Improving protein tertiary structure prediction by deep learning and distance prediction in CASP14

Jian Liu, Tianqi Wu, Zhiye Guo, Jie Hou, Jianlin ChengbioRxiv 2021 · DOI

TL;DR

This paper presents the MULTICOM system's advancements in protein tertiary structure prediction for CASP14, leveraging deep learning for improved inter-residue distance prediction. Their approach significantly enhances template-free (ab initio) prediction and achieves competitive rankings, demonstrating the power of distance-based methods in tackling complex protein folding challenges.

What Problem Does This Paper Address?

Predicting the 3D structure of a protein from its amino acid sequence is a fundamental challenge in biology. This is crucial for understanding protein function, designing drugs, and developing new therapies. Traditional methods fall into two categories: template-based modeling (TBM), which uses known protein structures as templates, and template-free (ab initio) modeling, which attempts to predict structure from scratch. While TBM is effective when suitable templates exist, it fails for proteins with no known homologs. Ab initio methods have historically struggled with accuracy, especially for larger proteins. The Critical Assessment of protein Structure Prediction (CASP) competitions serve as a benchmark for evaluating these methods. The authors aim to improve protein structure prediction accuracy, particularly for template-free cases, by integrating advanced deep learning techniques for predicting inter-residue distances.

Methods

The MULTICOM system for CASP14 was enhanced in three key areas:

  • Deep-learning based inter-residue distance predictor (DeepDist): This new component uses deep learning to predict the distances between amino acid residues. This is a crucial input for guiding the construction of protein structures, especially in template-free modeling.
  • Enhanced template-based tertiary structure prediction: The TBM pipeline was streamlined by reducing redundant template identification tools and utilizing deeper multiple sequence alignments (MSAs) for template searching.
  • Distance-based model quality assessment (QA): Deep learning was employed to develop QA methods that assess the quality of predicted protein models, leveraging predicted distance maps.

The overall pipeline involves parallel runs of template-based and template-free modeling. For template-free modeling, multiple MSAs are generated and fed into DeepDist to predict distance maps. These maps, along with other features, are then used by tools like DFOLD and trRosetta to generate structural models. The system also incorporates domain parsing to handle multi-domain proteins and a multi-stage model ranking and refinement process.

Datasets

The study primarily utilizes datasets from the CASP14 competition for evaluation. This includes:

  • 92 "all groups" domains: This encompasses both template-based (TBM) and template-free (FM) targets.
  • 54 TBM domains: These are targets with known structural templates in the Protein Data Bank (PDB).
  • 38 FM or FM/TBM domains: These are targets for which no suitable templates were found or recognized.
  • 10 multi-domain targets: Specifically used for evaluating inter-domain structure prediction.

The paper also mentions the use of large sequence databases like Uniclust30, UniRef90, Metaclust, and BFD for generating MSAs.

Key Insights and Ideas

  • Distance prediction as a powerful intermediate: The success of template-free modeling is heavily reliant on the accuracy of predicted inter-residue distances, which in turn depends on the quality of MSAs.
  • Deep learning revolutionizes distance prediction: The introduction of DeepDist, a deep learning-based distance predictor, significantly improved the accuracy of template-free structure prediction compared to previous CASP editions.
  • Template-free modeling is increasingly versatile: The study suggests that distance-based template-free modeling can perform well even on targets that have identifiable templates, blurring the lines between traditional TBM and FM approaches.
  • Model quality assessment remains a challenge: While improvements were made, the QA methods still struggled with targets where the distribution of model quality scores was highly skewed, indicating room for further development.
  • Domain prediction is critical for multi-domain proteins: Accurate identification of domain boundaries is essential for generating high-quality models of complex multi-domain proteins.

Key Results

In the CASP14 experiment, the MULTICOM system achieved notable rankings:

  • 7th out of 146 predictors for overall protein tertiary structure prediction.
  • 3rd out of 136 predictors for inter-domain structure prediction.

Specifically:

  • The MULTICOM human predictor correctly predicted the fold for 82.6% of domains (98% of TBM, 60.5% of FM/TBM) when considering the top model. This increased to 84.8% when considering the best of the top five models.
  • The server predictors showed strong performance, with average TM-scores for "all group" domains around 0.64, substantially higher than the 0.5 threshold for a correct fold.
  • The pure template-free server predictor, MULTICOM-DIST, achieved an average TM-score of 0.513 on FM/TBM domains, comparable to other advanced methods.
  • The performance on TBM targets was generally better than on FM/TBM targets, with QA methods showing a lower loss on TBM domains.

Limitations and Future Work

  • Quality assessment for skewed distributions: The paper acknowledges that the model quality assessment methods performed poorly when only a few good models were predicted for a hard target and the distribution of model quality scores was highly skewed. This suggests a need for more robust QA methods that can handle such scenarios.
  • Accuracy of domain prediction: For some multi-domain targets, incorrect domain prediction led to ranking failures, highlighting the importance of accurate domain boundary detection.
  • Comparison with end-to-end methods: While MULTICOM's distance-based approach showed significant improvements, it did not perform as well as AlphaFold2, which directly predicts 3D structures from MSAs using deep learning. This indicates that pursuing end-to-end prediction via deep learning is a promising future direction.
  • Further integration of distance information: The authors suggest that incorporating template-based distance information into template-free modeling might be beneficial, especially for large multi-domain proteins.

Summary updated

Abstract

Abstract Substantial progresses in protein structure prediction have been made by utilizing deep-learning and residue-residue distance prediction since CASP13. Inspired by the advances, we improve our CASP14 MULTICOM protein structure prediction system in the three main aspects: (1) a new deep-learning based protein inter-residue distance predictor (DeepDist) to improve template-free ( ab initio ) tertiary structure prediction, (2) an enhanced template-based tertiary structure prediction method, and (3) distance-based model quality assessment methods empowered by deep learning. In the 2020 CASP14 experiment, MULTICOM predictor was ranked 7 th out of 146 predictors in protein tertiary structure prediction and ranked 3 rd out of 136 predictors in inter-domain structure prediction. The results of MULTICOM demonstrate that the template-free modeling based on deep learning and residue-residue distance prediction can predict the correct topology for almost all template-based modeling targets and a majority of hard targets (template-free targets or targets whose templates cannot be recognized), which is a significant improvement over the CASP13 MULTICOM predictor. The performance of template-free tertiary structure prediction largely depends on the accuracy of distance predictions that is closely related to the quality of multiple sequence alignments. The structural model quality assessment works reasonably well on targets for which a sufficient number of good models can be predicted, but may perform poorly when only a few good models are predicted for a hard target and the distribution of model quality scores is highly skewed.
BibTeX
@misc{liu2021improving,
  author = {Jian Liu and Tianqi Wu and Zhiye Guo and Jie Hou and Jianlin Cheng},
  title = {{Improving protein tertiary structure prediction by deep learning and distance prediction in CASP14}},
  year = {2021},
  doi = {10.1101/2021.01.28.428706},
  howpublished = {bioRxiv},
}
Figure