Reverse-Complement Consistency Regularization

• RCCR is a regularization technique that enforces consistent outputs for a DNA sequence and its reverse complement.
• It uses an alignment operator and divergence metrics to reconcile predictions, leveraging the inherent symmetry of double-stranded DNA.
• Empirical results demonstrate that RCCR enhances predictive accuracy and orientation robustness across diverse genomics tasks with minimal architectural changes.

Reverse-Complement Consistency Regularization (RCCR) is a model-agnostic fine-tuning objective for DNA LLMs that enforces prediction consistency between a DNA sequence and its reverse complement. The method directly penalizes the divergence between a model’s prediction on a sequence and the aligned output on its reverse complement, thereby embedding double-stranded DNA’s fundamental biological symmetry into deep models. This approach yields models that are more robust to sequence orientation, with minimal changes to architecture or inference procedures, and is applicable to diverse sequence-based genomics tasks across varied model backbones (Ma, 23 Sep 2025).

1. Formalization of RCCR

The RCCR objective is defined for a DNA sequence $x$ of length $L$ and its reverse complement $\tau(x) = \mathrm{RC}(x)$. Given a model $f_\theta(x) = g_\theta \circ h_\theta(x)$, its task output on $x$ is $\hat{y} = f_\theta(x)$. The reverse complement output is realigned to the original sequence coordinates by an alignment operator $\Pi$: $\tilde{y} = \Pi f_\theta(\tau(x))$. Outputs are mapped into a common comparison space via a link function $\phi(\cdot)$ (e.g., softmax for classification, identity for regression). The divergence or distance $D(\cdot,\cdot)$ (e.g., symmetric KL for classification, squared error or Poisson-KL for regression) quantifies inconsistency.

The RCCR-augmented fine-tuning objective is: $L$0 where $L$1 is the task-specific loss, $L$2 is the regularization weight, and the expectation averages over the training set. $L$3 reduces to standard fine-tuning, while $L$4 enforces strict reverse-complement prediction invariance (Ma, 23 Sep 2025).

2. Alignment Operator Implementation

The alignment operator $L$5 adapts to the prediction head structure:

• For sequence-level heads (classification/regression), $L$6 is the identity.
• For bin-wise or profile outputs $L$7 (with $L$8 bins and $L$9 channels), the aligned output is given by:

$\tau(x) = \mathrm{RC}(x)$0

with $\tau(x) = \mathrm{RC}(x)$1 reversing bins and $\tau(x) = \mathrm{RC}(x)$2 permuting strand channels. If certain channels are known to be strand-specific, a binary mask $\tau(x) = \mathrm{RC}(x)$3 can suppress these before computing $\tau(x) = \mathrm{RC}(x)$4.

In practice, one computes forward passes for both $\tau(x) = \mathrm{RC}(x)$5 and $\tau(x) = \mathrm{RC}(x)$6 for each batch, then aligns and compares outputs via $\tau(x) = \mathrm{RC}(x)$7 and $\tau(x) = \mathrm{RC}(x)$8 (Ma, 23 Sep 2025).

3. Integration into Training and Inference

RCCR is integrated via a two-pass fine-tuning loop. For each mini-batch, compute standard task loss on $\tau(x) = \mathrm{RC}(x)$9 and RC consistency loss on $f_\theta(x) = g_\theta \circ h_\theta(x)$0 vs. RC$f_\theta(x) = g_\theta \circ h_\theta(x)$1:

$x$3 Ablation on $f_\theta(x) = g_\theta \circ h_\theta(x)$2 is advised, monitoring task accuracy and RC consistency on validation data. Inference requires only a single forward pass on $f_\theta(x) = g_\theta \circ h_\theta(x)$3, as the fine-tuned parameters yield RC-consistent predictions without test-time augmentation or ensembling (Ma, 23 Sep 2025).

