Clustering predicted structures at the scale of the known protein universe.

Barrio-Hernandez I, Yeo J, Jänes J, Mirdita M, Gilchrist CLM, Wein T, Varadi M, Velankar S, Beltrao P, Steinegger M. · 2023

Read Clustering predicted structures at the scale of the known protein universe. by Barrio-Hernandez I, Yeo J, Jänes J, Mirdita M, Gilchrist CLM, Wein T, Varadi M, Velankar S, Beltrao P, Steinegger M. (2023) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc10584675.txt.

Introduction

Barrio-Hernandez and co-authors grouped the known protein universe by shape in 2023. Shape decides what a protein does, yet lab reads of shape lag far behind sequence reads. The AlphaFold data bank holds over 214 million predicted shapes. No past tool could sort that mass at speed.

Their method, Foldseek cluster, compares shapes far faster than older tools. It sorted all bank shapes into 2.30 million groups with more than one member. Of those, 31% had no known label, marking likely new structures. Barrio-Hernandez found that most groups are ancient in origin. About 4% seem tied to one species, hinting at new genes or weak predictions. Shape kinship also surfaced between human and microbe proteins.

Catalog record

CREATORBarrio-Hernandez I, Yeo J, Jänes J, Mirdita M, Gilchrist CLM, Wein T, Varadi M, Velankar S, Beltrao P, Steinegger M.
YEAR2023
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC10584675

