Some of the strongest recent prognostic signals in blood cancers come from epigenetic disorder. The term means the degree to which neighboring DNA methylation marks disagree on the same DNA molecule, or the degree to which large, normally methylated stretches of the genome erode in one patient's tumor. Measuring disorder requires reading individual molecules across wide regions, which in practice means whole-genome bisulfite or enzymatic sequencing. That is too expensive for routine cohorts, so the published disorder scores rest on a few dozen to a few hundred patients. This proposal treats a patient's methylome the way a mobile robot treats an unmapped building. Sequenced regions are the known map. Unsequenced regions are unknown territory. The informative places are the frontiers: the boundaries, specific to each patient, where a methylated domain gives way to an eroded or disordered one. A sequencing controller that spends its read budget on these frontiers, instead of spreading reads uniformly, should recover boundary-localized prognostic scores at a fraction of full-depth cost.
The core claim is narrow and testable. For a fixed read budget, frontier-guided sampling estimates boundary-localized disorder scores more accurately, and stratifies patients more faithfully, than the four cheaper options: uniform shallow sequencing, a fixed panel built from reference domain maps, uncertainty sampling driven by a methylation imputation model, and that imputation model alone. The claim is deliberately limited to scores whose signal sits near domain boundaries. Global averages are probably estimated just as well by uniform reads, and the proposal says so up front.
What each side brings, and where the real gap is
Frontier-based exploration supplies the method. A frontier is the edge between mapped free space and unknown space, and driving toward it guarantees new information (Navigating to Objects Specified by Images builds exactly this over a 2D occupancy grid inside a modular navigation system). Two results from this literature constrain the design. First, Information Gain Is Not All You Need shows that greedily maximizing information gain is fine when the mission has a fixed budget but harmful when the mission requires complete coverage. In their office environment it was 23% worse in path length than going to the nearest frontier, because it leaves small isolated pockets that must be revisited later. Their "distance advantage" rule prefers frontiers that are cheap to reach now but isolated from everything else. Second, POMP++ handles a map that grows as the robot moves. When the map updates, it re-projects its belief particles onto the new map and injects fresh particles at newly exposed frontiers, so the planner does not stop looking at unexplored areas. We reuse both ideas: the budget-versus-coverage distinction tells us which acquisition rule to expect to win, and belief reinvigoration handles boundaries that appear only after early reads arrive.
The epigenetics side supplies the target. Lee et al. define local pairwise methylation discordance (LPMD), the rate at which adjacent CpG sites on one molecule disagree. In acute myeloid leukemia (AML) they link high LPMD to worse survival (hazard ratio 1.85, p = 0.0011). They also show that discordance restricted to bivalent chromatin domains (bivLPMD) predicts remission under hypomethylating therapy (p = 0.0066). Derrien et al. show that in multiple myeloma, partially methylated domains (PMDs) erode in patient-specific ways across roughly 65% of the genome. Their entropy-shift measure, eloci per million (EPM), associates with shorter relapse-free survival (p = 0.02) in 21 patients. Both papers motivate this proposal: the signal is regional, and the cohorts are small because full-molecule profiling is costly. Both analyze pre-sequenced data and contain no sampling strategy. MethylProphet predicts methylation at all 28 million CpGs from RNA-seq and DNA sequence, reaching a median across-sample Pearson correlation of 0.72 on ENCODE. It acts as a constraint and a prior rather than a competitor. It is one-shot, and by its authors' account it cannot capture patient-specific deviations, which are precisely what a disorder score measures.
The research file found no work that steers sequencing toward epigenetic boundaries. The closest algorithmic relative is Adaptive Sampling for Discovery, which chooses which chemistry experiments to run next under a small budget. Its candidate pool is fixed and has no spatial structure, so it has no notion of a boundary or of a state space that grows. Our setting has both.
Core Claim
Disorder signals that matter clinically concentrate at a patient's own domain boundaries: PMD edges, eroding bivalent domains, and transitions from ordered to discordant methylation. Because those boundaries move from patient to patient, a fixed panel misses them. Because they are a small fraction of the genome, uniform shallow reads cover them thinly. A controller that keeps a running belief over where each patient's boundaries lie, and sends reads to the most uncertain frontier segments, should reach full-depth accuracy on boundary scores with substantially fewer reads.
Research Program
The controller keeps a per-patient segmentation of the genome into domain states (highly methylated, partially methylated/eroded, and disordered) using a hidden Markov model over 1–10 kb windows. The model is initialized from a population prior: reference PMD and bivalent maps, plus MethylProphet predictions when the patient has RNA-seq. Each incoming batch of single-molecule reads updates two quantities for every window: the posterior over its domain label, and a per-window discordance estimate.
A window is a frontier if it is next to confidently labeled windows but its own label, or the exact position of a boundary inside it, is still uncertain. The controller ranks frontier windows by expected reduction in the variance of the target score per read spent. It then reinvigorates in the manner of POMP++: whenever an update moves a boundary, the newly exposed neighboring windows receive prior mass as candidate frontiers. This prevents the controller from locking onto the boundaries it believed in at the start.
Following Ericson et al., we implement two acquisition rules. The first maximizes information gain, which should suit fixed-budget estimation of one score. The second is a genomic distance-advantage rule that prefers short isolated boundary fragments and should suit tasks that require mapping every boundary. Running both turns their robotics finding into a prediction we can check on genomes.
The hardware target is real-time adaptive nanopore sequencing. In that setup the controller sees the first portion of each read and either keeps sequencing it or ejects it. All primary experiments, however, are retrospective simulations. We take existing deep single-molecule data and replay reads in random arrival order. The controller accepts or rejects each read, and a rejected read is charged a fixed fraction of its full cost. This keeps the study cheap and fully reproducible. The rejection cost is varied in a sensitivity analysis because the file reports no figure for it.
