TL;DR
This paper presents a hierarchical, multi-level statistical modeling approach to understanding protein folding pathways, moving beyond evolutionary considerations to focus on the physical process. By developing models for local structure initiation, propagation, pairwise condensation, and multibody packing, the authors offer a framework that aligns with theoretical models of protein folding and suggests a limited number of possible protein folds.
What Problem Does This Paper Address?
Proteins fold into specific three-dimensional structures, a process crucial for their function. While evolutionary relationships between protein sequences are well-studied, understanding the folding pathway itself – the dynamic sequence of events that leads to the final structure – is a significant challenge. Existing methods often implicitly model evolution rather than the physical folding process. The authors aim to bridge this gap by developing statistical models that capture sequence-structure correlations at different levels of complexity, reflecting the hierarchical nature of folding and providing insights into energetically stable intermediate states.
Methods
The paper introduces a suite of four statistical models, each addressing a different level of structural detail in the folding process:
- I-sites: This model identifies local sequence motifs that are likely to initiate folding into specific structures. It uses a library of 262 motifs, represented as position-specific scoring matrices and confidence curves, trained using reinforcement learning.
- HMMSTR: This model captures the "propagation" stage, where local structures extend along the protein chain. It's a non-linear Hidden Markov Model (HMM) trained on motif adjacencies, representing residue backbone angles as states.
- HMMSTR-CM: This model addresses the "condensation" stage by modeling pairwise interactions between local structure motifs. It uses knowledge-based Markov state pair potentials derived from contact maps (residue pairs within 8Å).
- SCALI: This model tackles the "packing" stage, focusing on non-sequential, three-dimensional arrangements of multiple motifs. It uses structure-based alignments to identify conserved core packing arrangements, which are then modeled as HMMs.
These models are presented hierarchically, mirroring the nucleation/condensation mechanism of protein folding, from local initiation to global structure.
Datasets
The models are trained and evaluated using the CATH database of protein domain structures. This database is used to derive statistical interaction potentials and identify recurrent sequence-structure motifs. The authors also mention using a non-redundant database of known protein structures for training and clustering.
Key Insights and Ideas
- Folding is Hierarchical: The paper strongly supports the idea that protein folding proceeds through distinct, hierarchical stages, from local initiation to global packing.
- Statistical Models Reflect Physical Processes: The authors argue that statistical correlations observed in protein databases can be interpreted as energetically stable states or interactions, providing a link between bioinformatics and physical folding models.
- Limited Folding Pathways: The hierarchical nature of the models, with a decreasing number of choices at each subsequent stage, suggests that the total number of possible protein folding pathways might be much smaller than previously thought, potentially explaining the observed recurrence of protein folds.
- Evolutionary vs. Folding Signals: The work emphasizes the importance of distinguishing between evolutionary signals (like sequence similarity) and signals directly related to the folding process.
Key Results
- I-sites: About one-third of all residues are found in high-confidence (>70%) I-sites motif regions, though nearly all residues (98%) belong to one or more I-sites motif structures.
- HMMSTR: Predictions achieve an average accuracy of 60% in predicting 8-residue fragments with RMSD < 1.4Å.
- HMMSTR-CM: For a hypothetical protein (HI0073), the contact map prediction achieved 42% contact coverage and 29% accuracy, which improved to 75% coverage and 57% accuracy when counting near misses (±1 residue).
- SCALI: Demonstrated the ability to align structures with different topologies but conserved core packing arrangements, even when components are not sequential. For example, it aligned 1ALK and 1VPT with an RMSD of 4.3Å.
Limitations and Future Work
- Approximation of Topological Constraints: HMMSTR-CM's use of simple rules to extract contacts from high-scoring ones only approximates complex topological constraints, suggesting the need for multibody models.
- Refinement of SCALI-HMMs: The paper notes that refinement and analysis of conserved core packing arrangements identified by SCALI are ongoing.
- Future Directions: The authors suggest that the absence of a combinatorial explosion in their physical folding model might explain the lack of sparse data problems in their statistical models. They also propose that the limited number of ways a protein chain can fold might be a consequence of convergent evolution, where similar topologies arise independently.
Summary updated