Polynomial algorithms for protein similarity search for restricted mRNA structures

In this paper we consider the problem of computing an mRNA sequence of maximal similarity for a given mRNA of secondary structure constraints, introduced by Backofen et al. in [BNS02] denoted as the MRSO problem. The problem is known to be NP-complete for planar associated implied structure graphs of vertex degree at most 3. In [BFHV05] a first polynomial dynamic programming algorithms for MRSO on implied structure graphs with maximum vertex degree 3 of bounded cut-width is shown. We give a simple but more general polynomial dynamic programming solution for the MRSO problem for associated implied structure graphs of bounded clique-width. Our result implies that MRSO is polynomial for graphs of bounded tree-width, co-graphs, $P_4$-sparse graphs, and distance hereditary graphs. Further we conclude that the problem of comparing two solutions for MRSO is hard for the class of problems which can be solved in polynomial time with a number of parallel queries to an oracle in NP.

💡 Research Summary

The paper addresses the mRNA Structure Optimization (MRSO) problem, originally introduced by Backofen et al., which asks for an mRNA sequence that respects a given secondary‑structure constraint while maximizing similarity to a target amino‑acid sequence. The secondary structure is modeled as a graph whose vertices correspond to nucleotides (or, more conveniently, to codons—triples of nucleotides) and whose edges connect complementary nucleotides according to the biological pairing set Γ = {(A,U),(C,G)}. The objective function is the sum of similarity scores f_i for each codon, where each f_i maps a triple of nucleotides to a rational value derived from PAM matrices.

Previous work (BFHV05) showed that MRSO‑d1 (the restriction where the structure graph has vertex degree at most one) is NP‑complete on planar graphs of maximum degree three, but becomes polynomial when the cut‑width of the underlying implied structure graph is bounded. This restriction, however, does not cover many biologically relevant RNA structures that contain more intricate bonding patterns.

The authors propose a far more general parameterization based on clique‑width, a graph invariant that measures how a graph can be constructed using four elementary operations on labeled vertices: creation of a single labeled vertex, disjoint union, relabeling, and insertion of all edges between two label classes. A graph has clique‑width k if it can be described by a k‑expression (a tree‑like term using those operations). Importantly, many graph families—trees, bounded‑tree‑width graphs, cographs, P₄‑sparse graphs, distance‑hereditary graphs—have bounded clique‑width, while planar graphs and grids do not.

The core technical contribution is an dynamic‑programming scheme that works directly on a given k‑expression of the implied structure graph. For each sub‑expression X the algorithm computes a set F(X) of pairs (L, f), where L records which label‑class of the expression is assigned to which codon‑label (a triple from Σ³) and f is the accumulated similarity score for the subgraph. The size of F(X) is bounded by O(|V|·|Σ|³·k), i.e., polynomial in the input size for fixed k. The algorithm respects the following recurrence:

1. Base case – a single vertex •a yields all possible codon assignments and their scores in constant time.
2. Disjoint union (⊕) – combine the sets from the two operands by Cartesian product, adding the scores.
3. Edge insertion (η_{a,b}) – filter out those pairs where any (label a, codon ℓ₁) and (label b, codon ℓ₂) violate the complementarity condition Γ.
4. Relabeling (ρ_{a→b}) – replace label a by b throughout the pair.

Because each operation can be performed in polynomial time and the expression tree has linear size, the whole procedure runs in polynomial time for any fixed clique‑width k. Consequently, Theorem 4.1 states that MRSO is solvable in polynomial time on any graph given together with a k‑expression for bounded k.

From this general result the authors derive several corollaries:

• Bounded tree‑width graphs (tree‑width tw ⇒ clique‑width ≤ 3·2^{tw}−1) admit polynomial‑time MRSO.
• Cographs (clique‑width ≤ 2) and distance‑hereditary graphs (clique‑width ≤ 3) are also covered.
• P₄‑sparse and related graph families, which are known to have bounded clique‑width, fall under the same tractability umbrella.

The paper also investigates the comparison problem: given two optimal solutions, decide whether they have the same objective value. By constructing a suitable reduction, the authors show this problem is P^{NP}_k‑complete, i.e., it belongs to the class of problems solvable in polynomial time with a bounded number of parallel queries to an NP oracle. This indicates that while finding a single optimal solution may be easy on bounded‑clique‑width graphs, verifying uniqueness or comparing multiple solutions remains computationally hard.

In biological terms, the work suggests that the “structural complexity” of an RNA molecule can be captured by its clique‑width. When this measure is small—something that can happen even for non‑planar, non‑outer‑planar structures—efficient design of high‑similarity mRNA sequences is feasible. This expands the practical applicability of MRSO algorithms beyond the previously studied outer‑planar or low cut‑width cases, potentially aiding synthetic biology applications where custom mRNA design under structural constraints is required.

The paper concludes with several avenues for future research: (1) developing heuristics or approximation schemes for graphs with large clique‑width, (2) integrating experimental data to estimate the clique‑width of real RNA structures, and (3) extending the dynamic‑programming framework to related problems such as RNA‑RNA interaction prediction or splicing site optimization. Overall, the study bridges graph‑theoretic parameterized complexity with computational biology, delivering a robust algorithmic toolkit for a broader class of mRNA structure optimization problems.