The label requirement in existing DMR tools is not administrative overhead. It decides in advance which differences the analysis is capable of finding: split a cohort into healthy and diseased, and the algorithm returns the regions that separate healthy from diseased, and nothing else [2]. For a cohort where the grouping is itself the open question, that is circular. The unsupervised mode in metilene3 reverses the order, segmenting the genome on the methylation signal first and assembling the sample groups out of whatever that segmentation produces [4].
What the inversion buys and what it costs are both visible in the test cases. In human blood cells the software recovered known developmental pathways among immune cell types from methylation alone [5], and picked out regulatory regions tied to the transcription factors that fix those cell identities [6]. On glioblastoma data it separated molecular subgroups and singled out individual samples with unusual biology [7]. Pancreatic tissue was the harder shape, and it traced the sequence from healthy tissue through precancerous lesions to tumour [8]. The blood result is graded against pathways the field already had [5], which is the right way to test a clustering method and also the reason the benchmark says little about the situation the tool exists for: where nobody knows the correct grouping, nobody can count the incorrect ones either.
The single result the authors present as new biology is a set of regions where NF-kappaB and NFAT binding sites occur together unusually often in pancreatic cancer, and they say plainly that the hypothesis still has to be verified experimentally [9][1]. That is the shape of most of what this method will emit. An unsupervised pipeline moves work from the front of a study, where someone had to label samples, to the back, where someone has to test the groups that came out. The labelling bottleneck was at least cheap to describe.
Helene Kretzmer of the Hasso Plattner Institute makes the point that medical questions impose particularly high requirements on the interpretability of predictions, and that the methylation changes the method finds support direct conclusions about disrupted transcription factors [10]. The design answer to that is traceability: every similarity and difference the software reports can be attributed back to specific methylation patterns [11]. A clustering that cannot name the regions driving it is hard to argue with and harder to act on, so this is the part of the paper that determines whether the unsupervised output reaches a clinical discussion at all.
Both modes ship in the same tool [3], which matters for adoption more than the novelty does: a lab can keep its predefined group comparisons and run the unlabeled pass on the same data. The claim to test is the one about heterogeneous tissue and clinical datasets where groups cannot be defined in advance [13], because that is where there is no ground truth to fall back on.
