Evolution as a Control for Faithfulness Metrics in Genomic Deep Learning

When a deep network predicts whether a transcription factor binds a stretch of DNA, researchers ask it "which bases mattered?" using feature attribution, meaning methods that score each input position by its influence on the output. They then judge those attribution maps with faithfulness metrics, which delete the "important" bases and check whether the prediction collapses. In vision and language, these metrics are known to disagree with each other, to reward artifacts, and to crown different winners depending on the metric. Genomics has barely been examined. It is also the one domain with a model-independent record of which positions matter biologically: evolutionary conservation. Bases that natural selection has kept unchanged across species are usually functional. Bases that drift freely usually are not.

This proposal uses conservation in two specific ways. The first is as a source of realistic perturbations: deleted bases are replaced with what neutral evolution would plausibly have put there, rather than with "N" characters or zeros that no genome contains. The second is as a decoy explanation that tells us which metrics measure the model and which merely reward biology. The central question is whether faithfulness-metric disagreement on genomic models is mostly an artifact of off-genome deletions. It also asks whether any current metric can tell "the model used this base" apart from "evolution cares about this base."

Core Claim

Conservation should not be treated as ground truth for faithfulness. That is the obvious move at this intersection, and it is a category error. A model can be accurate while ignoring conserved motifs, and a faithful explanation of such a model should ignore them too. The survey Interpretable machine learning for genomics draws this distinction between model-level and system-level explanation but offers no test for it.

Our claim is that conservation becomes useful for meta-evaluation once it is paired with models whose actual evidence is known by construction. On a model trained to rely on a planted, non-conserved signal, the conservation track is a biologically real but model-unfaithful "explainer." Any faithfulness metric that scores it well is measuring something other than the model. Separately, a neutral substitution process fitted to multi-species alignments gives graded, composition-preserving deletions. These let us test whether the disagreement among metrics shrinks once perturbations stay on the genomic manifold, which is the question that Faithfulness Under the Distribution and F-Fidelity answered for images, time series and text but not for DNA.

Relation to Prior Work

Three lines of work motivate the gap by showing that metric rankings are unstable. M4 benchmarks perturbation, consensus and synthetic-trigger metrics on ImageNet and movie reviews and finds that they do not correlate and that rankings flip across architectures. MetaQuantus scores metrics by stability under harmless perturbations and reactivity to destructive ones, without ground truth. A Comparative Study of Faithfulness Metrics measures how often metrics prefer a real explainer over random noise. There, Comprehensiveness (COMP, the prediction drop when top tokens are removed) and Sufficiency (SUFF, the prediction retained when only top tokens are kept) reach roughly 75–76% diagnosticity, while decision-flip metrics reach 8–10%. None of these works touches biological sequence. M4's synthetic ground truth relies on poisoned trigger patches, which we reuse in spirit as planted motifs but embed in real genomic background.

Two papers supply the method we adapt. They trace metric failure to out-of-distribution deletions. Faithfulness Under the Distribution shows that zero-masked images are detectable as off-manifold with nearly 90% accuracy, and it repairs deletions with diffusion in-painting, raising confidence-decay monotonicity (Kendall's $\tau$) above 0.80 from below 0.30. F-Fidelity instead fine-tunes the model on random masks up to a ratio $\beta$ and evaluates within that ratio. It reaches Spearman correlations of 1.0 or −1.0 when ranking deliberately degraded explainers. Our evolutionary in-filler plays the role of FUD's generator without training a DNA diffusion model, and F-Fidelity is our strongest competitor.

A constraint comes from Normalized AOPC. It shows that raw area-over-perturbation-curve (AOPC) scores have per-input, per-model bounds, and that normalizing them flips rankings. We therefore report normalized AOPC (NAOPC) whenever we compare across architectures. The work on Fourier-transform-based attribution priors supplies our datasets (SPI1, GATA2, Nanog/Oct4/Sox2 binding, K562 accessibility). It also shows that genomic attributions are unstable across random seeds; its prior raises cross-seed Jaccard overlap by roughly 40% on SPI1. That paper evaluates stability, not faithfulness, and we include its prior-trained models as an extra condition.