4. Experimental Protocol

RCCR was evaluated across three pre-trained DNA LM backbones with frozen encoders and fine-tuned heads:

• NT-v2 (Nucleotide Transformer v2): 50M param., 6-mer tokens
• DNABERT-2: 100M param., BPE encoding
• HyenaDNA-Medium-160k: single-nucleotide, 160 kb receptive field

Baselines included vanilla (no RC), RC-aug (training with random RC input per batch), and test-time averaging (TTA) across x/RC(x). Tasks:

• Sequence classification: 18 datasets (promoters, enhancers, splice sites, histone modifications)
• Bulk RNA regression: 4096 bp input, 218 outputs, MSE/Huber
• Binned profile regression: CAGE profiles, 4096 bp → 128 bins, log(1+count) stabilized

Negative control: DNA strand classification (intrinsically strand-specific) (Ma, 23 Sep 2025).

Training details:

• Optimizer: AdamW (lr: $f_\theta(x) = g_\theta \circ h_\theta(x)$4 for classification/bulk RNA, $f_\theta(x) = g_\theta \circ h_\theta(x)$5 for CAGE)
• Batch size: 256 (classification), 32 (regression/profiles)
• Epochs: 3–4
• $f_\theta(x) = g_\theta \circ h_\theta(x)$6: tuned per task (0.1–0.5)
• Hardware: single NVIDIA H100, bf16 precision

5. Quantitative Outcomes

RCCR consistently improved both RC robustness and predictive accuracy across models and tasks:

Task	Metric	RCCR	RC-Aug
Histone (CLS)	AUPRC	0.812	0.784
Splice (CLS)	AUPRC	0.994	0.992
Enhancer (CLS)	AUPRC	0.680	0.655
Promoter (CLS)	AUPRC	0.948	0.937
Bulk RNA (REG)	RMSE	0.680	0.709
	R²	0.488	0.444
	Spearman	0.756	0.733
CAGE Profile (REG)	RMSE	0.245	0.262
	Spearman	0.250	0.190

RCCR reduced the Symmetry Flip Rate (SFR) from ~0.15 to ~0.10 and improved RC_Corr from ~0.92 to ~0.93–0.98 for classification. For regression/profiles, both overall accuracy and RC consistency matched or exceeded RC-Aug and TTA. Lambda ablations indicated that low $f_\theta(x) = g_\theta \circ h_\theta(x)$7 ($f_\theta(x) = g_\theta \circ h_\theta(x)$80.1) suffices for standard classification, while higher values ($f_\theta(x) = g_\theta \circ h_\theta(x)$90.5) can further boost synergetic gains, particularly for complex profile prediction (Ma, 23 Sep 2025).

6. Computational Considerations

RCCR incurs a moderate increase in training cost (approximately 2× forward passes per mini-batch; empirical training time increase ≈1.7×), but there is no additional inference cost compared to test-time averaging, which always requires two passes. This yields a single intrinsically RC-consistent model and avoids specialized architectures or ensembling. The marginal cost for divergence computation is negligible (Ma, 23 Sep 2025).

7. Recommended Usage and Constraints

For effective RCCR application:

• Verify that target labels are RC-invariant (classification/regression) or RC-equivariant (profiles); avoid RCCR in strand-specific tasks or mask channels accordingly.
• Select $x$0 via grid search (0.1–0.3 for most tasks; up to 0.5 for synergy regimes). Smaller $x$1 is preferable when baseline accuracy is already high; higher $x$2 may benefit complex regression profiles.
• For classification, symmetric KL divergence provides stable gradients and Jensen–Shannon control; for regression, use MSE, Huber, or Poisson deviance metrics.
• Always report both standard metrics (AUPRC, MCC, R², Spearman, RMSE) and RC robustness metrics (SFR, RC_Corr).
• Integrate RCCR as a fine-tuning regularizer without modifying model backbones or heads.
• Include strand-specific negative controls to confirm appropriateness of the RC prior (Ma, 23 Sep 2025).

By encoding the structural symmetry of double-stranded DNA directly into the model fine-tuning process, RCCR substantially improves orientation robustness and can enhance predictive accuracy without increased inference cost.