Reverse-Complement Constraint in DNA Codes

• Reverse-complement constraint is a symmetry principle defined by the Watson–Crick involution that governs admissibility and reversibility in string systems.
• It is applied in formal language models, DNA-code design, and learning algorithms, influencing hairpin completions, duplication rules, and genomic sequence modeling.
• It underpins the design of duplication-correcting codes and assembly algorithms, ensuring computational and empirical invariance for error-correcting and sequence analysis.

In the literature considered here, the reverse-complement constraint denotes a family of admissibility, symmetry, and closure conditions induced by the Watson–Crick involution on a finite alphabet. For $\Sigma=\{A,C,G,T\}$ with $c(A)=T$, $c(T)=A$, $c(C)=G$, and $c(G)=C$, the reverse complement of $s=s_1\cdots s_n$ is $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$, and $RC(RC(s))=s$ (Boneh et al., 2024). Depending on the domain, the constraint appears as an exact string-rewriting rule, an algebraic closure condition for DNA codes, a probabilistic symmetry behind Chargaff’s second parity rule, or an invariance/equivariance prior for sequence models (Hart et al., 2011, Ma, 23 Sep 2025).

1. Formal definitions and principal variants

Several formalizations recur. In string systems and formal-LLMs, legality itself is constrained by reverse complementarity. In the reverse-complement constrained variant of hairpin completion, a right hairpin completion of length $k$ has the form

$S \mapsto S\cdot RC(pref_k(S)),$

and a left hairpin completion of length $c(A)=T$0 has the form

$c(A)=T$1

Only reverse complements of already-present prefixes or suffixes may be appended; arbitrary insertions are forbidden (Boneh et al., 2024).

In DNA-code design, the constraint is often stated as a distance condition. A DNA code $c(A)=T$2 satisfies the reverse-complement constraint with parameter $c(A)=T$3 when

$c(A)=T$4

for any $c(A)=T$5 and $c(A)=T$6; when $c(A)=T$7, the chapter on algebraic approaches refers to this simply as the RC constraint (Benerjee et al., 2 Oct 2025). In closure-based constructions, the same idea is expressed as $c(A)=T$8 for each $c(A)=T$9, so reverse-complemented codewords remain inside the image code (Benerjee et al., 2 Oct 2025).

In learning problems, the constraint becomes output-space alignment rather than exact equality of raw predictions. For a model $c(T)=A$0 and a task-aware alignment operator $c(T)=A$1, the relevant condition is

$c(T)=A$2

with $c(T)=A$3 for strand-invariant sequence-level tasks and a reversal-plus-strand-swap operator for profile prediction (Ma, 23 Sep 2025). Strand-specific tasks are explicitly excluded from this symmetry assumption, since enforcing it is harmful in that regime (Ma, 23 Sep 2025).

2. Reverse-complement constraints in string systems

Hairpin completion makes the reverse-complement constraint the defining biochemical rule of the dynamics. The induced hairpin completion distance

$c(T)=A$4

is directed rather than symmetric: the examples $c(T)=A$5 and $c(T)=A$6 show that completion can enlarge a string but not shorten it (Boneh et al., 2024). Boneh–Fried–Miclaus–Popa give an $c(T)=A$7 algorithm for the problem, and the lower-bound result shows that for every $c(T)=A$8 there is no $c(T)=A$9-time algorithm unless SETH is false (Boneh et al., 2024).

A second formal setting is the reverse-complement string-duplication system, where a fixed-length duplication rule inserts $c(C)=G$0 immediately after a factor $c(C)=G$1:

$c(C)=G$2

Its expressive power depends on duplication length parity and on parity-sensitive letter availability in the seed. For $c(C)=G$3, the system is fully expressive if and only if $c(C)=G$4 is odd and $c(C)=G$5, or $c(C)=G$6 is even and $c(C)=G$7; for $c(C)=G$8, full expressiveness occurs if and only if $c(C)=G$9 (Ben-Tolila et al., 2021).

