Gated Reverse Complement (GRC)

• GRC is a sequence-encoding mechanism that processes forward and reverse-complement DNA strands in parallel using a shared Transformer encoder.
• It adaptively fuses dual representations with a learned gating function to capture biologically meaningful strand symmetry.
• Empirical results show that integrating GRC reduces pretraining perplexity and enhances long-range genomic modeling in TrinityDNA.

Gated Reverse Complement (GRC) is a bio-inspired sequence-encoding mechanism introduced in TrinityDNA to make DNA modeling explicitly aware of reverse-complement symmetry, i.e., the fact that a biological signal on one strand may also be represented on the opposite strand as its reverse complement. In TrinityDNA, GRC is implemented as a shared-parameter dual-branch encoding mechanism in which a forward sequence and its reverse complement are processed in parallel by the same Transformer-based encoder and then fused by a learned gate. The module is positioned as one of TrinityDNA’s two main bio-inspired components, alongside Groove Fusion, and is used to improve representation learning for long DNA sequences by exploiting strand symmetry rather than treating sequence orientation as arbitrary (Yang et al., 25 Jul 2025).

1. Biological basis and modeling objective

DNA is antiparallel and complementary: if one strand contains a sequence $S=(s_1,\dots,s_N)$, the other strand carries the reverse-complement sequence

$S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$

where $s_i^C$ denotes the Watson–Crick complement under the rules $A\leftrightarrow T$ and $C\leftrightarrow G$. TrinityDNA motivates GRC by noting that many genomic functions are orientation-invariant or strand-symmetric: promoters, enhancers, and other regulatory motifs may appear on either strand; transcription factor motifs can be recognized in either orientation; and the same biological site should not be treated as different merely because the input was read from the opposite strand (Yang et al., 25 Jul 2025).

Within that framing, GRC is intended to preserve biologically meaningful symmetry at the representation level. The paper explicitly states that incorporating reverse-complement information “enhances [the model’s] ability to capture symmetrical and complementary patterns,” and describes the mechanism as crucial for “long bi-directional genomic modeling” (Yang et al., 25 Jul 2025).

A broader statistical-mechanical motivation for reverse-complement symmetry appears in work on Chargaff’s second parity rule. Under a translation-invariant Gibbs model whose energy is invariant under reverse complement, the probability of every short DNA word equals the probability of its reverse complement, yielding CSPR as a probabilistic equilibrium law (Hart et al., 2011). This suggests that reverse-complement symmetry is not merely a modeling convenience but can also be viewed as a distributional regularity tied to duplex structure.

2. Formal mechanism and gating rule

Conceptually, GRC is a shared-parameter dual-branch encoding mechanism. TrinityDNA describes the procedure in three steps: feed the forward DNA sequence $S$ through a Transformer-based encoder; feed the reverse-complement sequence $S^R$ through the same encoder; and combine the two representations with a learned gating mechanism (Yang et al., 25 Jul 2025).

The paper defines the reverse complement as

$S^R = (s_N^C, s_{N-1}^C, \ldots, s_1^C).$

The GRC output is then written as

$$\text{Output} = \text{GRC}(S, S^R) = f_\theta(S) + \sigma\!\left(W_G \cdot f_\theta(\text{Flip}(S^{R}))\right),$$

where $f_\theta(\cdot)$ is the shared SMWA-equipped Transformer encoder, $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$0 is a learned weight matrix for gating, and $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$1 is described as either a sigmoid or identity function. The operation $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$2 “means to reverse the sequence as the original order,” i.e., it restores orientation alignment when mapping the reverse-complement branch back into the same coordinate convention (Yang et al., 25 Jul 2025).

This formulation has several immediate consequences that are explicit in the source. The forward branch representation is $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$3, the reverse-complement branch representation is $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$4, and the reverse-complement branch is passed through a learned gate before being added to the forward representation. The paper does not present the mechanism as a strict equivariance constraint with exact averaging; instead, the gate allows the model to weight the contribution of the reverse-complement view in a learnable way. In that sense, GRC is symmetry-aware but adaptive, rather than hard-equivariant (Yang et al., 25 Jul 2025).

A common misconception is to equate GRC with reverse-complement data augmentation. TrinityDNA distinguishes the two explicitly: GRC does not merely augment the training set with reverse complements, but builds reverse-complement handling directly into the architecture through a shared encoder operating on both orientations in parallel, followed by gated fusion (Yang et al., 25 Jul 2025).

