perchain INT save dimer/interface if there are N numbers of residues per chain [4]

common:

--gpu INT Use GPU (CUDA) if possible [0]

--prostt5-model STR Path to ProstT5 model

--threads INT Number of CPU-cores used (all by default) [16]

-v INT Verbosity level: 0: quiet, 1: +errors, 2: +warnings, 3: +info [3]

--sub-mat TWIN Substitution matrix file [aa:3di.out,nucl:3di.out]

--max-seq-len INT Maximum sequence length [65535]

--db-load-mode INT Database preload mode 0: auto, 1: fread, 2: mmap, 3: mmap+touch [0]

--compressed INT Write compressed output [0]

--gpu-server INT Use GPU server [0]

--gpu-server-wait-timeout INT Wait for GPU server for 0: don't wait -1: no wait limit: >0 this many seconds [600]

--remove-tmp-files BOOL Delete temporary files [1]

--force-reuse BOOL Reuse tmp filse in tmp/latest folder ignoring parameters and version changes [0]

--mpi-runner STR Use MPI on compute cluster with this MPI command (e.g. "mpirun -np 42")

expert:

--chain-name-mode INT Add chain to name:

0: auto

1: always add

[0]

--model-name-mode INT Add model to name:

0: auto

1: always add

[0]

--write-mapping INT write _mapping file containing mapping from internal id to taxonomic identifier [0]

--write-foldcomp INT write _fcz Foldcomp database [0]

--coord-store-mode INT Coordinate storage mode:

1: C-alpha as float

2: C-alpha as difference (uint16_t) [2]

--write-lookup INT write .lookup file containing mapping from internal id, fasta id and file number [1]

--taxon-list STR Taxonomy ID, possibly multiple values separated by ','

--blocklen INT Block length for forward-backward [16]

--multimer-report-mode INT Complex report mode:

0: No report

1: Write complex report [1]

examples:

#Clustering of PDB files

foldseek easy-interfacecluster examples/ result tmp

# Cluster output

# - result_rep_seq.fasta: Representatives

# - result_cluster.tsv: Adjacency list

# Important parameter: --cov-mode and -c

# --cov-mode

# 0 1 2

# Q: MAVGTACRPA 60% IGN 60%

# T: -AVGTAC--- 60% 100% IGN

# -c 0.7 - + -

# -c 0.6 + + +

references:

- Kim, W., Mirdita, M., Levy Karin, E., Gilchrist, C.L.M., Schweke, H., Söding, J., Levy, E., and Steinegger, M. Rapid and sensitive protein complex alignment with Foldseek-Multimer. Nature Methods, doi:10.1038/s41592-025-02593-7 (2025)

> foldseek easy-interfacesearch -h

usage: foldseek easy-interfacesearch <i:PDB|mmCIF[.gz]> ... <i:PDB|mmCIF[.gz]>|<i:stdin> <i:targetFastaFile[.gz]>|<i:targetDB> <o:outputFileName> <tmpDir> [options]