Explaining Genomic Models in Variant Space, With Declared Baselines

When a genomic deep learning model predicts that a person's DNA near some gene raises its expression, the natural question is which of that person's genetic differences drove the prediction. Today's standard answers mutate every DNA letter in a window one at a time, or read the model's attention maps. Both methods ignore two facts that genomic storage systems already exploit. First, a genome is almost entirely identical to a shared reference, so the meaningful features are the few thousand differences ("variants") in a window, not the millions of letters. Second, nearby variants are inherited together in blocks. This co-inheritance, called linkage disequilibrium (LD), means that flipping one variant alone usually produces a genome no human carries. We propose to compute explanations directly over the variant differences that reference-based compression stores. Perturbations are drawn from real haplotypes (the stretches of DNA a person inherits together from one parent) in a population panel. Every choice that defines these features is recorded as part of the explanation, because each one silently changes what the explanation means: the reference build, the annotation release, the variant filters, and the panel ancestry.

The payoff is threefold. The cost of explaining one prediction scales with the number of variants in a window rather than its length. The perturbed genomes stay on the population's data distribution. And we can measure something no current genomic benchmark measures: how much an attribution changes when an undocumented preprocessing choice changes. That is the genomic version of the reproducibility collapse that Fell et al. documented in pathology, where re-implementing published methods from text alone dropped one algorithm's slide accuracy from 92% to 54.7%.

Core Claim

Model-agnostic explanation can be described as four steps: sample instances, intervene on features, query the model, and aggregate the outputs. This is the sampling–intervention–prediction–aggregation (SIPA) decomposition of Scholbeck et al.. For genomic models, the right instantiation is fixed by the data representation, not by the explainer. The features are the reference-relative variants that codecs such as LW-FQZip and tokenizers such as BioToken already isolate. The intervention distribution should be conditional on a haplotype panel. The obvious alternatives are marginal single-letter mutation, or "revert to reference", which quietly treats one specific reference genome as the neutral baseline. The resulting attributions are only reproducible if the intervention distribution and the definition of the features travel with them. We claim three things, each testable. At equal query budget, panel-conditional attributions over variants recover planted causal variants better than letter-level mutagenesis and than attribution over whole LD blocks. "Revert to reference" produces explanation quality that differs across ancestry groups, and panel conditioning removes much of that gap. Attribution rankings are at least as sensitive to plausible undocumented preprocessing changes as to the choice of explainer.

Where This Sits

Three strands of prior work motivate the gap without closing it.

The closest work is OmniGenBench, which standardizes evaluation of 31 genomic foundation models through declarative configuration files. Its interpretability module offers motif analysis and attention heatmaps but no quantitative faithfulness or stability scores. OpenXAI supplies exactly those scores (22 faithfulness, stability and fairness metrics, plus a synthetic generator with known ground-truth attributions), but it is limited to tabular data with independent features. We borrow OpenXAI's metric families and its planted-ground-truth design, and we replace its independent Gaussian features with real LD structure.

Among genomic models, BioToken and BioFM already feed variants and annotations to the model as tokens, reaching overall AUROC 0.780 on a five-part variant benchmark. However, it explains the model only through attention heads and UMAP projections (2-D visualizations of embeddings). Its authors also warn that annotation tokens may lend spurious importance to variants near transcript boundaries. Benchmarking DNA foundation models shows that zero-shot attention in NT-v2 misses chromatin domain boundaries. That result constrains any proposal that relies on attention as explanation.

On the theory side, Barbiero et al. argue that attributions over raw high-dimensional coordinates break "information invariance", the requirement to discard irrelevant variation before explaining. Reference-relative variant space is a concrete compression that satisfies this requirement for genomes. The Qi-framework writes an explanation as a mapping from sources to a target component and a user utility. We use it only as the bookkeeping for our provenance record: source, method, component explained, and intended use, extended with the preprocessing that defines the source.

On the scaling side, the coordinate-sorted, chunk-indexed designs of GTRAC and GVC supply the method we reuse to fetch haplotype blocks from a large panel by genomic window without decompressing whole files. MaSk-LMM shows how much compression buys in cohort statistics (88 minutes versus 15 hours for Regenie on 500,000 simulated samples). We do not explain sketched models; sketches mix samples and markers and so lack per-variant coordinates. We note this as a boundary of the proposal.

Research Program

The first component is the explainer. For an individual $x$ and a genomic window $W$, let $V(x,W)$ be the set of $x$'s variants relative to a declared reference. A coalition $S \subseteq V$ of kept variants is completed into a full genome in one of two ways. In the marginal completion, variants outside $S$ are reset to the reference allele. In the conditional completion, the complement is drawn from panel haplotypes that match $x$ on $S$ within the local LD block. The panel is 1000 Genomes; matching uses a Li–Stephens-style copying model, or exact haplotype matching when the panel is dense enough. Shapley values are then estimated over $V$ with a fixed number of model queries. For nucleotide-input models (HyenaDNA, Caduceus-Ph, NT-v2), completed genomes are rendered as sequence. For BioFM, the variant tokens are edited directly. The panel store is chunk-indexed by coordinate, so each explanation touches only the blocks overlapping $W$.

The second component is the benchmark, a genomic analogue of OpenXAI's synthetic generator. We plant labels on real 1000 Genomes haplotypes using known causal variants: additive, pairwise interacting, and ancestry-varying effects. We then provide two kinds of black box. The first is oracle models, whose output is the generating function itself, so attribution to the model is known exactly. The second is trained models, meaning BioFM and a nucleotide model fine-tuned on the planted labels, where the planted causal set is the data-level ground truth. Keeping both is deliberate. Explanations that are true to the model and explanations that are true to the data can legitimately disagree when a trained model relies on a correlated tag variant. Marginal and conditional interventions answer these two different questions. The benchmark reports both rather than hiding the choice.

