DNA basecaller for direct RNA sequencing and uses thereof in nanopores

Abstract

Disclosed is a method of training a complete DNA basecaller using high complex direct RNA sequencing (DRS) training datasets and applications of the trained complete DNA basecaller. Including a method to sequence single-cell RNA using capture oligonucleotides and oligonucleotides-DNA barcodes. The training of the DNA basecaller use a linked DNA-RNA oligonucleotide, in which the DNA part is complementary to the RNA part.

Description

[0001] DNA BASECALLER FOR DIRECT RNA SEQUENCING AND USES THEREOF IN NANOPORES

[0002] Technical Field

[0003] Provided is a computer implemented method to train a DNA basecaller for direct RNA sequencing (DRS) or to identify nucleotide sequences sequenced by direct RNA sequencing (DRS)

[0004] Background

[0005] Nanopore sequencing technologies have revolutionised our ability to study the transcriptome, by enabling direct sequencing of the native RNA molecules, offering insights into gene expression profiles while retaining the RNA modification and polyA tail length. Nanopore sequencing offers key advantages over traditional sequencing methods by obviating the need for PCR amplification and cDNA synthesis steps, thus preserving the native state of RNA molecules and providing a less-biased and comprehensive view of transcriptomes. In the process of library preparation, RNA molecules are extended twice: first with a DNA oligonucleotide (with an annealed oligonucleotide complementary to poly-A tail of mRNA), and subsequently with a DNA adapter (with a part complementary to the first DNA adapter) carrying a helicase needed for translocating the RNA through the nanopore one nucleotide at a time. The library preparation protocol further includes a step of complementary DNA strand (cDNA) synthesis in order to linearise RNA molecules, which greatly reduces pore blocking by folded RNA molecules. However only the RNA strand ligated to the DNA adapters of the RNA::cDNA duplex is sequenced during DRS. Moreover, only the RNA part of the RNA strand is analyzed: DNA barcode and DNA adapter are trimmed before or during basecalling. As of today, DRS is only possible on nanopore platform bringing a significant advantage to the RNA sequencing. However, there is still room to improve its accuracy and widen its implementation in different application addressing different needs. Here, we provide an advanced novel method, a novel solution to address these short comings of the platform, increasing DRS accuracy, and enabling it to be used in areas that it was not possible before.

[0006] Some of the key challenges persisting regarding direct RNA sequencing (DRS) comprises, basecalling of the DNA regions of the prepared RNA library, enabling barcoding to pool multiple RNA (e.g. low input) samples (also known as multiplexing) and demultiplexing sequencing data obtained from pooled samples. To date, there is no method or algorithm that can analyze the DNA that is sequenced by nanopore DRS or a commercial barcoding kit, technique or demultiplexing method, algorithm specific to DRS data. To overcome the latter, demultiplexing related limitations, community-driven efforts have been made, by including a shuffled set of nucleotides in the annealed sequence of the first DNA adapter. This sequence or “barcode” can then be demultiplexed by using community-developed algorithms such as DeePlexiCon (Smith, M. A. et al. Molecular barcoding of native RNAs using nanopore sequencing and deep learning. Genome Res.30, 1345–1353 (2020)) EP22382917 discloses the method of Smith et al, adapted to tRNA, the disclosed methods between the two document differs for example in the approach used for the identification of the barcode in sequenced molecules. In this document, an alignment-guided segmentation is used: signal position of the last aligned base of read. EP22382917 includes the entire poly-A tail signal, thus it works well only for protocols with fixed and short poly-A tails (Lucas, M. C. et al. Quantitative analysis of tRNA abundance and modifications by nanopore RNA sequencing. Nat. Biotechnol.42, 72–86 (2023)). The document SeqTagger, a rapid and accurate tool to demultiplex direct RNA nanopore sequencing datasets Leszek P Pryszcz, Gregor Diensthuber, Laia Llovera, Rebeca Medina, Anna Delgado-Tejedor, Luca Cozzuto, Julia Ponomarenko, Eva Maria Novoa bioRxiv 2024.10.29.620808; doi: https: / / doi.org / 10.1101 / 2024.10.29.620808 (Pryszcz et al, 2024) discloses SeqTagger, which is a highly accurate demultiplexing software, that achieves 99% precision with 95% recall, while also being ~9x faster than current demultiplexing software. SeqTagger has the following disadvantages: 1) It was trained with a limited training set in terms of both: sequence (fixed barcode sequences and lengths) and signals (chunks correspond to limited population of signals that start somewhere in the DNA adapter and end in the beginning of poly-A tail of RNA). SeqTagger is not able to classify more barcodes than the number of barcodes used for the training and this number cannot be above 100. 2) Training set consists of signal fragments of fixed length (chunks, for example 2,000 readouts per chunk) and corresponding sequence. In order to train SeqTagger demultiplexing model, a set of 100 DNA barcodes were ligated (replacing official RT adapter) to a set of 12 RNA molecules and sequenced in several batches (up to 12 barcodes at once). This approach is cost- and time-inefficient, due to custom: DNA barcode design and synthesis, library preparation for multiple sets of barcodes and numerous sequencing experiments (plus many other unforeseen challenges due to much larger barcode part than expected such as MinKNOW not recognising the reads, basecaller wrongly segmenting and basecalling them, etc). Such a limited training set is not complex enough to train a basecaller which is able to basecall any, arbitrary DNA sequence but limited to provide solutions for analysis of a limited set of short DNA fragments / barcodes that it's trained with. There are at least three problems that need to be addressed to enable to basecall arbitrary DNA sequenced by DRS: i) generating a DRS training set with high sequence complexity, ii) developing a method for aligning DNA signal-to-sequence for DRS so we can generate signal-to-sequence alignment, this is, signal-only segmentation, and iii) optimising neural network architecture for basecalling DNA signals from DRS. We should note that the need for training a complete DNA basecalling model rather than using existing trained DNA basecalling models (such as those produced by Haotian Teng, Minh Duc Cao, Michael B Hall, Tania Duarte, Sheng Wang, Lachlan J M Coin, Chiron: translating nanopore raw signal directly into nucleotide sequence using deep learning, GigaScience, Volume 7, Issue 5, May 2018, giy037, or Silvestre-Ryan, J., Holmes, I. Pair consensus decoding improves accuracy of neural network basecallers for nanopore sequencing. Genome Biol 22, 38 (2021)) arises from several key differences between DNA and RNA sequencing chemistries, which do not permit using existing DNA basecalling models on native RNA sequencing datasets: i) RNA is sequenced from its 3' end (in 3'>5' direction), while DNA is sequenced from its 5' end (in 5'>3' direction); ii) the RNA helicase is slower than the DNA helicase (70 bps for RNA002 and 130 bps for RNA004, compared to 400-450 bps for DNA, R9.4 and R10); and iii) RNA molecules are sampled with different frequency than DNA molecules (3-4kHz for RNA, compared to 4-5kHz for DNA, depending on the flowcell) MinKNOW samples RNA sequencing (3kHz for RNA002 or 4kHz for RNA004. Thus, the same DNA will generate different signals when sequenced with DNA and RNA kits, making existing DNA basecalling models unusable for basecalling of DNA fragments in DRS datasets. The invention discloses how to train a new complete DNA basecalling model for direct RNA sequencing. The method of the invention disclosed here solves the problems discussed above and provides a novel method for decoding (basecalling) DNA signals from DRS. The method further can be used in different applications, e.g. not limited to but including demultiplexing of complex RNA DRS data (for example single-cell RNA sequencing) or other applications such as improving RNA basecalling or modification detection by decoding both strands for RNA:: DNA duplex. Importantly, the DNA sequence may be originated from a sequence having an arbitrary length and composition. Description In the present specification basecalling is the process of assigning nucleotides to electrical current changes resulting from nucleotides passing through a nanopore. In the present specification the term “oligonucleotide” refers to RNA, DNA, RNA::DNA or RNA::cDNA oligonucleotides. The term RNA::cDNA molecule refers to RNA and its complementary DNA sequence. RNA::cDNA molecule can be a double-stranded hybrid, where the RNA strand is paired or bound to its complementary cDNA strand or RNA::cDNA molecule can be a single stranded where RNA is not paired or bound to its cDNA. The RNA strand can be linked to the cDNA strand in one end (5´ end or 3´ end) via ligation or similar. In the present specification RNA::cDNA molecule, RNA::cDNA fragment and RNA::cDNA chimera are synonyms. The term RNA::DNA refers to RNA and a DNA sequence, which may or may not be synthesized from an mRNA template through the process of reverse transcription (cDNA). Unless other way specified, all the characteristics, features, descriptions and embodiments disclosed in the present specification related to the molecule RNA::cDNA are applicable to RNA::DNA. In embodiments, DNA is complementary to RNA sequence in the RNA::DNA molecule or DNA is non-complementary to RNA sequence in the RNA::DNA molecule. In embodiments, DNA is a natural DNA or synthetically synthesized DNA or cDNA. In embodiments DNA can be cDNA which is synthesized from an mRNA template through the process of reverse transcription. In the present specification the term “nucleotide” can refer to both, ribonucleotide or deoxyribonucleotide, unless otherwise explained. The term "nanopore" is intended to mean a small hole that allows passage of nucleotide triphosphates across an otherwise impermeable barrier, this is pore-based single-molecule sequencing. The barrier is typically an electrically insulating layer and the nanopore typically permits ions to flow from one side of the barrier to the other, driven by an applied potential. The nanopore preferably permits nucleotides to flow through the nanopore lumen along the applied potential. The nanopore may also allow a nucleic acid, such as DNA or RNA, to be pushed or pulled through the lumen of the nanopore. However, in particular embodiments the nanopore need not allow passage of a double stranded or single stranded nucleic acid. A nanopore used in a particular embodiment can have a minimum lumen diameter of no more than 10 nm, 5 nm, 4 nm, 3 nm, 2 nm, 1 nm, 0.5 nm or less. One type of nanopore is a "protein nanopore" which is a polypeptide or a collection of polypeptides that forms the small hole to allow passage of nucleotide triphosphates across a barrier such as a lipid bilayer. Examples of protein nanopores include alpha hemolysin nanopore, mycobacterium smegmatis porin A (MspA) and variants thereof. Another type of nanopore is a "solid state nanopore" which is a small hole fabricated through a solid material. The solid material can be, for example, graphene or silicon. In the present specification, “Split-pool” or “Split-pool barcoding) is a single-cell analysis method that identifies individual cells by assigning them a unique combination of barcodes in successive rounds of splitting and pooling, such as SPLiT-seq. It is a technique known in the field. Concatenated-barcodes is a result of split-pool barcoding, meaning a set of, for example, 3 barcode sequences concatenated together (if 3 rounds of split-pool were perfomerd) As it will be shown later, to obtain the DNA basecaller for DRS of the invention the present invention discloses a DRS training dataset obtained from a biological sample (such as mRNA isolated from human or mouse cell-line). The first object of the invention relates to a method to obtain a Direct RNA Sequencing (DRS) training dataset for training a DNA basecaller carried out by a computer, comprising the following steps: a) providing a Direct RNA Sequencing data obtained from a sample, and b) establishing a correlation of each signal of the DRS data to a nucleotide by signal- segmentation to obtain a DRS training datasets. Another object of the invention relates to a method to obtain a trained a DNA basecaller for DRS carried out by a computer wherein the method comprises the use of the DRS training datasets previously disclosed. In a particular embodiment the sample is selected from DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA chimera. In another particular embodiment the DNA of RNA::DNA molecule is complementary to the RNA of the RNA::DNA molecule. In another particular embodiment the DNA of RNA::DNA molecule is a cDNA, in other embodiments DNA is synthetically synthesized DNA. In another particular embodiment, the DNA and RNA of the RNA::DNA molecule is linked. In a preferred embodiment, the sample is linked RNA::cDNA. Therefore, step a) of the method of the invention would be providing an DRS training dataset obtained from a sample in the form of RNA::cDNA molecules linked, the cDNA being complementary to the RNA. Another object of the invention relates to a method to train a DNA basecaller for Direct RNA Sequencing (DRS), carried out by a computer, comprising the following steps: a) providing a Direct RNA Sequencing data obtained from a sample, b) establishing a correlation of each signal of the Direct RNA Sequencing data to a nucleotide by signal-segmentation to obtain a DRS training dataset, and c) training the DNA basecaller for Direct RNA Sequencing (DRS) using the DRS training datasets of step b) In a particular embodiment the sample is selected from DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA chimera. In another particular embodiment the DNA of RNA::DNA molecule is complementary to the RNA of the RNA::DNA molecule. In another particular embodiment the DNA of RNA::DNA molecule is a cDNA, in other embodiments DNA is synthetically synthesized DNA In another particular embodiment, the DNA and RNA of the RNA::DNA molecule is linked. In a preferred embodiment, the sample is linked RNA::cDNA. Therefore, step a) of the method of the invention would be providing an DRS training dataset obtained from a sample in the form of RNA::cDNA molecules linked, the cDNA being complementary to the RNA and step b) would be to establish a correlation of each cDNA signal of the Direct RNA Sequencing data to a nucleotide by signal-segmentation to obtain a DRS training dataset. In embodiments, the method further comprises a step for optimising neural network architecture for basecalling DNA signals from DRS. The advantage of this optional step is to improve the accuracy of the trained DNA basecaller In embodiments, the method for training a DNA basecaller for DRS further includes a prior step to step a, wherein the prior step comprises preparation of DRS training library, said library being of DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA chimera, linked or not, preferably linked RNA::cDNA. Step a) providing an DRS training dataset obtained from a sample: Generating a DRS training set with high sequence complexity In embodiments the RNA molecules used to obtain the DRS training dataset is randomly generated RNA. In another embodiment, it can be artificially synthetized. In preferred embodiments, the RNA is from biological sample. In preferred embodiments, RNA is fragmented RNA (sheared), fragmentated RNA is preferably polyadenylated following to fragmentation. The fragmented RNA is mRNA such as transcriptome, or any other type of RNA isolated from a biological sample that is fragmented using methods that are known to the skilled person in the field. The RNA is be polyadenylated prior and / or post fragmentation. In embodiments, once the RNA is prepared it is reverse transcribed (RT) to generate a complementary cDNA. RT step is common for DRS library preparation. The inventors, in the embodiments of this method, linked the RNA to its complementary cDNA to allow cDNA strand to be pulled though the pore during the sequencing process following to the RNA that is linked to. Thus, allowing cDNA to be sequenced after RNA part that passed the pore which is a novel feature for nanopore sequencing techniques and protocols. The training library of this method is prepared to be sequenced using the technology of nanopore direct RNA sequencing. The steps required for preparing RNA libraries for DRS are known to the skilled person and kits and products are commercially available. Although may not been specified here, these steps are fully or partially included in the training library preparation of this method regardless