Symmetry as a Per-Task Dial: One Reverse-Complement-Aware DNA Network Steered by Feature-Wise Modulation

DNA is double-stranded, so a stretch of genome can be read in either direction. The reverse complement of a sequence is the same physical molecule read from the other strand. Networks that hard-wire this fact are called reverse-complement (RC) equivariant: flipping the input flips the internal features in a predictable way. These networks learn more from fewer examples than ordinary networks. The cost is that each one is built and trained for a single assay, with its own output head. Generalist genomic models go the other way. They solve many tasks with one network but discard the strand symmetry. We propose one small convolutional DNA network with a single output format that serves every task. The task identity steers the network through feature-wise linear modulation (FiLM): a small side network predicts a per-channel scale and shift that rescale the main network's activations. We derive the constraints these scales and shifts must obey so that the strand symmetry survives.

The central new ingredient is that the symmetry should not be a fixed property of the architecture. Some genomic labels really are strand-agnostic: a transcription factor binds a site regardless of which strand you read. Others are not: a splice donor read backwards is not a splice donor. A generalist network therefore needs the symmetry to be adjustable per task. We implement this as a learned gate for each task that mixes a symmetric pathway with a strand-selective pathway. A second content-aware gate chooses between a nucleotide-resolution stream and a coarse context stream. We expect the learned settings to recover which assays are strand-specific, without being told. We also expect one network to keep the low-data advantage of equivariance that single-task models currently enjoy.

Core Claim

A single RC-aware backbone, conditioned on task identity through symmetry-constrained FiLM and two gates, can match per-task equivariant specialists on transcription factor binding and base-resolution profile prediction. This should hold most clearly in the 1,000-sequence low-data regime. The same backbone should also beat unconstrained FiLM with RC data augmentation on those tasks. Separately, on orientation-dependent tasks it should beat a fully symmetric generalist.

The falsifiable prediction concerns the per-task symmetry gate. Its learned value should be near zero for binding tasks and large for orientation-dependent ones. If it is uniformly near zero, or if forcing exact symmetry everywhere costs nothing, the dial is unnecessary and the claim fails.

Where This Sits

The proposal draws on the four directions in different roles.

The prior. Reverse-Complement Equivariant Networks for DNA Sequences characterizes every linear RC-equivariant layer. It distinguishes two kinds of channels. "Regular" channels come in pairs that swap under the flip. "Irreducible" (irrep) channels are either even (unchanged by the flip) or odd (sign-flipped by it). The paper shows that odd channels need odd nonlinearities, and that equivariant networks beat augmentation and post-hoc averaging, most of all with 1,000 training sequences.

Two further papers supply the evidence that the prior pays off. Self-GenomeNet builds RC awareness into self-supervised pretraining and matches supervised baselines with roughly ten times fewer labels. Both papers, however, use one model or one linear head per task. We keep their constraint and remove that fragmentation.

The regime. Omni-DNA unifies genomic tasks by phrasing each one as an instruction followed by a token response. This is our contrasting design. It has no strand symmetry, and continuous base-resolution profiles would have to be broken into discrete tokens. The companion version, Omni-DNA, adds a warning we take seriously: lower pretraining loss did not track downstream biological accuracy. We therefore evaluate on task metrics only.

The steering mechanism. MMAML shows that FiLM driven by a task embedding lets one network serve task families that do not overlap. It reduced 5-mode regression error from 1.668 to 0.868 relative to unmodulated MAML. Its modulation is unconstrained, which is exactly what breaks strand symmetry.

The gates. Attentional Feature Fusion shows that content-aware gates beat static addition when merging streams. An iAFF-ResNet-50 reached 20.2% top-1 error, against 23.2% for a ResNet-101 that merges by plain addition. Gated Bridging uses gates to stop one task's noise from leaking into another. Neither work faces a symmetry constraint. We borrow their gate form and restrict it.

Research Program

Backbone. The backbone is a dilated convolutional stack in the style of GPN, made equivariant with the mixed regular and irrep layers of Mallet et al. It reads nucleotides one at a time. Mallet et al. found that k-mer tokens, which group several bases into one input symbol, blur base-pair resolution, so we avoid them.

Constrained FiLM. Every block is modulated by $\gamma_t \odot h + \beta_t$. Here $h$ is the block's activation, $t$ is the task, and $\gamma_t, \beta_t$ are produced from a learned task embedding. Because $t$ does not change when the sequence is flipped, equivariance reduces to tying constraints. Let $P$ be the permutation that swaps paired regular channels. Equivariance then requires

$$\gamma_t = P\gamma_t, \qquad \beta_t = P\beta_t \ \text{(regular channels)}, \qquad \beta_t^{\text{odd}} = 0 \ \text{(odd irrep channels)}.$$

The odd-channel condition matters for design. A shift on an odd channel changes how the following nonlinearity responds to that channel, and that is impossible without breaking the symmetry. FiLM can therefore rescale odd channels but cannot shift their baseline. We will measure how much task control is lost as a result.

Resolution gate. A gate merges the nucleotide-resolution stream $X$ with a pooled context stream $Y$:

$$\alpha \odot X + (1-\alpha)\odot Y.$$

For this merge to be equivariant, the gate $\alpha$ must itself flip consistently with the input. We compute it only from features that are unchanged by the flip: the even channels and the squares of the odd channels. Under the flip, $\alpha$ then simply reverses in position. The gate also receives FiLM conditioning from the task. The expected behavior is that profile tasks favor $X$ and classification tasks favor $Y$. This addresses the resolution-versus-context trade-off that Mallet et al. observed.

