Most genetic variants linked to disease sit outside genes, in stretches of DNA called enhancers that act as dimmer switches for nearby or distant genes. Many genes are controlled by several enhancers at once. When one enhancer is broken, the others can sometimes cover for it, so a mutation that clearly damages an enhancer may leave the gene's output unchanged. Current DNA sequence models are good at saying whether a variant damages its local enhancer. They are much worse at saying whether that damage reaches the gene. Neuroscience has a well-tested way to quantify this kind of covering: ask how many disjoint subsets of units can each reconstruct the system's output on their own. We propose to apply that measurement to the enhancers that control each gene, using their activity across hundreds of cell types. The resulting redundancy score would predict which enhancer variants are absorbed by the rest of the ensemble and which get through to change expression.
The opportunity is concrete. On gene-centric evaluation, which asks a model to identify the affected gene rather than just flag the variant, model performance drops. A basic local network nearly matches large DNA language models when scoring variants at their own location (Gene-centric evaluation of causal variant prediction for DNA models). TraitGym reports that every model class struggles specifically with distal enhancer variants (Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics). If part of this failure comes from buffering by other enhancers that the models never account for, then a cheap, interpretable redundancy correction should recover some of the lost accuracy. If it does not, that negative result is informative too.
Core Claim
An enhancer's measured contribution to its gene says little about how much a variant in it will matter. What matters is whether the gene's activity pattern across cell types can be rebuilt without that enhancer.
We borrow the measurement from retinal recordings. There, 6 to 7 disjoint subsets of ganglion cells each reconstructed a motion trajectory with over 85% of the whole population's accuracy, so the code was about 6-fold redundant (High Accuracy Decoding of Dynamical Motion from a Large Retinal Population). The theory behind why redundancy protects the output comes from work on overcomplete codes, meaning codes with more units than the signal needs. That work shows the extra units create directions of activity the downstream reader ignores, so a unit can drift or fail without changing the decoded result. Neighbors with overlapping tuning can also substitute for a failed unit (Over-complete representations on recurrent neural networks can support persistent percepts).
To carry this over to genes, let $y_g \in \mathbb{R}^C$ be gene $g$'s expression across $C$ cell types. Let $A_g \in \mathbb{R}^{C \times N}$ hold the activities of its $N$ candidate enhancers in those cell types. The candidates and activities come from the ENCODE-rE2G encyclopedia of more than 13 million enhancer-gene links across 352 biosamples (An encyclopedia of enhancer-gene regulatory interactions in the human genome). We define a substitution index for enhancer $j$:
$$R_j = \frac{\mathrm{fit}(y_g \mid A_{g,-j})}{\mathrm{fit}(y_g \mid A_g)},$$
Here $\mathrm{fit}$ is the cross-validated variance explained by a sparse non-negative regression, and $A_{g,-j}$ is the activity matrix with enhancer $j$ removed and the model refit. When $R_j$ is near 1, the other enhancers can rebuild the gene's pattern without $j$. When $R_j$ is well below 1, enhancer $j$ carries part of the pattern that no other enhancer supplies.
We also compute a gene-level redundancy count. This is the number of disjoint enhancer subsets that each reach 85% of the full fit, the retinal statistic applied directly to enhancers. The predicted effect of a variant in enhancer $j$ on gene $g$ is then the sequence model's predicted local activity change $\Delta a_j$, weighted by how irreplaceable $j$ is:
$$\widehat{\Delta y}_{g} = w_j \, \Delta a_j \, (1 - R_j)^{\gamma}.$$
In this formula, $w_j$ is $j$'s weight in the full regression and $\gamma \ge 0$ is a single exponent fit on training chromosomes. Setting $\gamma = 0$ recovers ordinary additive scoring.
This is a statistical proxy, and we do not claim enhancers literally implement neural-style recurrent substitution. Cross-cell-type covariation is one observable trace of overlapping tuning. The proxy can fail: two enhancers can be active in the same cell types and still not compensate for each other in any one of them. The experiments below are built to detect that failure.
Relation to Prior Work
The motivation comes from three sources. ENCODE-rE2G documents promoter buffering and non-additive enhancer interactions. However, it scores each enhancer-gene pair separately, adding only heuristic terms for neighboring enhancer activity, and never asks whether the rest of an ensemble can reconstruct a gene's output without a given element. The gene-centric benchmark and TraitGym locate the failure of current models in distal and gene-linking settings.
The method comes from the retinal decoding and overcomplete-code papers above. We take their measurement, disjoint-subset decodability, and their explanation of why redundancy protects outputs. We do not take their recurrent dynamics.
Two neighboring theory papers constrain the design. A normative enhancer model shows that a single highly specific enhancer necessarily produces bursty, noisy transcription, and it explicitly leaves multi-enhancer ensembles unmodeled (Normative models of enhancer function). That gives a biophysical reason to expect redundancy to be selected for, though it is not something we test. The quantitative omnigenic model propagates small variant effects through a known yeast regulatory network (Quantitative omnigenic model discovers interpretable genome-wide associations). That work concerns effects acting between genes through regulatory networks. Our work stays within a single gene's own regulatory neighborhood and is complementary to it.
Research Program
We first build per-gene redundancy profiles for all protein-coding genes with at least two ENCODE-rE2G-linked enhancers. We fit sparse non-negative regressions of expression on enhancer activity across cell types, then compute $R_j$ and the disjoint-subset count. Because correlated noise caps the benefit of adding units in neural populations (Scaling of sensory information in large neural populations shows signatures of information-limiting correlations), we also measure how quickly fit saturates as enhancers are added. This separates true redundancy from many near-duplicate enhancers that add no new information.