On the biology side, c-REDUCE and phylogenetic footprinting establish that cross-species conservation separates functional motifs from chance matches. c-REDUCE recovers 18 versus 10 of 37 hard yeast factors. This justifies treating conservation as a strong system-level signal. The same literature's warnings about motif turnover and lineage-specific elements explain why it cannot be a model-level truth.

Research Program

We would build three components.

The first is the neutral-evolution perturbation. From multi-species alignments (BLASTZ-style, as in Human–Mouse Alignments with BLASTZ) restricted to poorly conserved, likely neutral positions, we fit a context-dependent nucleotide substitution rate matrix $Q_c$ for each flanking context $c$. Fitting by context captures effects such as the fast decay of CpG dinucleotides. To "delete" a set of attributed positions, we resample each base $b$ from

$$P(b' \mid b, c, t) = \left[\exp(Q_c\, t)\right]_{b b'},$$

where $t$ is a branch length that sets perturbation strength. As $t$ grows, the replaced bases approach neutral-equilibrium DNA while keeping local composition. The result is a deletion curve indexed by evolutionary time instead of by mask ratio. The fitting and sampling cost is negligible next to generative in-painting, and each perturbation needs only one forward pass.

The second component is a set of models with known evidence. We build three classes for each dataset.

  • Real-label models are trained on the original binding or accessibility labels.
  • Shortcut models are trained on real genomic windows, but the label is set by a planted k-mer inserted only at non-conserved positions, with the real motifs balanced across classes so they carry no label information.
  • Mixed models use both signals at a controlled ratio.

For shortcut and mixed models, model-level ground truth is known at the planted positions, in the style of M4's synthetic score. We verify that the model actually uses the plant by checking its accuracy on held-out windows where the plant is moved. Architectures are a standard 1D convolutional network, a dilated convolutional network, and the DNA language model DNABERT.

The third component is the conservation decoy. Alongside ordinary explainers, we submit the per-base conservation track as if it were an attribution map. The ordinary explainers are input gradients, Integrated Gradients, DeepSHAP, in silico mutagenesis, a random map, and F-Fidelity-style degraded variants of each. On shortcut models, a metric passes the decoy test on an input if it scores the true planted-position map above the conservation map. The decoy rejection rate is the fraction of inputs passed. On real-label models, conservation overlap is reported only as a biological plausibility score, kept separate from faithfulness.

Every metric is then crossed with every deletion scheme and every model: COMP, SUFF, AOPC, NAOPC, deletion/insertion curves, and the MetaQuantus consistency scores. The deletion schemes are N-masking, zeros, uniform random bases, dinucleotide shuffle of the window, F-Fidelity fine-tuning, and the evolutionary sampler.

Baselines

The naive reference point is the standard genomic practice: deletion by N characters or all-zero one-hot vectors, scored with raw AOPC and COMP/SUFF. It shows how much disagreement exists before any fix. We expect it to reproduce, in DNA, the off-manifold detectability that FUD found in images. It shares the same models, explainers and forward-pass budget as every other condition.

The strongest existing method is F-Fidelity, ported to DNA. The model is fine-tuned once on windows with a random fraction up to $\beta$ of bases replaced by uniform random nucleotides, and evaluation deletions are kept within $\beta$. It is the strongest competitor because it is the only published method that removes the off-manifold problem across several modalities at modest cost. Its reported result is Spearman 1.0/−1.0 on degraded-explainer ranking in vision, time series and text. We found no reported result on Lacuna for F-Fidelity or any faithfulness metric on genomic models. To keep the comparison fair, we give F-Fidelity the same $\beta$ grid as our branch-length grid and tune both on the same validation windows with an equal number of forward passes. We also report its fine-tuning compute separately, since our sampler needs no retraining. Beating it means higher decoy rejection and lower cross-metric disagreement at matched perturbation strength, while matching its degraded-explainer ranking correlation.

