Critical Bands of the Methylome: Frequency-Specific Masking for Function-Guided MSP Design

Vision science has an old trick for finding out which scales of detail a perceiver actually relies on: bury the target in noise restricted to a narrow band of spatial frequencies — coarse blur or fine grain — and watch when recognition collapses. Noise inside the observer's "critical band" impairs performance; noise outside it does nothing. That is how the field established that human object recognition runs through a channel roughly one octave wide, while standard deep networks use channels two to four times broader and adversarially trained ones nearly six octaves wide, a distortion that accounts for most of the variance in their adversarial fragility Spatial-frequency channels, shape bias, and adversarial robustness. A promoter's DNA methylation profile is, structurally, the same kind of object: a one-dimensional signal varying across spatial scale along the genome, from broad multi-kilobase heterochromatin domains down to sharp transitions at the edges of CpG islands. Nobody has asked which of those scales the gene-regulation system — or our machine-learning models of it — actually reads.

The payoff of answering that lands on a clinical assay. Methylation-specific PCR (MSP) reads methylation by first treating DNA with sodium bisulfite, a chemical step that rewrites every unmethylated cytosine into a thymine while leaving methylated ones intact, and then amplifying a short promoter stretch with primers that bind only the methylated or only the unmethylated version — a cheap, sensitive, binary readout used throughout cancer diagnostics, for example in renal and lung adenocarcinoma marker studies Lei Li et al., Xuewei Hao et al.. But primer placement is essentially arbitrary: primers go wherever a designable 20–30 base stretch happens to sit, even though genome-wide evidence says the informative signal is the spatial shape of methylation across ±7 kb around the transcription start site, not any single local call. This proposal measures the functional bandwidth of promoter methylation using frequency-specific noise masking, and then redesigns MSP panels so that their primers sample exactly that bandwidth.

Core Claim

The claim is that promoter methylation has a measurable critical band: a limited range of spatial scales, in cycles per kilobase along the genome, over which methylation variation carries the information that predicts gene silencing and tumor identity — and that this band can be measured by the same narrowband-masking logic used in psychophysics, then converted directly into MSP panel geometry (amplicon span, CpG density, placement). The evidence that shape matters is already strong: replacing scalar average methylation with spatial basis-function features more than doubled expression-prediction accuracy in K562 cells, from Pearson's r = 0.31 to r = 0.70 Higher order methylation features for clustering and prediction in epigenomic studies. But that work used fixed spatial bumps in a regression model; it says which features predict, not which scales are load-bearing, and it offers no route to a wet-lab assay. The masking diagnostic answers the causal question a predictive feature set cannot, because band-limited noise selectively removes one scale at a time.

Three further strands motivate or constrain the design. The cross-platform tumor classifier crossNN achieves its robustness by hiding up to 99.75% of methylation features at random during training, reaching 99.1% precision on brain tumors and 97.8% on pan-cancer cohorts crossNN is an explainable framework for cross-platform DNA methylation-based classification of tumors; this constrains us, because it shows classifiers survive white noise — but i.i.d. dropout is spectrally flat, and in vision, spectrally flat perturbations routinely miss failure modes that narrowband noise exposes, whether in deep networks Spatial-frequency channels, shape bias, and adversarial robustness or in human lightness perception, where only noise in a specific 1–5 cycles-per-degree band suppresses White's illusion while out-of-band noise does nothing Noise masking of White's illusion exposes the weakness of current spatial filtering models of lightness perception. Finally, two results warn against over-claiming a linear-filter story: linear filter banks fail on texture segmentation and on lightness A computational model of texture segmentation, and channels interact nonlinearly under correlated broadband noise Contrast discrimination with pulse trains in pink noise. Our channel fits are descriptive diagnostics of trained models and of the regression task, not a claim that the biology is a linear filter bank.