3. Architectural placement in TrinityDNA

In TrinityDNA, GRC sits at the sequence-encoding level rather than as a preprocessing trick. The relevant pipeline is described as follows. DNA is tokenized at the character level with vocabulary size $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$5, namely $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$6, where $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$7 denotes an unknown symbol. Pretraining uses masked language modeling, with $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$8 of tokens masked. The raw sequence is first enriched with local structural patterns using Groove Fusion, which employs convolution kernels of sizes $S^R=(s_N^C,s_{N-1}^C,\dots,s_1^C),$9, $s_i^C$0, and $s_i^C$1 to model minor/major groove structure. The resulting representations are then passed to a Transformer encoder using Sliding Multi-Window Attention (SMWA), in which different attention heads are assigned different window sizes $s_i^C$2 so that some heads emphasize local motifs and others capture longer-range dependencies. GRC then runs the same encoder $s_i^C$3 on both $s_i^C$4 and $s_i^C$5 and merges the outputs by the gating rule above (Yang et al., 25 Jul 2025).

The architecture is summarized in the paper as “sequence, structure, and strategy.” Groove Fusion models DNA’s physical geometry through multi-kernel convolutions; SMWA handles long-context dependencies efficiently and avoids attention oversmoothing by assigning different attention windows to different heads; and GRC handles strand symmetry. These components are presented as complementary rather than interchangeable: GFM addresses local structural patterns, SMWA addresses long-range dependencies, and GRC addresses orientation and reverse-complement invariance (Yang et al., 25 Jul 2025).

This placement matters for interpretation. GRC does not operate on raw symbols alone; it fuses two contextual encodings that have already been enriched by local groove-aware features and multi-scale attention. A plausible implication is that the module is designed to exploit reverse-complement symmetry after the encoder has already extracted both local structural motifs and long-context dependencies.

4. Empirical behavior and architectural interpretation

The clearest isolated empirical evidence for GRC in TrinityDNA is the pretraining perplexity ablation. Without GRC, the reported PPL is $s_i^C$6; with GRC, the reported PPL is $s_i^C$7, a reduction of $s_i^C$8 perplexity points (Yang et al., 25 Jul 2025).

Setting	PPL	Difference
without GRC	2.731	baseline
with GRC	2.599	-0.132

The same ablation table reports that GFM further improves PPL from $s_i^C$9 to $A\leftrightarrow T$0, while SMWA slightly changes PPL to $A\leftrightarrow T$1 in that table. TrinityDNA therefore presents GRC as a meaningful contributor to the language-modeling objective, with a larger marginal effect than some other components in that specific ablation (Yang et al., 25 Jul 2025).

For downstream tasks, the paper does not provide a standalone GRC-only benchmark table, but it explicitly attributes the model’s strong performance on tasks requiring recognition of extended promoter regions or higher-order structural features to the architecture’s multi-window attention and GRC-based reverse-complement awareness. It also argues that GRC helps the model generalize better across genomic contexts, avoid strand-specific brittleness, and better capture complementary patterns relevant to genomic annotation tasks (Yang et al., 25 Jul 2025).

The interpretability claim is similarly specific. Because GRC explicitly models both orientations, the paper argues that biologically meaningful motifs can be represented consistently whether they are seen on the forward strand or as a reverse complement. This is not described as exact invariance in the formal equivariant sense; rather, the gate allows the model to decide how much to trust each orientation’s representation. That characterization is consistent with the paper’s claim that GRC is not simply symmetry-enforcing, but symmetry-aware and adaptive (Yang et al., 25 Jul 2025).

5. Relation to prior reverse-complement-aware methods

TrinityDNA contrasts GRC with prior reverse-complement handling strategies in DNA models. The paper notes that Caduceus/MambaDNA incorporates reverse-complement symmetry by using BiMamba and explicitly enforcing RC equivariance by taking a sequence and its reverse complement as inputs to the same SSM module. It also notes that earlier DNA models often relied on reverse-complement data augmentation, reverse-complement parameter sharing, or hard equivariant architectures (Yang et al., 25 Jul 2025).

Against that background, GRC is characterized by three features. First, it uses a Transformer encoder with SMWA rather than an SSM-only stack. Second, it combines forward and reverse-complement branches with a learned gate rather than only enforcing hard symmetry or averaging. Third, it is integrated with Groove Fusion, so reverse-complement processing occurs in a model already enriched with groove-aware local features (Yang et al., 25 Jul 2025).

