TL;DR
This paper introduces DEMO2, a novel method for assembling multi-domain protein structures. By combining deep learning-predicted inter-domain restraints with analogous template alignments, DEMO2 significantly improves the accuracy of full-length protein structure prediction, addressing a key limitation in current state-of-the-art methods.
What Problem Does This Paper Address?
Proteins are often composed of multiple independent folding units called domains. While recent advances like AlphaFold2 have revolutionized single-domain protein structure prediction, accurately modeling the relative orientations and interactions between these domains in multi-domain proteins remains a significant challenge. Many existing methods struggle to predict these inter-domain relationships, leading to inaccurate full-length structures, especially for proteins lacking homologous full-length templates in databases. This gap hinders our understanding of complex biological functions that rely on the precise arrangement of protein domains.
Methods
DEMO2 builds upon its predecessor, DEMO, by integrating several key advancements:
- Deep Learning for Inter-Domain Restraints: DEMO2 uses deep residual convolutional networks (specifically, DeepPotential) to predict crucial inter-domain spatial restraints. These include residue-residue distances, interface contacts, orientations, and hydrogen-bond networks. This is a significant departure from purely template-based approaches.
- Global Template Identification: Instead of just local templates for adjacent domains, DEMO2 identifies "global" templates that encompass multiple domains of the target protein. This provides a more holistic structural context.
- L-BFGS Assembly Simulation: Full-length models are assembled using an L-BFGS (Limited-memory Broyden–Fletcher–Goldfarb–Shanno) optimization algorithm. This method simultaneously optimizes the positions and orientations of all domains, avoiding the potential biases introduced by iterative, pairwise assembly.
- Hybrid Energy Function: The L-BFGS simulation is guided by a hybrid energy function that combines the deep-learning restraints with information from analogous templates and physics-based steric potentials.
- Cryo-EM Data Integration: The method can also incorporate cryo-EM density maps to assist in rigid-body domain assembly and subsequent refinement, further improving accuracy when experimental data is available.
- Quality Estimation: DEMO2 introduces estimated TM-scores (eTM-score) and RMSD (eRMSD) to provide users with a confidence measure for the predicted structures.
Datasets
- Benchmark Set: A comprehensive benchmark set of 461 non-redundant multi-domain proteins was used for evaluation. This set includes proteins with both continuous and discontinuous domains and was curated to exclude proteins with high sequence identity to the DEMO2 training set.
- CASP14 Targets: DEMO2 was also tested on multi-domain targets from the CASP14 (Critical Assessment of protein Structure Prediction) experiment, serving as a blind test.
- PDB: The Protein Data Bank (PDB) was used for searching analogous structural templates.
Key Insights and Ideas
- Synergy of Deep Learning and Templates: The core insight is that combining the predictive power of deep learning for inter-domain relationships with the structural context provided by analogous templates leads to superior multi-domain structure assembly.
- Simultaneous Domain Optimization is Crucial: Moving from iterative pairwise domain assembly to a simultaneous L-BFGS optimization of all domains significantly improves the ability to find optimal global arrangements.
- Global Templates Provide Better Context: Identifying templates that span multiple domains offers a more comprehensive structural guide than focusing solely on local pairwise alignments.
- Domain Decomposition Advantage: For proteins where individual domain models are already accurate (e.g., from tools like D-I-TASSER), DEMO2 excels at assembling them into correct full-length structures, highlighting the importance of accurate domain modeling as a prerequisite.
Key Results
- Significant Performance Improvement: DEMO2 achieved an average TM-score of 0.71 on the benchmark set, which is 12.7% higher than its predecessor DEMO (0.63) and 36.5% higher than the linker-based method AIDA (0.52).
- Improved Domain Orientations: The rTM-score, which specifically measures domain orientation accuracy, showed even more substantial improvements: 26.3% over DEMO and 108.7% over AIDA.
- High Accuracy on Discontinuous Domains: DEMO2 models achieved correct global topologies for 87.6% of proteins with discontinuous domains, outperforming DEMO and AIDA.
- Competitive with State-of-the-Art: DEMO2 was compared to deep-learning methods like DMPfold and trRosetta, achieving higher TM-scores and better domain orientations.
- CASP14 Performance: On the CASP14 blind test, DEMO2 achieved the highest average TM-score among the tested servers, demonstrating its robustness in real-world prediction scenarios.
- AlphaFold2 Comparison: While AlphaFold2 itself predicts full-length structures, DEMO2, when provided with accurate domain models (even those predicted by AlphaFold2), can achieve comparable or even better results, particularly in domain orientation.
Limitations and Future Work
- Dependence on Individual Domain Quality: The accuracy of DEMO2's full-length models is still dependent on the quality of the input individual domain models. Errors in domain prediction can propagate.
- Deep Potential Limitations: The current inter-domain restraints are predicted by DeepPotential, which was developed for full-length proteins. A more specialized deep-learning network for inter-domain restraints could further improve performance.
- Rigid Domain Assumption: The L-BFGS simulation treats domain models as rigid bodies, which might limit its ability to correct local errors within domains.
- Future Directions: The authors suggest exploring more specialized deep-learning networks for inter-domain restraints, integrating AlphaFold2's capabilities as new restraints, and developing flexible domain assembly to handle local errors within domains. They also propose leveraging protein-protein complex libraries to infer inter-domain orientations.
Summary updated