Knowledge-Based Potential

Knowledge-based potentials are constructed by mining high-resolution RNA structures from the Protein Data Bank (PDB). The frequency of observed pairwise atomic distances or residue contacts is converted into an energy-like score using the inverse Boltzmann relation: E = -kBT ln(P_obs/P_ref). Here, P_obs is the observed probability of a specific interaction in native structures, and P_ref is the expected probability in a random reference state where no specific interactions occur. This approach implicitly captures complex physical forces—hydrogen bonding, base stacking, electrostatic interactions, and solvation effects—without explicitly parameterizing them. The quality of the potential depends critically on the size and diversity of the training database and the choice of reference state, which defines the zero-energy baseline.

The choice of reference state is the most critical and debated aspect of constructing a knowledge-based potential. The reference state represents the expected distribution of interactions in the absence of specific folding forces. Common choices include:

• Averaged reference state: The mean interaction frequency across all atom types in the database, assuming uniform distribution.
• Finite ideal-gas reference state: Accounts for the spherical volume available to atom pairs within a finite protein or RNA structure, correcting for chain connectivity.
• Quasi-chemical approximation: Treats each interaction type independently, normalizing by the mole fractions of the interacting species. An improper reference state introduces systematic biases that can favor collapsed, non-native conformations or fail to discriminate near-native decoys from misfolded structures.

The most common formulation is the distance-dependent pairwise potential, which scores interactions between atom or residue pairs as a function of their spatial separation. For RNA, these potentials are often parameterized for specific interaction types:

• Base-base contacts: Capturing stacking and pairing preferences between nucleotides.
• Base-backbone interactions: Scoring the packing of nucleobases against the sugar-phosphate backbone.
• Backbone-backbone interactions: Modeling the conformational preferences of the RNA phosphodiester chain. The potential is typically discretized into distance bins (e.g., 0-2 Å, 2-4 Å) and smoothed using spline functions to avoid discontinuities during gradient-based optimization. These potentials are computationally efficient and can be evaluated rapidly during Monte Carlo or molecular dynamics simulations.

Advanced knowledge-based potentials incorporate angular and dihedral dependencies to capture the directional nature of RNA interactions. These potentials score the relative orientation of two bases using parameters such as the Leontis-Westhof classification edges (Watson-Crick, Hoogsteen, Sugar) and glycosidic bond orientation (cis/trans). By binning observed interactions by both distance and relative orientation angles, these potentials can distinguish between canonical Watson-Crick pairs, Hoogsteen pairs, and non-specific hydrophobic contacts. This granularity is essential for accurately modeling RNA tertiary motifs like A-minor interactions, ribose zippers, and pseudoknots, where the precise geometric arrangement of bases determines structural stability.

Knowledge-based potentials are frequently combined with physics-based energy terms in hybrid scoring functions. In the Rosetta energy function (used by FARFAR2 for RNA), the total score is a linear combination of knowledge-based terms (e.g., base pair statistics, backbone torsional preferences) and physics-based terms (e.g., van der Waals repulsion, implicit solvation). The weights of these terms are optimized to maximize the energy gap between native structures and decoy conformations. This hybrid approach leverages the strengths of both paradigms: the comprehensive sampling of experimentally observed interaction patterns from knowledge-based terms and the physical realism of explicit steric and electrostatic constraints from physics-based terms.

Despite their utility, knowledge-based potentials face fundamental limitations:

• Database bias: Over-representation of certain structural motifs (e.g., ribosomal RNA) can skew potentials toward specific architectures.
• Transferability: Potentials derived from crystallographic structures may not accurately score dynamic or partially folded states.
• Sparse data: Rare interaction types suffer from poor statistics, leading to noisy or unreliable energy estimates.
• Decoy discrimination: The ultimate test is the ability to identify the native structure among thousands of computationally generated decoys. Potentials that perform well on native structure recapitulation may fail to distinguish near-native conformations (RMSD < 4 Å) from non-native traps, a phenomenon known as the energy funnel problem.

An energy function derived from the inverse Boltzmann relationship, converting the frequency of observed structural features in a database into a pseudo-energy. Unlike physics-based force fields, a statistical potential captures the effective free energy implicitly, accounting for solvent effects and entropic contributions without explicit calculation.

• Derivation: -k_B T ln(P_obs / P_ref)
• Reference State: Defines the expected random distribution
• Key Advantage: Speed over physics-based simulations

A specific class of knowledge-based potential that scores interactions between atom or residue pairs as a function of their spatial separation. For RNA, this captures the propensity of specific nucleobase stacking and backbone phosphate geometries observed in crystallographic structures.

• Radial bins: Divide interatomic distances into discrete ranges
• Atom typing: C4', N1, O2' each have distinct interaction profiles
• Application: Used in Rosetta FARFAR2 for low-resolution folding

The theoretical baseline distribution against which observed interaction frequencies are compared to extract meaningful signals. The choice of reference state critically determines the potential's accuracy. A poor reference state introduces systematic biases that favor collapsed, non-native conformations.

• Averaged: Mean contact frequency over all atom types
• Finite ideal-gas: Accounts for spherical volume constraints
• Quasi-chemical: Corrects for stoichiometric effects in dense packing

A knowledge-based potential that scores the local conformation of the RNA backbone by analyzing the distribution of dihedral angles (α, β, γ, δ, ε, ζ) in known structures. This captures the intrinsic rigidity of the sugar-phosphate backbone and the discrete rotameric states of the ribose ring.

• Sugar pucker: C2'-endo vs. C3'-endo distributions
• γ angle: Gauche+/gauche-/trans preferences
• Integration: Combined with pair potentials for full scoring

A residue-level knowledge-based potential that quantifies the burial propensity of each nucleotide in folded RNA structures. Purines (A, G) show a stronger statistical preference for burial than pyrimidines (C, U), driving the hydrophobic core formation observed in large RNA assemblies like the ribosome.

• SASA: Solvent Accessible Surface Area as input feature
• Classification: Buried, partially exposed, fully exposed states
• Use: Guides initial collapse during folding simulations