AWSEM-Suite: a protein structure prediction server based on template-guided, coevolutionary-enhanced optimized folding landscapes

Shikai Jin, Vinícius G. Contessoto, Mingchen Chen, Nicholas P. Schafer, Wei Lu, Xun Chen, Carlos Bueno, Arya Hajitaheri, Brian J. Sirovetz, Aram Davtyan, Garegin A. Papoian, Min-Yeh Tsai, Peter G. WolynesNucleic Acids Res 2020 · DOI

TL;DR

Predicting protein 3D structures is crucial but challenging. AWSEM-Suite is a new web server that combines physics-based modeling with evolutionary information and template guidance to achieve accurate protein structure predictions, even for proteins lacking clear structural templates. It demonstrated strong performance in the CASP13 competition, highlighting its utility for researchers.

What Problem Does This Paper Address?

Despite advances in experimental techniques, determining the 3D structure of every protein remains a significant bottleneck in biological research. This "sequence-to-structure gap" hinders our understanding of protein function. While computational methods have improved, accurately predicting protein structures, especially for novel folds (free modeling), is still a major challenge. Existing template-based methods require significant sequence similarity to known structures, and purely physics-based simulations are computationally expensive and difficult to tune. The paper addresses the need for a robust, accessible tool that can predict protein structures effectively, particularly in scenarios where traditional template-based approaches fall short.

Methods

AWSEM-Suite employs a coarse-grained protein force field called AWSEM (Associative memory, Water-mediated, Structure and Energy Model). This model is rooted in neural network ideas and optimized using energy landscape theory. Key features include:

  • Coarse-Grained Model: Each amino acid is represented by a few "force centers," allowing for significantly faster simulations compared to all-atom models.
  • Physically Motivated Potentials: The model incorporates terms that reflect fundamental physical and chemical properties of proteins, including backbone geometry and hydrogen bonding.
  • Knowledge-Based Local Structure Biasing: An "associative memory" term imposes local structural biases based on overlapping fragments.
  • Template-Guided Restraints: The server can utilize information from homologous protein structures to guide predictions.
  • Co-evolutionary Restraints: It incorporates predicted residue-residue contacts derived from analyzing co-variation patterns in multiple sequence alignments. This is a powerful technique for inferring spatial proximity of residues.
  • Integrative Approach: AWSEM-Suite combines these elements, allowing users to provide template information, co-evolutionary data, or both, to enhance prediction accuracy.
  • Simulation Protocol: The process involves generating an extended conformation, followed by an initialization simulation, and then an annealing simulation to find stable structures.

Datasets

The paper evaluates AWSEM-Suite using data from the CASP13 (Critical Assessment of Structure Prediction) experiment. This involved predicting the structures of 111 protein domains across various categories: TBM-hard, TBM-easy, FM (Free Modeling), and TBM/FM. The performance is benchmarked against other state-of-the-art methods participating in CASP13.

Key Insights and Ideas

  • Synergy of Physics and Evolution: The paper strongly suggests that combining physics-based energy landscapes with evolutionary information (co-evolutionary restraints and templates) is more powerful than either approach alone. This is a recurring theme in modern protein structure prediction.
  • Effectiveness in Free Modeling: AWSEM-Suite shows particular promise in the "free modeling" category, where traditional template-based methods struggle due to a lack of suitable templates. This indicates its ability to predict novel folds.
  • User-Friendly Interface: The development of a web server with an intuitive interface and visualization tools is crucial for making advanced prediction methods accessible to a broader scientific community.

Key Results

  • CASP13 Performance: AWSEM-Suite ranked eighth among 39 groups in the TBM/FM category of CASP13. When considering all categories, it ranked 13th.
  • Target T0958-D1: For this specific target, AWSEM-Suite achieved the highest GDT-TS score (74.03) among all entries.
  • Target T0970-D1: AWSEM-Suite submitted the second-best prediction for this target, with a GDT-TS score of 64.41.
  • GDT-TS Correlation: The GDT-TS scores of AWSEM-Suite predictions showed a positive correlation (Pearson r = 0.35) with the percentage of residues predicted as helix and strand in the crystal structure.

Limitations and Future Work

  • Monomeric Proteins Only: The current version of AWSEM-Suite is limited to predicting the structures of monomeric proteins.
  • Future Work: The authors suggest that extending the server to handle protein complexes and incorporating more advanced deep learning techniques could further enhance its predictive capabilities. The paper also implies that continued refinement of the force field and simulation protocols will be beneficial.

Summary updated

BibTeX
@article{jin2020awsem,
  author = {Shikai Jin and Vinícius G. Contessoto and Mingchen Chen and Nicholas P. Schafer and Wei Lu and Xun Chen and Carlos Bueno and Arya Hajitaheri and Brian J. Sirovetz and Aram Davtyan and Garegin A. Papoian and Min-Yeh Tsai and Peter G. Wolynes},
  title = {{AWSEM-Suite: a protein structure prediction server based on template-guided, coevolutionary-enhanced optimized folding landscapes}},
  journal = {Nucleic Acids Res.},
  year = {2020},
  doi = {10.1093/nar/gkaa356},
}
Figure