Second, we score variants. For each variant we take $\Delta a_j$ from an existing sequence model and apply the correction. We use Enformer (Effective gene expression prediction from sequence by integrating long-range interactions), AlphaGenome (AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model), and the reporter-assay ensemble MPAC (Identifying non-coding variant effects at scale via machine learning models of cis-regulatory reporter assays). MPAC is attractive for this test because its reporter assay places each enhancer on a plasmid, separated by design from all other enhancers. Its predictions therefore isolate the local effect that our correction is meant to discount.
Third, we stratify. Every benchmark is split into strata by the redundancy count of the target gene. Our hypothesis predicts that baseline models lose more accuracy in high-redundancy strata, and that the correction recovers accuracy there and nowhere else.
Baselines
The naive reference point is nearest-gene assignment with the raw local variant score. It tells us how much any gene-linking structure helps at all.
The second baseline is additive, activity-by-contact-style scoring, which is our formula with $\gamma = 0$. It uses the same enhancer links and the same sequence-model $\Delta a_j$, so any gain from $\gamma > 0$ can only come from the redundancy term.
The strongest gene-linking competitor is ENCODE-rE2G itself, retrained on the same CRISPRi folds with its neighboring-enhancer features intact. For linking fine-mapped GWAS variants to target genes it reports a precision of 61%. We count it as beaten only if adding $R_j$ as a feature raises precision at matched recall on the 569-variant GWAS set, across held-out chromosomes.
For variant-to-gene ranking the strongest sequence model on the file's gene-centric benchmark is Enformer, at AUC 0.764 on GTEx SuSiE variants with PIP > 0.9 (Gene-centric evaluation of causal variant prediction for DNA models). We rerun Enformer and AlphaGenome scores ourselves under identical splits, and we count a gain only if it is consistent across chromosome folds, not a single number above 0.764.
On TraitGym, CADD reports AUPRC 0.85 on Mendelian traits (Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics). We do not expect to beat it there, because those variants are dominated by purifying selection. Our target is the complex-trait enhancer subset, where we found no per-subset reported number on Lacuna. For the CRISPRi enhancer-effect regression we found no reported correlation on Lacuna either, so we establish it ourselves for every baseline.
Three ablations isolate the key ingredient. The first replaces $R_j$ with the plain count of linked enhancers. The second replaces it with ENCODE-rE2G's neighbor-activity term. The third recomputes $R_j$ after shuffling cell-type labels within each enhancer, which destroys co-tuning while preserving activity distributions. All methods receive the same hyperparameter search budget, equal to the size of the ENCODE-rE2G configuration sweep scaled down per feature.
Why not just…?
Why not just count how many enhancers a gene has? This is the most serious alternative. Genes with many enhancers may simply be more buffered, regardless of which enhancers overlap in function. The enhancer-count ablation tests it directly. If the count does as well as $R_j$, the claim that cross-cell-type co-tuning matters is wrong, and we would report the simpler finding. Our reason to expect a difference comes from the retina, where redundancy depended on overlapping receptive fields rather than on cell count alone. The shuffled-label control separates these two explanations.
Why not just let ENCODE-rE2G's neighbor features absorb it? Neighbor activity measures how busy an enhancer's surroundings are in one cell type. It does not measure whether those neighbors reproduce the target gene's pattern across cell types. If adding $R_j$ to ENCODE-rE2G gives no gain on the CRISPRi folds, the neighbor features already capture redundancy, and we would drop the claim.
Why not just fine-tune AlphaGenome on CRISPRi effects? With a 1 Mb context, AlphaGenome can in principle learn buffering implicitly. However, the CRISPRi set contains about 10 thousand pairs from a single cell line, which is a thin signal for a model of that size. Fine-tuning also yields no explicit redundancy quantity that could be checked in other cell types. We will run this comparison anyway, using a fine-tuned head under the same folds. If it matches the correction, the remaining value of our approach is interpretability and transfer beyond K562, which we would test on GTEx tissues.
Decisive Checks
The first test uses the 10,411 CRISPRi enhancer-gene pairs in K562. Holding the enhancer's own activity and 3D contact fixed, does a higher $R_j$ predict smaller knockdown effects? A null partial correlation on held-out chromosomes falsifies the core claim.
The second test asks whether, on the gene-centric GTEx SuSiE benchmark, Enformer's and AlphaGenome's losses concentrate in high-redundancy strata, and whether the correction closes that gap while leaving low-redundancy strata unchanged. A uniform gain across strata would point to a generic recalibration effect rather than redundancy.
The third test checks whether the TraitGym complex-trait enhancer variants that every model misses are enriched in high-redundancy genes. If they are not, redundancy does not explain the distal-variant failure, and that failure should be sought elsewhere, for example in interactions between genes of the kind the omnigenic model addresses.
References
- Gene-centric evaluation of causal variant prediction for DNA models
- Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics
- High Accuracy Decoding of Dynamical Motion from a Large Retinal Population
- Over-complete representations on recurrent neural networks can support persistent percepts
- An encyclopedia of enhancer-gene regulatory interactions in the human genome
- Normative models of enhancer function
- Quantitative omnigenic model discovers interpretable genome-wide associations
- Scaling of sensory information in large neural populations shows signatures of information-limiting correlations
- Effective gene expression prediction from sequence by integrating long-range interactions
- AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model
- Identifying non-coding variant effects at scale via machine learning models of cis-regulatory reporter assays