One 0.23M-parameter model flags five DNA base modifications without an unmodified control, including 5hmU
Synopsis
RawMod, released by a University of Maryland-led team, stacks the raw nanopore signal of many reads at each site and compares it base by base with the expected signal of the reference, so a single 0.23M-parameter model can say whether a site carries a modification without naming its type. Across 15 datasets spanning 5mC, 5hmC, 4mC, 6mA and 5hmU in bacterial, phage, plant, mouse, human and synthetic DNA, it reaches a mean AUROC of 0.964, against 0.799 for the best released model, Dorado 4mC+5mC. It needs no unmodified control from the same sample, but a modification type absent from training transfers zero-shot only when another modification of the same base was seen.
Interpretation
A single model can tell whether a genomic site carries a base modification without knowing which modification it is, and without sequencing a separate modification-free control for each sample. Released supervised callers detect only the modification types they were trained on, while comparative callers rely on a baseline from the same sample after PCR or whole-genome amplification has removed the modifications. Scored on held-out genomic regions with the same 10 reads per site, RawMod reaches a mean AUROC of 0.964 over the five modification types, against 0.799 for Dorado 4mC+5mC and 0.899 for all four Dorado models combined, which lack 5hmU; RawMod's weakest type, 5hmU, is 0.930, while the weakest type of every released model is at most 0.503.
The method lays out the raw current signals of many reads around a site as a pileup, so the model reads how the signal departs from the reference sequence. A convolutional encoder compresses each row, including the reference row, into a 96-dimensional vector, cross-read attention combines these into a site vector, and a linear head gives the probability that the site is modified. Training combines classification, supervised contrastive and deep semi-supervised anomaly-detection objectives, first on pairwise sites whose modified and unmodified members share the same sequence and then on all data; the complete model has 230,697 parameters.
Perspective
The evaluation covers 5mC, 5hmC, 4mC, 6mA and 5hmU across 15 datasets of bacterial, phage, plant, mouse, human and synthetic DNA, and every scored site lies inside genomic regions withheld from training. The 5hmU map covers forward-strand thymines with at least 10 reads on contig_3762 (7.73 Mb) of Amphidinium carterae, scored by a model that never saw that genome.
Zero-shot transfer appears only when another modification of the same base stays in training: 5hmC, 5mC and 4mC are detected, while 6mA and 5hmU do not transfer, at AUROC 0.594 and 0.497. 5hmU reaches 0.82 only after retraining with about 3,000 labeled sites, or 0.72 with fine-tuning; scores are not calibrated modification fractions and cannot name the modification. The dinoflagellate map rests on a draft assembly with manually annotated genes, a median read error of 5.4%, and published profiles from a different strain, so it is compared by feature class rather than by position.