These distinctions delimit the scope of the method. GRC is best viewed as a gated, shared-parameter, bidirectional reverse-complement fusion layer. It is not equivalent to augmentation-only handling, and it is not identical to strict reverse-complement-equivariant design. The paper’s formulation instead occupies an intermediate position in which both orientations are represented explicitly, but their fusion remains learnable.

A separate terminological ambiguity arises in "Cached Transformers: Improving Transformers with Differentiable Memory Cache," where GRC denotes Gated Recurrent Cache rather than Gated Reverse Complement. There, GRC refers to a recurrently updated differentiable memory cache used to extend attention over long contexts in language and vision tasks (Zhang et al., 2023). The acronym collision is purely nominal; the mechanisms address different problems.

6. Reverse-complement constraints in coding theory and DNA sequence design

Outside TrinityDNA, reverse-complement symmetry also appears as a constraint in DNA coding theory. In the duplication-correcting setting, reverse-complement duplication means that a duplicated factor is not copied verbatim, but instead inserted after being reversed and complemented. Over an alphabet $A\leftrightarrow T$2 with an involutive complement operation $A\leftrightarrow T$3, a reverse-complement duplication of length $A\leftrightarrow T$4 at position $A\leftrightarrow T$5 is formalized by

$A\leftrightarrow T$6

with the note that the actual inserted block is the reverse-complement copy of the duplicated $A\leftrightarrow T$7-factor. For this channel, the asymptotic coding capacity for correcting any number of reverse-complement duplications is zero for all $A\leftrightarrow T$8 and even $A\leftrightarrow T$9:

$C\leftrightarrow G$0

The exact bound stated is

$C\leftrightarrow G$1

Only the length-$C\leftrightarrow G$2 case admits nontrivial behavior: for $C\leftrightarrow G$3, $C\leftrightarrow G$4 and $C\leftrightarrow G$5; for even $C\leftrightarrow G$6,

$C\leftrightarrow G$7

with an optimal code of size

$C\leftrightarrow G$8

The paper explicitly describes this reverse-complement setting as the one relevant to a GRC-style constraint (Yohananov et al., 2023).

Subsequent work studies explicit correction schemes for reverse-complement duplication. For arbitrary many disjoint reverse-complement duplications of length

$C\leftrightarrow G$9

there exists a code with one redundant symbol, with encoding and decoding in $S$0 average time. For arbitrary-length reverse-complement duplication correction, a Gilbert–Varshamov-type argument yields redundancy upper bounded by

$S$1

For $S$2, two explicit constructions correct $S$3 length-one reverse-complement duplications: the first has redundancy

$S$4

with encoding complexity $S$5 and decoding complexity $S$6; the second improves redundancy to

$S$7

with encoding and decoding complexities $S$8 (Sun et al., 1 Feb 2026).

Reverse-complement constraints also structure algebraic DNA code design. Reversible $S$9-codes constructed from group matrix rings can be specialized to $S^R$0 DNA codes satisfying Hamming distance, reverse, reverse-complement, and fixed GC-content constraints; one explicit example gives a length-$S^R$1 code with $S^R$2 codewords and Hamming distance $S^R$3, together with a subcode of $S^R$4 codewords satisfying reverse-complement and fixed GC-content $S^R$5 (Korban et al., 2021). Related work over $S^R$6 characterizes reversible-complementary codes module-theoretically as reverse-invariant submodules containing the repetition code, while double cyclic codes over $S^R$7 satisfy reverse-complement closure exactly when they are reversible and contain the all-ones word (García-Claro, 23 Jun 2025, Acharya et al., 15 Dec 2025). In a different constructive direction, kernel-code-based DNA storage imposes the reverse-complement constraint

$S^R$8

and introduces the quantity

$S^R$9

as a reverse complement distance (G et al., 2023).

Taken together, these results show that reverse-complement awareness has two distinct but connected meanings. In neural sequence modeling, Gated Reverse Complement is an architectural mechanism for representation learning under strand symmetry. In DNA coding theory, reverse-complement constraints define admissible codebooks, duplication channels, and correction guarantees. The common substrate is the same duplex symmetry of DNA, but the operational goals differ: representation consistency in one case, and combinatorial or algebraic control of admissible sequences in the other.