of whether the library includes or excludes the linking the cDNA molecules to the RNA. Such known steps of DRS library preparation include but not limited to extension of RNA molecules with a DNA barcode (with a part complementary to poly- A tail of mRNA), and subsequently with a DNA adapter (with a part complementary to DNA barcode) carrying a helicase needed for sequencing). Further details of library preparation are given in the examples provided within this application. The RNA can be synthetic or from a biological sample and selected from the group consisting of, messenger RNA (mRNA), ribosomal RNA (rRNA), transfer RNA (tRNA), long non-coding RNA (lncRNA), small nuclear RNA (snRNA), microRNA (miRNA), small interfering RNA (siRNA) and antisense RNA (aRNA), preferably messenger RNA (mRNA), and / or transfer RNA (tRNA). The DNA can be, synthetic or natural DNA, genomic DNA, cDNA, single or double stranded DNA. In a particular embodiment, the RNA sample obtained from a biological sample comprises a tail of polynucleotides, this tail can be naturally-occurring such as polyadenosine tail in mRNA or artificially added. This can be any combination of nucleotides that produce homogenous signal ie polyA (polyadenosine) polyC, polyG, polyU, or polyG / I, polyinosine (polyI) or 2’-O-methylated (polyNm) tails, the tail length ranges from 10 to 100 nucleotides. In embodiments the polyA tail or alternatives as described above can have between 1 to 100 nucleotides. Such as, 1 to 90, 2 to 80, 3 to 70, 4 to 60, 5 to 50, 5 to 40, 5 to 30, 5 to 20, 5 to 10 nucleotides, preferably nucleotide is an A. The polyI and polyNm tails can be added with yeast PAP enzyme or E.coli PAP enzymes. In another particular embodiment, the RNA has no tail but instead has a region that anneals with the sequence itself, for instance the sequence of the oligonucleotides uses for sequencing The sequencing datasets obtained from an RNA of a biological sample used in the methods described herein may be from any source including, human beings, animals, plants, bacteria and fungi or yeast. For example, a body fluid (blood or plasma), tissue sample, organ, organelle, or single cells obtained using methods known in the art. In preferred embodiments, the RNA is from biological sample, for example a living organism or material derived thereof. In embodiments, RNA is fragmented transcriptome of a cell. The cell can be, a human cell line, mouse cell line bacterial or plant cells or cell lines. RNA can be obtained from a tissue of an animal including human or from plant tissue. In a particular embodiment the DRS training dataset is obtained from a biological sample or from artificially generated, such as synthesized RNA sequences.in embodiments, the RNA::cDNA molecules of the DRS training library comprises one or more barcodes. However, the use of barcodes is not recommended as the use of designed barcodes is not practical in order to obtain a DRS training dataset with high sequence complexity. The DRS training dataset needs to be complex enough so the complete DNA basecaller is trained to be able to basecall any arbitrary DNA sequence which than can be used for a wider range of applications. Preferably, when the RNA molecules used to obtain the DRS training dataset are randomly generated RNA, this is, artificially synthetized RNA. In the present specification “barcodes” is a string of polynucleotides, RNA or DNA that attached to a molecules / fragment of RNA to allow its later identification, preferably DNA fragments. In particular embodiment the Direct RNA Sequencing training dataset is obtained from DRS runs, RNA nanopore sequencing runs, or from native RNA nanopore sequencing runs. The nanopore direct sequencing and the materials and protocols to Direct RNA Sequencing training dataset are known in the art. For instance, in EP0815438B1. In a particular embodiment the nanopore direct sequencing comprises a membrane, said membrane can be either solid-state or biological membranes. Any known nanopore direct sequencing method or product can be used to obtain the DRS suitable for the method of the present invention, for instance the one disclosed in EP0815438B1 or EP1238275B1. In embodiments the samples to obtain the training dataset are obtained by synthetizing cDNA strand of the RNA molecule and linking the RNA molecule to the synthesized cDNA molecule. The linking of RNA to its cDNA allows RNA and cDNA to pass through the sequencing pore and hence sequenced together. According to the method disclosed, the RNA part of the RNA::cDNA molecule forms ground- truth sequence for the cDNA part of the training input sample (cDNA is simply reverse complement of RNA). This enables correct decoding of the DNA sequence. In embodiments, when random RNA molecule is sequenced, complementary DNA sequence of each read is decoded based on the basecalling of RNA. In embodiments, where RNA is fragmented transcriptome each sequenced cDNA and RNA read sequence is aligned to the reference transcript and known the underlying sequence with high confidence. In embodiments of this disclosed method, biological material can be used as a starting point of generation of DRS training data set. Biological material or biological sample means that there is no barcode design added to the training data set. Therefore, training sets can be generated using just one round of library preparation and sequencing, which is cost- and time- effective. The linking of the DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA, preferably RNA::cDNA, can be either at the 3´ end or 5´ end of the RNA molecule. The DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA, preferably RNA::cDNA molecule can be double stranded (duplex) or single stranded (singlex). Preferably, the cDNA is linked via its 3’end to the 5´ end of the RNA molecule. In other embodiments, 5’ end of the cDNA is linked to the 3’ end of the RNA. Preferably, the RNA::cDNA molecule is double stranded and when passing through the sequencing pore it becomes single stranded. The DRS library preparation protocol can include a step of complementary DNA strand (cDNA) synthesis in order to linearise RNA molecules, which greatly reduces pore blocking by folded RNA molecules. In embodiments this cDNA in combination with its corresponding RNA counterpart can be used to generate the training data set. It is preferred that for every RNA::cDNA fragment, the cDNA is complementary to the linked RNA. It is important for the invention that both RNA and its cDNA is sequenced together. The inventors for the first time offer a solution to enable that by linking the RNA to its cDNA and using the linked RNA::cDNA molecule as the DRS libraries, both for training libraries and sample libraries to be analyzed. In order to generate a suitable DRS training dataset the cDNA and RNA strands of RNA::cDNA molecule share linked at the 3´ end or 5´end. The connection of both molecules is preferably a covalent link such as by using either a dedicated ligase (such as CircLigase, RtcB or anything else that connects RNA and cDNA strand of RNA::cDNA duplex) or by ligating a DNA-RNA hairpin, and sequence both strands simultaneously (similar to 2D protocol for DNA, Figure 6). Any methods mentioned or not mentioned herein but known to a skilled person for linking, joining, or connecting the end of an RNA strand to the end of a DNA strand may be used in the disclosed method to generate a link between RNA and cDNA strands for creating the training library. Disclosed method is enabled by the sequencing of cDNA along with RNA molecule in DRS. Alternatively, the methods allowing or improving the possibility of cDNA molecule passing through the sequencing pore after the RNA strand thus enabling its sequencing can be used to make the DRS library and generating the dataset to be used in this method. Not limiting but one example can be using adaptors ligated to both RNA and cDNA strand to increase the possibly of cDNA to pass through the pore. The length of the RNA::cDNA fragment / molecule can vary. In some embodiments, the length is between 100-10,000 nucleotides, preferably between 300 and 1,200 nucleotides. In embodiments, the length of RNA::cDNA ranges between 100 and 9000, 100 and 8000, 100 and 7000, 100 and 6000, 100 and 10000, 100 and 4000, 100 and 3000, 100 and 2000, between 100 and 1000, between 200 and 3000, between 200 and 2000, between 300 and 2000 or, between 300 and 1500, or between 300 and 1200 nucleotides. In the present specification sequence complexity means the number of different sequences used in the method of training the DNA basecaller for DRS. For example, the maximum complexity of SeqTagger is N =100, as a maximum of 100 barcodes, were used. However, as it will be shown later on, the method of the invention allows much higher complexity. For example, if the DNA basecaller is trained with a whole transcriptome, depending how it is processed. In a particular embodiment the complexity will be N>10000; in the case where only chunks from the beginning of the transcript are taken, or possibly N>100000 when the transcript is chunked into smaller pieces. In addition, the present specification, longer signal-sequence pairs are used for training (200- 400 bases) than in training models utilizing barcodes (20-37 bases), which additionally increases the sequence complexity of the training set 10 times. The terms Signal-sequence pairs or signal chunk-sequence pairs are used interchangeably and refers to signal chunks of obtained DRS signal corresponding to a specific sequence in DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA, preferably RNA::cDNA molecule. By using the DRS training dataset obtained from biological sample to train a DNA basecaller for DRS, said basecaller will be able to basecall any, arbitrary DNA sequence, without a limitation to the preselected set of barcodes that it's trained with. This is because of the difference in the training procedure of the basecalling model. In brief, the complexity of the training set built for methods trained based on DNA barcodes is too small in order (number of templates used for training, N=4, 12, 96 or 100, respectively) to allow the model to learn to basecall correctly any arbitrary DNA sequence. Because of that, we have to use dedicated laboratory and computational methods that will generate training sets with higher complexity (covering a much wider spectrum of DNA sequences than limited or preset number of barcodes, ideally in the range of at least N=10,000-30,000 templates, preferably with length between 200-400 bases can be used for training). When the DRS training dataset is obtained from a biological sample barcodes are not required because the underlying sequence can be identified as most organisms have a published reference transcript. Any ribonucleotide or deoxyribonucleotide sample or combination thereof can be used as training library input. The only limitation is that the ground-truth sequence we are generating the training from needs to be ascertained. For example, the randomly generated RNA sequences can be used as source followed by synthetizing their cDNA strand (RT) and linking them together and sequencing them. The advantage of using RNA from biological sample is that we can align each read to reference transcript and know the underlying sequence with high confidence. Further, when RNA::cDNA molecules are used for training datasets, RNA sequence sets the ground truth for the corresponding complementary cDNA adding an additional layer of confidence. In another particular embodiment sequence complexity (N) is at least 1000, preferably at least 2,000 or at least 3,000; or at least 5,000; or at least 5,000; or at least 6,000; or at least 7,000; or at least 8,000; or at least 9,000; or at least 10,000; or at least 15,000; or at least 20,000; or at least 25,000; or at least 30,000; or at least 40,000; or at least 50,000; or at least 100,000; or at least 1,000,000; or at least 2,000,000; or at least 3,000,000, or at least 4,000,000; or at least 5,000,000 Preferably between 10,000 and 3000000, even more preferably between 10,000 and 50,000. The higher the sequence complexity the better. However, we have observed that values above 3,000,000 or 4,000,000 give little improvement. As indicated above N is the number of templates, of DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA, preferably RNA::cDNA molecules in the training library. In another particular embodiment, the DRS training dataset has a high sequence complexity, preferably covering at least all possible 5-mer nucleotide sequences. As a summary for step a) the main differences between the method of Pryszcz et al, 2024 and the training method of a DNA basecaller for DRS disclosed in the present application lies in the training input samples and thus in the training dataset that is used. Further, the invention comprises features related to such as the method of translating signal from the cDNA to the nucleotides (signal-to-sequence alignment or signal-segmentation) and using RNA as ground truth for this translation. The DNA basecaller for DRS of the current invention is trained with datasets comprised of signals obtained by DRS sequencing of RNA molecules and, preferably, their complementary cDNA, linked. This allows generation of much more complex sets of sequences, this higher complexity enables to train a general DNA basecaller for DRS that can decode any arbitrary DNA sequence unlike the limitations of barcode dependent training sets. In preferred embodiments, the DNA basecaller for DRS is trained with the signals corresponding to cDNA strand that is covalently linked to RNA strand of RNA::cDNA duplex. Step b) establishing a correlation of each signal of the Direct RNA Sequencing training dataset of step a) to a nucleotide sequence by signal-segmentation Normally, specialized software is used to align reference sequence and read signal (tombo for RNA002 (https: / / github.com / nanoporetech / tombo) or remora for RNA002 and RNA004 (https: / / github.com / nanoporetech / remora)). Such programs require a table of expected signal values for every possible sequence context (k-mer) also known as kmer models. Those kmer models vary depending on molecules that are sequenced (DNA or RNA), sequencing chemistry (helicase, pore type, sequencing buffer) and even sequencing conditions (different voltages or temperature used during sequencing will require distinct kmer models, for example there are two kmer models for dna_r10.4.1_e8.2 for helicase sequencing with 260 bps and 400 bps speed). Oxford Nanopore Technologies (ONT) released several kmer models for DNA from DNA sequencing and RNA from RNA sequencing (https: / / github.com / nanoporetech / kmer_models). But no kmer model exists for DNA from RNA sequencing to date, so none of the existing softwares are applicable to the present invention, this is to a train a complete DNA basecaller from a DRS training dataset obtained from a biological sample. Therefore, the inventors developed a custom method for signal-to-sequence alignment. The state-of-the-art method of signal-to-sequence alignment cannot be used directly for the present invention because the complexity (N) is much here as there is a higher number of unique sequences and they are much longer compared for instance to the SeqTagger application, in which only 20-37 bases are used. In fact, there is no need to know the precise position of every base in the signal space. The inventors surprisingly found out that it is enough to know only the approximate position of every signal chunk in the sequence space. However, it is important to know precisely the start and end of each signal chunk. A signal chunk is a fixed number of signals (signal readouts) corresponding to a selected sequence length over the whole dataset. The signal readouts are raw pA measurements from the sequencing, Each DNA or RNA base generates several readouts. In embodiments a signal chunk is typically 10,000 signal readouts corresponding roughly to 325 bases) in the sequence space. Signal chunks together with their corresponding nucleotide sequence in the sequence space is called signal-sequence pair. It is known to skilled person, the correspondence varies based on multiple features such as the oligonucleotide type. In embodiments, the size of the signal chunk can vary due to mentioned features or different chunk sizes can be selected. In embodiments, according to the method a signal chunk can be selected between 500 and 100000 readouts, preferably, between 5000 and 50000 readouts, more preferably between 7500 and 20000 readouts. In embodiments the selected signal readouts can correspond to the 100-4000 bases, 200-2000 bases, 200-1000 bases, 200-700 bases, 300-500, 300-400 bases. In another particular embodiment the chunk size can be of 1,000 read samples, 1,500, or 2,000, or 2,500, or 3,000 or 3,500 or 4,000, or 4,500 or 5,000 or 10,000 or 20,000 or ever larger number of readouts (and proportionally shorter or larger corresponding cDNA sequence). According to the embodiments of this disclosure, RNA::cDNA