Abstract
Background
We study the sparsification of dynamic programming based on folding algorithms of RNA structures. Sparsification is a method that improves significantly the computation of minimum free energy (mfe) RNA structures.
Results
We provide a quantitative analysis of the sparsification of a particular decomposition rule, Λ^{∗}. This rule splits an interval of RNA secondary and pseudoknot structures of fixed topological genus. Key for quantifying sparsifications is the size of the so called candidate sets. Here we assume mfestructures to be specifically distributed (see Assumption 1) within arbitrary and irreducible RNA secondary and pseudoknot structures of fixed topological genus. We then present a combinatorial framework which allows by means of probabilities of irreducible substructures to obtain the expectation of the Λ^{∗}candidate set w.r.t. a uniformly random input sequence. We compute these expectations for arcbased energy models via energyfiltered generating functions (GF) in case of RNA secondary structures as well as RNA pseudoknot structures. Furthermore, for RNA secondary structures we also analyze a simplified loopbased energy model. Our combinatorial analysis is then compared to the expected number of Λ^{∗}candidates obtained from the folding mfestructures. In case of the mfefolding of RNA secondary structures with a simplified loopbased energy model our results imply that sparsification provides a significant, constant improvement of 91% (theory) to be compared to an 96% (experimental, simplified arcbased model) reduction. However, we do not observe a linear factor improvement. Finally, in case of the “full” loopenergy model we can report a reduction of 98% (experiment).
Conclusions
Sparsification was initially attributed a linear factor improvement. This conclusion was based on the so called polymerzeta property, which stems from interpreting polymer chains as selfavoiding walks. Subsequent findings however reveal that the O(n) improvement is not correct. The combinatorial analysis presented here shows that, assuming a specific distribution (see Assumption 1), of mfestructures within irreducible and arbitrary structures, the expected number of Λ^{∗}candidates is Θ(n^{2}). However, the constant reduction is quite significant, being in the range of 96%. We furthermore show an analogous result for the sparsification of the Λ^{∗}decomposition rule for RNA pseudoknotted structures of genus one. Finally we observe that the effect of sparsification is sensitive to the employed energy model.
Keywords:
Sparsification; Generating function; Dynamic programmingBackground
RNA structures, diagrams and genus filtration
An RNA sequence is a linear, oriented sequence of the nucleotides (bases) A,U,G,C. These sequences “fold” by establishing bonds between pairs of nucleotides. In this paper, we only consider the WatsonCrick base pair AU or GC and wobble base pairs UG. The global conformation of an RNA molecule is determined by topological constraints encoded at the level of secondary structure, i.e., by the mutual arrangements of the base pairs [1].
Secondary structures can be interpreted as (partial) matchings in a graph of permissible base pairs [2]. They can be represented as diagrams, i.e. graphs over the vertices 1,…,n, drawn on a horizontal line with bonds (arcs) in the upper halfplane. The length of an arc (ij) is denoted by j−i. Furthermore, we call two arc (ij) and (rs) (suppose i < r) cross if i<r<j<s holds. In this representation one refers to a secondary structure without crossing arcs as a simple secondary structure and pseudoknot structure, otherwise, see Figure 1.
Figure 1. RNA structures as planar graphs and diagrams. (A) an RNA secondary structure and (B) an RNA pseudoknot structure.
A diagram is a labeled graph over the vertex set [n]={1,…,n} in which each vertex has degree ≤ 3, represented by drawing its vertices in a horizontal line. The backbone of a diagram is the sequence of consecutive integers (1,…,n) together with the edges {{i,i + 1}∣1 ≤ i ≤ n−1}. The arcs of a diagram, (i,j), where i<j, are drawn in the upper halfplane. We shall distinguish the backbone edge {i,i + 1} from the arc (i,i + 1), which we refer to as a 1arc. A stack of length ℓ is a maximal sequence of “parallel” arcs, ((i,j),(i + 1,j−1),…,(i + (ℓ−1),j−(ℓ−1))) and is also referred to as a ℓstack, see Figure 2.
Figure 2. Diagram representation and irreducibility. A diagram over {1,…,55}. The arcs (1,21) and (11,33) are crossing and the dashed arc (9,10) is a 1arc which is not allowed. This structure contains 4 stacks with length 7, 4, 6 and 4, from left to right respectively. Irreducibility relative also to a decomposition rule. The rule Λ^{∗} splitting S_{i,j} to S_{i,k} and S_{k + 1,j}, S_{1,55} is not Λ^{∗}irreducible, while S_{2,40} and S_{43,55} are. However, for a specific decomposition rule Λ, which removes the outmost arc, S_{43,55} is not Λirreducible while S_{2,40} is.
We shall consider diagrams as fatgraphs, , that is graphs G together with a collection of cyclic orderings, called fattenings. Each fatgraph determines an oriented surface [3,4] which is connected if G is and has some associated genus g(G) ≥ 0 and number r(G)≥1 of boundary components. Clearly, contains G as a deformation retract [5]. Fatgraphs were first applied to RNA secondary structures in [6,7].
A diagram hence determines a unique surface (with boundary). Filling the boundary components with discs we can pass from to a surface without boundary. Euler characteristic, χ, and genus, g, of this surface is given by χ = v−e + r and , respectively, where v,e,r is the number of discs, ribbons and boundary components in , [5]. The genus of a diagram is that of its associated surface without boundary and a diagram of genus g is referred to as gdiagram.
A gdiagram without arcs of the form (i,i + 1) (1arcs) is called a gstructure. A gdiagram that contains only vertices of degree three, i.e. does not contain any vertices not incident to arcs in the upper halfplane, is called a gmatching. A diagram is called irreducible, if and only if it cannot be split into two by cutting the backbone without cutting an arc, see Figure 2.
Folding algorithms