Proposed Experiments
The data come from three sources. The Derrien et al. myeloma cohort (42 samples from 21 patients, with whole-genome bisulfite and eRRBS) is the testbed for EPM and PMD boundaries. The Li2016 AML cohort from Lee et al. (94 paired diagnosis/relapse samples, targeted EM-seq) is the testbed for LPMD and bivLPMD; its limitation is that the targeted design caps what can be explored. Deeply sequenced ENCODE methylomes serve as a mechanistic check with known high-depth ground truth.
Budgets sweep from 1% to 30% of each sample's full read set. The primary metric is the error of each score relative to its full-depth value: absolute error and patient-rank correlation for genome-wide LPMD, bivLPMD, EPM, and fraction of PMD eroded. The secondary metric is clinical. At each budget we recompute the published associations (the AML survival hazard ratio and remission association, the myeloma relapse-free survival split) and report how close the subsampled estimates come to the full-depth ones. The reference values are HR 1.85 and p = 0.0011 for LPMD survival, p = 0.0066 for bivLPMD remission, and p = 0.02 for EPM relapse-free survival. With 21 myeloma patients the clinical check is underpowered, so it is a consistency test, not a discovery claim. Boundary-mapping quality is measured as base-pair error of called PMD edges against full-depth calls.
Baselines
The naive reference is uniform shallow sequencing: accept every read until the budget runs out, then compute each score from whatever was read. This is a real competitor, because random reads give unbiased estimates of any genome-wide average. It is run on the identical read pool and budget. Beating it means lower error on the boundary-localized scores (bivLPMD, edge-restricted LPMD, PMD edge position) at matched budget, over many random read orderings, with paired significance tests across patients.
The fixed-panel baseline accepts only reads that fall in reference PMD and bivalent regions taken from public atlases, which is the static-panel position a skeptic would take. It receives the same budget and the same rejection-cost model. Beating it requires showing that patient-specific boundaries depart from reference maps often enough to matter. We measure that departure directly as the fraction of each patient's full-depth boundary length that lies outside the reference regions.
The strongest existing method is imputation-guided uncertainty sampling. Windows are ranked by the predictive variance of a MethylProphet-initialized per-window model, with no segmentation and no frontier structure. This is the closest analogue to Adaptive Sampling for Discovery and to standard active learning. It uses the same prior, the same per-window update, and the same tuning budget, with hyperparameters selected on held-out patients only. No reported result on Lacuna exists for adaptive methylation sequencing under any method. The only external anchor is MethylProphet's own imputation accuracy, a median across-sample Pearson correlation of 0.72 on ENCODE. We also report MethylProphet with zero reads as a floor for what imputation alone delivers on the disorder scores. Beating this baseline means matching its score error with at least 30% fewer reads, or achieving lower error at equal budget, on the boundary scores.
Two ablations isolate the key ingredients. Removing reinvigoration, so the frontier set is frozen after the prior, tests whether discovering new boundaries matters. Swapping information gain for distance advantage tests the Ericson et al. prediction that greedy information gain wins on single-score budgets but loses on complete boundary mapping.
Why not just…?
Why not just sequence everything shallowly and compute the score? For genome-wide LPMD or the overall eroded-PMD fraction, this may well be as good as anything adaptive, and we expect it to be. That is why the claim is restricted to boundary-localized scores. The uniform baseline settles the question directly. If it matches frontier sampling on bivLPMD and edge positions at every budget, the proposal's central claim is false, and we will report that.
Why not just use a fixed panel over known PMD and bivalent regions? Derrien et al. emphasize heavy heterogeneity both between patients and within a single patient. If boundaries were stable, reference maps would locate them for free. The panel baseline, together with the measured fraction of patient boundary length lying outside reference regions, tells us whether patient-specific discovery is needed. A small fraction would make the panel the right answer.
Why not just impute the methylome with MethylProphet? Its 0.72 correlation is a statement about mean methylation levels. Disorder scores depend on disagreement within single molecules, which a model that predicts a methylation level per site does not output at all. The zero-read imputation floor quantifies how much of each disorder score imputation recovers. We nonetheless use MethylProphet as the prior, so any benefit it offers is absorbed into our method rather than set against it.
Why not just do uncertainty sampling without frontiers? This is the most serious threat. A per-window variance criterion also tends to land near boundaries. The ablation that keeps the prior and update rule but drops the segmentation and reinvigoration answers it directly. If plain uncertainty sampling performs as well, the honest conclusion is that the robotics framing adds nothing beyond active learning, and the contribution shrinks to the adaptive-sequencing protocol itself.
Decisive Checks
The proposal succeeds only if three things hold. On held-out patients, frontier-guided sampling reaches the full-depth bivLPMD and PMD-edge estimates at a markedly lower budget than both uniform sampling and plain uncertainty sampling. The patient risk ordering it produces reproduces the published associations at that budget. And the acquisition-rule ablation behaves as the robotics result predicts: information gain is best for single scores, distance advantage is best for complete boundary maps. Failure on the first check, with uniform reads matching at every budget, would be a useful negative result. It would mean disorder biomarkers can scale with cheap shallow sequencing and do not need adaptive control at all.
References
- Navigating to Objects Specified by Images
- Information Gain Is Not All You Need
- POMP++: Pomcp-based Active Visual Search in unknown indoor environments
- Increased local DNA methylation disorder in AMLs with DNMT3A-destabilizing variants and its clinical implication
- The DNA methylation landscape of multiple myeloma shows extensive inter- and intrapatient heterogeneity that fuels transcriptomic variability
- A New Paradigm for Genome-wide DNA Methylation Prediction Without Methylation Input
- Adaptive Sampling for Discovery