hybrid is a single molecule that is generated from transcripts (mRNA) wherein the transcripts are intact or fragmented. Further shorter signal-sequence pairs can be generated from this RNA::cDNA molecule that are used in the training. In embodiments, signal-sequence pairs used for training have the same signal length. Although the signal length can be chosen arbitrary, in preferred embodiments, the signal length (signal chunk) is typically 10,000 signal points (readouts) corresponding to ~200- 400 DNA bases. In particular embodiments, RNA fragments used for preparation of the DRS library to generate the training dataset has a size of about 600 bases, meaning RNA::cDNA molecule will be about 1,200 bases. The signals generated by sequencing from these RNA::cDNA molecule can be later chunked computationally. In effect, from single 600 bases RNA fragment, 600 bases of cDNA is obtained and from this 600 bases cDNA several signal-sequence chunks that are ~325 bases (10,000 samples) long can be generated. For example, using a step of 50 bases, several chunks starting at base 1, 51, 101, 151, 201 and so on can be obtained. In embodiments the step can be chosen different than 50 for example, a step of 100, 150, 200 or, 250 bases etc. In embodiments, only one chunk (first chunk) from every RNA::cDNA chunk is used, preferably the first chunk starting at base 1. In other embodiments, more than one chunk from each RNA::cDNA is used. As indicated above, inventors developed a customed method that will detect significant changes in the continuous read signal (events). According to the method disclosed, the term event refers to a change in signal likely corresponding to changes in bases that pass through the pore. Such events can be detected using T-test statistics using two rolling windows (10 and 20 samples of signal) followed by peak detection (using Python scipy with peak prominence=9 and distance=5). In embodiments, the size of the two rolling windows are 10 and 20 readouts of signal. In other embodiments the size of the two windows can be selected from 2 to 10 and 20 to 100, respectively. In embodiments, the number of events roughly corresponds to number of bases in the sequence that is analyzed. Naturally, the number of events detected may be different from the number of bases, because some sequence context (i.e. homopolymers, see poly-A tail signal in the Figure 4) will not generate significant changes (events) in the signal. But overall, the number of events should roughly correlate to the number of bases and should be more or less constant along the read (except long homopolymers, longer than 3-5 bases). And therefore, we could use positions of events normalized by the total number of bases in the reference to get the approximate sequence for any arbitrary signal chunk. In embodiments according to the described method, the reference is corresponding transcriptome. In embodiments, T-test statistics using one rolling windows. In other embodiments more than 2 rolling windows are used, for example 3 rolling windows, 4 rolling windows, up to 5 rolling windows are used for T-test statistics. In embodiments, other methods known in the art can be used to detect events. As known in the field and mentioned in the background, standard DRS library preparation involves addition of DNA adapter and DNA barcode steps. The DNA basecaller for DRS trained or obtained according to the disclosed method of the present invention basecalls both, DNA adapter and / or DNA barcode, as signal chunks of arbitrary length can be used. In addition, it is able to basecall any DNA sequence, including cDNA, either at the extremes of RNA molecule (such as DNA adapter or DNA barcode), incorporated inside RNA body, preferably when it is covalently linked with RNA strand. More concretely, the step of establishing a correlation of each signal of the Direct RNA Sequencing training data of step a) to a nucleotide by signal-segmentation comprises the following steps: b.1 basecalling a read with a RNA basecalling model b.2 normalization b.3 identifying events in the signal, preferably DNA signal, more preferably cDNA signal b.4 generation of signal chunk, signal-sequence pairs and using the positions of events normalized by the total number of bases in a reference to predict the sequence of nucleotides for any signal chunk generated Below it is explained every substep b1 to b4 for a particular example in that a RNA::cDNA is used. b.1) basecalling the read with RNA basecalling model: both RNA part and cDNA part of the RNA::cDNA chimera are basecalled. A priori, it is difficult to know where cDNA part starts as the cDNA part of the molecule has lower signal values (picoamperes) on average than RNA part of the molecule. RNA basecaller is able to basecall only RNA part and reports random sequence for cDNA part. We can detect the end of the RNA part (and start of the cDNA part) by aligning the resulting RNA sequence to reference transcriptome – the last aligned base will be the last base of RNA part. And since we know the (approximate) position of every base of a read in the signal space (so called move table reported by the basecaller), we will know the (approximate) start position of the cDNA part in the signal space. In brief, in this step, we will get: i) ground-truth sequence of the RNA part (basecalling is error prone, but by aligning it to known transcript, we “fix” basecalling errors), ii) (approximate) start position of cDNA part in the signal space (from move table and alignment) and iii) ground-truth cDNA sequence, which is reverse complement of RNA sequence (which we know from the alignment). b.2) normalization of cDNA signal, preferably by median absolute deviation or standard deviation b.3) identifying events (signal changes corresponding to individual bases) in cDNA signal using T-test statistics with two rolling windows (10 and 20 samples of signal for example) followed by peak detection (using Python scipy with peak prominence=9 and distance=5 or default parameters). Typically, we detect fewer events than there are bases (some base transitions result in small signal changes), but we don't need to know the precise position of every base, we just need to know the (approximate) sequence for every signal chunk (typically 10,000 signal readouts). So if there are 100 events for 200 cDNA bases, we'll count each event as a change of 2 cDNA bases. of chunks for first cDNA basecaller by using a signal window of, for example, 10,000 readouts sliding every 5,000 readouts (for example for cDNA part with length of 20,000 readouts, we'll generate 3 chunks, starting at positions 0, 5,000 and 10,000). And if for this cDNA part we obtained a sequence of 600 bases (which is reverse- complementary to ground-truth sequence RNA part sequence from the alignment) and we detected 300 events in cDNA part, each chunk will start at roughly at base 0, 150 and 300 (suitable to some variation depending on the positions of consecutive events along the signal). Hence, each chunk will consist of cDNA signal (exactly 10,000 readouts) and corresponding cDNA sequence (roughly 325 bases, depending on positions of detected events). c) training the DNA basecaller for Direct RNA Sequencing (DRS) using the DRS training datasets of step b) The training of the DNA basecaller for DRS is also termed training of CTC-CRF DNA embodiment it is trained with bonito. training is typically performed for 50 epochs using 1-5 millions of cDNA signal chunk-cDNA sequence pairs.3% of training examples are kept for validation (not used for training). In a particular embodiment, in order to improve the DNA basecaller for DRS, it is possible to perform additional rounds of training (at least one) refining iteratively training set as follows : i) basecall cDNA signal chunks (from b.4) using DNA basecalling model for DRS (from step 5), align resulting cDNA sequences against reference transcriptome (the same from step 1) and replace cDNA sequences (from the first training step b.4-b.5) with ground-truth from trainscriptome alignment. This will generate improved training set in which each cDNA signal chunk will be paired more accurately with the underlying sequence (since event detection only gave us approximate sequence for each signal chunk). model for DRS performed with bonito using improved training set from step i). A skilled person would understand by the present disclosure that it is possible to perform a third round, a fourth round of training and so on by executing steps i)-ii) multiple times. The term ‘signal’ is referred to the measurements of current intensity during time that are measured in each nanopore. Alterations in the current intensity values occur as the RNA or DNA molecule is being translocated through the nanopore. This signal is one of the outputs produced by the nanopore sequencing technology. The term signal chunk is referred to a part of the signal of fixed length. The term sample is referred to any biological material or synthetically generated material from which RNA can be isolated. The extraction of the reads can be done by a commercial software, for instance MinKNOW or any software configured to analyze the sequencing results of the nanopore direct RNA sequencing. Furthermore, the raw signals can be extracted from propietary files using pod5 or ont-fast5-api Python libraries. In the present specification, “signal-segmentation” “signal-to-sequence alignments” and “signal-only segmentation” are synonyms. In a particular embodiment, the rolling window analysis technique is used for calculating the T- test score. As known to the skilled person, a rolling window is a technique used for analyzing time series data or sequential data points by applying a statistical measure over a moving subset of the data. When calculating T-test scores using a rolling window, you are essentially calculating T- test scores for subsets of the data rather than the entire dataset at once. In this disclosure the subsets of data that is used to calculate T-test with rolling window technique is defined as “read sample(s)” or “window size” which are signal chunks of fixed length over the whole dataset. In embodiments, the T-test score of signal chunks, the number of read samples (number of current intensity data points) are detected using T-test statistics using two rolling windows. It is not essential for the method to know the precise position of every base in the signal space but it is important to know the approximate position of every signal chunk (typically 10,000 samples corresponding roughly to 325 bases) in the sequence space. In embodiments, DNA signal is decoded into sequence space by first basecalling RNA part of the RNA::cDNA molecule, aligning RNA sequence to transcriptome and taking reverse complement of RNA part. The signal changes identified in the continuous read signal are called events in the context of this disclosure. In embodiments, the events are detected using T-test statistics using two rolling windows followed by peak detection. In some embodiments, the peak detection is done using Python scipy with peak prominence=9 and distance=5. In embodiments the events correlate with the number of the bases in reference RNA or sequence cDNA. In some embodiments, the number of events are less than the number of the bases in reference RNA or sequenced cDNA. It is expected to the skilled person that, the number of events detected this way will be different from the exact number of bases, because some sequence context won't generate significant changes in the signal. But overall, the number of events should roughly correlate to the number of bases and should be more or less constant along the read (except long homopolymers). In embodiments the positions of events normalized by the total number of bases in the reference to obtain the approximate sequence for any arbitrary signal chunk. Importantly, although the chunk size is not an essential feature, in a particular embodiment, the signal in question that will be used for DNA basecaller for DRS tranining consists of signals of 10,000 readouts (and corresponding cDNA sequences, roughly 325 bases per each signal chunk). In a particular embodiment a signal-only segmentation approach comprises T-test statistics that determine the T-test score by at least one window size rolling over a number of read samples, preferably two window size rolling window For example, with one or two windows sizes, preferably two window sizes, preferably two window sizes with a size from 1 to 100, preferably the rolling window is 10 and 20 readouts of signal. In other embodiments the size of the two windows can be selected from 2 to 10 and 20 to 100 respectively In embodiments, the oligonucleotide fragments further comprise at least one overhang sequences that enables the second ligation to the DNA oligonucleotide that contains the helicase which is required for nanopore sequencing technology. In embodiments, the overhangs are known from Oxford Nanopore Technologies (ONT) as RTA adapters. In the present specification a trained model or trained algorithm is synonym of a DNA Basecaller, which is a trained basecaller according to the method disclosed here for basecalling / decoding DNA sequenced by direct RNA sequencing (DRS). In general practice, DRS is used for sequencing RNA and the basecallers available for decoding of output of the DRS molecules are only available for the RNA molecule but not for DNA. DNA can be sequenced by DRS and there are several advantages of being able to decode this DNA sequences as discussed above. For example, in the context of demultiplexing, such a basecaller simplifies classification of DNA barcodes (no signal trimming is required) and is flexible toward an unlimited number of barcodes (with varying lengths and covering complete sequence space) and allows complex barcode structure ie random barcode concatenations that can be used to increase the available barcode pool. We took advantage of the latest CTC-CRF (Connectionist temporal classification- conditional random field; https: / / github.com / nanoporetech / bonito / ?tab=readme-ov-file#references) model architecture that is achieving over 99% basecalling accuracy for DNA basecalling (https: / / nanoporetech.com / platform / accuracy / ). We use a window of 31 and stride of 10, similarly to older RNA flip-flop models. We tested three model versions with increasing number of features (and parameters): 96 (fast with 519,880 parameters), 384 (hac with 6,499,048 parameters) and 768 (sup with 24,793,192 parameters). Those versions differ in basecalling accuracy of DNA sequences: starting from 92% for fast and ending with 99% for sup. But we found only marginal differences in their accuracy for barcode basecalling, thus we use the lightest version (fast). However, when high accuracy is needed (ie modification detection), we envision using hac or sup models. The CTC-CRF models can be trained with bonito (https: / / github.com / nanoporetech / bonito / ). Similarly, bonito can be used for basecalling of Nanopore signals. Bonito is written in Python and uses pytorch (https: / / pytorch.org / ) internally, therefore it's straightforward to incorporate it in a larger Python program. This gives us complete freedom: we can select which part of the signal we want to basecall, normalize it according to our needs, etc. (These are not easily available when using other basecallers such as guppy or dorado). Therefore, Bonito software is preferred. In a particular embodiment the trained DNA basecalling model (DNA basecaller) is trained with Bonito software. The number of chunks used to train the algorithm can differentiate between 100,000 and 100,000,000, preferably between 1 and 5 million chunks. In particular embodiment, the trained algorithm described herein is used to demultiplexing direct RNA sequencing reads from Oxford Nanopore Technologies (ONT), samples having any arbitrary barcodes. It is particularly providing advantage by not requiring specific training for the barcode sequences used and lifting the limitation of the barcode numbers compared to the methods known in the art. In the present specification the term "complementary" and "complementarity" are interchangeable and refer to the ability of polynucleotides to form base pairs with one another. Base pairs are typically formed by hydrogen bonds between nucleotide units in antiparallel polynucleotide strands or regions. Complementary polynucleotide strands or regions can base pair in the Watson-Crick manner (e.g., A to T, A to U, C to G).100% (or total) complementary refers to the situation in which each nucleotide unit of one polynucleotide strand or region can hydrogen bond with each nucleotide unit of a second polynucleotide strand or region. Less than perfect (or partial) complementarity refers to the situation in which some, but not all, nucleotide units of two strands or two regions can hydrogen bond with each other and can be expressed as a percentage. In the present specification the expression “ONT RTA adapter A” is synonym to “first adapter DNA nucleotide” and “Oligo A” and can be used interchangeably. In the present specification the expression “ONT RTA adapter B” is synonym to “second adapter DNA oligonucleotide” and “Oligo B” and can be used interchangeably. According to the disclosed method, by using different DRS training dataset obtained from biological samples, a complete DNA basecaller can be obtained. In the present specification we will show applications of such basecaller obtained by using the disclosed method and the results regarding the use of said basecallers in the following applications. 