The third component is the provenance audit. Every attribution is stored with a record of the reference build, annotation release, variant quality filter, window definition, panel and its ancestry composition, completion rule, and query budget. The audit recomputes attributions under plausible alternative settings of each field and reports how much the top-$k$ variants and the rank correlation change. This parallels Fell et al.'s component-by-component decomposition, but here it is executable and not a checklist.

Baselines

The naive reference point is letter-level in-silico mutagenesis: substitute each position in the window and record the change in prediction. It is the working default in genomics and costs roughly three queries per base. We run it on the same windows and models, truncated to the same query budget as our method, and also at full cost as an upper-resource variant. The strongest attention-based competitor is BioFM's attention on variant tokens, as analyzed in BioToken and BioFM. It is scored with the same recovery and faithfulness metrics as every other method.

The strongest existing post-hoc competitors come from the OpenXAI families: KernelSHAP, Integrated Gradients and SmoothGrad applied to one-hot sequence and summed within each variant's span. SmoothGrad led OpenXAI's predictive faithfulness comparison by an average of +43.03% on some metrics, and LIME led signed agreement by +61.6% (OpenXAI). No reported result was found on Lacuna for the faithfulness of any explainer on genomic foundation models. The bar is therefore set inside this study: the strongest of these methods, tuned with the same hyperparameter search budget.

The most dangerous simple competitor is Shapley attribution over LD blocks, where correlated variants are clumped into groups and each group is attributed as a unit. This method has no citation in the file, so we implement it ourselves.

The ablations isolate our key ingredients. The first varies the feature set: variants versus letters, with the same Shapley estimator and budget. The second varies the completion rule: revert to reference versus panel-conditional, with the same features. The third removes provenance by fixing every preprocessing field, which lets us read how much variance the audit exposes.

Beating the baselines means three things at a matched query budget: a higher area under the precision–recall curve (AUPRC) for recovering planted causal variants at single-variant resolution, a smaller faithfulness gap across the five 1000 Genomes super-populations, and a larger prediction drop when the top-$k$ attributed variants are removed on real tasks. All three must hold with confidence intervals from resampling individuals.

Why not just…?

Why not just run in-silico mutagenesis on the decompressed sequence? In-silico mutagenesis is a marginal intervention, and its costs grow with window length. Most of its queries go to invariant positions, and each substitution creates a genome that breaks local LD, which is the off-distribution sampling that Barbiero et al. identify as a failure of raw-coordinate attribution. It is still the true-to-model answer for oracle models, so we expect it to perform well there. The benchmark reports that result honestly. Our claim concerns trained models and matched budgets, and the feature-set ablation tests it directly.

Why not just attribute to whole LD blocks? This is the real threat. Grouping may match our method on block-level recovery at lower cost. Our advantage has to appear within blocks, in distinguishing the causal variant from its correlated tags. That is only possible because conditional completion varies each variant against panel haplotypes that happen to break the correlation. If block-level Shapley matches us at single-variant resolution on the planted benchmark, we will report it. The contribution then narrows to the benchmark and the provenance audit.

Why not just read attention maps? Attention has already failed a structural test: NT-v2's zero-shot attention misses chromatin domain boundaries (Benchmarking DNA foundation models). BioFM's own analysis finds heads locked onto periodic repeats. We still include attention as a baseline, because BioFM's variant tokens give it the best chance any attention method will get.

Why not just ask authors to fill in a reporting checklist? A checklist records that a choice was made but does not show whether it matters. Fell et al.'s failures arose from under-specified preprocessing, not unreported preprocessing. The audit quantifies how far attributions move per field. That tells a checklist which fields are load-bearing.

Decisive Experiments

The first decisive test runs on the planted-label benchmark built on 1000 Genomes haplotypes. We compare variant-space conditional Shapley against the baselines at budgets of 500, 2,000 and 10,000 queries per explanation, scoring single-variant AUPRC against the causal set and exact attribution error for oracle models. The central claim fails if the conditional method does not beat both letter-level mutagenesis and block-level Shapley on AUPRC for trained models at single-variant resolution.

The second test repeats this by super-population with the reference-reversion and panel-conditional completions, using an OpenXAI-style fairness gap. If panel conditioning does not shrink the gap, the ancestry-bias argument against "revert to reference" fails.

The third test moves to real tasks: the AlphaMissense, BEND, Borzoi, DeepSea and GRASP components of the BioFM variant suite, where BioFM reports overall AUROC 0.780 and NT-500M-v2-MS reports 0.727 (BioToken and BioFM), together with the GTEx expression tasks from the 57-dataset benchmark. There is no ground truth here, so we measure deletion faithfulness, stability under resampling the panel, and wall-clock cost with and without the chunk-indexed panel store. We specifically check whether variants near transcript boundaries receive inflated attribution under BioFM, using distance-matched controls, as its authors warn they might.

Finally, the provenance audit reports, for every model and task, the rank correlation of attributions across alternative preprocessing settings next to the rank correlation across explainers. The transparency claim holds if preprocessing variance is comparable to or larger than explainer variance on a substantial share of tasks. It fails if preprocessing variance is small everywhere, in which case a lightweight record suffices and the audit is unnecessary.