Minimum Free Energy (MFE)

MFE calculations rely on the Turner Energy Model, which assigns free energy values to adjacent base pairs rather than individual nucleotides. The total free energy of a structure is the sum of stacking interactions between neighboring pairs, plus destabilizing contributions from loops (hairpin, bulge, internal, and multibranch). Each motif's energy is empirically derived from optical melting experiments.

• Base pair stacks: The dominant stabilizing force (e.g., GC/CG stack = -3.4 kcal/mol)
• Hairpin loops: Destabilizing, with penalties scaling by loop size and sequence
• Bulge loops: Penalized based on the number of unpaired nucleotides and flanking stacks
• Internal loops: Modeled with asymmetry and mismatch parameters
• Multibranch loops: Penalized with a linear function of unpaired nucleotides and branching helices

MFE prediction is computed using a recursive dynamic programming approach, typically the Zuker-Stiegler algorithm, which runs in O(N³) time and O(N²) memory for a sequence of length N. The algorithm recursively calculates the optimal folding for all subsequences, building from shorter to longer segments.

• Fill step: Computes minimum free energy for all subsequences using recurrence relations
• Traceback step: Reconstructs the optimal structure by following pointers from the full sequence
• V(i,j) matrix: Stores the minimum energy for subsequence i..j given that i and j form a base pair
• W(i,j) matrix: Stores the minimum energy for subsequence i..j with no pairing constraint
• Complexity: O(N³) time and O(N²) space, limiting practical use to sequences under ~10,000 nucleotides without heuristics

MFE prediction returns exactly one structure—the global free energy minimum—ignoring the fact that RNA molecules exist as a Boltzmann ensemble of interconverting conformations at physiological temperatures. This is a fundamental limitation: the MFE structure may not be the biologically active conformation.

• Ensemble vs. MFE: The MFE structure can represent a negligible fraction of the ensemble at 37°C
• Suboptimal structures: Often within 5-10% of the MFE energy and functionally relevant
• Kinetic traps: RNA may fold into a local minimum rather than the global MFE during transcription
• Partition function: Complements MFE by calculating base-pairing probabilities across all structures
• Accuracy: MFE prediction alone achieves ~65-70% base-pair accuracy for sequences under 700 nt

The accuracy of MFE prediction is critically dependent on the quality and completeness of the thermodynamic parameter database. The Turner rules, last comprehensively updated in 2004, contain known gaps for certain motifs.

• Missing parameters: Non-canonical pairs, pseudoknots, and complex tertiary interactions are not modeled
• Salt corrections: Standard parameters assume 1M NaCl; deviations require empirical adjustments
• Temperature dependence: Parameters are measured at 37°C; extrapolation to other temperatures introduces error
• Dangling ends: Unpaired terminal nucleotides contribute stabilization that must be explicitly modeled
• Coaxial stacking: Helices stacking end-to-end in multibranch loops require special energetic treatment

MFE prediction accuracy improves significantly when experimental constraints from chemical probing experiments (SHAPE, DMS, CMCT) are incorporated as pseudo-energy penalties. These reactivity data report on nucleotide flexibility and solvent accessibility.

• SHAPE-directed folding: Converts reactivity values into pseudo-free energy bonuses/penalties
• Soft constraints: Penalize base-paired nucleotides with high reactivity (flexible, unpaired in reality)
• Hard constraints: Force specific nucleotides to be unpaired or paired based on strong experimental evidence
• Accuracy gain: Constrained MFE prediction can improve accuracy by 10-20 percentage points
• Integration: Tools like RNAstructure and ViennaRNA support constraint files in standard formats

Standard MFE algorithms using dynamic programming cannot predict pseudoknots—tertiary interactions where bases within a loop pair with bases outside that loop. This is a fundamental algorithmic limitation, not an energetic one.

• Computational barrier: Pseudoknot prediction is NP-complete in the general case
• Heuristic extensions: Algorithms like ProbKnot and IPknot use iterative refinement to add pseudoknots
• Biological significance: Pseudoknots are critical for ribosomal frameshifting, telomerase, and ribozyme catalysis
• Energy models: Specialized parameter sets exist for pseudoknot loop energies (e.g., Dirks-Pierce model)
• Alternative approaches: Maximum expected accuracy and partition function methods can identify pseudoknot probabilities

A statistical mechanics calculation that sums the Boltzmann-weighted free energies of all possible RNA secondary structures. Unlike MFE which predicts a single optimal structure, the partition function yields base pairing probabilities and ensemble diversity metrics. It uses the same Turner energy parameters but computes the full thermodynamic ensemble via dynamic programming, enabling the calculation of the ensemble free energy which includes entropic contributions from the density of states.

The standard nearest-neighbor empirical model that assigns thermodynamic parameters to RNA base pair stacks and loops. Parameters are derived from optical melting experiments on short synthetic oligomers. The model decomposes the total free energy into additive terms: stacking interactions between adjacent base pairs, loop penalties (hairpin, bulge, internal, and multi-branch), and dangling end contributions. Updated in 2004, it forms the energetic foundation for MFE algorithms like RNAfold and mfold.

The Zuker algorithm uses dynamic programming to find the MFE structure in O(n³) time and O(n²) space. It recursively computes optimal substructures by filling a triangular matrix, considering four cases at each step: (1) i and j are paired, (2) i is unpaired, (3) j is unpaired, or (4) the structure bifurcates into two independent substructures. This divide-and-conquer approach guarantees finding the global free energy minimum under the Turner model, assuming no pseudoknots.

The specific computational challenge of identifying pseudoknots—tertiary motifs where bases within a loop pair with bases outside that loop. Standard MFE dynamic programming algorithms cannot predict pseudoknots because they violate the nested structure assumption. Specialized algorithms like ProbKnot, IPknot, and HotKnots extend MFE frameworks using heuristic search, integer programming, or grammar-based approaches, but at increased computational cost and reduced accuracy compared to nested structure prediction.

Experimental techniques like SHAPE (Selective 2'-Hydroxyl Acylation analyzed by Primer Extension) measure nucleotide flexibility. SHAPE reactivity values are converted into pseudo-energy penalties and added to the Turner energy model during MFE calculation. This integration dramatically improves prediction accuracy—from ~65% to >90% for secondary structure—by biasing the algorithm toward structures consistent with experimental data. The pseudo-energy term penalizes base-paired nucleotides with high reactivity and rewards unpaired nucleotides with low reactivity.

MFE predicts a single structure, but RNA molecules exist as dynamic ensembles. The RNAsubopt algorithm generates all structures within a specified free energy range above the MFE, while stochastic sampling draws structures from the Boltzmann distribution. These approaches reveal alternative conformations that may be functionally relevant, such as riboswitch ligand-binding states. The ensemble view acknowledges that the MFE structure is a thermodynamic prediction, not necessarily the biologically active conformation.