Exploiting the language of the transcriptome for direct RNA sequencing

Abstract

Nanopore direct RNA sequencing (DRS) enables the measurement of full-length RNA molecules in their native state at single-molecule resolution. The raw electric current signals recorded during DRS open a new modality for interrogating RNA biology. Software that analyses RNA in DRS signal space is emerging, yet decoding DRS signals remains a challenging computational problem. One unexplored aspect of DRS signal analysis is the utilisation of domain knowledge about the "language" of RNA as an additional source of information. Since RNA is always sequenced in the 3' to 5' direction of the molecule, the nucleotide and biochemical modification patterns that constitute RNA language get implicitly encoded in the same order in each DRS signal. The aim of this PhD thesis in computational RNA biology was to investigate whether the interpretation of RNA molecules from DRS signals can be enhanced by exploiting the language of RNA.

The first objective was to explore whether RNA language could be exploited as an additional input to DRS basecalling, which remains an error-prone task. The first language- informed basecalling software architecture was developed, which incorporates a probabilistic model of messenger RNA (mRNA) language, a modified sequence-to-sequence neural network decoding algorithm and a time-efficient approach to decode variable length nanopore signals. The utilisation of mRNA language provided a modest benefit to basecalling accuracy and was able to guide DRS signal decoding toward the correct nucleotide when the signal on its own could not be confidently assigned to a nucleotide.

By exploiting the differences in language between different classes of RNA, the second objective was to develop a model that can accurately and efficiently distinguish between RNA classes directly from DRS signals. It was found that it is possible to discriminate between protein-coding and non-coding RNA molecules from just the first four seconds of DRS signals, which corresponds to roughly the first 280 nucleotides at the 3' end of RNA transcripts. Multiple convolutional neural network architectures were systematically optimised and evaluated with the dual objectives of high accuracy and speed, before the final selected model was evaluated on a test set from an independent cell line.

The third objective was to further exploit the implicit encoding of RNA language in DRS signals for the development of a novel targeted sequencing technology. The RNA class model was integrated with the ONT sequencing hardware to enable the real-time rejection of off-target molecules per nanopore, thus conserving sequencing time for the target RNA class. The biochemical-free enrichment and depletion of protein-coding and non-coding RNAs was then demonstrated in both a simulated sequencing environment and in real-time during live sequencing of multiple cell lines. This research outcome was a major advance in RNA sequencing, since DRS sequencing capacity is limited by the working time of the nanopores, thus sequencing time gets wasted on unwanted RNAs. Although biochemical treatments can filter the sequencing library to target sequencing towards selected RNAs, this is known to compromise RNA integrity and bias the resultant reads, besides being resource-intensive.

This PhD thesis establishes that RNA language is a valuable input to DRS signal analysis. Explicitly modelling RNA language to integrate with DRS signal decoding is a promising new avenue of research, while the implicit encoding of RNA language in DRS signals can also be exploited for novel signal analyses. Importantly, the biochemical-free targeted sequencing technology developed in this study opens up a broad range of possible applications for the in silico real-time analysis of RNA, to empower RNA researchers with biochemical-free, real-time enrichment or depletion of classes of RNA molecules. Show more