Moreover, for m1Y and Y, I am not aware of a non-enzymatic reaction that could convert one into the other during any downstream processing (leading to contamination). Unlike for m1A which can be converted to m6A at high temperatures, something that could happen easily at a heat inactivation step.

Hi James,
thanks for the nice words! I am speculating here but I think it is the latter. The training data should be synthetic sequences containing each of the modifications, and therefore quite clean.

Huge thanks to all the incredible people @novoalab.bsky.social that helped make this work possible! #rnasky #nanopore

Finally, we established that a combination of alignment and current features still presents a robust way of detecting RNA modifications with RNA004, as demonstrated on E.coli strains knockout for certain mod writers. (11/12)

While IVTs helped remove sequence-specific FPs two classes of FPs remained: I) cross-reactive mods found on the same base, and II) mods at adjacent positions (+/- 1nt) from identified FP sites, which cannot be accounted for with existing methods. (10/12)

Considering the substantial amount of FPs observed at unannotated sites we reasoned that the use of IVT controls might be necessary to help remove sequence specific FPs. (9/12)

From there we moved to rRNAs for benchmarking since accurate mod. maps exist for well-studied species. We could recapitulate most Ψ and m5C sites while also observing substantial FP, especially at base methylations and unannotated sites. (8/12)

Moving to per-site predictions (removing per read estimates that fall below the default cutoff), we observe that most models slightly underpredict the per-site stoichiometries with the pseU model performing the most accurate (particularly in the 0-25% modified range). (7/12)

Digging a bit deeper, we can identify particular sets of sequences (5mer) that produce lower probability estimates than the remaining ones, suggesting a strong impact of the sequence context on predictions. (6/12)

Key observations (per-read): 1) in general, models perform well. 2) Ψ and m1Ψ are indistinguishable. 3) m5C and hm5C cross-react in particular sequence contexts as indicated by the bimodal distribution of hm5C called with the m5C model. (5/12)

But first, a quick word on mod-aware basecallers and the information they provide. The models make predictions on a per-read and position basis (per-read) which can then be collapsed into per-site predictions, using downstream tools which consider a certain filter-threshold (per-site).

Next we went ahead and sequenced synthetic oligos containing all possible 5mers (n=1024) to assess model performance, cross-reactivities and potential sequence biases. (3/12)

First things first, we updated SeqTagger to support 96 barcodes with the newest sequencing chemistry (RNA004). If you are interested in multiplexing your own DRS runs the new model is openly available here -> github.com/novoalab/Seq.... Feedback is very welcome! (2/12)