The binary $c(G)=C$0 case exhibits a notable separation between combinatorial reachability and probabilistic typicality. Over $c(G)=C$1 with seed $c(G)=C$2, the system has full capacity,

$c(G)=C$3

yet zero entropy-rate. The almost-sure limiting 2-factor frequencies satisfy

$c(G)=C$4

so the random process concentrates on a semiconstrained alternating regime despite exponential combinatorial growth (Ben-Tolila et al., 2021).

3. Duplication-correcting codes and DNA storage

In in-vivo DNA storage, reverse-complement duplication is modeled as the adjacent insertion of $c(G)=C$5 after $c(G)=C$6. One central design rule is to forbid short adjacent reverse-complement repeats. An $c(G)=C$7-reverse-complement-duplication root is a word with no adjacent reverse-complement repeats of length $c(G)=C$8. If $c(G)=C$9 is an $s=s_1\cdots s_n$0-RCD root and $s=s_1\cdots s_n$1 with $s=s_1\cdots s_n$2, then the leftmost newly created adjacent RC repeat uniquely identifies the duplication block. Taking $s=s_1\cdots s_n$3 yields a disjoint $s=s_1\cdots s_n$4-duplication-correcting code that corrects an arbitrary number of reverse-complement duplications using just one redundant symbol, provided $s=s_1\cdots s_n$5, with average encoding and decoding complexity $s=s_1\cdots s_n$6 (Sun et al., 1 Feb 2026).

For a single reverse-complement duplication of arbitrary length, a Gilbert–Varshamov argument gives

$s=s_1\cdots s_n$7

For $s=s_1\cdots s_n$8, two explicit constructions correct $s=s_1\cdots s_n$9 length-one reverse-complement duplications, with redundancies

$RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$0

and

$RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$1

respectively (Sun et al., 1 Feb 2026).

The asymptotic capacity picture is sharply split by duplication length. For reverse-complement duplication-correcting codes capable of correcting any number of duplications, the capacity is $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$2 for $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$3 and even $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$4, but $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$5 for every $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$6. For palindromic duplication, the corresponding $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$7 capacity is $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$8, and it is likewise $RC(s)=c(s_n)c(s_{n-1})\cdots c(s_1)$9 for $RC(RC(s))=s$0 (Yohananov et al., 2023). This suggests that unbounded correction of longer reverse-complement duplications is fundamentally different from the length-one regime.

4. Algebraic and combinatorial realizations

Over $RC(RC(s))=s$1, with $RC(RC(s))=s$2, $RC(RC(s))=s$3, $RC(RC(s))=s$4, and $RC(RC(s))=s$5, the Watson–Crick complement becomes the coordinate-wise map $RC(RC(s))=s$6. A linear code $RC(RC(s))=s$7 satisfies the reverse-complement constraint precisely when it is invariant under the reverse permutation and contains the repetition code $RC(RC(s))=s$8 (García-Claro, 23 Jun 2025). Reversible codes are exactly the $RC(RC(s))=s$9-submodules of $k$0, and the paper gives explicit generator matrices, isomorphism types of the form $k$1, and counting formulas for all such codes (García-Claro, 23 Jun 2025).

Comparable criteria recur over larger rings. For double cyclic codes over $k$2 with $k$3, reverse-complement closure is equivalent to reversibility together with

$k$4

(Acharya et al., 15 Dec 2025). For cyclic DNA codes over $k$5, the criterion is that $k$6 be reversible and contain $k$7, where $k$8 (Mostafanasab et al., 2016). Over $k$9, the analogous condition is membership of $S \mapsto S\cdot RC(pref_k(S)),$0, while over $S \mapsto S\cdot RC(pref_k(S)),$1 and $S \mapsto S\cdot RC(pref_k(S)),$2 the odd-length theory is organized by reciprocal and self-reciprocal generator polynomials (Zhu et al., 2015, Dertli et al., 2016).