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Clustering predicted structures at the scale of the known protein universe By Inigo Barrio-Hernandez, Jingi Yeo, Jürgen Jänes, Milot Mirdita, Cameron L M Gilchrist, Tanita Wein, Mihaly Varadi, Sameer Velankar, Pedro Beltrao, Martin Steinegger Published in: Nature DOI: 10.1038/s41586-023-06510-w ABSTRACT Abstract Proteins are key to all cellular processes and their structure is important in understanding their function and evolution. Sequence-based predictions of protein structures have increased in accuracy1, and over 214 million predicted structures are available in the AlphaFold database2. However, studying protein structures at this scale requires highly efficient methods. Here, we developed a structural-alignment-based clustering algorithm—Foldseek cluster—that can cluster hundreds of millions of structures. Using this method, we have clustered all of the structures in the AlphaFold database, identifying 2.30 million non-singleton structural clusters, of which 31% lack annotations representing probable previously undescribed structures. Clusters without annotation tend to have few representatives covering only 4% of all proteins in the AlphaFold database. Evolutionary analysis suggests that most clusters are ancient in origin but 4% seem to be species specific, representing lower-quality predictions or examples of de novo gene birth. We also show how structural comparisons can be used to predict domain families and their relationships, identifying examples of remote structural similarity. On the basis of these analyses, we identify several examples of human immune-related proteins with putative remote homology in prokaryotic species, illustrating the value of this resource for studying protein function and evolution across the tree of life. Subject terms: Structural biology, Protein analysis, Software MAIN Proteins are the major actors in all cellular processes, from the generation of energy to the division of the cell. Knowing their structure is relevant for studying their function, their evolution and potentially for the design of drugs. Although our knowledge of protein sequences has grown considerably over the past years, reaching over hundreds of millions of sequences, the knowledge of their 3D structures has lagged behind owing to the lack of highly scalable experimental methods. Improvements in methods for predicting structure from sequences1,3,4 now enable the scalable prediction of protein structures for the known protein universe. The AlphaFold Protein Structure Database (AFDB) is a publicly available data repository of protein structures and their confidence metrics, predicted using the AlphaFold2 AI system1,2. The AlphaFold-predicted structures have been generally assessed to be of high quality when the predicted local distance difference test (pLDDT) confidence metrics are accounted for, despite remaining inferior to experimentally determined structures5. AlphaFold2 and its predicted structures have now been used for diverse applications, including studies of protein pockets6, prediction of structures of complexes7,8, studies of structural similarity9, novel fold predictions10 and even improvement of genomic annotation11. The large increase in available predicted protein structures has spurred the development of more efficient computational approaches, including structural data file compressions12, methods for pocket predictions13,14 and comparison of protein structures through structural alignments. For the latter, Foldseek has been developed. Foldseek can increase the speed of comparisons of structures by four to five orders of magnitudes relative to previous approaches while maintaining sensitivity15, making it possible to perform structural comparisons at a large scale. Clustering proteins by their structure is a crucial tool for analysing structural databases as it enables the grouping of remotely related proteins. Identifying distant relationships might provide valuable insights into protein structure evolution and function. For example, protein family analysis of the initial release of about 365,000 structures10,16, covering the proteomes of humans and 20 model organisms, suggested that 92% of predicted domains within this set match existing domain superfamilies. However, comparing all 214 million structures against each other using current methods would take approximately 10 years on a 64-core machine. To speed up the process of clustering amino acid sequences, a linear time algorithm, Linclust17, has been proposed to reduce the computational time significantly. However, such methods have yet to be applied to clustering by protein structural similarity. Here, we analysed the AlphaFold Protein Structure Database, which contains predicted structures for 214 million proteins across the tree of life. To be able to explore this resource, we developed a highly scalable structure-based clustering algorithm based on Linclust17 (Methods and Extended Data Fig. 1) that structurally aligns and clusters 52 million structures in 5 days on 64 cores. We clustered the AlphaFold structural database into 2.30 million clusters with 31% of clusters—representing 4% of protein sequences—not matching previously known structural or domain family annotations. We found that 532,478 clusters have representatives present in all of the tree of life and we found several species-specific structural clusters that could contain examples of de novo gene birth events. Finally, we used structural comparisons to predict domain families and their relationships identifying putative remote homologies that expand the evolutionary coverage of previously known families. [Figure: The five-step clustering pipeline for efficiently clustering millions of protein structures using Foldseek’s 3Di alphabet.(1) Protein structures are converted to 3Di sequences and processed through the Linclust workflow. (2) For each sequence, 300 min-hasing k-mers are extracted and sorted. (3) The longest structure is assigned to be the centre of each k-mer cluster. (4) Structural alignment is performed in two stages: first an ungapped alignment based on shared diagonal information is performed, hits are pre-clustered and second the remaining sequences are aligned using Foldseek’s structural Smith-Waterman. (5) The remaining structures meeting alignment criteria are clustered using MMseqs2’s clustering module. After the Linclust step the centroids are successively clustered by three cascaded steps of prefiltering, structural Smith-Waterman alignment and clustering using Foldseek’s search.] STRUCTURE-BASED CLUSTERING OF THE AFDB The AFDB covers over 214 million predicted protein structures and has grown in several stages (Fig. 1a). The initial release focused on 20 key model organisms, while subsequent updates provided predictions for the Swiss-Prot dataset of the Universal Protein Resource18 (UniProt) and proteomes relevant to global health, taken from priority lists compiled by the World Health Organisation. The current update covers most of the TrEMBL dataset of UniProt. The AFDB parses and archives these data and makes them accessible through bulk download options, programmatic access end points and interactive web pages. The programmatic access, in particular, facilitated the integration of AlphaFold models into other biological data repositories, such as Protein Data Bank Europe (PDBe)19, UniProt18, Pfam20, InterPro21 and Ensembl22. [Figure: The AFDB, structural clustering workflow and summary of the clusters.a, The AFDB started as a collaborative effort between EMBL-EBI and DeepMind in 2021. The database grew in multiple stages, with the latest version of 2022 containing over 214 million predicted protein structures and their confidence metrics. b, A two-step approach was used to cluster proteins in the database. First, MMseqs2 was used to cluster 214 million UniProtKB protein sequences (AFDB) on the basis of 50% sequence identity and 90% sequence overlap, resulting in a reduction of the database size to 52 million clusters (AFDB50). For each cluster, the protein with the highest pLDDT score was selected as the representative. Next, using Foldseek, the representative structures were clustered into 18.8 million clusters (Foldseek clusters) without a sequence identity threshold, but still enforcing a 90% sequence overlap and an E-value of less than 0.01 for each structural alignment. As the last step, we removed all sequences labelled as fragments from the clustering, ending up with 2.30 million clusters with at

Read the complete text