RNA molecules carry more information than just their sequence. Many RNAs contain chemical modifications, often called the epitranscriptome, that influence how RNA is processed, translated, and regulated. Mapping these modifications across the transcriptome is essential for understanding development, stress responses, and disease, but doing so at scale has been technically difficult.

Researchers at the University of Bari Aldo Moro, Italy describe a new computational framework designed to overcome key limitations in direct RNA sequencing with Oxford Nanopore technology. Direct RNA sequencing can, in principle, detect RNA modifications because modified bases alter the electrical signal as RNA passes through the nanopore. In practice, most existing tools rely on software that is unaware of modifications and require multiple separate training steps, each tailored to a single modification type.

The researchers introduce two complementary tools, NanoSpeech and NanoListener. NanoSpeech is a modification-aware basecaller built on a transformer model that can simultaneously detect multiple modified and unmodified RNA bases from raw nanopore signals. Instead of treating each modification separately, the model uses an expanded base vocabulary to recognize different chemical variants directly during basecalling.

Overview of NanoListener and NanoSpeech

(A) Raw signals in fast5/pod5 files are basecalled via a modification-unaware tool and mapped against reference sequences. Alignments are filtered, and ionic current signals are split into events and re-squiggled onto the reference using f5c eventalign to obtain their precise mapping position. (B) NanoListener uses re-squiggled signals to retrieve context information from eventalign tables. It extracts random chunks of electric measurements from fast5, trying to avoid perturbed alignment regions (shown in red) due to modified bases. For each read, a variable number of pA-scaled current chunks is dynamically extracted, along with their associated k-mer annotated for modified bases, when needed. A padding is added to make both of them uniform in length. (C) Providing a set of bona fide modified sites, NanoListener extracts positional and per-read information from various sources, such as basecalling features of IVT or KO samples, which are used to train an isolation Forest model. Such iForest model was used to identify kmers with modified bases and their positions, creating a training dataset that embeds this metadata and feeds it to NanoListener. (D) Annotated, balanced and filtered NanoListener datasets were used to train NanoSpeech models with a user-preferred vocabulary. The current chunks were pre-processed via STFT into padded spectrograms, and several NanoSpeech models, through a classical encoder-decoder transformer, were trained to predict annotated output kmers for every input spectrogram. In inference mode, NanoSpeech decodes per-base probabilities into nucleotide sequences and prints out Fasta/Fastq files. In downstream procedures, position mapping with a per-read resolution of all the basecalled modified bases can be easily retrieved using accessory NanoSpeech scripts and aggregated to the genome-space level.

NanoListener supports this approach by generating robust training datasets using a simulated randomers strategy. This makes it possible to train accurate models even when experimental datasets are limited. Importantly, both tools are independent of specific nanopore chemistries, meaning the same framework can be reused as sequencing platforms evolve.

Once trained, a single NanoSpeech model can accurately read standard RNA sequences and detect multiple RNA modifications at the same time. This simplifies analysis workflows and reduces the need for repeated retraining or modification-specific pipelines.

Together, these advances move direct RNA sequencing closer to routine, transcriptome-wide analysis of RNA modifications. By enabling ab initio detection of multiple epitranscriptomic marks in a single run, the approach provides a powerful foundation for studying how RNA chemistry shapes gene regulation in complex biological systems.

Availability – NanoListener and NanoSpeech, along with models and auxiliary scripts, are available at the following GitHub pages https://github.com/F0nz0/NanoListener and https://github.com/F0nz0/NanoSpeech_basecaller, respectively.