ModiDeC

Nanopore direct RNA sequencing reads native RNA rather than cDNA copies, so chemical modifications survive into the raw data as distortions of the ionic current as each nucleotide passes through the pore. Turning those distortions into confident, base-resolution calls is the hard part. Most callers target a single modification, and the field has been repeatedly disrupted by chemistry changes — models trained on Oxford Nanopore's RNA002 kit do not transfer to the RNA004 chemistry that replaced it.

ModiDeC is a deep learning classifier that identifies and distinguishes five RNA modifications from direct RNA sequencing signal: N6-methyladenosine (m6A), inosine, pseudouridine (Ψ), 2′-O-methylguanosine (Gm), and N1-methyladenosine (m1A). Developed by Nicolò Alagna, Susanne Gerber, Mark Helm, and colleagues at the University Medical Center Mainz and Johannes Gutenberg University Mainz, it was posted to bioRxiv in January 2025 and published in Nucleic Acids Research.

Two things distinguish it. First, it was trained on a purpose-built ground-truth resource: a database of in-vitro-transcribed and chemically synthesized oligonucleotides carrying known modifications at known positions, generated on both the RNA002 and RNA004 chemistries and released alongside the model. Second, it is explicitly built to be retrained. Three graphical interfaces and matching Epi2ME workflows cover data curation, training, and analysis, so a lab can extend the classifier to additional modifications or motifs without touching the source code — positioning ModiDeC less as a fixed transcriptome-wide scanner than as a customizable, highly specific classifier for the motifs a study cares about. Broader-panel detectors such as WattmaMod instead pursue coverage across many modification types in one pretrained model.

#Key Features

• Multi-modification classification: Each nucleotide is labeled unmodified or as one of m6A, inosine, Ψ, or Gm in a single pass, with m1A added through the same retraining path, so chemically similar marks are distinguished rather than detected one at a time.
• Two-input architecture: The network consumes re-squiggled raw current and the one-hot-encoded reference sequence in parallel, so sequence context disambiguates signal patterns shared across motifs.
• Purpose-built synthetic ground truth: Splint-ligated oligonucleotides with known modification positions supply exact labels, avoiding the circularity of training on inferred sites, with separate models for the RNA002 and RNA004 chemistries.
• Retrainable without coding: Three GUIs and three Epi2ME pipelines cover data curation, training, and analysis, letting users add modification classes and motifs.

#Technical Details

ModiDeC is a two-input neural network of roughly 20M parameters combining LSTM layers with a custom Inception-ResNet block. Raw pod5 signal is basecalled with Dorado v0.7.3 using --emit-moves, re-squiggled against the reference with Remora, Z-score normalized, and split into chunks of 400 measurement points, each paired with its one-hot-encoded reference sequence of up to 40 nucleotides. Training used 40 splint-ligated sequences carrying one modified oligonucleotide each, at 30,000 reads per modification-motif combination, plus in-vitro-transcribed HEK293 reads — more than 40 million labeled chunks in total.

On a held-out test set covering the trained motifs at 5,000 reads per oligonucleotide, the network identified the position and type of modification correctly for all 40 sequence-modification combinations, with 81% accuracy on quantifying modification frequency; per-motif accuracy varies substantially, and frequency is underestimated by about 10% on average. Virtual titration gave an average limit of detection of 1.1% and limit of quantification of 3.2%. False-positive rates on unmodified controls ranged from 0% to 1% for RNA004 and 0% to 5.5% for RNA002, and against Remora models trained on identical data ModiDeC recovered the expected inosine and Gm positions with no additional false positives where Remora produced several. It was then applied to engineered pseudouridylation in HEK293T cells, HEK293T rRNA, and human peripheral blood, and extended with m1A and 2′-O-methyladenosine to demonstrate retraining.

#Applications

ModiDeC targets epitranscriptomics groups asking focused questions: quantifying a specific modification at a specific motif across conditions, validating engineered RNA-modifying systems, or profiling modifications in clinical material such as peripheral blood. Its retraining path suits motifs no general-purpose caller covers, and its false-positive characterization matters where a call drives downstream experiments. The Epi2ME pipelines put the whole workflow inside Oxford Nanopore's platform; source code, GUIs, and tutorials are on GitHub with a Zenodo archive, and sequencing data sits at the ENA under PRJEB88778.

#Impact

ModiDeC contributes two things the epitranscriptomics field has been short of: a labeled synthetic benchmark spanning both nanopore RNA chemistries, and a classifier designed to be extended by users rather than frozen at publication. The accompanying analysis of which motif-modification combinations are learned easily and which are not — some remain difficult even with 30,000 training reads — is a candid account of where signal-based detection succeeds. Its scope is narrow: performance is characterized on the 40 trained motifs, frequency quantification is systematically low, and extending coverage requires generating new labeled data. It is best understood as a customizable instrument for targeted questions rather than a transcriptome-wide survey.