TL;DR
This paper reviews "pathway models" for ab initio protein structure prediction, a challenging task of determining a protein's 3D shape solely from its amino acid sequence. It introduces HMMSTR-CM, a novel method that uses a fragment library and rule-based nucleation/propagation to build protein contact maps, demonstrating promising results in CASP5.
What Problem Does This Paper Address?
Predicting a protein's three-dimensional structure from its amino acid sequence alone (ab initio prediction) is a grand challenge in biology. This problem is typically broken down into two parts: developing a scoring function that can accurately distinguish native-like structures from incorrect ones, and devising a search strategy that can efficiently explore the vast conformational space to find the native fold. Current methods struggle with both aspects, lacking reliable scoring functions that always guide searches to the correct structure and efficient search algorithms that guarantee exploration of near-native conformations. This paper focuses on "pathway models," which aim to integrate scoring and search by simulating or inferring the physical process of protein folding.
Methods
The paper reviews existing approaches and introduces a new method, HMMSTR-CM.
- Review of Existing Methods: The authors discuss various ab initio prediction strategies, including:
- Molecular Dynamics (MD) and Monte Carlo (MC): These simulate the physical folding process, often using physics-based energy functions. However, they are computationally expensive.
- Simplified Models and Lattice Simulations: These reduce conformational space by representing proteins on lattices or with simplified residue representations, making exhaustive simulations more feasible for shorter chains.
- Fragment Libraries: These leverage pre-computed preferred conformations for short sequence segments to "leapfrog" the initial stages of folding. Methods like ROSETTA and FRAGFOLD utilize this approach.
- HMMSTR-CM (New Method): This approach is not a direct simulation but rather a knowledge-based method for building protein contact maps.
- Contact Potential Map: It starts by calculating a contact potential map for a given sequence using HMMSTR, which uses hidden Markov models to capture local sequence-structure correlations. A low contact potential indicates a higher likelihood of contact.
- Nucleation/Propagation Rules: The core of HMMSTR-CM is a set of rules that encode nucleation/condensation-type folding pathways. The process involves:
- Predicting secondary structure.
- Selecting a "nucleation site" by assigning local contacts.
- Propagating these contacts (or removing them) based on a set of "physicality and propagation rules" (detailed in Table 1).
- Goal: The prediction is complete when all residue pairs are assigned a contact or non-contact status, and all rules are satisfied.
Datasets
The paper does not explicitly list datasets used for training or evaluation of the reviewed methods. However, it mentions that HMMSTR-CM was used for ab initio predictions as part of the CASP5 (Critical Assessment of protein Structure Prediction) experiment, implying that CASP5 targets were used for evaluation. The paper also refers to Figure 1, which shows a predicted contact map and the true contact map for a specific protein (T0130), suggesting this protein was used as a case study.
Key Insights and Ideas
- Pathway Models Integrate Scoring and Search: The central idea is that explicitly modeling or inferring the process of folding (the pathway) can be more effective than treating scoring and search as separate, independent problems.
- Fragment Libraries Accelerate Folding: Using pre-defined fragments significantly speeds up the prediction process by bypassing the initial, often difficult, steps of forming local structures.
- Contact Maps as a Useful Representation: Contact maps offer a lower-resolution, yet informative, representation of protein structure that is amenable to data mining and can be projected into 3D.
- Rule-Based Prediction for Tractability: HMMSTR-CM's rule-based approach, rather than full simulation, makes the prediction process more computationally tractable, especially for ab initio tasks.
- Nucleation Site is Critical: The success of rule-based pathway models like HMMSTR-CM heavily relies on the correct identification of the initial nucleation site. A wrong choice can lead to an incorrect topology.
- Long-Range Contacts are Challenging: The paper highlights that predicting long-range contacts, which are crucial for overall fold stability, is difficult for purely statistical methods and benefits from more physically informed approaches.
Key Results
- HMMSTR-CM Performance on T0130: For the target protein T0130 (116 residues, α/β sandwich), HMMSTR-CM achieved:
- 42% contact coverage and 29% accuracy.
- When "near-misses" (±one residue) were included, coverage rose to 75% and accuracy to 57%.
- The method correctly predicted long-range contacts between β1 and β4.
- General Success of Fragment Libraries: The paper notes that fragment library simulations have had broad success in ab initio prediction in recent CASP meetings.
- Computational Cost: While not quantified for HMMSTR-CM, the paper generally contrasts the high computational cost of detailed all-atom simulations with the improved feasibility of simplified models and fragment-based approaches.
Limitations and Future Work
- Nucleation Site Selection: The primary limitation of HMMSTR-CM is the critical dependence on the correct choice of the nucleation site. If the wrong site is chosen, the algorithm fails. The authors suggest that retrospective analysis can identify the correct topology, but this needs cross-validation.
- Computational Expense of Physics-Based Methods: Detailed all-atom simulations (MD, MC) remain computationally expensive, limiting their practical application for long proteins.
- Accuracy of Scoring Functions: The paper acknowledges that no reliable scoring function exists that can always drive a search to the native fold.
- General Search Methods: There is a lack of general search methods that can guarantee significant sampling of near-native conformations.
- Future Work: The paper implies that further development of robust nucleation site selection strategies and refinement of the rule sets in HMMSTR-CM would be beneficial. The broader field continues to seek more efficient and accurate scoring functions and search algorithms for ab initio prediction.
Summary updated