Single output interface. Every task reads out the same per-position, two-strand output track. The readout is

$$o_t = (1-\lambda_t)\,(o^{+} + o^{-}) + \lambda_t\, o^{+}.$$

Here $o^{+}$ and $o^{-}$ are the paired strand outputs, and $\lambda_t \in [0,1]$ is the learned symmetry-breaking gate for task $t$.

  • When $\lambda_t = 0$, the readout sums the two strands and is exactly strand-invariant.
  • When $\lambda_t = 1$, it keeps only the forward strand and is fully strand-selective.

Binary tasks take a log-sum-exp pool of the track, a smooth version of the maximum over positions. Profile tasks use the track directly. No task gets its own head and nothing is converted to tokens.

Where symmetry breaks. A stronger variant allows a $\lambda$-scaled shift on odd channels at intermediate depths. This tests whether breaking the symmetry only at the readout is enough or whether it must happen earlier.

Baselines

Naive reference. The do-nothing reference is a standard per-task CNN trained with RC augmentation. On Genomic Benchmarks, Grešová et al. report such CNNs at:

  • 93.3% accuracy on human_ensembl_regulatory,
  • 94.2% on demo_human_or_worm,
  • 58.6% on drosophila_enhancers_stark.

We will rerun these models in our own pipeline rather than rely only on the quoted figures.

Strongest specialist. The strongest competitor on our central question is the "best equivariant" per-task network of Mallet et al., with one model per ChIP-seq or ChIP-nexus target. It is the strongest because it already contains the prior whose advantage we hope to keep, and it pays no cost for sharing. The file states that its per-target numbers appear only as plotted curves, so there is no reported result on Lacuna for per-target Mallet et al. scores. We will retrain it ourselves.

Beating it means one generalist network reaches its score within seed noise at 1,000 sequences. It must do so with a total parameter count no larger than the sum of the specialists' parameters, and preferably with far fewer.

Large generalist reference. For broad context we cite Omni-DNA: it reports a 0.767 average on the Nucleotide Transformer benchmark, against 0.709 for NT-2.5B. Our models are orders of magnitude smaller, so this is a reference point rather than a target.

Ablations. Each ablation isolates one ingredient:

  • a symmetric multi-task trunk with separate per-task heads, which is the fragmented design;
  • unconstrained FiLM with RC augmentation;
  • the task embedding concatenated to the input instead of FiLM;
  • static addition in place of the resolution gate;
  • $\lambda_t$ fixed at 0 for all tasks, giving full symmetry;
  • $\lambda_t$ fixed at 1 for all tasks, giving full breaking.

Fair comparison. All models are trained from scratch on identical splits. Parameter counts are matched within 5%. Each model gets the same hyperparameter-search budget of 20 trials and is run with 5 seeds.

Why not just…?

Why not just use an ordinary network and average its forward and reverse-complement predictions? Averaging makes the outputs exactly symmetric at no design cost, so it is a serious alternative. Mallet et al. report that equivariant networks beat post-hoc averaging, and that the gap widens at 1,000 sequences. Averaging also hard-codes strand invariance, so by construction it cannot serve the strand-specific tasks. We include averaging as a baseline on both kinds of task.

Why not just share an equivariant trunk and attach a linear head per task? This is the most dangerous alternative. If it matches our model, then FiLM and the gates are unnecessary, and we will report that outcome. Our expectation that it falls short rests on two pieces of evidence. MMAML's gains came from modulating internal layers rather than only the output layer. Gated Bridging found that indiscriminate sharing transfers noise from one task into another. The per-task-head ablation, run on the mixed binding, profile, and orientation task set, decides the question.

Why not just add the task as a prompt token in an Omni-DNA-style model? That design abandons the symmetry, and it would have to spell out base-resolution profiles as quantized token sequences. Omni-DNA's own finding that pretraining loss tracks biology poorly also argues against relying on next-token loss. The concatenated-embedding ablation tests the weaker form of this alternative directly.

Why not just add the two streams instead of gating them? On ImageNet, AFF's gated fusion beat plain addition. Still, DNA streams differ from image streams, so we test static addition on profile tasks rather than assume the result transfers.

Proposed Experiments

Data. The primary testbed is Mallet et al.'s data:

  • ChIP-seq binding classification for Max, Ctcf, and Spi1, scored by area under the ROC curve (AUROC);
  • ChIP-nexus base-resolution profiles for Oct4, Sox2, Nanog, and Klf4, scored with their profile metrics.

Both are trained jointly at full size and at 1,000 sequences per task. We then add the 9 Genomic Benchmarks datasets and the 18 Nucleotide Transformer tasks. The latter include splice-site and promoter tasks, which are presented in a fixed orientation.

Decisive checks. Three results determine whether the claim holds.

  1. Low-data parity. At 1,000 sequences the generalist must match the per-task equivariant specialists, and it must beat unconstrained FiLM with augmentation on binding and profile tasks.
  2. Symmetry dial. On orientation-dependent tasks, learned $\lambda_t$ must exceed fixed $\lambda_t = 0$. We will confirm orientation dependence independently by measuring how much a plain CNN's predictions change when its input is flipped. The learned $\lambda_t$ values should rank-correlate with that measured sensitivity.
  3. Measured equivariance. On tasks with learned $\lambda_t \approx 0$, the gap between the model's prediction on a sequence and its flipped prediction on the reverse complement should be at floating-point level. A model that only approximates the symmetry would show a measurable gap.

The odd-channel restriction on shifts is quantified by comparing against a non-equivariant FiLM model that may shift every channel. Long-range 450 kbp tasks are left out of scope.