Using evolutionary context to classify difficult protein folds

Jimin Pei, R. Dustin Schaeffer, Qian Cong, Nick V. GrishinPLOS Computational Biology 2025 · DOI

TL;DR

This paper introduces DPAM, a new computational tool that combines structural and evolutionary information to better classify protein domains, especially those predicted to have novel folds. By comparing DPAM's classifications with those from the TED project, the study reveals that many "novel" folds are actually highly diverged relatives of known protein families, offering a more nuanced understanding of protein evolution.

What Problem Does This Paper Address?

The rapid advancement of protein structure prediction, particularly with tools like AlphaFold2, has generated millions of predicted protein structures. This has led to the identification of a vast number of potential protein domains that appear to have "novel folds" – structures unlike any previously characterized. Classifying these novel folds is a significant challenge for existing systems that rely heavily on sequence or structural similarity to known folds. This ambiguity makes it difficult to infer function and understand evolutionary relationships. The Encyclopedia of Domains (TED) project has cataloged many such candidate novel-fold (CNF) domains, but their evolutionary context remains unclear. This paper aims to bridge this gap by applying a new method, DPAM, to provide a more comprehensive classification of these challenging domains.

Methods

The core of this work is the application of DPAM (Domain Parser for AlphaFold Models), a novel domain parsing tool. Unlike purely structure-based methods like TED, DPAM integrates multiple sources of information:

  • AlphaFold-derived metrics: Predicted distance maps (distograms) and Predicted Aligned Error (PAE) matrices are used to infer domain boundaries and assess confidence.
  • Sequence profiles: Standard sequence-based information is incorporated.
  • Evolutionary context: DPAM leverages HHpred searches to find distant sequence homologs.
  • Sensitive structural comparison: Tools like DaliLite are used for more robust structural similarity detection compared to faster methods.

DPAM then assigns the parsed domains to the ECOD (Evolutionary Classification of protein Domains) hierarchy, which emphasizes evolutionary relationships. The study compares DPAM's assignments for TED CNF domains with existing ECOD classifications to understand the evolutionary relationships of these potentially novel folds.

Datasets

  • AlphaFold Structure Database (AFDB): Used as the source for predicted protein structures.
  • TED CNF domains: A dataset of 7427 candidate novel-fold domains identified by the TED project.
  • ECOD database: Used as the reference for structural classification and evolutionary relationships.
  • PDB (Protein Data Bank): Used for sequence and structural similarity searches.
  • Pfam database: Used for sequence similarity searches.

The study does not introduce a new benchmark dataset but rather uses existing datasets to perform a comparative analysis.

Key Insights and Ideas

  • "Novel" folds are often diverged relatives: A significant finding is that many TED CNF domains, initially thought to be entirely new folds, are actually highly diverged homologs of known ECOD domains. This suggests that evolution can create significant structural changes within existing fold families.
  • Complementarity of structural and evolutionary approaches: The study highlights that purely structural classification (like TED) and evolutionary classification (like DPAM/ECOD) offer complementary views. TED excels at identifying structurally distinct modules, while DPAM/ECOD can reveal deeper evolutionary connections, even when sequence similarity is low.
  • Domain boundary ambiguity: The comparison between DPAM and TED reveals that defining domain boundaries is inherently challenging. Sometimes TED identifies smaller, potentially functional units within a larger DPAM-defined domain, and vice versa. This reflects the complex, hierarchical nature of protein architecture.
  • Functional refinement through insertions/extensions: Many TED CNF domains embedded within larger DPAM domains appear to be insertions or extensions into existing folds, potentially conferring new functions or modifying existing ones. These are often evolutionarily selected, not random.
  • Consistently defined domains are robust: Domains where both DPAM and TED agree on boundaries are likely evolutionarily stable and represent well-defined protein families with diverse functions.

Key Results

  • 8044 DPAM domains overlap with TED CNF domains: Out of the 7232 proteins analyzed, 8044 DPAM domains showed significant overlap with TED CNF domains.
  • 2490 DPAM domains were well-assigned to ECOD: This subset represents domains with clear evolutionary relationships to known ECOD entries.
  • ~31% of DPAM domains are well-assigned: This indicates a substantial portion of TED CNF domains have identifiable evolutionary links.
  • ~17% of well-assigned DPAM domains show consistent boundaries: 432 domains had high overlap fractions with both DPAM and TED, suggesting robust definitions.
  • ~34% of well-assigned DPAM domains have TED CNF embedded in DPAM: 888 domains showed a high TED overlap fraction and low DPAM overlap fraction, meaning TED CNF domains are often substructures within larger DPAM domains.
  • ~8% of well-assigned DPAM domains have DPAM embedded in TED CNF: 202 domains showed a low TED overlap fraction and high DPAM overlap fraction, meaning DPAM identified smaller, distinct units within TED domains.
  • Sequence vs. Structure: A significant fraction of well-assigned domains (around 60%) showed strong sequence similarity (Best_Seq_Rel > 0.95), but another substantial portion had low sequence similarity (<0.8), underscoring the importance of structural comparisons for detecting distant homology.

Limitations and Future Work

  • Unassigned domains: A large fraction of DPAM domains (around 60%) remained unassigned to ECOD, highlighting the limits of current classification methods and the need for further computational innovation or experimental validation.
  • Ambiguity in domain definition: The study acknowledges the inherent difficulty in defining domain boundaries, as structural, evolutionary, and functional criteria can sometimes conflict.
  • Future work: The authors suggest that integrating DPAM's evolutionary context into TED could significantly improve domain annotation. Further investigation into the functional roles of the identified domain segments is also proposed to understand their contributions to protein dynamics and molecular recognition. The study also points to the need for new strategies to accurately annotate protein domains, especially in regions of the protein universe where current methods fall short.

Summary updated

BibTeX
@article{pei2025using,
  author = {Jimin Pei and R. Dustin Schaeffer and Qian Cong and Nick V. Grishin},
  title = {{Using evolutionary context to classify difficult protein folds}},
  journal = {PLOS Computational Biology},
  year = {2025},
  doi = {10.1371/journal.pcbi.1013773},
}
Figure