The key ablation is a composition-matched shuffle: a dinucleotide-preserving shuffle of the attributed positions within their window. It isolates whether any benefit comes from alignment-derived, context-specific evolutionary rates or merely from preserving base composition. A second ablation replaces $Q_c$ with a single context-free rate matrix. We also include a FUD-like condition only if a DNA generative model can be trained within budget; otherwise we state that it was omitted.

Why not just…?

Why not just score attributions by their overlap with the conservation track? On a real-label model, this measures whether the model learned biology, not whether the explanation describes the model. The shortcut models make the failure concrete. A good explainer of a shortcut model should overlap poorly with conservation, and this proposal's decoy test measures exactly how badly overlap-based evaluation misranks it. The survey Interpretable machine learning for genomics warns of this conflation, and our experiment quantifies it.

Why not just use a dinucleotide shuffle, the usual genomic reference, instead of an evolutionary model? This is a real threat. Shuffling already preserves much of the composition that N-masking destroys, and it may close most of the gap. We include it as an ablation. If it matches the evolutionary sampler on detectability, monotonicity and disagreement, we will report that composition matching suffices. The contribution would then rest on the measurement and the decoy test. What shuffling cannot do is give a graded, context-aware perturbation strength, and shuffling a window can recreate motifs by chance; the branch-length curve tests whether these differences matter.

Why not just port F-Fidelity and stop? F-Fidelity may well fix off-manifold deletions in DNA. Even so, a metric can be perfectly on-manifold and still reward biological plausibility over model reliance, because a model fine-tuned on random masks can still respond to deleting conserved motifs it partly uses. The decoy test is orthogonal to the off-manifold fix, and we run it under F-Fidelity too. Whichever deletion scheme wins, the question of which metric measures the model remains open until this test is run.

Why not just plant synthetic motifs, as M4 does? We do plant motifs, but into real genomic windows whose conserved motifs remain present and uninformative. The pure synthetic background criticized in the research file would lack the competing, biologically real signal. That competing signal is precisely the one a metric might wrongly reward, so the decoy test cannot exist without it.

Proposed Experiments

The first check measures off-manifold detectability. We train a small classifier to separate each deletion scheme's outputs from untouched genomic windows at matched fraction of bases changed and report its accuracy. The comparison point is the roughly 90% FUD reported for zero-masked images. We expect N-masking to be near-trivially detectable and the evolutionary sampler and shuffle to approach chance.

The second check measures disagreement. For each dataset, model and deletion scheme, we compute the mean pairwise Kendall's $\tau$ between the explainer rankings produced by COMP, SUFF, NAOPC and deletion curves, together with confidence-decay monotonicity as in FUD. The hypothesis is decisive in both directions. If rank agreement rises substantially when moving from N-masking to on-manifold schemes, genomic metric disagreement is largely an off-manifold artifact. If it stays low, the disagreement reflects genuinely different properties, as Consistent Saliency Benchmarking and spectral analysis of explanation quality argue for images.

The third check is decoy rejection on shortcut and mixed models. For each metric and scheme we report the decoy rejection rate against planted-position ground truth. We also report Chan et al.'s diagnosticity (explainer beats random) as a secondary score. A metric earns trust here only if it both rejects the conservation decoy on shortcut models and ranks the degraded explainers correctly. In mixed models, the rate at which it prefers the conservation map should track the planted-to-real signal ratio.

The fourth check covers robustness. We repeat the analysis on Fourier-prior-trained models, which may shift attributions toward motif-length structure and alter decoy scores. We also compare architectures under NAOPC versus raw AOPC to test whether reported convolutional-network versus DNABERT explainer rankings survive normalization.

Throughout, performance is reported separately for attributed positions inside and outside conserved elements. This matters because the evolutionary sampler perturbs conserved and neutral positions differently, and that difference is itself a finding about where metric disagreement originates.