Research Program

The program has four components. First, we build methylation trajectories $M(x)$ over ±7 kb windows around transcription start sites from ENCODE reduced-representation bisulfite sequencing in K562, GM12878, and H1-hESC cells (the same data as Higher order methylation features for clustering and prediction in epigenomic studies), plus the classification cohorts used by crossNN (the Heidelberg v11b4 brain-tumor set and TCGA pan-cancer). We compute Fourier and wavelet power spectra of these profiles to establish whether the epigenome has stable band structure at all — for instance, whether power concentrates at CpG-island-shore transition scales.

Second, the masking diagnostic. We corrupt clean profiles with band-limited Gaussian noise,

$$\tilde{M}(x) = M(x) + \mathcal{B}_{f_0,\Delta f}\{\varepsilon(x)\}, \qquad \varepsilon(x)\sim\mathcal{N}(0,\sigma^2),$$

where $\mathcal{B}_{f_0,\Delta f}$ is a bandpass filter keeping only frequencies within $\Delta f$ octaves of center $f_0$ (in cycles per kilobase), and beta values are clipped to $[0,1]$. For each center frequency we find the noise level $\sigma$ that halves the clean expression-prediction correlation — the genomic analog of the 50%-accuracy threshold used to fit channels in Spatial-frequency channels, shape bias, and adversarial robustness — and fit the same Gaussian channel model to extract center frequency, bandwidth in octaves, and peak sensitivity. We do this for several predictor families (scalar mean, a BPR reproduction, a 1D convolutional network); if all families yield the same band, the band belongs to the task itself, i.e., to the biology, not to one architecture. For classifiers, we corrupt the underlying profile before crossNN's binarization at its 0.6 threshold and ternary encoding, so the only change from the published pipeline is the noise spectrum. We additionally model realistic platform missingness as band-limited corruption, since read-level coverage gaps in sequencing are spatially contiguous, not i.i.d.

Third, from channel to assay. If function concentrates at high frequencies (sharp, localized transitions), band-matched panels should place short amplicons densely across those transitions; if at low frequencies, panels should composite methylation across wide regions via multiplexed amplicons. We build a panel designer that samples the measured critical band at its fitted spacing, and compare it against a conventional designer following standard practice (primers on CpG-dense promoter stretches).

Fourth, bisulfite-aware amplification noise. Any designed panel must actually amplify after bisulfite conversion, which depletes cytosines and collapses sequence complexity toward a three-letter alphabet. The only published sequence-dropout predictor — a 1D convolutional network reaching AUROC 0.88 and AUPRC 0.44, which found that motifs like CGTG near priming sites cause self-priming that halves abundance every 14 PCR cycles Predicting sequence-specific amplification efficiency in multi-template PCR with deep learning — was trained on native, composition-balanced synthetic oligos; no reported result found on Lacuna for bisulfite-converted sequences. We test that model zero-shot on in-silito bisulfite-converted versions of its 12,000-template, 90-cycle protocol design, then retrain with conversion-aware augmentation, and use the retrained model to filter candidate panels for self-priming dropout. A small wet-lab pilot closes the loop: band-matched and conventional panels for roughly thirty promoters in K562 and GM12878, where ENCODE RRBS provides ground truth, scored on concordance of MSP calls with measured methylation states.

Baselines

The do-nothing reference is scalar average promoter methylation as an expression predictor; Higher order methylation features for clustering and prediction in epigenomic studies reports Pearson's r = 0.31 for it in K562, and it defines the floor any spatial representation must clear. The strongest competitor is the same paper's BPR model, spatial radial-basis-function features in a probit regression, which reports r = 0.70 in K562 — the best expression-from-methylation number in the corpus. We will reproduce it on identical ENCODE RRBS windows, splits, and tuning budget, and count beating it as matching or exceeding r = 0.70 while using only the CpG sites the diagnostic marks as in-band (reporting the CpG fraction retained), or matching r with graceful degradation under critical-band noise where BPR collapses. For classification, the strongest competitor is crossNN, reporting 99.1% precision on the Heidelberg v11b4 set (2,801 samples, 82 tumor types) and 97.8% on TCGA (8,382 cases, 178 tumor types), while surviving random subsampling down to 0.5% of its feature space crossNN is an explainable framework for cross-platform DNA methylation-based classification of tumors; we retrain it on the same cohorts with the same ternary pipeline and compute, and beat it only if band-selected features retain that precision under spatially contiguous dropout at budgets where the i.i.d.-trained version degrades — no reported result found on Lacuna for crossNN under correlated dropout, so we must measure it ourselves. For amplification, the baseline is the published 1D-CNN applied unchanged (AUROC 0.88, AUPRC 0.44 on native pools) Predicting sequence-specific amplification efficiency in multi-template PCR with deep learning. The key ablation runs the entire diagnostic with i.i.d. masking at matched total noise power instead of band-limited masking; if that produces the same conclusions, the frequency structure is doing no work and the proposal's ingredient is inert.

