DaReUS-Loop: a web server to model multiple loops in homology models

Yasaman Karami, Julien Rey, Guillaume Postic, Samuel Murail, Pierre Tufféry, Sjoerd Jacob de VriesNucleic Acids Res 2019 · OpenReview

TL;DR

Modeling loops in protein homology models is a persistent challenge due to their inherent variability. This paper introduces DaReUS-Loop, a web server that automates the modeling of multiple loops within homology models, offering improved accuracy and a confidence score for the predicted loops. This advancement is crucial for generating more reliable protein structure predictions.

What Problem Does This Paper Address?

Proteins are the workhorses of biology, and understanding their three-dimensional structure is key to understanding their function. While we have millions of protein sequences, the number of experimentally determined structures is significantly smaller. Homology modeling bridges this gap by predicting a protein's structure based on known structures of similar proteins. However, a major hurdle in homology modeling is accurately predicting the "loop regions" – flexible segments of the protein chain that often play critical roles in protein function, interactions, and drug design. These loops are notoriously difficult to model because they are more variable in both sequence and structure compared to the more rigid parts of the protein. Existing methods often struggle with loop modeling, especially within the context of imperfect homology models, leaving this as an "open problem."

Methods

The DaReUS-Loop web server is built upon a previously developed data-based protocol. The core innovation lies in its ability to model multiple loops simultaneously within a homology model, a significant step beyond previous methods that typically focused on single loops or had strict limitations on the number of loops.

Key methodological aspects include:

  • Data-Based Approach: The server leverages existing protein structures to find suitable loop conformations.
  • Parallel Modeling: Unlike previous methods that modeled loops one by one, DaReUS-Loop can model up to 20 loops concurrently. This is a major advancement for complex proteins with multiple flexible regions.
  • Three Modes of Operation:
    • Remodeling: Takes an existing homology model and refines specified loops.
  • Modeling: Fills in missing loop regions in a gapped PDB file using the full protein sequence.
  • Advanced Modeling: Similar to modeling mode but models each loop independently, potentially improving accuracy at the cost of introducing gaps in the final output.
  • GROMACS for Minimization: The final minimization step, which was previously done with the proprietary MODELLER software, has been replaced with the open-source GROMACS package.
  • Confidence Score: The server provides a confidence level for each predicted loop, which correlates with its expected accuracy. This helps users assess the reliability of the predictions.
  • Interactive Visualization: Results are presented through an interactive graphical interface using the NGL Viewer, allowing users to easily visualize and inspect the modeled loops.

Datasets

The DaReUS-Loop web server was validated using benchmark datasets from the CASP11 and CASP12 protein structure prediction competitions. These datasets consist of homology models and their corresponding native loop conformations. The authors specifically used subsets of these targets (setai and setdb) that allowed for fair comparison with other state-of-the-art methods, ensuring that the loops could be predicted by the compared tools.

Key Insights and Ideas

  • Simultaneous Multi-Loop Modeling is a Game-Changer: The ability to model multiple loops in parallel is a significant practical improvement. This addresses the reality that many proteins have several flexible regions that need to be modeled together for a complete and accurate structure.
  • Confidence Scores Enhance Usability: Providing a confidence level for each loop prediction is invaluable. It moves beyond simply presenting models to offering a way to gauge their likely accuracy, empowering users to prioritize and interpret results more effectively.
  • Data-Based Approach Outperforms in Perturbed Scenarios: The success of DaReUS-Loop, a data-based method, in modeling loops within homology models (which are inherently "perturbed" compared to crystal structures) highlights the effectiveness of leveraging existing structural information for these challenging regions.
  • Performance is Competitive: The server demonstrates competitive or superior performance compared to several leading loop modeling tools, particularly in terms of accuracy (flanked RMSD).

Key Results

  • High Accuracy Predictions: DaReUS-Loop achieves high-accuracy loop models (<1 Å flanked RMSD) for 20% of cases and medium-accuracy models (<2 Å) for 53% of cases in the setai subset of the CASP benchmarks.
  • Improved Accuracy: The server's performance is better than several ab initio methods (Rosetta NGK, GalaxyLoop-PS2, RCD+) and data-based methods (LoopIng, Sphinx), with improvements of up to 0.94 Å in flanked RMSD.
  • Effective for Long Loops: The original DaReUS-Loop protocol, which this server implements, showed significant improvements in predicting long loops (15-30 residues).
  • Computational Time: Modeling a single loop takes approximately 20-30 minutes. For proteins with 10 or more loops, the runtime can range from 40 to 120 minutes, depending on loop sizes and cluster load.

Limitations and Future Work

  • Consensus Method: DaReUS-Loop does not explicitly rank the 10 predicted candidates for a given loop. The authors suggest that the Cartesian combination of all loop candidates should be considered, implying that further work is needed to develop effective scoring functions for selecting the best combination of candidates when multiple loops are modeled.
  • Flexibility Score: The authors propose developing a "flexibility score" to reflect the predicted conformational diversity of each loop, which could be useful for assessing the impact of loop conformation on protein complex modeling.
  • Integration of Experimental Data: Future research could integrate experimental data or residue contact predictions as filters to refine loop candidate combinations.
  • Disulfide Bridges: Incorporating knowledge about disulfide bridges into the protocol is another avenue for future development.
  • Template-Based Docking: Combining DaReUS-Loop with template-based docking methods for protein-protein complex prediction is suggested as a promising research direction.

Summary updated

BibTeX
@inproceedings{karami2019dareus,
  author = {Yasaman Karami and Julien Rey and Guillaume Postic and Samuel Murail and Pierre Tufféry and Sjoerd Jacob de Vries},
  title = {{DaReUS-Loop: a web server to model multiple loops in homology models}},
  booktitle = {Nucleic Acids Res},
  year = {2019},
  url = {https://openreview.net/forum?id=w6GC2Eip0G},
}
Figure