Folded configurations are energetically somewhat optimal. Here energy is obtained by adding contributions of loops [8] contained in RNA secondary and pseudoknot structures. Any RNA structure has a unique and disjoint decomposition into such loops which are really stems from the fatgraph [9,10] interpretation of such structures in which loops correspond to boundary components [11]. Additional constraints imply further properties, like for instance certain minimum arclength conditions [12] and the nonexistence of isolated bonds. An mfeRNA structure can be predicted in polynomials time by means of dynamic programming (DP) routines [12,13].
The most commonly used tools predicting simple RNA secondary structure mfold[13] and the Vienna RNA Package[14], require O(n^{2}) space and O(n^{3}) time. In the following we omit “simple” and refer to secondary structures containing crossing arcs as pseudoknot structures.
Generalizing the matrices of the DProutines of secondary structure folding [13,14] to gapmatrices [15], leads to a DPfolding of pseudoknotted structures [15] (pknot‐R&E) with O(n^{4}) space an O(n^{6}) time complexity. The following references provide a certainly incomplete list of DPapproaches to RNA pseudoknot structure prediction using various structure classes characterized in terms of recursion equations and/or stochastic grammars: [9,1526]. The most efficient algorithm for pseudoknot structures is [22] (pknotsRG) having O(n^{2}) space and O(n^{4}) time complexity. This algorithm however considers only a restricted class of pseudoknots.
Note that RNA secondary structures are exactly structures of topological genus zero [27]. The topological classification of RNA structures [10,11,28] has recently been translated into an efficient DPalgorithm [9]. Fixing the topological genus of RNA structures implies that there are only finitely many types, the so called irreducible shadows [11].
Sparsification
Let us have a closer look at sparsification and the results of [2931]. Sparsification is a method tailored to speed up DPalgorithms predicting mfesecondary structures [29,31]. The idea is to prune certain computation paths encountered in the DPrecursions, see Figure 3A. Let us consider the case of RNA secondary structure folding. Here sparsification reduces the DPrecursion paths to be based on so called candidates. A candidate is in this case an interval, for which the optimal solution cannot be written as a sum of optimal solutions of subintervals. This implies the structure over a candidate is an “irreducible” structures when tracing back from the optimal solution. Considering only these candidates gives the same optimal solution as considering all possible intervals. The crucial observation here is that if these irreducibles appear only at a low rate we have a significant reduction in time and space complexity.
Figure 3. (A) Sparsification of secondary structure folding. Suppose the optimal solution L_{i,j} is obtained from the optimal solutions L_{i,k}, L_{k + 1,q} and L_{q + 1,j}. Based on the recursions of the secondary structures, L_{i,k}and L_{k + 1,q} produce an optimal solution of L_{i,q}. Similarly, L_{k + 1,q} and L_{q + 1,j} produce an optimal solution of L_{k + 1,j}. Now, in order to obtain an optimal solution of L_{i,j} it is sufficient to consider either the grouping L_{i,q} and L_{q + 1,j} or L_{i,k} and L_{k + 1,j}. (B) General idea of sparsification: L_{v} is alternatively realized via and , or and . Thus it is sufficient to only consider one of the computation paths.
Sparsification has been also applied in the context of RNARNA interaction structures [30] as well as RNA pseudoknot structures [32]. In difference to RNA secondary structures, however, not every decomposition rule in the DPfolding of RNA pseudoknot structures is amendable to sparsification. By construction, sparsification can only be applied for calculating mfeenergy structures. Since the computation of the partition function [20,33] needs to take into account all substructures, sparsification does not work.
Sparsification [29,31,32] can be described as follows: let V={v_{1}v_{2},…} be a set whose elements v_{i}are unions of pairwise disjoint intervals. Let furthermore L_{v} denote an optimal solution (here optimal means to maximize the scores) of the DProutine over v. By assumption L_{v} is recursively obtained. Suppose we are given a decomposition rule Λ_{1}, for which the optimal solution L_{v} is , where . Then, under certain circumstances, the DProutine may interpret L_{v} either as or as , see Figure 3B. To be precise, this situation is encountered iff
there exists an optimal solution for a substructure over where via Λ_{2} and L_{v} is obtained from and via Λ_{1},
there exists an optimal solution for a substructure over where via Λ_{3} and L_{v} is obtained by and via Λ_{1}.
Given a decomposition
we call Λ_{2}scompatible to Λ_{1} if there exists a decomposition rule Λ_{3} such that
Note that if Λ_{2} is scompatible to Λ_{1}then Λ_{3} is scompatible to Λ_{1}. To summarize
Definition 1
(scompatible) Suppose L_{v} is the optimal solution for S_{v} over v, under decomposition rule Λ_{1}. is obtained from two optimal solutions and under rule Λ_{2}. Then Λ_{2} is called scompatible to Λ_{1} if there exist some rule Λ_{3} such that and .
Figure 3B depicts two such ways that realize the same optimal solution L_{v}. Sparsification prunes any such multiple computations of the same optimal value. Note that by symmetry, Λ_{2} and Λ_{3} are both scompatible to Λ_{1}.
We next come to the important concept of candidates. The latter mark the essential computation paths for the DProutine.
Definition 2
(Candidates) Suppose L_{v} is an optimal solution in a sense of maximizing. We call v is a Λcandidate if for any v_{1} ⊊v obtained by Λand , we have
and we shall denote the set of Λcandidates set by Q^{Λ}.
By construction a Λcandidate v is a union of disjoint intervals such that its optimal solution L_{v} cannot be obtained via a Λsplitting. This optimal solution allows to construct a nonunique arcconfiguration (substructure) over v[13,14] and the above Λsplitting consequently translates into a splitting of this substructure. This connects the notion of Λcandidates with that of substructures and shows that a Λcandidate implies a substructure that is Λirreducible.
Lemma 1
[29,32] Suppose L_{v} is obtained by selecting the optimal solution from the decomposition rules Λ_{1}Λ_{2},…,Λ_{n}. If Λ is scompatible to all Λ_{i},∀1 ≤ i ≤ n, then L_{v} can be obtained via Λcandidates.
