Structural Bioinformatics of Proteins: predicting the tertiary and quaternary structure of proteins from sequence

Joan Planas-Iglesias, Jaume Bonet, Manuel Alejandro Marín-López, Elisenda Feliu, Attila Gürsoy, Baldomero OlivaResearch 2012 · OpenAlex

TL;DR

This paper provides a comprehensive roadmap for bridging the "sequence-structure gap" in proteomics. It details how computational methods—ranging from homology modeling to integrative assembly—can predict the 3D shapes of individual proteins and the complex "social networks" of interactions they form, even when experimental data is sparse.

What Problem Does This Paper Address?

In the post-genomic era, we are drowning in sequence data but starving for structural information. While sequencing a genome is now relatively cheap and fast, determining the 3D structure of a protein experimentally (via X-ray crystallography or NMR) remains expensive and time-consuming. This creates a massive "gap" because a protein's function is dictated by its shape.

Furthermore, proteins rarely act in isolation. They form intricate quaternary structures—complexes and macro-assemblies—to perform biological work. Predicting how these pieces fit together is a monumental challenge, especially when the individual components might change shape upon binding. The authors tackle the dual challenge of predicting individual folds (tertiary structure) and multi-protein assemblies (quaternary structure) from sequence data alone.

Methods

The paper categorizes structural prediction into three main tiers based on the available information:

  1. Tertiary Structure Prediction:
    • Homology Modeling: The "gold standard" when a similar protein (template) with a known structure exists. It assumes that structure is more conserved than sequence.
  2. Threading (Fold Recognition): Used when sequence similarity is low. It "threads" a sequence through a library of known structural folds to find the best fit.
  3. Ab Initio (De Novo): Predicting structure from first principles (physics and statistics) without a template. This is computationally expensive and used for entirely new folds.
  4. Quaternary Structure Prediction (Docking):
    • Template-based: Modeling a complex based on a known homologous interaction (an "interolog").
  5. Protein-Protein Docking: Computationally "fitting" two unbound proteins together. This involves a global search (often using Fast Fourier Transforms) followed by a "ranking problem" to distinguish true interactions from "decoys."
    • PRISM: A novel approach that focuses on interface similarity. It predicts interactions if the complementary sides of a known interface are structurally similar to the target proteins, even if the proteins themselves aren't globally related.
  6. Integrative Modeling:
    • For massive macro-complexes (like the 450-protein nuclear pore complex), the authors describe a "puzzle-solving" approach. It combines heterogeneous data—low-resolution Electron Microscopy (EM) maps, mass spectrometry, and cross-linking—into a single scoring function to find the most likely assembly.

Datasets

The authors rely on several foundational structural biology resources:

  • Protein Data Bank (PDB): The primary repository for all known 3D structures.
  • Weng Benchmark Decoy Dataset: A "gold standard" set of 124 binary complexes used to test how well docking algorithms can rank near-native structures against thousands of false positives.
  • Specialized Interface Databases: PiBASE, SCOWLP, 3DID, and SCOPPI, which provide curated data on how protein domains interact at the atomic level.

Key Insights and Ideas

  • The "Twilight Zone": The authors highlight that homology modeling becomes highly unreliable when sequence identity drops below 30%. In this "twilight zone," misalignment of gaps becomes a major source of error.
  • Interfaces as Reusable Modules: A key insight is that the number of possible protein interfaces is much smaller than the number of possible protein interactions. Interfaces are often "reused" by proteins with entirely different global folds, which is the logic behind the PRISM method.
  • Split Statistical Potentials: The paper argues that breaking down energy scores into specific components (e.g., "local" environment vs. "3D" distance vs. "interface" specific terms) allows for much better ranking of docking poses than traditional, all-atom physics models, especially when proteins are flexible.
  • Integrative Modeling as a Constraint Satisfaction Problem: For huge assemblies, the method doesn't just "build" the structure; it uses experimental data as "restraints" (e.g., "Protein A must be within 10nm of Protein B"). The final model is the one that violates the fewest restraints.

Key Results

  • Docking Performance: The authors show that their "MixRank" strategy—which alternates between different scoring functions—and split statistical potentials (like $E_{pair}$ and $E_{S3DC}$) outperform atomistic-detailed functions (like ZRank) in "difficult" docking cases where proteins undergo large conformational changes.
  • Macro-complex Success: The integrative approach successfully elucidated the structure of the nuclear pore complex and chromatin at the megabase scale, tasks that are impossible for traditional crystallography.

Limitations and Future Work

  • The Flexibility Problem: Most docking algorithms still struggle with "induced fit," where proteins change shape significantly upon binding. Rigid-body docking often fails in these cases.
  • Template Dependency: Despite the growth of the PDB, many proteins still lack suitable templates, leaving researchers reliant on less accurate ab initio methods.
  • Data Quality in Integrative Modeling: The accuracy of large-scale models is strictly limited by the quality of the "restraints" provided by experimentalists. If the input data (like cross-linking) is noisy, the resulting model may be ambiguous.
  • Refinement: There is a persistent need for better "model refinement" techniques to fix small structural distortions that occur during the initial modeling phase.

Summary updated

Abstract

may overcome the lack of experimental structural data.Models obtained using these approaches may be further assessed, in order to ensure that the inferred 3D structure contains no errors (see section 4).In case of persistence, such errors would damper the deduction of further biological conclusions such as the mode the modeled protein can interact with others.Figure 2 summarizes the different steps and strategies that can be exploited to achieve these objectives. Homology modelingHomology or comparative modeling techniques are those devoted to the prediction and construction of the 3D conformation of proteins.These methods are based on the assumption that structural features in proteins are more conserved than its sequences.Thus, two proteins with enough sequence similarity will fold in a similar way and share the same conformation in space.The process through which a tertiary structure is assigned to a given sequence is carried out in three steps, namely: template identification, template alignment, and model building.Finally, the produced model should be assessed (see section 4).
BibTeX
@inproceedings{planasiglesias2012structural,
  author = {Joan Planas-Iglesias and Jaume Bonet and Manuel Alejandro Marín-López and Elisenda Feliu and Attila Gürsoy and Baldomero Oliva and Baldo Oliva},
  title = {{Structural Bioinformatics of Proteins: predicting the tertiary and quaternary structure of proteins from sequence}},
  booktitle = {Research},
  year = {2012},
}

Loading related content...

Figure