Why not just…?

Why not just use average promoter methylation? Because averaging demonstrably destroys the signal: it yields r = 0.31 against r = 0.70 for spatial features in the same cell line, and it obliterates the U-shaped active-promoter profile (TSS hypomethylation flanked by hypermethylated shores, p < 2.2e-16) that carries the regulatory information Higher order methylation features for clustering and prediction in epigenomic studies. A scalar also cannot say anything about scale, which is the entire question here.

Why not just use random feature masking, as crossNN already does? This is the most serious objection, and crossNN's numbers make it a real threat. The answer is twofold. First, i.i.d. dropout injects white noise, and the visual literature shows spectrally flat perturbations miss band-specific structure: networks look fine under broadband tests while relying on channels 2.3 to nearly 6 octaves wide Spatial-frequency channels, shape bias, and adversarial robustness, and only noise within 1–5 cycles per degree attenuates White's illusion while out-of-band noise has zero effect Noise masking of White's illusion exposes the weakness of current spatial filtering models of lightness perception. Second, real coverage gaps in bisulfite sequencing are spatially contiguous, not i.i.d., so i.i.d. masking likely overstates deployed robustness. Still, if our experiments show band-selected features do not outperform random selection at matched budgets under contiguous dropout, we will report that the epigenome has no exploitable band structure for classification — that outcome would itself settle the question, and the ablation at matched noise power is designed to detect exactly this.

Why not just design primers with melting-temperature and GC rules? Two reasons. Even a perfectly amplifying primer at the wrong locus measures nothing, and only the masking diagnostic says where the informative locus-scales are. And even for amplification itself, thermodynamic rules fail: in GC-normalized synthetic pools, severe "dark dropout" persisted, driven by positional motifs causing adapter-mediated self-priming that halves abundance every 14 cycles, and only a deep sequence model predicted it (AUROC 0.88) Predicting sequence-specific amplification efficiency in multi-template PCR with deep learning. After bisulfite conversion the sequence statistics shift so far (cytosine depletion, T-rich homopolymers) that GC-calibrated rules are aimed at a different distribution. One could instead abandon bisulfite entirely for bisulfite-free site-specific PCR, which shows clinical promise in colorectal cancer detection Early Diagnosis of Colorectal Cancer Based on Bisulfite‐free Site‐specific Methylation Identification PCR Strategy, but bisulfite conversion remains the basis of essentially all genome-wide profiling and of every model in this corpus, so the band diagnostic must first be established on the dominant chemistry; the diagnostic itself transfers to any readout.

Decisive Checks

The proposal dies in a well-defined way if the masking curve is flat: if halving the expression-prediction correlation requires the same noise power at every center frequency, there is no critical band and no basis for band-matched design. It survives in weak form if band-matched feature subsets match full-window accuracy at reduced CpG counts, and in strong form if band-selected features beat random selection under contiguous dropout in the classification cohorts and the wet-lab pilot's band-matched panels show higher concordance with RRBS ground truth than conventionally placed panels. The bisulfite dropout question is independently decisive for the assay leg: if the published amplification model transfers zero-shot to converted sequences with negligible loss, retraining is unnecessary and we say so; if it degrades sharply, that measured gap is itself a finding, since no reported result found on Lacuna covers amplification noise in the chemistry that nearly every methylation assay in the clinic still depends on.