In summary, as for the impact of sparsification, [29] claims that sparsification reduces the time complexity by a linear factor. This claim is based on the assumption that RNA molecules satisfy the polymerzeta property[29]. Subsequent studies draw a slightly different picture [31] concluding that that sparsification requires O(nZ) time, where n denotes the length of input sequence, and Z is a sparsity parameter satisfying n ≤ Z<n^{2}. Recently, it has been shown in [34] that an asymptotic time complexity of a sparsified RNA folding algorithm using standard energy parameters remains O(n^{3}) under a wide variety of condition.
Sparsification of RNA secondary structures
Here we recall some results of [29,31] on the sparsification of RNA secondary structures. Secondary structures satisfy a simple recursion which gives the optimal (maximum) solution over ij by , where V_{i,j} denotes the optimal solution in which (ij) is a base pair, and W_{i,j} denotes the optimal solution obtained by adding the optimal solutions of two subsequent intervals, respectively. Note that the optimal solution over a single vertex is denoted by L_{i,i}. We have the recursion equation for V_{i,j} and W_{i,j}:
where w(ij) is the energy contribution of (ij) forming a base pair, see Figure 4. In case two positions, i, j in the sequence are incompatible then we have w(ij)=−∞.
Figure 4. The recursion solving the optimal solution for secondary structures.
An interval [i,j] is a Λ^{∗}candidate if the optimal solution over [i,j] is given by L_{i,j}=V_{i,j}>W_{i,j}. Indeed, [i,j] is a candidate iff [i,j] is in the candidate set of Λ^{∗}, and we denote the set by Q. Suppose the optimal solution W_{i,j} is given by W_{i,j} = L_{i,q} + L_{q + 1,j} and suppose we have L_{i,q} = L_{i,k} + L_{k + 1,q}. Then since [i,q] is not a candidate, Lemma 1 shows that we can compute W_{i,j} = L_{i,k} + L_{k + 1,j}, where [i,k] is a candidate.
Sparsification on RNA pseudoknot structures
Sparsification can also be applied to the DPalgorithm folding RNA structures with pseudoknots [32]. In contrast to the decomposition rule Λ^{∗} that spliced an interval into two subsequent intervals, we encounter in the grammar for pseudoknotted structures additional more complex decomposition rules [15]. As shown in [32] there exist some decomposition rules which are not scompatible and which can accordingly not be sparsified at all, see Figure 5B. For instance, given a decomposition rule Λ in pknot‐R&E subsequent decomposition rules which are scompatible to Λ are referred to as split type of Λ [32].
Figure 5. Decomposition rules for pseudoknot structures of fixed genus (decomposed into three colors). (A) three decompositions via the rule Λ^{∗}, which is scompatible to itself. (B) three decomposition rules Λ_{1},Λ_{2},Λ_{3}where Λ_{2},Λ_{3} are scompatible to Λ_{1}. (C) three decomposition rules Λ_{1},Λ_{2},Λ_{3} where Λ_{2},Λ_{3} are not scompatible to Λ_{1}.
In the following we will study RNA pseudoknot structures of fixed topological genus, see RNA structures, diagrams and genus filtration for details. An algorithm folding such pseudoknot structures, gfold, has been presented in [9]. The decomposition rules that appear in gfold are reminiscent to those of pknot‐R&E but as they restrict the genus of substructures, the iteration of gapmatrices is severely restricted and the effect of sparsification of these decompositions is significantly smaller.
In the following, we restrict our analysis in pseudoknotted structures to only the decomposition rule Λ^{∗}, which splices an interval into two subsequent intervals. Put differently, Λ^{∗} cuts the backbone of an RNA pseudoknot structure of fixed genus g over one interval without cutting a bond.
Efficiency of sparsification
By construction, the fewer candidates the DProutine encounters, the more efficient the sparsification. Thus it is of utmost importance to analyze the number of candidates. In the case of sparsification of RNA secondary structures we have one basic decomposition rule Λ^{∗} acting on intervals, namely Λ^{∗}splices an interval into two disjoint, subsequent intervals. The implied notion of a Λ^{∗}irreducible substructure is that of a substructure nested in a maximal arc, where maximal refers to the partial order of two arcs (ij) ≤ (i^{′}j^{′}) iff i^{′} ≤ i∧j ≤ j^{′}. This observation relates irreducibility to nesting of arcs and following this line of thought [29] identifies a specific property of polymerchains introduced in [35,36] to be of relevance for the size of candidate sets:
Definition 3
(Polymerzeta property) Let denote the probability of a structure over an interval [i,j] under some decomposition rule Λ. Then we say Λ follows the polymerzeta property if for some constant b, c > 0 and m = j−i.
Polymerzeta comes from modeling the 2Dfolding of a polymer chain as a selfavoiding walk (SAW) in a 2D lattice [37]. It implies that the probability of a base pair (ij) depends only on the length of the arc, i.e. , where m = j−i. In [29] stipulate that RNA molecules satisfy the polymerzeta property and approximate by [29] using 50,000 mRNA sequences of an average length of 1992 nucleotides [38]. They find b≈2.11 and c≈1.47. The average probability is displayed in Figure 4, Page 865 [29] for increasing m. Furthermore, it is implied via Figure six, Page 867 [29] that the average number of candidates converges to a constant, implying that sparsification of DProutine folding secondary structure takes Λ(n^{2}) time complexity.
These findings have been questioned by [34], where it has been observed that the time complexity of a sparsified RNA folding algorithm based on energy minimization remains O(n^{3}) independently of the energy function used and the base composition of the RNA sequence. [34] argues that the significant effect of sparsification on the DProutine is largely a finitesize effect. Namely, when the sequence length is below some threshold, the algorithm is dominated by the quadratic time factor. In this context, it may be worth pointing out that In [31] noticed that the improvement of a sparsified basepairing maximization algorithm depends heavily on the base composition of the input. Backofen parameterizes explicitly the cardinality of candidate sets in [31].
Contribution
In this paper we study the sparsification of the decomposition rule Λ^{∗}[31,32] for RNA secondary and RNA pseudoknot structures of fixed topological genus. Based on Assumption 1 below our paper provides a combinatorial framework for quantifying the effects of sparsification of the Λ^{∗} rule.
