RNA Secondary Structure Prediction

The foundational thermodynamic assumption that an RNA molecule folds into the single structure with the lowest Gibbs free energy (ΔG) . Algorithms recursively decompose the sequence and sum empirically derived energy parameters for each structural motif—stacking base pairs, hairpin loops, bulges, internal loops, and multibranch loops—to identify the optimal configuration.

• The Turner Energy Model provides the nearest-neighbor parameters that quantify the stabilizing contribution of adjacent base pairs.
• Dynamic programming algorithms like Zuker's algorithm solve this minimization efficiently in O(n³) time.
• The MFE structure represents a single point estimate and does not capture the ensemble of suboptimal folds that may be biologically relevant.

Rather than predicting a single structure, the partition function (Q) sums the Boltzmann-weighted free energies of all possible secondary structures to calculate the equilibrium probability of any given base pair. This statistical mechanics approach reveals the full thermodynamic ensemble.

• Base pairing probabilities are visualized as dot plots, where the probability of each pair (i,j) is represented by a dot of proportional size.
• The ensemble diversity metric quantifies how many distinct low-energy structures exist, indicating whether the RNA is a rigid switch or a flexible molecule.
• Algorithms like McCaskill's algorithm compute the partition function in O(n³) time, enabling centroid structure estimation and Shannon entropy calculations per nucleotide.

The standard string representation for encoding predicted or experimentally validated secondary structures. Matching parentheses denote canonical base pairs—opening '(' for the 5' nucleotide and closing ')' for the 3' partner—while dots '.' represent unpaired nucleotides.

• Example: (((...))) represents a simple stem-loop with three base pairs and three unpaired loop nucleotides.
• This format serves as the training target for deep learning models that predict structure directly from sequence.
• Extended notations like WUSS notation add symbols for pseudoknots ([ ], { }) and non-canonical interactions, though standard dot-bracket cannot represent pseudoknots without crossing parentheses.
• Tools like ViennaRNA and RNAstructure output dot-bracket as their primary format, enabling interoperability across the field.

The algorithmic backbone of thermodynamic prediction relies on dynamic programming that decomposes the RNA sequence into smaller, overlapping substructures. The nearest-neighbor model assumes that the stability of a base pair depends only on the identity of the immediately adjacent base pair and the type of intervening loop.

• The ViennaRNA package implements Zuker's algorithm with energy parameters measured from optical melting experiments.
• RNAstructure provides an alternative implementation with a graphical user interface and SHAPE constraint integration.
• The recursion considers four cases: stacking on a previous pair, starting a hairpin loop, bifurcating into a multibranch loop, or leaving a nucleotide unpaired.
• This framework cannot predict pseudoknots because the recursion assumes non-crossing base pairs, a limitation addressed by more complex algorithms like maximum weighted matching.

Experimental reactivity data from SHAPE (Selective 2'-Hydroxyl Acylation analyzed by Primer Extension) or DMS (Dimethyl Sulfate) probing can be incorporated as pseudo-energy constraints to dramatically improve prediction accuracy. Flexible, unpaired nucleotides react preferentially, providing a per-nucleotide signal of local structural dynamics.

• SHAPE reagents acylate the 2'-hydroxyl of flexible nucleotides; the resulting reactivity is inversely correlated with base pairing probability.
• The reactivity profile is converted to a pseudo-free energy term and added to the folding algorithm's objective function, penalizing structures that pair highly reactive nucleotides.
• Integration reduces prediction error from ~35% to below 15% for many RNAs, making it essential for accurate modeling of long, biologically active sequences.
• Tools like ShapeKnots extend this approach to pseudoknotted structures.

When multiple sequence alignments of homologous RNAs are available, covariance models leverage evolutionary information to predict structure with higher accuracy than single-sequence methods. Compensatory base pair mutations—where both nucleotides in a pair change but maintain complementarity—provide strong evidence for structural interactions.

• A stochastic context-free grammar (SCFG) is trained on aligned sequences to capture both sequence conservation and correlated mutations.
• Infernal software implements covariance models for database searching and structure prediction, powering the Rfam database of RNA families.
• The mutual information score between two alignment columns quantifies the degree of covariation; high mutual information strongly predicts a base pair.
• This approach fails for sequences without known homologs, motivating the development of de novo deep learning methods.

The thermodynamic principle predicting the single most stable RNA secondary structure by minimizing the sum of empirically derived loop and stacking energy parameters. Algorithms like Zuker's dynamic programming efficiently compute the MFE structure by recursively finding the optimal configuration of base pairs, internal loops, and hairpins. Key insight: MFE assumes the molecule is at equilibrium and that the lowest energy state dominates the ensemble, which may not hold for kinetically trapped or multi-stable RNAs.

The standard nearest-neighbor empirical model assigning thermodynamic parameters to RNA base pair stacks and loops. Parameters are derived from optical melting experiments on short synthetic oligomers. Key contributions: Stacking energies for all 10 Watson-Crick and wobble nearest-neighbor combinations, loop initiation penalties, and coaxial stacking terms. This model forms the energetic foundation for virtually all free energy minimization and partition function calculations in tools like RNAfold and mfold.

A standard string representation where matching parentheses denote canonical base pairs and dots represent unpaired nucleotides. Example: (((...))) represents a simple stem-loop with three base pairs and three unpaired loop nucleotides. Usage: Serves as the primary training target for deep learning models like SPOT-RNA and UFold. Extended notations incorporate pseudoknots using brackets [] and braces {}.

A statistical mechanics calculation summing the Boltzmann-weighted free energies of all possible RNA secondary structures to derive base pairing probabilities. Unlike MFE which returns a single structure, the partition function yields an ensemble view—the probability that any two nucleotides form a pair. This is visualized as a dot plot and is critical for identifying regions of structural ambiguity or alternative conformations.

A chemical probing method that acylates the 2'-hydroxyl of flexible nucleotides, providing per-nucleotide data correlating with local structural dynamics. Mechanism: Unpaired or conformationally flexible nucleotides react preferentially with SHAPE reagents (e.g., 1M7, NMIA), while base-paired nucleotides are protected. Reactivity profiles are integrated as pseudo-energy restraints into folding algorithms, dramatically improving prediction accuracy for long RNAs.

The specific computational challenge of identifying pseudoknots—tertiary structural motifs where bases within a loop pair with bases outside that loop. Standard dynamic programming algorithms cannot handle these crossing base pairs without exponential complexity. Solutions: Specialized algorithms like ProbKnot, IPknot, and deep learning approaches (e.g., SPOT-RNA) use iterative refinement or graph neural networks to predict pseudoknots, which are critical for telomerase and viral frameshifting elements.