1. DRS demultiplexing via DNA barcode basecalling The DNA basecaller for DRS is trained on the arbitrary training set. This allows the basecaller obtained based on the method disclosed here to identify any given number or sequences of DNA barcodes and to be used in demultiplexing of DNA barcoded DRS RNA data. The basecaller disclosed in EP24 termed SeqTagger is a special case of 96- or 100-barcode basecaller for polyA+ RNAs (mRNA or polyadenylated rRNA). For the SeqTagger to function, the software first needs to be trained with a training set that has the barcodes that will be used in the sample to be sequenced. This means that the software cannot be used for random barcoded samples and the sample preparation has to be specially prepared with limited number of set sequence of barcodes to be possible demultiplexed by the SeqTagger method. Different than the any available demultiplexing method, the method of the present invention is not restricted to any set of barcodes, this is because of the training of the basecaller, particularly due to the use of arbitrary complex biological samples. In a particular embodiment, DNA basecaller used for the demultiplexing is trained with a DRS training datasets obtained from the samples comprising at least a first adapter DNA oligonucleotide and a second adapter DNA oligonucleotide, optionally each adapter oligonucleotide comprising a complementary barcode. Furthermore, the DNA basecaller of the invention can be used to demultiplex samples comprising at least a first adapter DNA oligonucleotide and a second adapter DNA oligonucleotide, each adapter oligonucleotide comprising a complementary barcode. In specific embodiments these barcodes are chosen from the sequences listed in table 3. In other embodiments, the first and second adapter DNA oligonucleotide are chosen from the sequences listed in table 2 and / or table 3, The oligonucleotides disclosed in Table 1 and 2 comprises the barcodes of Table 3 as part of the sequence. In a particular embodiment, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33,34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, or 96 barcodes are used. In another particular embodiment the first adapter DNA oligonucleotide and the second DNA oligonucleotide are selected from one or more of the group comprising of SEQ ID No 1 or SEQ ID No 2 and either SEQ ID No 3 or SEQ ID No 4, or a variant having at least 80%, preferably at least 85%, more preferably at least 90 % and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with respect to SEQ ID No 1 and / or SEQ ID No 2 and / or SEQ ID No 3 and / or SEQ ID No 4. 2. single-cell native RNA sequencing To achieve single cell direct RNA sequencing, the DNA basecaller for DRS must be able to demultiplex hundreds of different barcodes. However, classical approaches to demultiplex (i.e train a model to predict a fixed number of barcodes) will lead to decreased algorithm accuracy with the increase in the number of barcodes, which would be problematic to achieve single cell sequencing. The basecaller trained and generated based on the method disclosed can address this need and provide solution to the problem when used in DNA basecalling of barcode regions to demultiplex single cell barcodes, enabling scaling the number of barcodes with minimal bias in the ligation. There are known methods to barcode single cell samples and the absolute DNA basecaller generated based on the method of this invention can be successfully used to analyze barcoded single cell DRS RNA independent of the method of barcoding used. In Figure 5 is disclosed an example protocol for single-cell barcoding based on splint-ligation. The state of the art already comprises known methods to prepare those samples, for example: Gierahn, T., Wadsworth, M., Hughes, T. et al. Seq-Well: portable, low-cost RNA sequencing of single cells at high throughput. Nat Methods 14, 395–398 (2017). https: / / doi.org / 10.1038 / nmeth.4179; Klein, Allon M et al. “Droplet barcoding for single-cell transcriptomics applied to embryonic stem cells.” Cell vol.161,5 (2015): 1187-1201. doi:10.1016 / j.cell.2015.04.044; Scaling single cell transcriptomics through split pool barcoding Alexander B. Rosenberg, et al. bioRxiv 105163; Svensson, V., Natarajan, K., Ly, LH. et al. Power analysis of single-cell RNA- sequencing experiments. Nat Methods 14, 381–387 (2017). https: / / doi.org / 10.1038 / nmeth.4220 Another object of the invention relates to a method of single-cell RNA sequencing comprising the following steps: 1. Obtaining a single-cell RNA sample by using at least one capture oligonucleotide and at least two oligonucleotide-DNA barcodes, wherein said RNA sample comprises sequentially concatenated DNA barcodes arranged in multiple splints, by performing one or more rounds of split- pool barcoding and 2-obtaining a DRS signal from the single-cell RNA sample with at least sequential concatenated DNA barcodes and 3. demultiplexing the DRS signal obtained from the sequential concatenated DNA barcodes using a DNA basecaller, the demultiplexing comprising: a. detecting a subset of sequentially concatenated DNA barcode combinations generated during split-pool barcoding b. assigning basecalled reads to the detected subset of barcode combinations. Step 1) of the method above refers to a Split-seq method. This technique is commonly used in the field of single cell sequencing. An advantage of this method is that the single-cell RNA sample is not amplified before carrying out the steps method thereby reducing amplification bias and preserving native transcript representation. The use of DNA barcodes provides enhanced molecular stability, higher indexing accuracy, and compatibility with standard sequencing workflows. RNA barcodes, are more fragile, prone to degradation or mis-priming during reverse transcription, which risks losing cell identity information where DNA barcodes are chemically stable. RNA barcodes would require additional reverse transcription steps before being incorporated into sequencing libraries, increasing complexity and error risk. DNA barcodes can be directly ligated or synthesized into primers used for reverse transcription or amplification. Further DNA barcodes are replicated faithfully by polymerases, providing more accurate combinatorial indexing than RNAs. The sequencing of step of step 2) can be performed by any method that allows the obtention of Direct-RNA-Sequencing, such as Nanopore In a particular embodiment, step 3(a) comprises: – aligning each basecalled DNA barcode sequence to a reference set of barcode sequences arranged in multiple splints, – retaining only the best match at each splint when the difference (mindiff) between the best and second-best match is above a threshold, and / or – identifying valid barcode combinations that occur in at least 3% of the expected number of reads per barcode; In another particular embodiment, step 3(b) comprises: – aligning each basecalled barcode sequence to a reference sequence while retaining all matches within a predefined maximum edit distance (maxdiff) from the best match, – generating possible barcode chains from the retained matches, restricted to combinations observed in step 3 (a), and – selecting a final barcode chain for each read based on the highest alignment score, provided the difference (mindiff) between the best and second-best score is above a defined threshold. In another particular embodiment, the threshold value mindiff is equal to or greater than 7. In another particular embodiment, the maximum edit distance maxdiff is equal to or less than 5. Mindiff and Maxdiff are calculated following know approaches in the field. In another particular embodiment, the barcode combinations are represented as combinations of sequential barcodes over three or more splints, each barcode chain being identified by its component barcodes at each splint position. Each “splint position” refers to a defined location within the concatenated barcode sequence corresponding to a round of split-pool barcoding. Each unique cell RNA is identified by the specific set of barcodes assigned at each splint position, wherein the positional order of these splints leads to the combinatorial uniqueness of the resulting barcode chain. Split-pool ligation uses splints to join consecutive barcodes, so in the present discloseure each splint means each barcode concatenation or barcode iteration In another particular embodiment, the DNA basecaller is trained according to the method disclosed above in the preceding object of the invention. In another preferred embodiment the method of single-cell RNA sequencing uses one or more of the following combination of oligonucleotides-DNA barcodes that increase the sensitivity and accuracy of the method and reduced the error, When one of more of the following oligonucleotides are used there is no need to calculate mindiff or maxdiff Those oligonucleotides-DNA barcodes are based on those sequences that in the method to train a DNA basecaller provided the most accurate predicitions for (30mer sequences) 288 barcodes. These sequences are then randomly mixed and created, one, two or three plates of more, preferably of 96 barcodes so that the average accuracy is the same across all the plates: In a particular embodiment it can be used from 3 to 288 barcodes, of those listed in Tables 4 to 6. A skilled person would understand there is no need to split them in the same groups as disclosed in the application. For example, you can use 144 barcodes and split them in three plates of 48 wells or, 192 barcodes and split them in two 96 well plates. Every barcode must have at the 5´ end a further “5´ sequence” said sequence may be from 10 to 20 nucleotides preferably from 12 to 18 nucleotides, and may have a 5´ Pho group at the opposite end of the union with the barcodes. The 5´sequence of each group of barcodes may have a different size. SEQ ID 301 (TCCGAGACGCGTCAA); SEQ ID 302 (CTCAAGCACGTGGAT); SEQ ID 303 (CGAAACATCGGCCAC) disclose 3 examples of 5´sequences, it is important that one of said sequences is used for all the barcodes of each plate or group. In another particular embodiment “5´ sequence” is selected from one or more of the group comprising of SEQ ID No 301 or SEQ ID No 302 and / or SEQ ID No 303 or a variant having at least 80%, preferably at least 85%, more preferably at least 90 % and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with respect to SEQ ID No 301 or SEQ ID No 302 and / or SEQ ID No 303 Every barcode must have at the 3´ end a further “3´ sequence” said sequence may be from 10 to 20 nucleotides or from 12 to 18 nucleotides. SEQ ID 304 (CGAATGCTCTGGCCT); SEQ ID 305 (AGTCGTACGCCGATG); SEQ ID 306 (TAGTAGGTTC) disclose 3 examples of 3´sequences, it is important that one of said sequences is used for all the barcodes of each plate or group. The 5´sequence of each group of barcodes may have a different size In another particular embodiment “3´ sequence” is selected from one or more of the group comprising of SEQ ID No 304 or SEQ ID No 305 and / or SEQ ID No 306 or a variant having at least 80%, preferably at least 85%, more preferably at least 90 % and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with respect to SEQ ID No 304 or SEQ ID No 305 and / or SEQ ID No 306. In a particular embodiment, the oligonucleotide-DNA barcode comprises a 5´sequence, a barcode and a 3´ sequence, the barcode linked to the 3´end of the 5´sequence and the 3´sequence linked to the 3´end of the barcode, being - the 5´sequence a sequence that partially or fully anneals with the 5´ sequence of the capture oligonucleotide, preferably SEQ ID No 301 or a variant thereof having at least 90% or 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No 301, or the 5´sequence a sequence that partially or fully anneals with the 3´ sequence of the oligonucleotide-DNA barcode used in a previous round of the split-pool barcoding, preferably SEQ ID No 302 or a variant thereof having at least 90% or 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No 302, or SEQ ID No 303 or a variant thereof having at least 90% or 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No 303; and - the barcode being one of SEQ ID No 309-596 or a variant thereof having at least 90% or 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No: 309-596; and - the 3´ sequence a sequence that anneals with the 5´ sequence of the oligonucleotide- DNA barcode used in a next round of the split-pool barcoding, preferably SEQ ID No 304 or a variant thereof having at least 90% or 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No:304, or SEQ ID No 305 or a variant thereof having at least 90% or 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No:305, or the 3´ sequence a sequence that anneals with oligonucleotide with a helicase enzyme for RNA sequencing. A “capture oligonucleotide” is also required. Said “capture oligonucleotide” comprises 2 regions, a “5´ sequence” which is the partially or fully reverse complement of the oligonucleotide-DNA barcode, preferably it anneals with the “5´ sequence” oligonucleotide- DNA barcode. For instance, for SEQ ID NO 301 The capture oligonucleotide is 5'- TTGACGCGTCTCGGA-3' (SEQ ID No 307) or a variant having at least 80%, preferably at least 85%, more preferably at least 90 % and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with respect to SEQ ID No 307. In the present specification a sequence that anneal with other sequence means that they are complementary to each other. The other region, next to the 3´ end of the “5´ sequence” of the “capture oligonucleotide” is specific to whatever target is need mRNA, or a particular RNA of interest such as 18S or 28S. In a particular embodiment, to sequence poly-A RNA we have capture oligos with the full sequence of 5'-TTGACGCGTCTCGGATTTTTTTTTT-3' (SEQ ID No 308) where the poly-T stretch anneals to the poly-A tails of target mRNA. The “5’ sequence” anneals to one barcode in a set of barcodes, bringing the barcode sequence and the RNA molecule next to each other. It is important that the “5´ sequence” of the capture oligo only anneals with one of the 5´ sequences of the oligonucleotide-DNA barcode. So, in an example with three 96 well plates, the capture oligo must anneal only with the 5´ sequences of the oligonucleotide DNA barcode of one plate. The “3' sequences” are needed for splint ligation with the subsequent Plate. In the same manner that the “5´sequence” of the capture oligo anneals with the “5´sequence” of the oligonucleotide-DNA barcode of the 1stplate, the 3´sequence of the oligonucleotide-DNA barcode anneals with the 5´ sequence of the oligonucleotide-DNA barcode of the second plate and so on (Splint ligation) the olignonucleotide-DNA barcode of the last plate must have a 3´ sequence that is complementary to an oligonucleotide with an helicase enzyme for RNA sequencing. In a particular embodiment the capture oligonucleotide comprises a 5´sequence and a 3´ sequence, the 3´sequence linked to the 3´end of the 5´ sequence, being the 5´sequence a sequence that partially or fully anneals with one of the 5´sequence of the at least two oligonucleotide DNA-barcode, preferably SEQ ID No 307 or a variant thereof having at least 90% or 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No 307 and the 3´ sequence a sequence that anneals with an RNA target of the single-cell RNA, preferably selected from a poly A tail, preferably SEQ ID No 599 or a variant thereof having at least 90% and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No 599; or an 18s rRNA, preferably SEQ ID No 597 or a variant thereof having at least 90% and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No 597; or an 28S rRNA, preferably SEQ ID No 598 or a variant thereof having at least 90% and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % identity with SEQ ID No 598. Steps 3a) and b) relate to the multiplexing, generated concatenated barcodes in the split-pool are detected, and then reads are assigned. In another embodiment the single-cell is coupled to a semi-permeable capsule, this is, the cell is encapsulated in a microcapsule with an aqueous core precursor solution, and a semipermeable shell precursor solution. The aqueous core precursor solution and / or the semipermeable shell precursor solution can be those disclosed in US12360024B2; EP4565207A1 and / or EP4436707A1 while they are photo-inducible chemicals such as the one disclosed in Doméjean, H. et al. Controlled production of sub-millimeter liquid core hydrogel capsules