We shall prove that the candidate set [29,31,32] is indeed small. We compute the probability of an interval being a candidate for two different energy models. For both models, this is facilitated via computing the generating function (GF) of structures and the generating function of irreducible structures. By studying the asymptotics of coefficients in these generating functions, we can compute the expected number of candidates of a uniformly random input sequence for large n. We show similar results for RNA pseudoknot structures of fixed topological genus. This provides new insights into the improvements of the sparsification of the concatenationrule Λ^{∗} in the presence of cross serial interactions. Our observations complement the detailed analysis of Backofen [31,32]. We show that although for pseudoknot structures of fixed topological genus [10,11] the effect of sparsification on the global time complexity is still unclear, the decomposition rule that splits an interval can be sped up significantly.
Methods
Suppose w is an energy function for RNA structures. Let w_{δ}(σ) denote the energy of an RNA structure σover a sequence δ. The partition function of δ is given by
where R is the universal gas constant and T is the temperature. (Here we consider w_{δ}(σ) as a positive score.) The partition function induces a probability space in which the probability of a structure σ is
The concept of a partition function is close to that of a generating function. In case of , i.e., each structure contributes equally regardless the underlying sequence and the partition function equals [z^{n}]G(z), where G is the generating function and [z^{n}]G is the coefficient of the term z^{n}.
Two important energy models are arcbased [39] and loopbased [8], respectively. The loopbased energyfiltration is different from the notion of “stickiness” [40]. The compatibility of two positions by folding random sequences is considered to be 6/16, reminiscent of the probability of two given positions to be compatible by WatsonCrick and Wobble base pairs rules.
Assumption 1
Let
where η > 1 is a constant, w(σ) is the energy value assigned to σ based on a given energy model and ℓ is the number of arcs contained in σ. Then the probability of a particular structure σ to be the mfestructure of a uniformly random input sequence is
Asymptotics
In this section we compute two generating functions and their singular expansions [11]. Let c_{g}(n) and d_{g}(n) denote the number of gmatchings and gstructures having n arcs and n vertices, respectively, with GF
The GF C_{g}(z) has been computed in the context of the virtual Euler characteristic of the modulispace of curves in [41] and D_{g}(z) can be derived from C_{g}(z) by means of symbolic enumeration [11]. The GF of genus zero diagrams C_{0}(z) is wellknown to be the GF of the Catalan numbers, i.e., the numbers of triangulations of a polygon with (n + 2) sides,
As for g ≥ 1 we have the following situation [11]
Theorem 1
Suppose g ≥ 1. Then the following assertions hold
(a) D_{g}(z) is algebraic and
In particular, is the only dominant singularity of D_{g}(z). we have for some constant a_{g}depending only on g and γ≈2.618:
(b) The bivariate GF of gstructures over n vertices, containing exactly m arcs, E_{g}(z,t), is given by
Irreducible gstructures
In the context of Λ^{∗}candidates we observed that irreducible substructures are of key importance. It is accordingly of relevance to understand the combinatorics of these structures. To this end let denote the GF of irreducible gstructures.
Lemma 2
For g ≥ 0, the GF satisfies the recursion
For a proof of Lemma 2, see Section Proofs.
Theorem 2
For g ≥ 1 we have
(a) the GF of irreducible gstructures over nvertices is given by
where , U_{g}(z) and V_{g}(z) are both polynomials with lowest degree at least 2g, and U_{g}(1/4), V_{g}(1/4) ≠ 0. In particular, for some constant and γ≈2.618:
(b) the bivariate GF of irreducible gstructures over n vertices, containing exactly m arcs, , is given by
We shall postpone the proof of Theorem 2 to Section Proofs.
The main result
Nussinovlike energy model
In the following we mimic some form of mfegstructures: inspired by the Nussinov energy model [39] we consider the weight of a gstructure over n vertices σ_{g,n} to be given by w(σ_{g,n}) = cℓ, where c is a constant contribution of a single arc and ℓ is the number of arcs in σ_{g,n}[40]. Then by Assumption 1, we have the weight function . Note that the case (6/16)η^{c}=1 corresponds to the uniform distribution, i.e. all gstructure have identical weight.
This approach requires to keep track of the number of arcs, i.e. we need to employ bivariate GF. In Theorem 1(b) we computed this bivariate GF and in Theorem 2(b) we derived from this bivariate GF , the GF of irreducible gstructures over n vertices containing ℓ arcs.
The idea now is to substitute for the second indeterminant, t, some fixed . This substitution induces the formal power series
which we regard as being parameterized by τ. Obviously, setting τ = 1 we recover D_{g}(z), i.e. we have D_{g}(z) = D_{g,1}(z) = E_{g}(z,1). Note that for τ>1/4, the polynomial τz^{2}−z + 1 has no real root. Thus we have for τ>1/4 the asymptotics
with identical exponential growth rates as long as the supercritical paradigm [42] applies, i.e. as long as γ_{τ}, the real root of minimal modulus of
is smaller than any singularity of . In this situation τ affects the constant a_{g,τ} and the exponential growth rate γ_{τ} but not the subexponential factor . The latter stems from the singular expansion of C_{g}(z). Analogously, we derive the τparameterized family of GF . We set the contribution of a single arc c=1 and the constant η = e, where e is the Euler number. Then we have the parameter τ = (6/16)e^{1}≈1.0125. By abuse of notation we will omit the subscript τassuming τ = (6/16)e^{1}.
The main result of this section is that the set of Λ^{∗}candidates is a small proportion of all entries. To put this size into context we note that the total number of entries considered for the Λ^{∗}decomposition rule is given by
Theorem 3
Suppose an mfegstructure over an interval of length m is irreducible with probability , then the expected number of candidates of gstructures for sequences of lengths n satisfies
