Nanopore direct RNA sequencing (DRS) lets scientists read native RNA molecules directly, preserving important features such as chemical modifications and avoiding biases introduced by copying RNA into DNA. This makes it a powerful tool for studying the transcriptome, the full set of RNA molecules in a cell. However, one long-standing challenge has been a relatively high error rate when converting the raw electrical signal from the nanopore into readable RNA sequence, a process called basecalling.

Researchers at Shenzhen University have developed a new basecaller named Coral that significantly improves the accuracy of nanopore DRS. Coral uses a modern deep learning architecture with a Transformer-based encoder-decoder design. This allows it to learn contextual information not only from the raw signal data but also from the sequence context itself, giving it a more complete understanding of how RNA signals correspond to real sequences.

Overview of Coral framework. a, Workflow of basecalling based on Coral

RNA transcripts are first sequenced using nanopore DRS to generate raw signal data. Then the signals are normalized and converted into temporal features by the signal encoder. Finally, the decoder uses these temporal features and the contextual information of the decoded k-mers to predict the next k-mer. b, Framework of encoder. The signal encoder extracts temporal features through several CNN and RNN layers. c, Framework of decoder. The decoder receives the initial symbol y0 = representing the start of the sequence. The gray box illustrates the process for decoding the first k-mer. The range of temporal features needed for current k-mer decoding is determined by the alignment position of the previously decoded k-mer. d, Likelihood estimation in decoding a k-mer (gray box in c, taking the first k-mer as an example). Signal temporal features are firstly concatenated with k-mer contextual features, and then the linear layer with softmax activation is applied to obtain predicted probabilities for 1024 possible k-mers (45 , for 5-mers) at all valid alignment positions. The alignment probabilities are determined through vector dot-product and the sigmoid function. Finally, the k-mer prediction probability and alignment probability are multiplied element-wise to estimate the k-mer likelihood.

When tested on human RNA samples, Coral outperformed the existing Oxford Nanopore Technologies basecaller called Dorado, boosting basecalling accuracy by up to 6.17%. That increase may sound modest, but it has meaningful downstream benefits. Coral’s improved accuracy enables detection of about 26% more annotated transcript isoforms, which are different versions of RNA produced from the same gene. Identifying these isoforms correctly is crucial for understanding gene regulation, alternative splicing, and cellular function.

Coral also enhances haplotype phasing, the process of linking multiple genetic variants on the same RNA molecule. Thanks to its accuracy improvements, Coral reduces errors in phasing and can phase more single nucleotide variants than before. This makes it much easier to track how different versions of RNA relate to genetic differences between individuals.

Overall, Coral represents a significant step forward in making nanopore direct RNA sequencing more reliable and informative. As nanopore technologies continue to mature, tools like Coral will help researchers unlock more of the transcriptome’s complexity with confidence.

Availability – The code package for this study, along with a guide to use the code, is available on GitHub at [https://github.com/BioinfoSZU/Coral]