for parallelized 3D cell culture. Lab Chip 17, 110–119 (2016); and Leonaviciene, G., Leonavicius, K., Meskys, R. & Mazutis, L. Multi-step processing of single cells using semi- permeable capsules. Lab Chip 20, 4052–4062 (2020) Another object of the invention relates to an oligonucleotide comprising a RNA and a DNA covalently linked in the 5´ end or the 3´ end of the RNA with the 3´ end or the 5´ end of the DNA, wherein the DNA is complementary to the RNA and between the DNA a the RNA there are at least 2 concatenated barcodes. These barcodes may be two or more of those disclosed in Tables 4 to 6 3. improving accuracy of state of the art DRS sequencing, and modification detection, by sequencing and combining sequence information from both strands, RNA and cDNA. Similar to the DRS training set generation, RNA and cDNA strands of the RNA samples to be prepared can be connected, sequenced together to be analyzed by the DNA basecaller trained and generated according to the disclosed method to improve the accuracy of the decoding of the DRS. Once RNA and cDNA strands are connected for example, either via hairpin ligation or direct ligation using a ligase, both strands of RNA::cDNA duplex will be streamed through the pore allowing basecalling of both strands (Figure 7). The accuracy of RNA basecalling can be improved by consensus decoding of both strands. Such pair consensus decoding can be performed either on: i) the aligned signal of RNA and cDNA strands (no methods exist so far, likely not feasible computationally), ii) probability profiles of both strands (similar to PoreOver, computationally expensive) or iii) on the basecalled sequence (the fastest of the three, using the DNA Basecaller for DRS of the invention). In embodiments, DRS library preparation included a step where RNA and cDNA strands of the RNA::cDNA double strands are ligated to each other in one end to form a single or double RNA::cDNA strand. In embodiments, the disclosed method is used to improve the RNA sequencing accuracy of DRS samples by basecalling RNA::cDNA molecules. According to the method, 1) the RNA samples and its cDNA counterpart is connected at the oligo nucleotide ends to form a RNA::cDNA molecule and the molecule is sequenced together using DRS. The resulting signal generated from the molecule of RNA::cDNA chimera is basecalled using the basecaller trained according to the disclosed method. In embodiments, RNA::cDNA molecule is basecalled with two different basecallers. In embodiments, only DNA (or cDNA) part of the signal generated form the RNA::cDNA molecule is basecalled with the DNA basecaller trained according to the disclosed method. In embodiments, both DNA (or cDNA) and RNA parts of the signal generated form the RNA::cDNA molecule is basecalled with the basecaller trained according to the disclosed method. In other embodiments, RNA part of the signal generated form the fragment of RNA::cDNA molecule is basecalled with a second basecaller, preferably the second basecaller is an RNA basecaller Similar to the generation of the DRS training dataset, RNA::cDNA strands obtained from the sample to be sequenced using DRS can be connected via ligation. . In embodiments, the ligation can be done using specific ligation enzymes that is able to ligate single stranded polynucleotides and / or complementary polynucleotides, preferably ligation enzymes that ligates RNA to DNA, RNA to cDNA, preferably by using CircLigase ssDNA Ligase or RtcB. In embodiments, RNA::cDNA strands can be connected by a DNA-RNA hairpin ligation. Once ligated the fragment comprising both strands are sequenced together. Another object of the invention relates to a method of basecalling the bases of a nucleic acid sequenced using DRS, comprising a DNA basecaller for DRS trained as disclosed above: a) providing a Direct RNA Sequencing data obtained from a sample, b) establishing a correlation of each signal of the Direct RNA Sequencing data to a nucleotide by signal-segmentation to obtain a DRS training dataset, and c) training the DNA basecaller for Direct RNA Sequencing (DRS) using the DRS training datasets of step b) and d) using trained basecaller to basecall the nucleic acid sequenced using DSR. The DNA basecaller of the invention is universal it can be applied to any nucleic acid sequence originating to any source, preferably to same type of nucleic acid as used for generating the DRS training datasets: DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA, preferably RNA::cDNA. In a particular embodiment, the nucleic acid is DNA molecule, RNA molecule or a hybrid of two, preferably RNA::DNA or RNA::cDNA sequenced with Direct RNA Sequencing. Another object of the invention relates to a DNA basecaller for Direct RNA Sequencing (DRS) obtained by the method as described above. Another object of the invention relates to the use of the method of sequencing and basecalling of RNA::cDNA samples according to the disclosed method will provide advantages and accuracy in several applications of nanopore DRS. Nonlimiting examples of the applications by the DNA basecaller for DRS obtained by the method of the invention are disclosed herein 3.1, 3.2, 3.3, 3.4, 3.5. 3.1. and mapping of small, highly modified RNA molecules In embodiments, RNA::cDNA chimera basecalling according to the disclosed method is used to accurately basecall non-coding RNA such as rRNA, rRNA, lncRNA, snRNA, miRNA, siRNA, aRNA and tRNA, preferably short (miRNA) and highly modified molecules (tRNAs).Nanopore signals are affected by modifications. This provides a treasure of information about modification status of individual molecules. But signal deviations caused by modification typically result in basecalling errors, resulting in more challenging alignments for highly modified molecules. It makes Nanopore not well suited to analyze short (miRNA) and highly modified molecules (tRNAs), because their sequence decoding will be inaccurate and we won't be able to decode 10-15 bases from 5'-end of the molecule. To alleviate this, we had to develop a custom protocol to sequence tRNA (Nano-tRNAseq) that ligates RNA oligos to both ends of tRNA molecules, this protocol is for example disclosed in EP22382917. Those oligos facilitate better alignment (since tRNA body has many basecalling errors) and allow us to basecall the 5'-end of tRNA. Still, we have to rely on reduced tRNA reference sets, because we are not able to align multiple reads uniquely. Here, we propose to sequence and basecall simultaneously both strands of RNA::cDNA duplex, allowing for the first time analysis of various types of small RNAs without dedicated library preparation and analysis protocols (such as Nano-tRNAseq). This will be feasible because: i) only some modifications alter Watson-Crick pairing, thus we can obtain more accurate basecalling from cDNA strand than from RNA strand, allowing more accurate alignment of highly modified tRNAs, ii) we we'll basecall 5'-terminal bases of RNA, because it'll be connected physically to cDNA allowing alignment of very short molecules such as miRNA or tRNA fragments. 3.2 Another layer of information about RNA modification status from cDNA strand In embodiments, RNA::cDNA chimera basecalling according to the disclosed method is used to identify RNA modifications. Since the complete DNA basecaller of the invention is able to sequence both strands, RNA and cDNA, of RNA::cDNA duplex, it is able to detect the modification status of nucleotides in RNA molecules, especially for modifications affecting Watson-Crick base pairing including but not limited to inosine (I), pseudouridine (Y), N6-methyladenosine (m6A), N6,2’O- dimethyladenosine (m6Am), N1-methylpseudouridine (m1Y), N7-methylguanosine (m7G), 2’O-methyladenosine (Am), 2’O-methylcytosine (Cm), 2’O-methylguanosine (Gm), or2’O- methyluridine (Um). For example, Inosine, a modification of Adenine, is complementary to Cytosine, and it results in A-to-G mismatch in cDNA relative to RNA. As a result, Inosine (and many other RNA modifications such as m1A, m3C, m1G, m22G, m3U), besides altering signals on the RNA strand, will also affect incorporation of bases in the cDNA strand which can be identified when both the sequences of cDNA and RNA are identified by using the disclosed method. 3.3 Generating synthetic DRS training sets for training modification-aware basecallers Obtaining a very good DRS training set is a bottleneck for training of modification-aware basecallers, such as m6ABasecaller (https: / / github.com / novoalab / m6ABasecaller). One possibility is to design a DRS training dataset obtained from a biological sample embedding a modified base (and its unmodified equivalent) in a random sequence context. This would provide a perfect training set, but the underlying random sequence context has to be known (ground truth). If only the RNA strand is basecalled it will be inaccurate, because many RNA modifications cause basecalling errors at the modified base and in its proximity. ONT disclosed synthetic oligonucleotides that have modifications incorporated in random sequence context. For example, unmodified (NNNNNNN A NNNNNNN) and modified (NNNNNNN m6A NNNNNNN) oligos are sequenced independently and resulting reads are used to train modification callers. Similarly, to basecaller training, we need to know not only the modification status, but also the underlying sequence. Approximate sequence can be decoded using basecalling. But this will be inaccurate, especially for modifications that affect sequencing signals. And because sequences are synthetic and random, we have no reference to align to. ONT solved the problem of lacking the ground truth for DNA modifications not RNA, by sequencing the complementary DNA strand. This is possible, because suitable protocol (1D2) and basecaller (general DNA basecaller is fine, since both strands are DNA) are available (Fig.6). This system does not rely on covalent binding of two strands of DNA duplex with a hairpin ligation. Instead, they altered the original 1D protocol in order to increase the probability of sequencing the second strand after the first one passed the pore. Instead, we could sequence both strands of RNA::cDNA duplex and take cDNA sequence as a ground truth (only position corresponding to modified base will be affected if given modification affects Watson-Crick pairing) or use pair consensus basecalling giving more weight to cDNA basecalls. Therefore, we propose our RNA::cDNA duplex sequencing protocol and our complete DNA basecaller for DRS will be suitable to generate high quality synthetic training sets for modification calling of any RNA modification that is available for oligo synthesis. In embodiments, molecules of synthetic RNAs that have modifications incorporated in random sequence context and their corresponding cDNA counterparts are used as training dataset in disclosed method to basecall and detect RNA modifications. In this embodiment RNA::cDNA molecules are connected to each other. 3.4 Analysis of 5'-cap of RNA Nanopore sequencing is not able to basecall / analyze the last 10-15 bases of the molecule (from the 5'-end of RNA in the case of DRS, because it's being sequenced from 3'-end). This is because once the molecule exits from the helicase (that sits on top of the pore, roughly 10- 15 bases away from its center), it starts to pass through the pore with much higher speed (~10,000 bps) than nominal one (70 bps for RNA002 and 130 bps for RNA004). Because of that, standard ONT protocols are not suitable to study the biology of 5'-end of RNAs (i.e. 5'- cap). Naturally, if RNA and cDNA strands are physically connected, it'll allow to basecall and analyze 5-end of RNA molecule. 3.5 Analysis of rRNA from microbiome The DNA basecaller of the invention can be used for rRNA microbiome profiling. It produces much more accurate results because it basecalls the cDNA strand. This strand w is not affected by modifications. While rRNA strand will provide rich RNA modification information. Each of the terms “comprising,” “consisting essentially of,” and “consisting of” may be replaced with either of the other two terms. The term “a” or “an” can refer to one of or a plurality of the elements it modifies (e.g., “a reagent” can mean one or more reagents) unless it is contextually clear either one of the elements or more than one of the elements is described. The term “about” as used herein refers to a value within 10% of the underlying parameter (i.e., plus or minus 10%; e.g., a weight of “about 100 grams” can include a weight between 90 grams and 110 grams). Use of the term “about” at the beginning of a listing of values modifies each of the values (e.g., “about 1, 2 and 3” refers to "about 1 , about 2 and about 3"). When a listing of values is described the listing includes all intermediate values and all fractional values thereof (e.g., the listing of values "80%, 85% or 90%" includes the intermediate value 86% and the fractional value 86.4%). When a listing of values is followed by the term "or more," the term "or more" applies to each of the values listed (e.g., the listing of "80%, 90%, 95%, or more" or "80%, 90%, 95% or more" or "80%, 90%, or 95% or more" refers to "80% or more, 90% or more, or 95% or more"). When a listing of values is described, the listing includes all ranges between any two of the values listed (e.g., the listing of "80%, 90% or 95%" includes ranges of "80% to 90%", "80% to 95%" and "90% to 95%"). While the disclosure has been set out herein in reference to specific aspects, features and illustrative embodiments, it will be appreciated that the utility of the disclosure is not thus limited, but rather extends to and encompasses numerous other variations, modifications and alternative embodiments, as will suggest themselves to those of ordinary skill in the field of the present disclosure, based on the description herein. Correspondingly, the invention as hereinafter claimed is intended to be broadly construed and interpreted, as including all such variations, modifications and alternative embodiments, within its spirit and scope. The invention further relates to the following clauses 1. A method to train a DNA basecaller for Direct RNA Sequencing (DRS), carried out by a computer, comprising the following steps: a) providing a Direct RNA Sequencing data obtained from a sample, b) establishing a correlation of each signal of the Direct RNA Sequencing data to a nucleotide by signal-segmentation to obtain a DRS training dataset, and c) training the DNA basecaller for Direct RNA Sequencing (DRS) using the DRS training datasets of step b) 2. The method according to clause 1, wherein the sample is selected from DNA, RNA, cDNA or any combination therein, such as RNA::DNA, DNA::cDNA, RNA::cDNA chimera. 3. The method according to clause 2, wherein the DNA of RNA::DNA molecule is complementary to the RNA of the RNA::DNA molecule. 4. The method according to clause 3, wherein, the DNA and RNA of the RNA::DNA molecule is linked. 5. The method according to clause 1, wherein the signal of the Direct RNA Sequencing data of step b) is signal obtained from DNA. 6. The method according to any of the preceding clauses that comprises the following steps: a) providing a Direct RNA Sequencing data obtained from a sample in the form of linked RNA::DNA or RNA::cDNA molecules, wherein the DNA or cDNA is complementary to the RNA, b) establishing a correlation of each DNA or cDNA signal of the Direct RNA Sequencing data to a nucleotide by signal-segmentation to obtain a DRS training dataset, and c) training the DNA basecaller for Direct RNA Sequencing (DRS) using the DRS training datasets of step b) 7. The method according to any of the preceding clauses, wherein before step a) the method comprises preparation of DRS training library from which the DRS data is obtained. 8. The method according to any of the preceding clauses 6 or 7, wherein in step a) the RNA of the linked RNA::DNA or RNA::cDNA molecules are obtained from a biological sample. 9. The method according to any of the preceding clauses 6 or 7, wherein in step a) the RNA of the linked RNA::DNA or RNA::cDNA molecules is obtained from synthetic random RNA molecules, preferably comprising modifications, for example inosine, m1A, m3C, m1G, m22G, m3U and combinations and / or a poly-nucleotide tail, selected from the group consisting of a poly A tail, polyC tail, polyG tail, polyU tail, or polyG / I tail, polyinosine (polyI) tail or 2’-O- methylated polyNm tail. 10. The method according to any of the preceding clauses 8 or 9, wherein the biological sample is an RNA selected among the group consisting of mRNA, rRNA, rRNA, lncRNA, snRNA, miRNA, siRNA, aRNA and tRNA. 11. The method according to any of the preceding clauses 8 to 10, wherein the RNA of the linked RNA::cDNA molecules is obtained from a transcriptome. 12. The method according to any of the preceding clauses 8 to 11, wherein the RNA of the linked RNA::DNA or RNA::cDNA molecules is fragmented and optionally each fragment is polyadenylated before the synthesis of cDNA. 13. The method according to any of the preceding clauses 8 to 12, wherein the linked RNA::DNA or RNA::cDNA molecules has a length of about 100 to 9000 nucleotides. 