and furthermore, setting we have
where b_{g} > 0 is a constant.
We provide an illustration of Theorem 3 in Figure 6.
Figure 6. The expected number of candidates for secondary and 1structures from an random input with a simplified arcbased energy model,and: we compute the expected number of candidates obtained by folding 100 random sequences for secondary structures (A)(solid) and 1structures (B)(solid). We also display the theoretical expectations implied by Theorem 3 (A)(dashed) and (B)(dashed).
Proof
We proof the theorem by quantifying the probability of [i,j] being a Λ^{∗}candidate. In this case any (not necessarily unique) substructure, realizing the optimal solution L_{i,j}, is Λ^{∗}irreducible, and therefore an irreducible structure over [i,j].
Let m = (j−i + 1), by assumption, the probability that [i,j] is a candidate conditional to the existence of a substructure over [i,j] is given by
Note that does not depend on the relative location of the interval but only on the intervallength. Let , then according to Theorem 1,
for m ≥ m_{0} where m_{0} > 0 and 0 < ε < 1 are constants. On the one hand
where is a constant. On the other hand, we have
Setting , we can conclude that , see Figure 7.
Figure 7. The probability distribution of(A) and(B) on a simplified arcbased energy model.
We next study the expected number of candidates over an interval of length m. To this end let
The expected cardinality of the set of Λ^{∗}candidates of length m = (j−i + 1) encountered in the DPalgorithm is given by
since there are n−(m−1) starting points for such an interval [i,j]. Therefore, by linearity of expectation, for sufficiently large m > m_{0}, with ε being a small constant. Thus we have
Consequently, the expected size of the Λ^{∗}candidate set is Λ(n^{2}). We proceed by comparing the expected number of candidates of a sequence with length n with ,
For sufficient large n ≥ n_{0}, . Furthermore
from which we can conclude and the theorem is proved. □
Loopbased energy model
In this section we discuss the loopbased energy model of RNA secondary structure folding. To be precise we evoke here trivariate GFs F(ztv) and F^{∗}(ztv) whose coefficients counting the numbers of secondary structures and irreducible secondary structures over n vertices having ℓ arcs and energy j, respectively. This becomes necessary since the loopbased model distinguishes between arcs and energy. The “cancelation” effect or reparameterization of stickiness [40] to which we referred to before does not appear in this context. Thus we need both an arc as well as an energyfiltration.
A further complication emerges. In difference to the GFs E_{g}(z,t) and the new GFs are not simply obtained by formally substituting into the power series D_{g}(z) and as bivariate terms. The more complicated energy model requires a specific recursion for irreducible secondary structures.
The energy model used in prediction of secondary structure is more complicated than the simple arcbased energy model. Loops which are formed by arcs as well as isolated vertices between the arcs are considered to give energy contribution. Loops are categorized as hairpin loops (no nested arcs), interior loops (including bulge loops and stacks) and multiloops (more than two arcs nested), see Figure 8. An arbitrary secondary structure can be uniquely decomposed into a collection of mutually disjoint loops. A result of the particular energy parameters [8] is that the energy model prefers interior loops, in particular stacks (no isolated vertex between two parallel arc), and disfavors multiloops. Based on this observation, we give a simplified energy model for a loop λcontained in secondary structure which only depends on the loop types by
• w(λ)=0.5 if λ is a hairpin loop,
Figure 8. Diagram representation of loop types in secondary structures: (A) hairpin loop, (B) interior loop, (C) multiloop.
• w(λ)=1 if λ is an interior loop,
• w(λ)=−5 if λ is a multiloop,
where λ is a loop in a structure. The energy for a secondary structure σaccordingly is given by
Let and F_{0}(z) be the energyfiltered GFs obtained by setting t=6/16 and v=η=e in F^{∗}(z,t,v) and F(z,t,v), where e is the Euler number. Then
where σ is an arbitary and σ^{′} is an irreducible secondary structure. Along these lines, ℓ, ℓ^{′} denote the number of arcs in σand σ^{′}. In other words, what happens here is that we find a suitable parameterization which brings us back to a simple univariate GF whose coefficients count the sum of weights of structures over n vertices.
Lemma 3
The energyfiltered generating function of RNA secondary structures, , satisfies the recursion
and F^{∗}(z) is uniquely determined by the above equation. Furthermore
Proof
We first consider the GF whose coefficient of z^{n} denotes the total weight of irreducible secondary structures over n vertices, where (1,n) is an arc. Thus it gives a term 6/16z^{2}. Isolated vertex lead to the term
where p denotes the minimum number of isolated vertices to be inserted. Depending on the types of loops formed by (i,n), we have
• multiloops: there are at least two irreducible substructures, as well as isolated vertices, thus
Considering the contributions from the energy model we compute
which establishes the recursion. The uniqueness of the solution as a power series follows from the fact that each coefficient can evidently be recursively computed.
An arbitrary secondary structure can be considered as a sequence of irreducible substructures with certain intervals of isolated vertices. Thus
□
Lemma 4
and F_{0}(z) have the same singular expansion.
where α≈0.24 and β≈2.88 are constants and γ≈2.1673
Proof
Solving eq. 14 we obtain a unique solution for whose coefficient are all positive. Observing the dominant singularity of is ρ≈0.4614. F_{0}(z) is a function of and we examine the real root of minimal modulus of is bigger than ρ. Then by the supercritical paradigm [42] applying, F_{0}(z) and have identical exponential growth rates. Furthermore, and F_{0}(z) have the same subexponential factor , hence the lemma. □
Theorem 4
Suppose an mfesecondary structure over an interval of length m is irreducible with probability , then the expected number of candidates from a random sequence of length n with a simplified loopbased energy model is
and furthermore, setting , we have
where b=α/β≈0.08.
Proof
By Lemma 4 we have where b is a constant. The proof is completely analogous to that of Theorem 3. □
