Side-Chain Packing

The accuracy of side-chain packing is fundamentally constrained by the rotamer library—a curated set of statistically observed, low-energy side-chain conformations. Libraries are categorized as backbone-dependent (conditioned on local phi/psi angles) or backbone-independent. Modern libraries, such as the Dunbrack or Penultimate libraries, discretize continuous torsional space into manageable states, trading completeness for computational tractability. The choice of library directly impacts the resolution of the final model.

Packing algorithms rely on a physics-based or statistical potential energy function to evaluate the favorability of a given rotamer arrangement. Key components include:

• Van der Waals repulsion: A harsh penalty for atomic overlaps to prevent steric clashes.
• Lennard-Jones attraction: Rewards optimal packing density.
• Hydrogen bonding and electrostatics: Orientation-dependent terms that stabilize polar networks.
• Solvation models: Implicitly account for the hydrophobic burial of non-polar side chains.

Finding the global minimum energy configuration is an NP-hard combinatorial explosion. For a protein of N residues, each with R possible rotamers, the search space is R^N. Exhaustive enumeration is impossible. Algorithms must navigate this rugged energy landscape using heuristics. The challenge is to avoid local minima traps where a single buried clash prevents the system from reaching the true ground state.

A provable, deterministic algorithm that prunes the search space by iteratively eliminating rotamers that cannot possibly be part of the Global Minimum Energy Conformation (GMEC). The Goldstein criterion compares the best-case interaction of a rotamer against the worst-case interaction of an alternative. If the former is always worse, the rotamer is eliminated. DEE drastically reduces the combinatorial complexity before a final search.

When DEE cannot reduce the space sufficiently, Monte Carlo simulated annealing and genetic algorithms are employed. Simulated annealing mimics physical annealing by accepting high-energy states early to escape local minima, then gradually cooling to settle into a low-energy basin. Genetic algorithms evolve a population of rotamer configurations through mutation and crossover, preserving beneficial packing motifs across generations.

Modern methods formulate packing as a graphical model or Markov Random Field (MRF). Residues are nodes, and edges represent pairwise interaction energies. Belief propagation or A search* with tree decomposition can find the exact GMEC for sparse interaction graphs. This approach is particularly effective for re-packing small, localized regions like enzyme active sites or protein-protein interfaces.

A curated collection of statistically preferred side-chain conformations (rotamers) for each amino acid type, derived from high-resolution experimental structures in the Protein Data Bank (PDB). These libraries discretize the continuous conformational space into a finite set of low-energy states, making the combinatorial search for optimal packing computationally tractable.

• Backbone-dependent libraries: Rotamer probabilities conditioned on local backbone dihedral angles (phi/psi), providing higher accuracy.
• Penultimate libraries: Include the effect of the preceding residue's conformation.
• Libraries are essential for reducing the search space from infinite continuous angles to a manageable set of discrete possibilities.

A provable, deterministic algorithm that systematically prunes rotamers that cannot be part of the global minimum energy conformation (GMEC). DEE applies mathematically rigorous elimination criteria to iteratively remove unfavorable rotamers before an exhaustive search.

• Goldstein criterion: A powerful pairwise elimination condition that identifies rotamers always energetically worse than another alternative when considering all possible interacting partners.
• DEE drastically reduces the combinatorial complexity, often making an otherwise intractable NP-hard problem solvable for core redesign tasks.

A stochastic optimization heuristic widely used for side-chain packing that probabilistically samples conformational space to find low-energy states. The algorithm accepts or rejects random rotamer changes based on the Metropolis criterion, allowing occasional uphill moves to escape local minima.

• Temperature schedule: The probability of accepting unfavorable moves decreases over time, gradually converging toward a minimum.
• Unlike DEE, this method does not guarantee the global optimum but scales efficiently to very large systems and is central to the Rosetta macromolecular modeling suite.

A key component of the molecular mechanics energy function that penalizes steric overlap between non-bonded atoms. During side-chain packing, the clash score quantifies unfavorable atomic overlaps using a Lennard-Jones 6-12 potential.

• A high clash score indicates physically impossible atomic interpenetration.
• MolProbity uses an all-atom clashscore to validate the physical realism of packed models, reporting the number of serious steric overlaps per 1000 atoms.
• Packing algorithms aggressively minimize this term to ensure physically plausible geometries.

A message-passing neural network that solves the inverse problem: given a fixed backbone structure, predict an amino acid sequence that will stably fold to that conformation. ProteinMPNN implicitly performs side-chain packing by learning the sequence-structure relationship from data.

• Encodes local and global geometric features of the backbone as a graph.
• Predicts residue identity probabilities at each position, capturing the optimal sequence for the target fold.
• Demonstrates that learned models can outperform physics-based rotamer sampling for sequence design tasks.

A post-packing computational procedure that adjusts atomic coordinates to find the nearest local minimum on the potential energy surface. After discrete rotamer selection, gradient-based minimization relieves residual bond geometry strain and subtle steric clashes.

• Uses algorithms like L-BFGS or conjugate gradient descent.
• Refines both side-chain and backbone atoms, allowing small induced-fit adjustments.
• Essential for producing physically realistic models suitable for downstream docking or design calculations.