14. The method according to any of the preceding clauses, 9 to 13, wherein the DRS training dataset has a high sequence complexity, preferably of at least 1000 and or 15. The method according to any of the preceding clauses 6 to 14, wherein the linked RNA::DNA or RNA::cDNA molecules of step a) are - linked covalently, preferably from the 3’ end of the cDNA to the 5´ end of the RNA molecule and / or, - double stranded or single stranded. 16. The method according to any of the preceding clauses 6 to 15, wherein the cDNA of the linked RNA::cDNA molecule is reverse transcribed from the RNA of the RNA::cDNA molecule. 17. The method according to any of the preceding clauses 14 or 16, wherein the linked RNA::DNA or RNA::cDNA is double stranded and is linked by ligation with a dedicated ligase or a hairpin ligation, or by using adapter sequences. 18. The method according to the preceding clause, wherein the dedicated ligases are CircLigase ssDNA Ligase or RtcB. 19. The method according to any of the preceding clause, wherein the Direct RNA Sequencing data are obtained by nanopore sequencing. 20. The method according to any of the preceding clauses, wherein step b) comprises the following stages: b.1 basecalling a read with RNA basecalling model b.2 normalization b.3 identifying events in the signal b.4 generation of signal chunks, signal-sequence pairs and using the positions of events normalized by the total number of bases in a reference to predict the sequence of nucleotides for any signal chunk generated 21. The method according to the preceding clause, wherein step b.1) comprises aligning resulting cDNA sequences against a reference. 22. The method according to any of the preceding clauses 20 to 21, wherein step b.4) comprises signal chunks of between 500 and 10000 readouts. 23. The method according to any of the preceding clauses 20 to 22 wherein step b.4) comprises generating signal-sequence pairs, preferably of signal chunks-cDNA sequence pairs. 24. The method according to any of the preceding clauses 19 to 22, that wherein the sample is RNA::cDNA step b1) comprises establishing the ground-truth sequence of cDNA from complementary RNA strand. 25. The method according to the preceding clause, wherein step b.4) comprises the replacement of nucleotide sequences, preferably cDNA sequences, with ground-truth from transcriptome alignment, 26. The method according to any of the preceding clause, wherein step c) comprises the training of CTC-CRF DNA basecaller for DRS using the obtained training dataset of step b). 27. The method according to the preceding clause, comprising basecalling cDNA signal chunks using DNA basecalling model for DRS. 28. The method according to the preceding clauses 26 to 27, comprising at least an additional round of training with the following steps: i) basecalling signal chunks, preferably, cDNA signal chunks, from step b.4 of clause 19, using the trained CTC-CRF DNA basecaller for DRS of clause 25, and ii) training of CTC-CRF DNA basecaller for DRS using the obtained training dataset of step i) 29. The method according to any of the preceding clauses 26 to 28, wherein step c) comprises optimization of the neural network architecture. 30. The method according to the preceding clauses 26 to 29, comprising the replacement of nucleotide sequences, preferably cDNA sequences with ground-truth from transcriptome alignment, 31. The method according to any of the preceding clauses 1 to 25, wherein the signal- segmentation of step b) comprises identifying events according to T-test score, optionally followed by peak detection. 32. The method according to the preceding clause, wherein the T-test score is determined using two rolling windows wherein the two window sizes are from 2 to 100, preferably 10 to 20 readouts of signal, preferably the size of the two windows is selected from 2 to 10 and 20 to 100, respectively. 33. A method of basecalling the bases of a nucleic acid sequenced using DRS, comprising a DNA basecaller for DRS according to any of the preceding clauses 6 to 32. 34. The method according to clause 33, wherein the nucleic acid is DNA molecule, RNA molecule or a hybrid of two sequenced with Direct RNA Sequencing. 35. A DNA basecaller for Direct RNA Sequencing (DRS) obtained by the method of any of the preceding clauses 6 to32. 36. The use of the DNA basecaller for DRS according to the preceding clause, for demultiplexing nucleic acids sequenced by using DRS. 37. The use of the DNA basecaller for DRS according to the preceding clause, for basecalling single-cell native RNA sequence, and / or demultiplexing single-cell native RNA sequences. 38. The use according to the preceding clause, wherein the RNA comprises one or more barcodes, concatenated sequential barcodes and / or randomly generated barcodes, wherein the barcodes are oligonucleotides. 39. The use according to any of the preceding clauses 36 to 38, wherein the number of barcodes is higher than 96 barcodes. 40. The use of the DNA basecaller for DRS according to clause 35, for basecalling coding RNA such as mRNA and / or non-coding RNA such as rRNA, rRNA, lncRNA, snRNA, miRNA, siRNA, aRNA and tRNA. 41. The use of the DNA basecaller for DRS according to clause 35, for modification calling in an RNA sample. 42. The use according to the preceding clause, wherein the modifications are selected from inosine, pseudouridine, N1-methylpseudouridine, m6A, m6Am, m1Y, m7G, Am, CM, Gm, Um m1A, m3C, m1G, m22G, m3U and combinations thereof. 43. The use of the DNA basecaller for DRS according to clause 35, for basecalling rRNA microbiome profiling. 44. A data processing apparatus comprising means for carrying out the steps of the method of clause 5. 45. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method of clause 5. 46. A computer-readable data carrier having stored thereon the computer program of clause 45. Brief description of the drawings Figure 1. Schematic workflow to generate substrates of high-sequence complexity. HEK293T cells in culture are used to extract total RNA which are then further enriched for poly- (A)-tail containing RNAs. These RNAs are fragmented to ~750nt to obtain a library with a narrow size distribution. Following the removal of 3’ phosphoryl groups and the addition of 5’Phospates the library is poly-(A)-tailed making internal fragments accessible for library preparation. Figure 2. Schematic workflow to generate an RNA004 to sequence mRNA::cDNA chimeras. (A) single stranded RNA, ssRNA (~750nt) containing a 5’-monophosphate and 3’- OH will be ligated to the RT adapter which is provided with the RNA004 kit using T4 DNA Ligase according to the manufacturers instructions. Next reverse transcription is carried out using Maxima H minus RT which will add untemplated nucleotides at the 3’-end (indicated by N). CircLigase II ssDNA Ligase will utilize the RNA 5’-P with the overhand produced from the reverse transcription step to generate an mRNA::DNA chimera. Finally, we will ligate the RL adapter containing the helicase enzyme to make the library amenable to sequencing on an RNA004 flowcell (FLO-PRO004RA). Figure 3. Library preparation for direct RNA sequencing (DRS). (A) DRS library preparation starts with poly-A+ mRNA (or any other polyadenylated RNA). First, DNA barcodes are ligated. Then, cDNA strand is synthesized. Finally, the DNA adapter is ligated, carrying RNA helicase. Only RNA strands pass through the pore. Since sequencing adapters are ligated to 3' end of RNA, the RNA library is always sequenced from 3' end. (B) Only signal corresponding to the RNA part of the read is basecalled (left, white background). The RNA molecule is depicted above its signal: starting from poly-A tail and ending with 5'-cap (from left- to-right). A first step of RNA read basecalling is trimming of the DNA adapter and DNA barcode (top, marked in red). DNA on average generates lower signal values (picoamperes, pA) than its RNA counterpart. In addition, because mRNA ends with a poly-A tail, the signal corresponding to RNA will start with a long, flat region (in green), whose duration is correlated with the length of the poly-A tail. Figure 4: Identification of barcode signal (segmentation) by T-test statistics and barcode classification using DNA basecaller (A) Nanopore signal (blue) corresponding to direct RNA read and the absolute T-test statistics (yellow) using a window of 600 samples are shown. Barcode end is defined as the position with the maximum T-test statistics. Barcode signal is marked in gray. Only the first 30,000 read-outs are shown. (B) Read boundaries detection of ) Nanopore signal (blue) corresponding to RNA::cDNA signal split into RNA and cDNA part; RNA start: signal position (from move table) of the first aligned base (or polyA end detected from T-test) cDNA start: signal position (from move table) of the last aligned base cDNA end: polyT start detected from T-test (C) Event detection zommed view over DNA adaptor - polyA tail (long flat signal) and RNA part of the signal; the lower panel is an even higher magnification. (D) Confusion matrix of the initial model acc=0.464 (E) Confusion matrix of the initial model using high quality calls (best1-best2 >= 3); acc=0.820 (F) Confusion matrix of the refined model all reads; acc=0.569 (G) Confusion matrix of the refined model with high quality calls (best1-best2 >= 3); acc=0.882. Figure 5. Schematic overview of the direct RNA sequencing (DRS) demultiplexing workflow Figure 6: DNA sequencing protocols Typically, only one strand of double stranded DNA (dsDNA) or RNA::cDNA duplex passes through the pore (1D). Both strands can be sequenced if they are physically connected either by hairpin (2D) or using CircLigase. ONT developed an alternative strategy of sequencing both strands that increases the probability of sequencing both strands of the library simultaneously without physically connecting them (1D2) - both strands are sequenced in roughly 40% of molecules. Figure 7: RNA::cDNA chimera sequencing improves basecalling accuracy and modification detection (A) Pair consensus decoding of RNA and cDNA strands will improve basecalling accuracy of RNA. Example from DNA duplex sequencing from DOI:10.1186 / s13059-020-02255-1. (B) By sequencing both strands of RNA::cDNA duplex, features of RNA modification from two sources can be detected, those specific to: cDNA strands (incorporation of "unexpected" bases in cDNA due to altered Watson-Crick pairing of Inosine) and RNA strand (altered Nanopore signal, here depicted as "Current", around modified RNA bases). Combining the two will increase the power and accuracy of RNA modification calling. Figure 8. Sequencing mRNA::cDNA chimeric reads to generate high-complexity DNA basecaller training data. (A) Schematic representation of the library preparation workflow. Starting from a library of short (~750nt) RNAs carrying a 5’-P and 3’-poly-A-tail, a direct RNA sequencing workflow is performed using Maxima H minus RT. The RT enzyme adds several untemplated nucleotides to the 3’-end (commonly, 5’-CCC-3’) which is used in the subsequent reaction to ligate the mRNA to the cDNA strand. Following the addition of the RL adapter the chimeric reads are amenable to sequencing. (B) Annotated raw current signal obtained from a chimeric read generated by the approach of the invention. (C) Scatterplots of cDNA signal versus RNA signal, demonstrating the linear relationship between the two in the approach of the invention(Left), and the absence of a relationship in a control standard DRS run (Right). Figure 9. Barcode Optimization of the DNA basecaller (A) Histogram showing the distribution of edit distances calculated between all 96 barcodes of the optimized barcode set. The set achieves a minimum distance of 16 and a median edit distance of 18. (B) Boxplots showing the Gibbs Free energy (kcal / mol) obtained from the optimized barcode set compared to the control set corresponding to the 96 barcodes used in SeqTagger. Individual barcodes are indicated as points. Values on top of boxplots correspond to the median. Figure 10. Schematic representation of the Split-seq barcoding. SSBC is synonym to SCBC Figure 11 Single-Cell DRS simulation. Histogram showing the distribution of basecalling accuracy for simulated reads with a normally distributed basecalling error of 0.3. Colors indicate two different settings for standard deviation (blue = 0.05, orange = 0.1) Figure 12. Two-pass alignment strategy using a full DNA basecaller on simulated data. (A) Schematic representation of the two-pass alignment strategy implemented to obtain high accuracy predictions of single-cell barcodes from combinatorial barcoding experiments. In the first pass, barcodes are aligned against the entire universe of possible barcode combinations (i.e 963 for a 3x96 barcoding approach) leading to low recovery. Next, the remaining barcodes are aligned against a reduced set of references (i.e. those barcodes observed in the initial pass), yielding high recovery with high precision. (B) Distribution of identified barcodes and their corresponding assigned reads for the first-pass (Left), and second-pass (Right) alignment on a simulated dataset Figure 13. Coupling semi-permeable capsule technology (SPCs) to single-cell compartmentalization to direct RNA-sequencing. (A) Schematic representation of the workflow used to obtain DRS libraries from encapsulated single-cells. (B) Representative fluorescent confocal images of K562 cells encapsulated in SPCs with different stainings (Left). The scale is indicated on the bottom right. (C) Barplot representing the overall occupation efficiency obtained from a single encapsulation experiment (see Methods). (D) Boxplot representing the overall size distribution of reads obtained for human 18s and 28s ribosomal RNA following SPC encapsulation and library preparation. The median size is indicated to the left of each plot. Sample size (n) is indicated at the bottom of each plot. Figure 14. Scheme of barcoding design CPT means capture. Examples Materials and Methods Generating a DRS training dataset with high sequence complexity To generate a complete DNA basecaller we need training data that provides large sequence complexity. To achieve this, we use human poly-(A)-selected material from HEK293T cells. Subsequently, the material is fragmented to ~650 to 750 nt to achieve a narrow library size and ensure more even coverage across different sequences. The fragmented library is dephosphorylated using Fast-AP in order to remove 3’ phosphoryl groups that otherwise prevent the subsequent ligation of the RT adapter and genomic DNA is removed using Turbo- DNAseI. Next, T4 Polynucleotide Kinase will be used to obtain 5’-Phosphates required for mRNA::cDNA chimera formation. Finally, the library is treated with Poly-(A)-Polymerase to achieve short poly-(A)-tails also making internal fragments amenable to library preparation (Fig 1). Generating mRNA::cDNA chimeras to train a complete DNA basecaller Once a library of ssRNA molecules with a 5’-P and a 3’poly-(A)-tail with high sequence complexity is produced, we proceed with generating mRNA::cDNA chimeras followed by direct RNA sequencing library preparation. To this end, we will first add the standard RT adapter used in the DRS workflow and ligate it to our fragmented RNA library (using concentrated T4 DNA Ligase). Next, we will perform a same tube reverse transcription reaction using Maxima H minus reverse transcriptase as this enzyme has a high rate of incorporating untemplated nucleotides at the 3’-end (preferentially a CCC overhang is generated). The overhang will be exploited in the subsequent step using CircLigaseII ssDNA Ligase to generate the mRNA::cDNA hybrid molecule. Finally, the chimeras will be ligated to the RL adapter containing the helicase enzyme followed by sequencing using an RNA004 flowcell (FLO-PRO004RA) (Fig 2). In summary, In the process of library preparation to generate a DRS training dataset with high sequence complexity, RNA molecules are extended twice: first with a DNA barcode (with a part complementary to poly-A tail of mRNA), and subsequently with a DNA adapter (with a part complementary to DNA barcode) carrying a helicase needed for sequencing. This is followed by complementary DNA strand (cDNA) synthesis in order to linearise RNA molecules, which greatly reduces pore blocking by folded RNA molecules (Fig 3A). Normally, only the RNA strand of RNA::cDNA duplex is sequenced during DRS. Moreover, only the RNA part of the RNA strand is analyzed: DNA barcode and DNA adapter are trimmed before basecalling (Fig 3B). Unlikely the traditional method, after the cDNA synthesis we further proceed with ligating the ends of RNA and cDNA creating one hybrid RNA::cDNA molecule, allowing passing of the cDNA through the pore thus, enabling the sequencing of the cDNA along with RNA. Next, a correlation of each cDNA signal of the Direct RNA Sequencing data to a nucleotide by signal- segmentation to obtain a DRS