We show the distribution of and in Figure 9.
Figure 9. The distribution of(A) andobtained by folding 100 random sequences on the loopbased model (B)(solid), as well as the theoretical expectation implied by Theorem 4 (B)(dashed).
Conclusion
In this paper we quantify the effect of sparsification of the rule Λ^{∗}. This rule splits intervals and separates concatenated substructures. The sparsification of Λ^{∗} alone is claimed to provide a speed up of up to a linear factor of the DPfolding of RNA secondary structures [29]. A similar conclusion is drawn in [30] where the sparsification of RNARNA interaction structures is shown to experience also a linear reduction in time complexity. Both papers [29,30] base their conclusion on the validity of the polymerzeta property. However, [34] comes to a different conclusion reporting a mere constant reduction in time complexity. While Λ^{∗}is the key for the time complexity reduction of secondary structure folding, it is conceivable that for pseudoknot structures there may exist nonsparsifiable rules in which case the overall time complexity is not reduced.
In any case, the key is the set of candidates and we provide an analysis of Λ^{∗}candidates by combinatorial means. In general, the connection between candidates, i.e. unions of disjoint intervals and the combinatorics of structures is actually established by the algorithm itself via backtracking: at the end of the DPalgorithm a structure is being generated that realizes the previously computed energy as mfestructure. This connects intervals and substructures.
So, does the condition c>1 in polymerzeta apply in the context of RNA structures? In fact this condition would follow if the intervals in question are distributed as in uniformly sampled structures. This however, is far from reasonable, due to the fact that the mfealgorithm deliberately designs some mfestructure over the given interval. What the algorithm produces is in fact antagonistic to uniform sampling. We here wish to acknowledge the help of one anonymous referee in clarifying this point.
Our results imply that polymerzeta does not hold. Our framework critically depends on a specific distribution of mfe structures within irreducible and arbitrary structures, explicated in Assumption 1. We have crosschecked Assumption 1 with the number of candidates in DPprograms (using the same energy model), see Figure 7 and Figure 9. With this conclusion we are in accord with [31,34] but provide an entirely different approach.
The non validity of polymerzeta has also been observed in the context of the limit distribution of the 5’3’ distances of RNA secondary structures [43]. Here it is observed that long arcs, to be precise arcs of lengths O(n) always exist. This is of course a contradiction to the polymerzeta property in case of c>1.
The key to quantification of the expected number of candidates is the singularity analysis of a pair of energyfiltered GF, namely that of a class of structures and that of the subclass of all such structures that are irreducible. We show that for various energy models the singular expansions of both these functions are essentially equal–modulo some constant. This implies that the expected number of candidates is Λ(n^{2}) and all constants can explicitly be computed from a detailed singularity analysis. The good news is that depending on the energy model, a significant constant reduction, around 96% can be obtained. This is in accordance with data produced in [31] for the mfefolding of random sequences. There a reduction by 98% is reported for sequences of length ≥ 500.
Our findings are of relevance for numerous results, that are formulated in terms of sizes of candidate sets [32]. These can now be quantified. It is certainly of interest to devise a full fledged analysis of the loopbased energy model. While these computations are far from easy our framework shows how to perform such an analysis.
Using the paradigm of gapmatrices Backofen has shown [32] that the sparsification of the DPfolding of RNA pseudoknot structures exhibits additional instances, where sparsification can be applied, see Figure 5B. Our results show that the expected number of candidates is Λ(n^{2}), where the constant reduction is around 90%. This is in fact very good new since the sequence length in the context of RNA pseudoknot structure folding is in the order of hundreds of nucleotides. So sparsification of further instances does have an significant impact on the time complexity of the folding.
Proofs
In this section, we prove Lemma 2 and Theorem 2.
Proof for Lemma 2: let D(z,u) and D^{∗}(z,u) be the bivariate GF , and . Suppose a structure contains exactly j irreducible structures, then
and
as well as . Let . Then F(z,u)D(z,u)=1, whence for g ≥ 1,
and F_{0}(z)D_{0}(z)=1, where . Furthermore, we have and
Proof for Theorem 2 Let [n]_{k}denote the set of compositions of n having k parts, i.e. for σ∈[n]_{k} we have σ=(σ_{1},…,σ_{k}) and .Claim.
We shall prove the claim by induction on g. For g=1 we have
whence eq. (22) holds for g=1. By induction hypothesis, we may now assume that for j ≤ g, eq. (22) holds. According to Lemma 2, we have
We next observe
and setting h=g_{1}−j we obtain,
and setting j=g−h
Consequently, the Claim holds for any g ≥ 1.
For any g ≥ 1, we have [11]
where P_{g}(u) is a polynomial with integral coefficients of degree at most (3g−1), P_{g}(1/4)≠0, u^{2g}P_{g}(u)≠0 and u^{h}P_{g}(u)=0 for 0 ≤ h ≤ 2g−1. Let , the Claim provides in this context the following interpretation of
and
As 0 ≤ j ≤ g−2 and g−j ≤ s ≤ 2g + 1−2j, we have s ≥ 2. Consequently we arrive at
where
and
We have for σ∈[g]_{k}, k ≥ 1
where , h_{i} ≥ 0. Then we obtain that
Since , h_{i}≤2σ_{i}−1, and . Thus for 0 ≤ h ≤ 2g−1,
As shown in [11] we have
and we obtain U_{g}(1/4)=P_{g}(1/4)/4. Furthermore,
We can recruit the computation of [11] in order to observe . In order to compute the bivariate GF, , we only need to replace in eq. (22) D_{g}(z) by E_{g}(zt) and the proof is completely analogous.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
FWDH and CMR contributed equally to research and manuscript. Both authors read and approved the final manuscript.
Acknowledgements
We want to thank Thomas J.X. Li for discussions and comments. We want to thank an anonymous referee for pointing out an incorrect assumption of first version of this paper.
References

