Fragment Assembly

The process begins by constructing a fragment library from the Protein Data Bank (PDB). For every overlapping window of 3 or 9 residues in the target sequence, a set of candidate fragments is retrieved from non-homologous experimental structures. Selection is based on sequence similarity and secondary structure prediction profiles. These libraries constrain the conformational search space to physically realistic local geometries, preventing the model from exploring non-protein-like torsion angles.

Assembly proceeds via a stochastic optimization protocol. A starting extended chain is perturbed by replacing a random segment with a fragment from the library. The new conformation is accepted or rejected based on the Metropolis criterion, which probabilistically accepts energy-increasing moves to escape local minima. The temperature of the system is gradually lowered, allowing broad exploration of the energy landscape early in the trajectory before settling into a low-energy native-like basin.

Conformations are evaluated using a composite scoring function that combines physics-based terms with statistical potentials. Key components include:

• Van der Waals repulsion to prevent atomic overlap
• Lazaridis-Karplus solvation to model hydrophobic burial
• Hydrogen bonding geometry satisfaction
• Residue pair statistics derived from PDB frequencies These terms guide the search toward compact, protein-like topologies without requiring explicit solvent simulation.

Fragment assembly typically operates in a centroid representation where side chains are reduced to a single interaction center. This coarse-grained mode enables rapid sampling of the backbone fold. Once a low-resolution model is identified, a full-atom refinement stage reconstructs explicit side-chain rotamers and performs gradient-based energy minimization to resolve steric clashes and optimize hydrogen bonding networks. This hierarchical strategy balances computational efficiency with physical accuracy.

The simulation generates tens of thousands of decoys (candidate models). The lowest-energy structures are not always the most accurate due to imperfections in the energy function. To address this, models are clustered by Cα RMSD, and the centers of the largest clusters are selected as final predictions. This consensus approach leverages the principle that native-like conformations reside in broad energy minima, while incorrect folds are scattered across the landscape.

Fragment assembly was the core engine of the Rosetta software suite for over two decades, achieving landmark results in CASP blind prediction experiments. It demonstrated that the native fold is determined by local sequence-structure correlations and global packing constraints. While largely superseded by end-to-end deep learning methods like AlphaFold2, the fragment concept remains influential in loop modeling, de novo design, and scenarios where multiple sequence alignments are sparse or unavailable.