training dataset, was established according to steps b.1; b.2, b.3 and b.4 disclosed above, briefly: b.1 basecalling a read with the RNA reference, in this case HEK293T cells mRNA b.2 normalization b.3 identifying events in cDNA signal b.4 generation of signal chunks, signal-sequence pairs and using the positions of events normalized by the total number of bases in a reference to predict the sequence of nucleotides for any signal chunk generated, so the DRS training dataset is generated Training the DNA basecaller for DRS The training algorithm, bonito (https: / / github.com / nanoporetech / bonito / ), was used to train the DNA basecaller for DRS model. Bonito requires as input signal chunks of fixed length and their corresponding sequences. The CTC-CRF models were trained with bonito v0.7.2. We used a window of 31 and stride of 10. We tested three model versions with an increasing number of features used by the encoder (and model parameters): 96 (fast with 519,880 parameters), 384 (hac with 6,499,048 parameters) and 768 (sup with 24,793,192 parameters). For DNA basecaller for DRS we use sup architecture. Bonito is written in Python and uses pytorch internally, therefore it's straightforward to incorporate it in a larger Python program. This allows us to select which part of the signal we want to basecall, normalise it according to particular needs, etc. (These are not easily available when using other basecallers such as guppy or dorado). A complete DNA basecaller was trained with a DRS training set. When training RNA basecaller with remora aligned signal-to-sequence, the resulting model achieves 93.2% (fast) and 96.4% (sup) median identity on the validation set. When training a RNA basecalling model with the same set of reads, but signal-to-sequence aligned using our event detection method, we obtained 90.9% (fast) and 91.5% median identity on the validation set. (Fig.4 A-C) Evaluation of DNA basecaller in demultiplexing using SCBC barcodes 1-24 from SeqTagger An initial DNA basecaller was trained with RNA::cDNA originating mRNA isolated from HEK293T cells. The initial model was refined as described in paragraph above. In order to evaluate the demultiplexing accuracy of DNA basecaller, we used SCBC barcodes (Single Cell Barcodes) 1-24 (10,112 reads in total) as disclosed in Pryszcz et al, 2024 and materials and methods as disclosed therein. We followed standard Pryszcz et al, 2024 demultiplexing procedure (read segmentation, signal normalisation, barcode basecalling, barcode alignment and filtering), but used DNA basecaller instead of SeqTagger model for barcode basecalling. The design of the 24 barcodes used in the current experiment was performed as disclosed in Pryszcz et al, 2024. However, a skilled person would understand that many other barcodes can be used, for example, below we disclose the 96 barcodes that were used for the training of the SeqTagger, but others barcodes are covered under the scope of the invention while they have a similar structure to them and taking into account the description of Pryszcz et al, 2024. SEQ ID No 1: N1N(15-45)TAGTAGGTTC. The N(15-45) is the barcode region, that may be as well, 15 to 37 nucleotides, or 20 to 37 nucleotides or longer as indicated above. N1 can be A, C, G or T, preferably A or G. It is advantageous that all the first adapter DNA oligonucleotide with different barcodes used in the same experiment comprise the first N1 nucleotide to avoid ligation bias difference across the barcodes. This is, to make a fixed final nucleotide in the first adapter DNA oligonucleotide that can be ligated to the RNA molecule The first adapter DNA oligonucleotide requires 5Phos (5´monophosphate group) to do efficient ligation to the RNA molecule. The 5´N1is added to all oligos to make ligation equally efficient (ligation bias is known to occur based on the identity of the last base). SEQ ID No 2 GN(15-45)TAGTAGGTTC. (SEQ ID No 2) The N(15-45) is the barcode region, that may be as well, 15 to 37 nucleotides, or 20 to 37 nucleotides or longer as indicated above. Wherein “N” is a nucleotide selected from the group consisting of A, C, T, G and will contain the barcode. SEQ ID1 and SEQ ID 2 are the general structures of Oligo A The general structure of the second adapter DNA oligonucleotide for multiplex run is: GAGGCGAGCGGTCAATTTTN(15-45) N2(10-20). (SEQ ID No 3) The N(15-45) is the barcode region, that may be as well, 20 to 37 nucleotides or longer as indicated above. Wherein “N” is a nucleotide selected from the group consisting of A, C, T, G and will contain the barcode. The barcode region of the first and second adapter DNA oligonucleotide needs to be complementary The N2(10-20) is a string of nucleotides complementary to the polyN tail added to the RNA molecule, this is, for example, complementary to as polyinosine (polyI) tails, this is a poly C tail or complementary to polyA (polyadenosine) polyC, polyG, polyU, or polyG / I, polyinosine (polyI) or 2’-O-methylated (polyNm) tails In a particular embodiment, Oligo B comprises a string of 10 “T” which is the poly A hybridization region. In a particular embodiment, the second adapter DNA oligonucleotide is SEQ ID No 4 GAGGCGAGCGGTCAATTTTN(15-45)TTTTTTTTTT SEQ ID 3 and SEQ ID 4 are the general structure of oligo B. In Table 1 the set of 24 pairs of oligonucleotides that are used in the current experiment are disclosed. However, a skilled person would have no problem to design an unlimited number of barcodes pair having regarded to the sequences SEQ ID No 1, SEQ ID No 2, SEQ ID No 3, and SEQ ID No 4, and the particular examples disclosed in Table 1. The first 8 barcodes (8pairs of oligonucleotides A / B) of the 24 set of barcodes (24 pairs of oligonucleotides A / B) disclosed below are the same sequences that were disclosed in the b12_RNA002 experiment in Pryszcz et al, 2024, together with the BC-01, BC-02, BC-03, BC-04 disclosed therein. Table 1 oligonucleotides pairs with barcodes used in the SCBC The following table 2 discloses a set of 72 pairs of Oligo A and Oligo B barcodes that can be used for demultiplexing. These 72 pairs can be used together with the 24 pairs of Table 2 to obtain a 96 barcodes model. However, any other combination is possible. In embodiments, the Oligo A and Oligo B used in the method are having at least 80%, preferably at least 85%, more preferably at least 90 % and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % sequence identity with respect to the sequences listed in table 1 and 2, while maintaining the complementarity between the sequences of the barcodes. A skilled person would understand this is applicable to the method disclosed in Pryszcz et al, 2024. Table 2 Examples of oligonucleotides pairs disclosed The following table 3 discloses a set of 96 barcode sequences that can be used for demultiplexing. These barcode sequences are integrated into oligo A whereas their antisense sequences integrated in to oligo B as described in SEQ ID 1, 2 and SEQ ID 3, 4 respectively. In embodiments, the barcodes used in the method are having at least 80%, preferably at least 85%, more preferably at least 90 % and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % sequence identity with respect to the sequences listed in table 3. The following barcodes are those disclosed in Oligo A sequences of Tables 1 and 2, the complementary sequence of these barcodes is disclosed in Oligo B sequences of said Tables Table 3 Examples of Barcodes Initial DNA basecaller achieved 0.464 accuracy when demultiplexing SCBC (Single Cell Barcode) barcodes 1-24 (Fig.4D). The accuracy is 0.820 when only highly confident predictions are taken into account (Fig.4E), but at a cost of losing filtering. The filtering used was a follow: keep only barcode predictions (classifications) for which the best barcode alignment has an alignment score at least 3 or more higher than the alignment score of the second the best barcode. For example, if a given read has a predicted barcode sequence that matches SCBC1 with a score of 15 and SCBC2 with score of 12, we would keep this prediction. But if it matches SCBC1 with score 15 and SCBC2 with score of 13 it would not be kept. We observed a further improvement of the demultiplexing accuracy after training set and DNA basecalling model refinement (4E-F) to 0.569 when all reads were classified, and to 0.889 when only high quality predictions were kept (at the cost of losing over half of the reads). The refinement comprises one iterative training of the basecaller performed as described in the description. Importanty, SeqTagger can demultiplexed only aforementioned barcodes, while DNA basecaller are able to demultiplex any arbitrary barcode sequences. Refined model is able to classify all reads with accuracy of 0.569 or over 40% of reads with 0.889 accuracy when only high-quality results are considered. The DNA basecaller for DRS can be used to further refine a training set by basecalling training chunks with the first DNA basecaller for DRS, align resulting sequences onto reference transcripts and replace original training sequences with those matched by basecalled chunks. This certainly results in better signal-to-sequence alignment than from the initial event detection algorithm. DRS demultiplexing via DNA barcode basecalling In Pryszcz et al, 2024, the training dataset used was comprising the Oligo A and B sequence pairs given in table 1 and table 2 and the method was able to identify only these barcodes and thus demultiplex the RNA samples that are barcoded only these preselected 96 barcodes. In embodiments, the oligo sequence pairs used in Seqtagger method are having at least 80%, preferably at least 85%, more preferably at least 90 % and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % sequence identity with respect to the sequences listed in table 1 and 2 of the present application. The barcode sequences of the oligoA and oligoB sequences are listed in table 3. The training dataset of method disclosed in Pryszcz et al, 2024 comprises one or more barcode sequences listed in table 3. The barcodes used in the method can have at least 80%, preferably at least 85%, more preferably at least 90 % and even more preferably 91 % or 92 % or 93 % or 94 % or 95 % or 96 % or 97 % or 98 % or even up to 99 % sequence identity with respect to the sequences listed in table 3. In a particular embodiment, the method of Pryszcz et al, 2024 can use 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33,34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, or 96 barcodes of those listed in Table 3 of the present application. This application is the similar of the Pryszcz et al, 2024, but the trained DNA basecaller of the inventions allows us to use any arbitrary DNA sequence as a barcode (or combination of multiple barcodes concatenated) instead of using a limited set of barcode (4, 12, 96 and so on). This is important for single-cell sequencing, where a given set of barcodes (let’s say 96) can be randomly concatenated resulting in exponentially increasing pool of potential barcodes (for example, when concatenating randomly 1 out of 96 barcodes 4 times we will generate a pool of nearly 85 million combinations). And while It’s impractical to train a demultiplexing models for 85 million potential combinations of 96 barcodes, the DNA basecaller for DRS is able to demultiplex such cases. Generating mRNA::cDNA chimeras to train a full DNA basecaller: training a DNA basecaller for Direct RNA Sequencing Currently, there are no manufacturer-provided barcoding kits for direct RNA sequencing.. To this end, we have developed a method of training a DNA basecaller for Direct RNA Sequencing by using a training dataset of high sequence complexity, and the matched ground-truth RNA sequence. Therefore, we first extracted poly-A-selected RNA from HEK293 cells and performed several biochemical preprocessing steps to generate a dataset of high sequence complexity (see Methods, below). Next, we developed a custom direct RNA sequencing library preparation workflow, which generates chimeric mRNA::cDNA reads providing matched RNA::DNA nanopore sequencing data required for training (see Methods, below), (Fig. 8A). We first qualitatively validated the generation of chimeric reads by inspecting the raw current intensity signal produced. This revealed that the majority of the reads ended in a 5-’sequence with a lower average current intensity signal (pA) than the RNA read which corresponds to the cDNA sequence, as well as a lowly variable region at the 5’-extremity corresponding to the poly-dT sequence generated through reverse transcription of the poly-A-tail (Fig. 8B). Next, we quantitatively assessed the relationship between cDNA and mRNA signal, demonstrating that ~60% of the sequenced reads carried a cDNA sequence, as compared to a control standard DRS library preparation in which less than 1% were identified as chimeric (Fig.8C). Moreover, a linear relationship between the mRNA and cDNA signal length was observed in Chimera-seq, which was absent from the control sample, and suggests that the DNA signal indeed corresponds to the complementary DNA ligated to RNA (Fig.8C). Following a stringent filtering step the generated data was used for training of a full DNA basecaller (see Methods, below). We tested the resulting model on previously obtained sequencing data of SCBC1-24, Table 1 and observed moderate accuracy (= 0.569) when considering all reads and good accuracy (= 0.882) when only high confidence predictions were considered which came at the cost of sacrificing >50% of reads (Fig.4F). 2. Single-cell demultiplexing Design of distinct DNA barcodes The DNA-basecaller of the invention show improved performance with DNA barcodes that maximize the sequence difference between each other. We performed a greedy search on a limited number of input barcodes (10,000) resulting in a set of 96 barcodes with 18 median Levenstein distance (~40% difference). A Second and third set of 96 barcodes were generated by reversing and taking every other base of the first set, respectively. Fig 9 (A, B) Thereby creating three independent sets of 96 barcodes that can be used in a sequential barcoding approach. An alternative simplified barcode design is also used to barcode single cell samples as shown in figure 14 and detailed under the section “Sequencing mRNA::cDNA chimeras using direct RNA sequencing”. Instead of computational generation, we used three predefined sets of 96 barcodes (30 nt each) that had been experimentally validated for optimal performance with the DNA basecaller described herein. Two-pass demultiplexing approach to sequence single-cell RNA We then used a two-pass strategy to demultiplex single-cell (combinatorial) barcodes after the split-pool following the divide-and-conquer principle. DNA basecaller for DRS has low accuracy (70%) therefore it is challenging to assign most of the basecalled barcodes (reads) uniquely to one of 884,736 (96^3) possible barcode combinations. However, we expected to see only a small fraction of theoretically possible barcodes actually sequence: the number of observed barcodes is similar to the number of sequenced cells (typically 10,000-20,000). As indicated above, step 3 of the method comprised demultiplexing the DRS signal obtained from the sequential concatenated DNA barcodes using a DNA basecaller, the demultiplexing comprising: a. detecting a subset of sequentially concatenated DNA barcode combinations generated during split-pool barcoding b. assigning basecalled reads to the detected subset of barcode combinations. In the first step (step 3(a)) of the method of the invention relates to single-cell RNA sequencing, we only detect a set of barcode combinations generated during split-pool barcoding and in the second step (step 3(b)) we assign basecalled reads to the observed subset of all possible barcode combinations. In the first step, basecalled barcode sequence is aligned onto every set of barcode references (3*96), keeping only the best match at every splint and reporting the best combination of splints if best at every splint is above a certain threshold (mindiff) than the second best match for that splint (Figure 10A. Here we found mindiff equal or above 7 to work best. For example, read 1 matching barcode 1 at the first splint, barcode 9 at the second splint and barcode 94 at the third splint would be depicted as 1.1|2.9|3.94. After all reads are mapped, we plot a histogram of barcode combinations occurrences and keep only those that were detected in at least N reads where N is 3% of expected number of reads per barcode. So if we expect to sequence 1,000 reads per cell / barcode, N=30. (In general, N>=3 works well for 100 as well as for 1,000 reads.) Mindiff a difference in edit distance between basecalled read and barcode sequence. In the second step, we again align every basecalled barcode onto every set of barcode references (3*96) keeping all matches at every splint with a certain edit distance to the best match (maxdiff). We found here that edit distance equal or smaller than 5 works very well (Figure 