WattmaMod

Chemical modifications of RNA—m6A, m5C, pseudouridine, and dozens of others that together form the epitranscriptome—regulate splicing, translation, stability, and localization, yet mapping them transcriptome-wide remains difficult. Nanopore direct RNA sequencing is uniquely suited to the problem because it reads native RNA and preserves modifications as subtle perturbations in the raw ionic-current signal as each molecule translocates through the pore. Turning those perturbations into accurate, per-site modification calls is hard: most computational tools target a single modification type, depend on large labeled training sets, and generalize poorly across sequencing chemistries and organisms.

WattmaMod is a deep-learning framework that addresses these limitations with a pretrain-then-adapt design. It first learns general representations of nanopore signal through self-supervised pretraining on event-level features, then specializes through supervised contrastive fine-tuning, and finally uses low-label incremental adaptation to extend to additional modification types with only minimal labeled data. Developed by Bin Yu, Xinghui Sun, Xiao Li, Junhai Qi, Ting Yu, Xin Gao, and Renmin Han across Shandong University, Qingdao University of Science and Technology, the University of Science and Technology of China, and King Abdullah University of Science and Technology, it was posted to bioRxiv in July 2026 under a CC BY 4.0 license.

What distinguishes WattmaMod is the combination of a broadly reusable signal backbone with an extensible detection head: a single pretrained model resolves eleven modification types and can be adapted to new, low-resource modifications rather than being retrained from scratch for each target. The authors report that the framework generalizes across sequencing chemistries and species and can reveal patterns of coordinated modification along individual RNA molecules.

#Key Features

• Self-supervised signal pretraining: Learns transferable representations directly from raw, event-level nanopore current before any modification labels are introduced, reducing dependence on large annotated datasets.
• Eleven modifications in one model: Detects m6A, m5C, m1A, A-to-I, m7G, hm5C, m1Ψ, f5C, ac4C, m5U, and Ψ within a single unified framework.
• Low-label extensibility: Incremental adaptation lets the model extend to new, low-resource modification types using only a small amount of labeled data.
• Wavelet-guided multi-scale encoding: A wavelet-guided encoder captures signal structure across multiple temporal scales, matching the varied kinetic footprints of different modifications.
• Dynamic cross-attention fusion: Combines complementary signal and sequence-context features so that evidence is weighted adaptively per site.
• Cross-chemistry and cross-species generalization: Operates across different nanopore chemistries and organisms rather than being tied to a single dataset.

#Technical Details

WattmaMod processes event-level features derived from nanopore direct RNA sequencing through a wavelet-guided multi-scale encoder and fuses representations with a dynamic cross-attention mechanism. Training proceeds in stages: self-supervised pretraining establishes a general signal representation; supervised contrastive fine-tuning sharpens the separation between modified and unmodified sites; and low-label incremental adaptation extends the model to additional modification types without full retraining. This staged recipe is what supports the model's coverage of eleven modification types and its extensibility to low-resource targets. The preprint is released under CC BY 4.0.

#Applications

WattmaMod is aimed at researchers studying the epitranscriptome—RNA-modification biology, gene-regulatory mechanisms, and disease processes in which aberrant modification plays a role. Because it reads native RNA rather than amplified cDNA, it suits workflows that must preserve labile modifications, and its extensibility makes it attractive for surveying rare or newly characterized modification types without assembling a large labeled dataset for each. Its reported ability to resolve coordinated modifications along single molecules supports investigations into how multiple marks co-occur and interact on the same transcript.

#Impact

WattmaMod illustrates how self-supervised pretraining on raw nanopore signal, paired with contrastive fine-tuning and low-label adaptation, can turn direct RNA sequencing into a broad, extensible epitranscriptome assay rather than a set of single-modification detectors. As a July 2026 bioRxiv preprint, it awaits peer review, and its performance rests on the authors' own reported results without independent benchmarking or community validation. No public code repository, model weights, or hosted inference package has been released alongside the preprint, so reproduction and downstream adoption currently depend on materials not yet available.