Bailor MH, Sun X, AlHashimi HM: Topology Links RNA Secondary Structure with Global Conformation, Dynamics, and Adaptation.
Science 2010, 327:202206. PubMed Abstract  Publisher Full Text

Tabaska JE, Cary RB, Gabow HN, Stormo GD: An RNA folding method capable of identifying pseudoknots and base triples.
Bioinformatics 1998, 14:691699. PubMed Abstract  Publisher Full Text

Loebl M, Moffatt I: The chromatic polynomial of fatgraphs and its categorification.
Adv. Math. 2008, 217:15581587. Publisher Full Text

Penner RC, Knudsen M, Wiuf C, Andersen JE: Fatgraph models of proteins.
Comm Pure Appl Math 2010, 63:12491297. Publisher Full Text

Massey WS: Algebraic Topology: An Introduction. SpringerVeriag, New York; 1967.

Penner RC, Waterman MS: Spaces of RNA secondary structures.
Adv. Math 1993, 101:3149. Publisher Full Text

Penner RC: Cell decomposition and compactification of Riemann’s moduli space in decorated Teichmü, ller theory. In Woods Hole Mathematicsperspectives in math and physics. Edited by Tongring N, Penner RC. Singapore; World Scientific 2004:263301.
[ArXiv: math. GT/0306190]

Mathews D, Sabina J, Zuker M, Turner D: Expanded sequence dependence of thermodynamic parameters improves prediction of RNA secondary structure.
J. Mol. Biol 1999, 288:911940. PubMed Abstract  Publisher Full Text

Reidys CM, Huang FWD, Andersen JE, Penner RC, Stadler PF, Nebel ME: Topology and prediction of RNA pseudoknots.
Bioinformatics 2011, 27:10761085. PubMed Abstract  Publisher Full Text

