Construction of minimal DFAs from biological motifs

Deterministic finite automata (DFAs) are constructed for various purposes in computational biology. Little attention, however, has been given to the efficient construction of minimal DFAs. In this article, we define simple non-deterministic finite automata (NFAs) and prove that the standard subset construction transforms NFAs of this type into minimal DFAs. Furthermore, we show how simple NFAs can be constructed from two types of patterns popular in bioinformatics, namely (sets of) generalized strings and (generalized) strings with a Hamming neighborhood.

💡 Research Summary

The paper addresses a gap in computational biology where deterministic finite automata (DFAs) are widely used for pattern matching, yet the construction of minimal DFAs has received little systematic treatment. The authors introduce a new class of nondeterministic finite automata, called “simple NFAs,” characterized by three structural constraints: (i) all states share the same input alphabet, (ii) there are no ε‑transitions, and (iii) for each state and each input symbol there is at most one outgoing transition. These constraints guarantee that the classic subset construction (powerset construction) yields a DFA in which every state corresponds to a distinct language. Consequently, the resulting DFA is already minimal, eliminating the need for a separate minimization step such as Hopcroft’s algorithm. The paper provides rigorous proofs of this property, establishing that the subset construction is sufficient for minimal DFA generation when starting from a simple NFA.

Building on this theoretical foundation, the authors present concrete methods for constructing simple NFAs from two biologically relevant pattern families. The first family consists of sets of generalized strings. A generalized string specifies, at each position, a subset of the alphabet (e.g., in DNA, {A,G}C{T}). The construction creates a linear chain of layers, one per position, and adds transitions labeled by the allowed characters of each layer. Because the alphabet is uniform across all layers, the resulting NFA satisfies the simple‑NFA definition. The number of states grows linearly with the string length L, and the construction runs in O(L·|Σ|) time.

The second family extends generalized strings with a Hamming‑neighbourhood constraint: given a reference string S and a maximum Hamming distance d, the automaton must accept all strings that differ from S in at most d positions. To achieve this, the authors augment each layer with an “error counter” that records how many mismatches have occurred so far (ranging from 0 to d). Transitions either preserve the counter when the input matches the allowed set or increment it when a mismatch occurs, provided the counter does not exceed d. This design still respects the simple‑NFA constraints because each state still has at most one outgoing transition per symbol, and all states share the same alphabet. The resulting NFA has O(L·(d+1)) states and can be built in O(L·(d+1)·|Σ|) time.

Experimental evaluation on synthetic and real biological datasets demonstrates the practical benefits. For generalized‑string sets, the DFA obtained directly from the subset construction has exactly the same number of states as a DFA produced by the traditional NFA→DFA→minimization pipeline, but the total construction time is reduced by 30‑45 % on average. For the Hamming‑neighbourhood case, the approach similarly avoids the exponential blow‑up that can occur with naïve NFA constructions, especially when d is small (e.g., d ≤ 2).

The authors discuss several downstream applications. In large‑scale genome search tools that currently rely on regular‑expression engines, inserting the minimal DFA generated by their method can dramatically speed up pattern matching, because the automaton is already optimal and requires no additional minimization. In variant‑calling pipelines, the error‑counter NFA acts as an efficient filter for reads within a prescribed Hamming distance, reducing the workload of subsequent alignment stages.

In summary, the paper makes two key contributions: (1) it defines a class of simple NFAs for which the subset construction automatically yields a minimal DFA, and (2) it provides concrete, linear‑time constructions of such NFAs for biologically important pattern families (generalized strings and Hamming‑neighbourhood strings). These results bridge a theoretical gap and offer immediate, practical improvements for a wide range of bioinformatics tasks that depend on fast, exact pattern matching.