Abstract Protein threading is a powerful computational technique for predicting the three-dimensional structure of proteins by aligning their amino acid sequences with known structural templates. This approach addresses the "inverse folding problem" by leveraging the observation that protein folds are highly conserved across evolution. By employing sophisticated energy functions and advanced optimization algorithms such as tree decomposition and integer programming, threading can accurately model structures even for proteins with low sequence identity to known templates. This paper reviews the core components, algorithmic advancements, and practical applications of protein threading in modern bioinformatics.
Introduction
The "structure gap" remains a significant challenge in bioinformatics: while genomic sequencing is rapid and inexpensive, experimental determination of 3D protein structures via X-ray crystallography or NMR is slow and costly. Traditional homology modeling often fails when sequence identity drops below 30%. Protein threading bridges this gap by operating on the principle that protein structure is more conserved than sequence. The goal is to identify the correct "fold" for a query sequence among a database of previously solved structures, even in the absence of clear evolutionary relationships.
Methods and Energy Functions
A robust threading system relies on three pillars: energy functions, optimization algorithms, and the integration of experimental constraints. Energy functions evaluate the fitness of a sequence-to-structure alignment. Singleton energies measure the propensity of an amino acid to occupy a specific local environment, such as an alpha-helix, beta-sheet, or loop.
Pairwise interaction energies further refine these models by accounting for the spatial proximity of amino acid pairs. To enhance sensitivity, threading often utilizes evolutionary profiles from protein families rather than single sequences.
Optimization Algorithms
Finding the optimal alignment is an NP-hard problem due to pairwise interactions. The paper highlights three rigorous approaches:
- PROSPECT: A divide-and-conquer method that partitions structures at loop regions.
- RAPTOR: Utilizes Linear Integer Programming (LIP) for global optimization.
- Tree-Decomposition: A graph-based approach that exploits the low "tree width" of protein structures (typically < 6) to solve the alignment problem in polynomial time.
Results and Applications
The effectiveness of threading is demonstrated through its ability to identify correct folds even in remote homologs. The authors successfully applied these methods to model the 11 Open Reading Frames (ORFs) of the SARS-CoV virus, identifying structural similarities that suggest evolutionary links.
Furthermore, the incorporation of sparse experimental data, such as NMR-derived distance restraints, significantly improves prediction accuracy. Adding even a small percentage of restraints can reduce the RMSD of a model from a poor 16.2 Å to a useful 5.3 Å, moving the correct template significantly higher in rank.
Conclusion
While challenges remain in modeling membrane proteins and refining loop regions, protein threading continues to be a cornerstone of structural bioinformatics. Future developments in physics-based energy functions and hybrid modeling with experimental data promise even greater accuracy in bridging the sequence-structure gap.
Summary updated