Bon M, Vernizzi G, Orland H, Zee A: Topological Classification of RNA Structures.
J Mol Biol 2008, 379:900911. PubMed Abstract  Publisher Full Text

Andersen JE, Penner RC, Reidys CM, Waterman MS: Topological classification and enumeration of RNA structrues by genus.
J. Math. Biol 2011.
[Prepreint]
Publisher Full Text 
Zuker M: On finding all suboptimal foldings of an RNA molecule.
Science 1989, 244:4852. PubMed Abstract  Publisher Full Text

Hofacker IL, Fontana W, Stadler PF, Bonhoeffer LS, Tacker M, Schuster P: Fast Folding and Comparison of RNA Secondary Structures.
Monatsh Chem 1994, 125:167188. Publisher Full Text

Rivas E, Eddy SR: A Dynamic Programming Algorithm for RNA Structure Prediction Including Pseudoknots.
J Mol Biol 1999, 285:20532068. PubMed Abstract  Publisher Full Text

Uemura Y A Hasegawa, Kobayashi S, Yokomori T: Tree adjoining grammars for RNA structure prediction.
Theor Comp Sci 1999, 210:277303. Publisher Full Text

Akutsu T: Dynamic programming algorithms for RNA secondary structure prediction with pseudoknots.
Discr Appl Math 2000, 104:4562. Publisher Full Text

Lyngsø RB, Pedersen CN: RNA pseudoknot prediction in energybased models.
J Comp Biol 2000, 7:409427. Publisher Full Text

Cai L, Malmberg RL, Wu Y: Stochastic modeling of RNA pseudoknotted structures: a grammatical approach.

Dirks RM, Pierce NA: A partition function algorithm for nucleic acid secondary structure including pseudoknots.
J Comput Chem 2003, 24:16641677. PubMed Abstract  Publisher Full Text

Deogun JS, Donis R, Komina O, Ma F: RNA secondary structure prediction with simple pseudoknots.
Proceedings of the second conference on AsiaPacific bioinformatics (APBC 2004), Australian Computer Society 2004, 239246.

Reeder J, Giegerich R: Design, implementation and evaluation of a practical pseudoknot folding algorithm based on thermodynamics.
BMC Bioinformatics 2004, 5:104. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Li H, Zhu D: A New Pseudoknots Folding Algorithm for RNA Structure Prediction. In COCOON 2005, Volume 3595. Edited by Wang L. Springer, Berlin; 2005:94103.

Matsui H, Sato K, Sakakibara Y: Pair stochastic tree adjoining grammars for aligning and predicting pseudoknot RNA structures.
Bioinformatics 2005, 21:26112617. PubMed Abstract  Publisher Full Text

Kato Y, Seki H, Kasami T: RNA Pseudoknotted Structure Prediction Using Stochastic Multiple ContextFree Grammar.

Chen HL, Condon A, Jabbari H: An O(n5) Algorithm for MFE Prediction of Kissing Hairpins and 4Chains in Nucleic Acids.
J Comp Biol 2009, 16:803815. Publisher Full Text

Waterman MS: Secondary structure of singlestranded nucleic acids.

Orland H, Zee A: RNA folding and large N matrix theory.
Nuclear Physics B 2002, 620:456476. Publisher Full Text

Wexler Y, Zilberstein C, ZivUkelson M: A study of accessible motifs and RNA complexity.
J Comput Biol 2007, 14(6):856872. PubMed Abstract  Publisher Full Text

Salari R, Möhl M, Will S, Sahinalp C, Backofen R: Time and space efficient RNARNA interaction prediction via sparse folding.

Backofen R, Tsur D, Zakov S, ZivUkelson M: Sparse RNA folding: Time and space efficient algorithms.
J Disc Algor 2011, 9(1):1231. Publisher Full Text

Möhl M, Salari R, Will S, Backofen R, Sahinalp SC: Sparsification of RNA structure prediction including pseudoknots.
Algorithms Mol Biol 2010, 5:39. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

McCaskill JS: The equilibrium partition function and base pair binding probabilities for RNA secondary structure.
Biopolymers 1990, 29:11051119. PubMed Abstract  Publisher Full Text

Dimitrieva S, Bucher P: Practicality and time complexity of a sparsified RNA folding algorithm.
J Bioinfo Comput Biol 2012, 10(2):1241007. Publisher Full Text

Kafri Y, Mukamel D, Peliti L: Why is the DNA Denaturation Transition First Order?
Phys Rev Lett 2000, 85:49884991. PubMed Abstract  Publisher Full Text

Kabakcioglu A, Stella AL: A scalefree network hidden in the collapsing polymer.

Vanderzande C: Lattic models of polymers. Cambridge University Press, New York; 1998.

NCBI database [ http://www.ncbi.nlm.nih.gov/guide/dnarna/#downloads_ webcite]

Nussinov R, Piecznik G, Griggs JR, Kleitman DJ: Algorithms for Loop Matching.
SIAM J Appl Math 1978, 35:6882. Publisher Full Text

Nebel ME: Investigation of the Bernoulli model for RNA secondary structures.

Zagier D: On the distribution of the number of cycles of elements in symmetric groups.

Flajolet P, Sedgewick R: Analytic Combinatorics. Cambridge University Press, New York; 2009.

Han HSW, Reidys CM: The 5’3’ distance of RNA secondary structures.
J Comput Biol 2012, 19(7):867878. PubMed Abstract  Publisher Full Text