A mapping-free natural language processing-based technique for sequence search in nanopore long-reads.

Background: In unforeseen situations, such as nuclear power plant's or civilian radiation accidents, there is a need for effective and computationally inexpensive methods to determine the expression level of a selected gene panel, allowing for rough dose estimates in thousands of donors. The new generation in-situ mapper, fast and of low energy consumption, working at the level of single nanopore output, is in demand. We aim to create a sequence identification tool that utilizes natural language processing techniques and ensures a high level of negative predictive value (NPV) compared to the classical approach.

Results: The training dataset consisted of RNA sequencing data from 6 samples. Multiple natural language processing models were examined, differing in the type of dictionary components (word length, step, context) as well as the encoding length and number of sequences required for algorithm training. The best configuration analyses the entire sequence and uses a word length of 3 base pairs with one-word neighbor on each side. For the considered FDXR gene, the achieved mean balanced accuracy (BACC) was 98.29% and NPV was 99.25%, compared to minimap2's performance in a cross-validation scenario. The next stage focused on exploring the dictionary components and attempting to optimize it, employing statistical techniques as well as those relying on the explainability of the decisions made. Reducing the dictionary from 1024 to 145 changed BACC to 96.49% and the NPV to 98.15%. Obtained model, validated on an external independent genome sequencing dataset, gave NPV of 99.64% for complete and 95.87% for reduced dictionary. The salmon-estimated read counts differed from the classical approach on average by 3.48% for the complete dictionary and by 5.82% for the reduced one.

Conclusions: We conclude that for long Oxford nanopore reads, a natural language processing-based approach can reliably replace classical mapping when there is a need for fast, reliable and energy and computationally efficient targeted mapping of a pre-defined subset of transcripts. The developed model can be easily retrained to identify selected transcripts and/or work with various long-read sequencing techniques. Our results of the study clearly demonstrate the potential of applying techniques known from classical text processing to nucleotide sequences.