The same symmetry also appears in de Bruijn theory as the condition

$S \mapsto S\cdot RC(pref_k(S)),$3

equivalently $S \mapsto S\cdot RC(pref_k(S)),$4 for all $S \mapsto S\cdot RC(pref_k(S)),$5 in a binary sequence of period $S \mapsto S\cdot RC(pref_k(S)),$6. For odd order $S \mapsto S\cdot RC(pref_k(S)),$7 there always exist order-$S \mapsto S\cdot RC(pref_k(S)),$8 binary de Bruijn sequences satisfying this condition, thereby settling Fredricksen’s forty-year-old open problem; for even $S \mapsto S\cdot RC(pref_k(S)),$9, such CR de Bruijn sequences do not exist (Chang et al., 2024).

5. Probabilistic genomics and empirical asymmetry

On a single DNA strand, the reverse-complement constraint is formalized by Chargaff’s second parity rule,

$c(A)=T$00

or, in the Gibbsian model, by exact equality of cylinder probabilities. If the single-strand potential satisfies

$c(A)=T$01

then the unique translation-invariant Gibbs measure $c(A)=T$02 is invariant under the reverse-complement transformation $c(A)=T$03 and therefore complies with CSPR for all $c(A)=T$04 (Hart et al., 2011). A dinucleotide test based on

$c(A)=T$05

with asymptotic $c(A)=T$06 null law was applied to 1049 complete bacterial genomes: the null was accepted in 410 genomes and rejected in 639, and no relationship was found between GC content, genome length, and rejection of the null (Hart et al., 2011).

At the level of genomic word organization, reverse-complement symmetry of frequencies does not imply symmetry of spatial arrangement. In the human genome, reverse complementary word pairs for $c(A)=T$07 were compared through Euclidean distance, Jeffreys divergence, and peak dissimilarity on inter-occurrence distance distributions up to $c(A)=T$08; peak dissimilarity worked best in this setting (Tavares et al., 2017). The study reports reverse complementary word pairs with very dissimilar distance distributions, as well as pairs with very similar distance distributions even when both distributions are irregular and contain strong peaks, and concludes that some asymmetries in the human genome go far beyond Chargaff’s rules (Tavares et al., 2017).

6. Algorithmic assembly and reverse-complement-aware learning

In shortest common superstring with reverse complements, each input string may be used either as given or reverse-complemented, so the overlap graph carries a bidirectional orientation constraint. A refined analysis of RC-MGREEDY and RC-TGREEDY established approximation ratios $c(A)=T$09 and $c(A)=T$10, respectively (Yamano et al., 22 Jan 2026). A later gadget-based algorithm computes an optimal constrained cycle cover via a reduction to maximum-weight perfect matching and improves the approximation ratio to

$c(A)=T$11

the same work proves that it is NP-hard to approximate SCS-RC within a factor better than

$c(A)=T$12

even for the DNA alphabet (Yamano et al., 27 Mar 2026).

For DNA LLMs, the reverse-complement constraint becomes a training prior. Reverse-Complement Consistency Regularization augments the task loss by penalizing divergence between a prediction on $c(A)=T$13 and the aligned prediction on $c(A)=T$14:

$c(A)=T$15

with task-specific alignment $c(A)=T$16 for sequence classification, scalar regression, and profile prediction (Ma, 23 Sep 2025). Across Nucleotide Transformer, HyenaDNA, and DNABERT-2, RCCR substantially improves RC robustness by dramatically reducing prediction flips and errors while maintaining or improving task accuracy. Representative results include NT-v2 splice-family mean, where RCCR achieved AUPRC $c(A)=T$17 versus $c(A)=T$18 for RC-Aug and SFR $c(A)=T$19 versus $c(A)=T$20, and a strand-classification negative control where the method was appropriately detrimental because the labels were orientation-dependent (Ma, 23 Sep 2025).

Across these settings, the reverse-complement constraint is not a single theorem or algorithm but a unifying symmetry principle. It governs which string operations are legal, which codewords are admissible, which genomic distributions are expected at equilibrium, which approximations are possible in assembly, and which inductive biases are appropriate in modern sequence models.