Sequence-based Clustering


The amino acid sequence of a protein directs its folding into shapes that enable specific functions. For most of the proteins in cells, protein folding is a rapid and in most cases repeatable process (Anfinsen, 1973) suggesting that protein sequences have the necessary information to fold into functional proteins and that each protein sequence forms a characteristic structure. While local regions of the protein may adopt slightly different conformations in different biological contexts, the overall structure remains the same. In a few exceptions a protein may adopt a completely different shape in the presence of a specific environment or binding partner(s). Research directed towards predicting protein structure from sequence has been ongoing for more than 50 years. Recently, our ability to compute the 3D shapes of proteins using their amino acid sequence has made tremendous progress by applying machine learning techniques to the archived experimental structural data in the PDB (Baek et al., 2021, Jumper et al., 2021).

When exploring PDB structures, the level of similarity between the amino acid sequences of two or more proteins can be used to infer their structural and functional similarity (Sander and Schneider, 1991). Protein sequences that are 100% identical to each other belong to the same protein, but high levels of sequence identity (e.g., >90%) is also indicative of the same protein, perhaps with a few mutations or variations due to different sources of the protein. Lower levels of sequence similarity between protein sequences may indicate some relationship between their structures and functions. The threshold of sequence similarity that indicates structural homology depends on the length of the alignment. As a rule of thumb for protein sequences that are longer than 100 amino acids, >25% sequence identity indicates similar structure and function (Sander and Schneider, 1991).

Sequences and Sequence Clusters

As the single worldwide repository for macromolecular structures, the Protein Data Bank holds many structures with the same or similar sequence and structures. This redundancy enables deep understanding of the biology of these proteins. However, some bioinformatics analyses may benefit from grouping these redundant sequences and structures. For example, all protein structures of the same protein have the exact same sequence. These may be grouped together. Protein sequences where 90% of the sequence is identical is said to have a 90% sequence identity, while proteins whose sequences are only 30% identical have a 30% sequence identity. Grouping proteins into clusters by sequence identity is a way to reduce/remove redundancy in 3D structures (including experimental structures and Computed Structure Models or CSMs). The sequences in a particular cluster are expected to share structural and functional properties depending on the level of sequence identity.

What are Sequence Clusters?

The amino acid sequences of all proteins, whose 3D structures are available from (including experimental structures and CSMs) are grouped at different levels of sequence identity (e.g., 100%, 95%, 90%, 70%, 50% and 30%) to yield sequence clusters. These pre-computed sequence groups are available for exploring the PDB archive and grouping search results.

Why use Sequence Clusters?

Instead of using all sequences of the 3D structures available from for analysis, representative sequences from each of the sequence clusters can be used. Depending on the level of sequence similarity, properties and features of the representative proteins can be extended to other members in the cluster. Using sequence clusters has the following advantage:

  • It reduces the size of the sequence data set of all 3D structures available and can help simplify, optimize, and make their analysis more efficient.
  • Monitoring growth in the non-redundant sequence clusters enables monitoring the variety of structures being deposited to the PDB
  • It can be used to organize sequences from both experimental structures and CSMs to explore evolutionary relationships between specific proteins.


How are protein sequences in the PDB clustered?

Sequence clusters are calculated using the MMseqs2 software (Steinegger and Söding, 2017). Currently, only protein sequences are subject to clustering. The rationale for clustering considers the following points:

  • All protein chains of at least 10 amino acids are included in the clusters.
  • Sequence identity is defined as the percentage of identical residues between the two amino acid sequences in the alignment.
  • The sequence clustering process begins with an all by all comparison of protein sequences in the PDB.
  • Only alignments with sequence identity scores above the threshold (100%, 95%, 90%, 70%, 50% and 30%) and covering at least 80% (-c 0.80) of both sequences are retained.
  • The clustering is run with the following parameters of the MMseqs2 software:
    • The clustering uses --cluster-mode 1, which corresponds to the connected component algorithm.
    • 50% sequence identity and above: computed with easy-linclust
    • Below 50%: computed with easy-cluster and sensitivity is set to (-s 8) for MMseqs2's highest alignment sensitivity for clustering
  • For more details on the procedure, please refer to the mmseqs2 user guide.

Note: The sequence clusters are subject to change over time as new protein sequences continue to be added to the archive.

How to use sequence clusters to explore the 3D structures in

Each week, RCSB PDB computes sequence clusters for all protein sequences available from [including experimental (PDB) structures, and available CSMs]. You can use these pre-computed clusters the following ways:


  • Anfinsen, C. (1973), Science, 181, 223-230; doi: 10.1126/science.181.4096.223
  • Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G. R., Wang, J., Cong, Q., Kinch, L. N., Schaeffer, R. D., Millán, C., Park, H., Adams, C., Glassman, C. R., DeGiovanni, A., Pereira, J. H., Rodrigues, A. V., van Dijk, A. A., Ebrecht, A. C., Opperman, D. J., Sagmeister, T., Buhlheller, C., Pavkov-Keller, T., Rathinaswamy, M. K., Dalwadi, U., Yip, C. K., Burke, J. E., Garcia K. C., Grishin, N. V., Adams, P. D., Read, R. J., Baker, D. (2021). Accurate prediction of protein structures and interactions using a three-track neural network. Science (New York, N.Y.), 373, 871–876; doi: 10.1126/science.abj8754
  • Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S., Ballard, A. J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., Back, T., Petersen, S., Reiman, D., Clancy, E.,, Zielinski, M., Steinegger, M., Pacholska, M., Berghammer, T., Bodenstein, S., Silver, D., Vinyals, O., Senior, A. W., Kavukcuoglu, K., Kohli, P., Hassabis, D. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583–589; doi: 10.1038/s41586-021-03819-2
  • Sander, C., Schneider, R. (1991). Database of homology-derived protein structures and the structural meaning of sequence alignment. Proteins, 9, 56–68; doi: 10.1002/prot.340090107
  • Steinegger, M., Söding, J. (2017). MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat Biotechnol 35, 1026–1028.

Please report any encountered broken links to
Last updated: 5/1/2023