Ab Initio Prediction

The core of ab initio methods is a knowledge-based or physics-based energy function that approximates the free energy of a protein conformation. These functions model:

• Van der Waals forces and steric repulsion to prevent atomic clashes
• Electrostatic interactions and hydrogen bonding patterns
• Solvation effects using implicit solvent models like Generalized Born
• Torsion angle potentials derived from Ramachandran statistics

The goal is to identify the global free energy minimum, which corresponds to the native state under the thermodynamic hypothesis proposed by Anfinsen.

Unlike template-based modeling, ab initio prediction does not rely on known homologous structures from the Protein Data Bank (PDB). Instead, it performs a large-scale conformational search to explore the energy landscape. Key strategies include:

• Monte Carlo simulations with simulated annealing to escape local minima
• Molecular dynamics in reduced representation (e.g., united-atom models)
• Fragment assembly, where short backbone segments from the PDB are combinatorially recombined (pioneered by ROSETTA)
• Genetic algorithms that evolve a population of decoy structures This independence from global templates makes it essential for predicting orphan proteins and novel folds.

To make the conformational search computationally tractable, ab initio methods often simplify protein geometry:

• Coarse-grained models: Represent each residue by 1–3 pseudo-atoms (e.g., Cα, Cβ, and a side-chain centroid)
• Lattice models: Confine residues to discrete grid points to drastically reduce the search space
• United-atom force fields: Collapse non-polar hydrogens into their parent carbon atoms These simplifications allow the exploration of folding pathways on realistic timescales, after which all-atom refinement restores full atomic detail using physics-based force fields like AMBER or CHARMM.

The search phase generates thousands to millions of candidate structures (decoys), but the native-like conformation must be identified. Decoy discrimination uses:

• Statistical potentials (e.g., DFIRE, DOPE) that score residue-pair distances based on observed frequencies in the PDB
• Clustering algorithms (e.g., SPICKER) that identify the centers of highly populated conformational basins—the assumption being that the native state occupies the widest free energy well
• Consensus scoring across multiple energy functions to reduce single-function bias The final selected model is the centroid of the largest cluster, not necessarily the single lowest-energy decoy.

Modern ab initio methods are dramatically enhanced by residue coevolution analysis. By analyzing correlated mutations in a Multiple Sequence Alignment (MSA), direct coupling analysis (DCA) identifies residue pairs in spatial contact. These contacts are used as:

• Distance restraints that drastically constrain the conformational search
• Scoring terms that penalize models violating evolutionary coupling This hybrid approach, exemplified by AlphaFold and trRosetta, blurs the line between pure ab initio and template-based methods, achieving atomic accuracy for single-domain proteins without requiring a homologous structural template.

Predicted models must pass rigorous stereochemical validation to be considered physically plausible:

• Ramachandran plot analysis: Backbone dihedral angles (φ, ψ) must fall within allowed regions; >98% of residues in favored regions is expected for high-quality models
• Packing scores: Assess complementary surface burial and cavity volume using tools like MolProbity
• Per-residue energy profiles: Identify local frustration where the sequence is incompatible with the predicted fold These metrics distinguish physically impossible conformations from merely inaccurate ones, providing confidence even when no experimental structure exists for comparison.

The core of ab initio methods relies on force fields that approximate the free energy of a protein conformation. These functions model:

• Bonded interactions: bond lengths, angles, and dihedral angles
• Non-bonded interactions: van der Waals forces and electrostatic potentials
• Solvation effects: implicit models like Generalized Born (GB) or Poisson-Boltzmann to account for water

The thermodynamic hypothesis states that the native structure is the global minimum of the free energy landscape. The challenge is that the energy surface is astronomically complex, with a vast number of local minima that can trap search algorithms.

Protein structure prediction methods exist on a spectrum defined by their reliance on experimental data:

• Template-Based Modeling (TBM): uses a homologous structure from the PDB as a scaffold; highly accurate when a close template exists (>30% sequence identity)
• Free Modeling (FM) / Ab Initio: no global template is used; relies solely on physicochemical principles
• Hybrid approaches: modern methods like AlphaFold2 blur this distinction by using deep learning to extract coevolutionary signals from Multiple Sequence Alignments (MSAs), effectively learning a template-free energy function from data

Efficiently exploring the energy landscape is the central computational challenge. Key algorithms include:

• Monte Carlo with Minimization (MCM): combines random moves with local energy minimization to focus sampling on low-energy regions
• Molecular Dynamics (MD): simulates atomic motion over time using Newtonian physics; limited by femtosecond timesteps
• Replica Exchange MD: runs multiple simulations at different temperatures in parallel, periodically swapping conformations to overcome energy barriers
• Genetic Algorithms: evolves a population of conformations using crossover and mutation operators
• Coarse-graining: reduces the representation from all-atom to one bead per residue to accelerate sampling

In 1969, Cyrus Levinthal noted that if a protein were to randomly sample all possible conformations, folding would take longer than the age of the universe. Yet proteins fold in milliseconds to seconds. This paradox implies:

• Folding is not a random search but a directed process
• The energy landscape is funnel-shaped, guiding the chain toward the native state
• Ab initio prediction algorithms must replicate this funnel by biasing search toward native-like topologies
• The paradox motivates the use of fragment insertion, coarse-graining, and knowledge-based potentials to reduce the effective search space