Metagenomic-Scale Analysis of the Predicted Protein Structure Universe

Jingi Yeo, Yuanyuan Han, Nicola Bordin, Andy M. Lau, Shaun M. Kandathil, Hyunbin Kim, Eli Levy Karin, Milot Mirdita, David T. Jones, Christine Orengo, Martin SteineggerSSRN 2025 · DOI

TL;DR

This paper tackles the challenge of understanding the vast, uncharted territory of protein structures predicted from metagenomic data. By combining and clustering over 800 million predicted protein structures, the authors reveal that while entirely new protein folds are rare, the real novelty lies in the unprecedented combinations of existing domains, offering a deeper understanding of protein evolution and function in diverse environments.

What Problem Does This Paper Address?

The advent of powerful protein structure prediction tools like AlphaFold2 and ESMFold has led to an explosion of predicted protein structures, creating massive databases like the AlphaFold Protein Structure Database (AFDB) and the ESM Metagenomic Atlas (ESMatlas). While these databases offer unprecedented coverage, the sheer volume and the inclusion of structures from uncultured microbes (metagenomics) present a significant challenge: how do we organize, interpret, and extract meaningful biological insights from this deluge of data? Specifically, the authors aim to explore the structural diversity within this vast dataset, identify novel protein structures and domain combinations, and understand their distribution across different environments and taxa. The ESMatlas, in particular, represents a largely underexplored "dark matter" of the protein universe, and this work seeks to illuminate its potential.

Methods

The core of this study involves a multi-step computational pipeline:

  1. Dataset Combination: The researchers combined the AlphaFold Protein Structure Database (AFDB) and the ESM Metagenomic Atlas (ESMatlas) to create a comprehensive dataset of 821 million predicted protein structures, termed AFESM.
  2. Sequence Clustering: MMseqs2 was used for sequence-based clustering to reduce redundancy and handle the high proportion of protein fragments in the ESMatlas. This step identified representative proteins for each cluster.
  3. Structure Clustering: Foldseek was employed for structure-based clustering of the representative proteins. This allowed for the identification of structurally similar proteins, even if their sequences were not highly similar (the "twilight zone").
  4. Annotation and Analysis: The resulting clusters were annotated with taxonomic labels (using MMseqs2 taxonomy) and biome metadata (from the MGnify database). This enabled the researchers to analyze the distribution of protein structures across different organisms and environments.
  5. Domain Identification and Novelty Assessment: For ESM-only clusters, domains were identified using a consensus-based approach. These domains were then searched against databases like CATH and Alldoms to identify novel folds and novel domain combinations.
  6. Multi-Domain Protein (MDP) Analysis: The study specifically investigated multi-domain protein architectures to identify novel combinations of known domains.

Datasets

  • AlphaFold Protein Structure Database (AFDB): Provided over 214 million predicted protein structures.
  • ESM Metagenomic Atlas (ESMatlas): Contributed over 600 million predicted protein structures from uncultured microbes.
  • Combined Dataset (AFESM): An 821-million-entry dataset created by merging AFDB and ESMatlas.
  • CATH Database: Used for structural similarity searches and domain classification.
  • Alldoms Database: A comprehensive database of CATH, SCOP, and ECOD representatives.
  • TED (Encyclopedia of Domains): Used for domain identification and classification.
  • MGnify Database: Provided biome annotations for metagenomic sequences.
  • GTDB (Genome Taxonomy Database) and UniRef90: Used for taxonomic labeling.

Key Insights and Ideas

  • Novelty in Combinations, Not Folds: The study strongly suggests that the primary source of structural novelty in metagenomic data is not the discovery of entirely new protein folds, but rather the unprecedented combinations of existing protein domains.
  • Vast Underexplored Protein Universe: The ESMatlas, in particular, reveals a significant number of protein structures and domain combinations not observed in existing databases, highlighting the vastness of the uncharacterized protein universe.
  • Environmental Adaptation: Protein structures show clear adaptations to specific environmental niches, especially in extreme environments like thermal springs, hypersaline conditions, and glaciers, with distinct taxonomic groups dominating these clusters.
  • Limitations of Current Predictors: While ESMFold is fast and accurate for known structures, it struggles to produce high-quality models for structurally uncharacterized regions, suggesting limitations in its ability to explore novel sequence space.
  • Importance of Metagenomics: Metagenomic data is crucial for illuminating underexplored regions of the protein structural universe, revealing diversity that would be missed by focusing solely on culturable organisms.

Key Results

  • The AFESM dataset, comprising 821 million predicted protein structures, was clustered into 5.12 million non-singleton structural clusters.
  • 11,941 unique multi-domain architectures were identified, meaning combinations of domains not previously observed.
  • Of these, 5,203 novel multi-domain protein (MDP) architectures were found to contain at least two distinct H-level CATH categories.
  • 431 out of 530 significantly overrepresented H-level categories in the "Novel" MDP set were found to be overrepresented, indicating greater domain diversity in ESM-only proteins.
  • Four membrane-associated domains were among the top 10 significantly overrepresented H-level domains in the Novel set, suggesting their importance in novel functional combinations.
  • While AlphaFold2 could rescue some low-quality ESMFold predictions, only one non-symmetrical new fold was identified that was not already present in TED.

Limitations and Future Work

  • Fragment Handling: The AFESM dataset contains a high fraction of fragments (56%), which can lead to missed entries, singletons, and misassigned members. Future work could focus on improving the handling of fragmented proteins.
  • Taxonomic Prediction Discrepancies: Discrepancies between reference databases (GTDB vs. UniRef90) and metagenomic sampling biases can affect taxonomic assignments, particularly for Eukaryota and Viruses.
  • Biome Annotation Completeness: Inconsistent or incomplete biome metadata, and sampling biases in certain biomes, hinder the discovery of structures specific to particular environments.
  • CATH Coverage: The analysis of multi-domain protein architectures is limited by incomplete CATH coverage, potentially excluding MDPs with fewer than two CATH annotations.
  • Domain Order and Copy Number: The study did not consider domain order, copy number, or spacing, which can influence domain function in multi-domain contexts.
  • Future Work: The authors suggest leveraging genomic context and multimer structure prediction methods to extend analyses to multimeric complexes. They also propose implementing biome search functionality in their webserver. The large pool of fragments and low-quality structures may harbor additional structural novelty discoverable with more robust prediction methods or full-length sequences.

Summary updated

BibTeX
@misc{yeo2025metagenomic,
  author = {Jingi Yeo and Yuanyuan Han and Nicola Bordin and Andy M. Lau and Shaun M. Kandathil and Hyunbin Kim and Eli Levy Karin and Milot Mirdita and David T. Jones and Christine Orengo and Martin Steinegger},
  title = {{Metagenomic-Scale Analysis of the Predicted Protein Structure Universe}},
  year = {2025},
  doi = {10.2139/ssrn.5281078},
  howpublished = {SSRN},
}
Figure