10B). Subsequently, we generate all possible combinations of barcodes from matched barcodes at every splint, but keeping only those that were observed in the previous step (Figure 10A). Finally, we align basecalled barcode onto full barcode combinations and report the one with the highest score only if the best score is above certain edit distance than the second best score (mindiff). Here we found mindiff equal or larger than 7 to work best (Figure 10C). Simulation #1: 20k cells with 1k reads per cell First, we assigned 20,000 cells to 884,736 (963) possible barcode combinations obtaining 19,784 unique barcodes. We observed 1.08% collisions, meaning 216 barcodes were assigned to more than one cell (this is expected to happen by chance in splint-seq). Subsequently, we simulated 1,000 reads per cell, generating in total 20 million reads (20k cells x 1k reads) and introducing 30% errors distributed normally with standard deviation of 0.1 or 0.05. First-pass of demultiplexing recovered all expected barcode combinations, while second-pass correctly assigned 72% of reads to those barcodes with 0.984 precision. The method is fast, taking 15 minutes and 4 hours for first and second pass on 16 cores, respectively. Simulation #2: 2M cells with 100 reads per cell In order to test if the method could be scaled-up for demultiplexing of more cells / reads, we repeated the simulation increasing the number of cells to 2 million (factor of 100) and decreasing the number of reads per cell to 100. To handle more cells, we introduced a fourth splint barcode (reverse sequence of third barcode), increasing the number of possible barcode combinations to 84,934,656 (964). Assigning 2 million cells to these barcodes resulted in 1,976,754 unique barcodes (1.12% collision). This time 200M reads were simulated introducing 30% error (with standard deviation of 0.1). First pass detected all expected barcodes, while second pass assigned 66% of reads at 0.986 precision. First pass is still relatively fast (9h), but the second pass becomes a bottleneck (10k reads in 12 minutes). With the trained DNA basecaller of the invention and using refined training sets the accuracy of the event detection is improved. The refinded training set can be obtained by using new signal-to-sequence alignments. This improvement leads to Kmer model with a basecalling accuracy similar (or superior) to RNA basecalling (~97%). A full DNA basecaller accurately demultiplexes simulated single-cell direct RNA sequencing data (scDRS) In order to test whether the full DNA basecaller could indeed be used for demultiplexing scDRS data, we first had to generate a set of barcodes that could be used in a combinatorial barcoding approach such as the one described in SPLiT-seq (Rosenberg, A. B. et al. Single-cell profiling of the developing mouse brain and spinal cord with split-pool barcoding. Science (2018) doi:10.1126 / science.aam8999). Therefore, we generated a set of 96 barcodes, starting from the existing barcodes used by SeqTagger, by maximizing their edit distance, and minimizing secondary structure formation. This yielded a set of barcodes with a median edit distance of 18, corresponding to a median sequence difference of ~40%, with reduced secondary structure formation (Fig 9). A Second and third set of 96 barcodes were generated by reversing and taking every other base of the first set, respectively. Thus, creating three independent sets of 96 barcodes that can be used in a sequential barcoding approach. Furthermore, since real-life scDRS is currently unavailable we had to rely on in silico simulations to test the demultiplexing capacity of our full DNA basecaller. To this end, we assigned 20,000 cells to 884,736 (963) possible barcode combinations obtaining 19,784 unique barcodes. We observed 1.08% collisions, meaning 216 barcodes were assigned to more than one cell which is expected in combinatorial barcoding approaches. Next, we simulated a sequencing run with 1,000 reads per cell, generating a total of 20 million reads (20k cells x 1k reads) while introducing 30% basecalling error, corresponding to the expected error rate of the DNA basecaller, distributed normally with standard deviation of 0.1 or 0.05 (Fig. 11). Finally, we performed a two-pass alignment approach to account for the high basecalling error obtained by the DNA basecaller, in which the first alignment is performed against all possible barcode combinations (963), and a second-pass in which non-aligned reads are remapped against the observed barcodes (see Methods, below), (Fig.12A). The result showed that the first-pass of demultiplexing recovered all expected barcode combinations (precision = 1.00) while discarding 75% of reads, with the second-pass correctly assigned 72% of reads to those barcodes with 0.984 precision (Fig. 12B). Together, these results demonstrate the feasibility of our general DNA basecaller to perform demultiplexing of scDRS data with high precision and recovery. Semi-permeable capsules provide single-cell compartmentalization compatible with DRS Currently, no experimental protocols exist for performing combinatorial barcoding of single- cells coupled to DRS. Most of the sequencing-by-synthesis based approaches rely on compartmentalization through permeabilized cells Rosenberg, A. B. et al.,. To this end, cells are fixed and permeabilized with reagents such as formaldehyde, which cause extensive crosslinking of the cellular RNA and therefore negatively affect long-read sequencing protocols. Therefore, we explored other ways of compartmentalization that retain full-length reads while providing a single-cell compatible workflow. For this purpose, we tested the compatibility of direct RNA sequencing with a system relying on semi-permeable capsules (SPCs), which are formed around single-cells using photo-inducible chemicals (Doméjean, H. et al. Controlled production of sub-millimeter liquid core hydrogel capsules for parallelized 3D cell culture. Lab Chip 17, 110–119 (2016); and Leonaviciene, G., Leonavicius, K., Meskys, R. & Mazutis, L. Multi-step processing of single cells using semi-permeable capsules. Lab Chip 20, 4052–4062 (2020)). (Fig.13A). First, we tested the encapsulation efficiency of SPCs on a human lymphocytic cell line (K562). To this end, we performed the encapsulation workflow coupled to live cell staining using Syto- 13 and Propidium Iodide to stain live and dead cells, respectively (Fig.13B). Furthermore, to more systematically assess the encapsulation efficiency a random set of ten images was analysed and the amount of occupied versus empty cells was determined. This yielded an overall encapsulation efficiency of 4% which was slightly below the expected efficiency of 10% suggested by the manufacturer. Next, to test whether 20.000 single-cells encapsulated in SPCs could be sequenced with DRS we modified the library preparation protocol such that both the ligation of the RT adapter and the reverse transcription reaction are performed in encapsulated cells. Furthermore, as poly-A-tailed RNA will be sparse in 20.000 cells, and no amplification reaction is performed, we first used sequence-specific RT-adapters targeting the highly abundant 18s and 28s ribosomal RNAs. Said RT-adapter comprised a 3´sequence that anneals with the target RNA, GACAAACCCT for 18s (SEQ ID No 597) and TAATGATCCTT for 28s (SEQ ID No 598) To our surprise, this protocol yielded libraries that were sequenced efficiently on a PromethION flowcell. Moreover, analyzing the size distribution of the two investigated transcripts showed that long reads could be retrieved from both 18s and 28s, with a median size distribution of approximately 880nt for both (Fig. 13D). Furthermore, a substantial fraction of the 18s transcripts were full length reads (Fig. 13D). Collectively, these results demonstrate the compatibility of DRS with semi-permeable capsule technology to obtain single-cell compartmentalization, while providing long-read information. Furthermore, they showcased the feasibility of sequencing the two major ribosomal RNA molecules from as little input as 20.000 cells. Materials and methods for the DNA basecaller, single-cell demultiplexing experiments and Semi-permeable capsules Mammalian cell culture Human embryonic kidney cells (HEK-293, ATCC #CRL-1573) were cultured in DMEM (Thermo Fisher, 10566016) media supplemented with 10% fetal bovine serum (FBS) and 1X penicillin / streptomycin (ThermoFisher, #15140122). All cells were grown at 37 C with 5% CO2. Cells were passaged every 2-3 days using a 1:8 dilution, following treatment with 1X Trypsin- EDTA (Sigma Aldrich, T4174) for one minute. Cell counts and viability were regularly assessed using the Countess Automated Cell Counter (Invitrogen). Human lymphoblast cells (K562, ATTC #CCL-243) were cultured in RPMI 1640 (Thermo Fisher Scientific, 11875093) supplemented with 10% fetal bovine serum (FBS) and 1X penicillin / streptomycin (Thermo Fisher Scientific, #15140122). Cells were passed every 2-3 days and kept in suspension using T-75 flasks (Thermo Fisher Scientific, #156800). Total RNA extraction After removing the culture media from each plate (15cm, corresponding to ~20*10^6 cells), cells were resuspended in 10 mL Trizol (Invitrogen,15596018) and aliquoted into 1.5 mL DNA LoBind tubes (1 mL each). Following incubation for 5 minutes at room temperature, each sample was vigorously vortexed for ~10 seconds. Next, 200 μL of chloroform (Sigma, C2432) was added to each sample, mixed, and incubated for 5 minutes at room temperature followed by centrifugation at 12,000 g for 5 minutes. The resulting upper aqueous phase (~400 μL) was transferred to a new tube and mixed with 500 μL of molecular grade 2-propanol (Sigma, I9516). Samples were vortexed and incubated for 15 minutes on ice, followed by centrifugation at 20,000 g for 15 minutes at 4ºC to precipitate the RNA. The pellet was washed twice with 1ml of 70% ethanol and centrifuged at 20,000 g for 5 minutes at 4ºC. After removing residual ethanol, the pellet was air-dried for 3 minutes and resuspended in nuclease-free H2O (Invitrogen, 10977035). Concentration was measured using the Qubit 4 Fluorometer (Invitrogen, Q33226), and RNA integrity was assessed using the TapeStation 4150 (Agilent, G2992AA). A-selection 75 μg of total RNA (in multiple replicates) was subjected to double poly-A-selection using Dynabeads Oligo(dT)25 (Invitrogen, 61002) following manufacturer’s protocol and eluted in nuclease-free H2O (Invitrogen, 10977035). Concentration was measured using the Qubit 4 Fluorometer (Invitrogen, Q33226), and the distribution of poly-A-selected RNA was assessed using the TapeStation 4150 (Agilent, G2992AA). Generating RNA libraries containing high sequence complexity Next, 1.5 μg of poly-A-selected material (in multiple replicates) was fragmented to a median size of approximately 750 nt using 1X RNA Fragmentation Buffer (New England Biolabs, E6150S). Samples were incubated at 95 C for 1 minute followed by the addition of 1X RNA Fragmentation Stop Solution (New England Biolabs, E6150S). RNA samples were cooled to 4 C before column clean-up using the RNA Clean and Concentrator-5 kit (Zymo Research, R1014). The fragmented poly-A-selected RNA was treated with fast alkaline phosphatase (Thermo Scientific, EF0651) for 15 minutes at 37 C in a total volume of 37 μL to remove any 3’-phosphoryl groups generated during fragmentation. Additionally, the reaction was supplemented with 1 μL of TURBO DNAse (Thermo Scientific, AM2238) to remove any residual genomic DNA followed by column purification using the RNA Clean and Concentrator- 5 kit. Next, RNA samples were treated with 1 μL of T4 Polynucleotide Kinase (New England Biolabs, M0201S) in a total volume of 50 μL following the manufacturer’s protocol, skipping the heat inaction step to avoid further RNA fragmentation. Samples were column purified using the RNA Clean and Concentrator-5 kit. Finally, a short poly-A-tail was added to the fragmented RNA using 1 μL of E. coli Poly(A) Polymerase (NEB, M0276S), following the manufacturer’s protocol but reducing the incubation time at 37 C to 2 minutes. The final RNA library consisting of poly-A-selected material fragmented to 750 nt containing a 5-phosphate group and a 3’-OH was column-purified using the RNA Clean and Concentrator-5 kit (Figure 1). Sequencing mRNA::cDNA chimeras using direct RNA sequencing A previous report has demonstrated the ability of CircLigase II ssDNA Ligase (Biosearch Technologies, CL9021K) to efficiently ligate annealed RNA / DNA strands generating mRNA::cDNA chimeras (Despic, V. & Jaffrey, S. R. mRNA ageing shapes the Cap2 methylome in mammalian mRNA. Nature 614, 358–366 (2023).). We envisioned that mRNA::cDNA chimera formation could be coupled to direct RNA sequencing enabling the simultaneous sequencing of mRNA and its corresponding cDNA strand, providing matched signal-to- seqeunce data (Fig.8A-B). To this end, 700ng of RNA library generated in the previous step were mixed with NEBNext Quick Ligation Reaction Buffer (New England Biolabs, B6058S), 1.5 μL of T4 DNA Ligase, concentrated (New England Biolabs, M0202T), 1 μL of RNase Inhibitor, Murine (New England Biolabs, M0314S) and 1 μL of 1.4 μM pre-annlead sequencing oligos (SCBC-01, oligoA+B) in a total volume of 15 μL. The reaction was incubated at room temperature for 15 minutes before adding 13 μL of nuclease-free H2O, 2 μL of 10mM dNTPs (New England Biolabs, N0447S), 8 μL of Maxima RT Buffer and 1 μL of Maxima H Minus Reverse Transcriptase (Thermo Fisher, EP0751). The reverse transcription reaction was incubated at 50C for 15 minutes without subsequent heat inactivation. RNA samples were cleaned up using 1.2X Agencourt RNAClean XP beads (Beckman Coulter, A63987) and washed using freshly prepared 70% molecular-grad ethanol (Merck). For elution, the RNA was incubated with 18.5 μL of CircLigase Reaction Mix (generated according to the manufacturer’s protocol) for 5 minutes. Following the addition of 1 μL of CircLigase II ssDNA Ligase and 0,5 μL of RNase Inhibitor, Murine, the reaction was incubated for 1 hour at 60C. The Generated mRNA::cDNA chimeric library was cleaned up using 1.2X Agencourt RNAClean XP beads, washed using freshly prepared 70% molecular-grade ethanol, and eluted in 20 μL of nuclease- free H2O. Finally, the RNA ligation adapter (Oxford Nanopore Technologies, SQK-RNA004) was attached following the manufacturer’s protocol for RLA ligation (direct-rna-sequencing- sqk-rna004-DRS_9195_v4_revB_20Sep2023-minion). Subsequent bead clean-up was performed using 0.8X Agencourt RNAClean XP beads and all remaining steps were identical to the aforementioned protocol by Oxford Nanopore Technologies. Chimeric mRNA::cDNA libraries were loaded onto a primed RNA flowcell (FLO-MIN004RA) and sequenced using a MinION sequencer with MinKNOW acquisition software version >24.06.12. Barcode design and diversity optimization To generate a library of barcode sequences with high sequence diversity and constrained base composition, we developed a custom R script. The algorithm produced sequences of defined length (= 30nt) while avoiding homopolymeric stretches (> 3nt) and maximizing pairwise dissimilarity based on the Levenshtein (edit) distance. Each sequence was assembled by randomly sampling and shuffling repeated k-mer elements to match the desired sequence length. To ensure maximal pairwise sequence diversity, an iterative selection process was employed (greedy approach). All sequence generation and analysis steps were performed in R (v4.3.2) with a fixed seed for reproducibility. In other experiments, we used simplified barcode sequence design approach. Instead of generating barcodes computationally, we used a predefined three sets of 96 sequences containing 30-nucleotide barcodes that have been validated for optimal performance with the DNA basecaller of the disclosure. These barcodes correspond to the most accurately predicted 30-mer sequences from the DNA basecaller of the disclosure, distributed evenly across the three sets (lists) (PLATE-1, PLATE-2, and PLATE-3) to ensure balanced performance within and between plates. The most accurately prediction means in this context that the bases in these barcode sequences are basecalled correctly compared to the other barcode sequences tested using the DNA basecaller of the disclosure. 30 nt barcodes comprises additional sequences to accommodate the target capture (capture oligos) and in general split-pool seq barcoding. The 3'end sequences are needed for splint ligation with the subsequent Plate. To give an example, Plate-1 has a 3-end sequence (15nt), and Plate-2 has a 5-end sequence of (15nt), we then use so called splints which are complementary oligos (30nt in length) That span this region (3-end of Plate-1 and 5-end of Plate-2) to enable efficient ligation of those two sequences. The same concept is used to fuse Plate-2 with Plate-3 barcodes. Finally, Plate-3 carries a 3-end sequence that is complementary to the RLA adapter that is the oligo containing the helicase enzyme provided by ONT with their dRNA-seq kits. Besides the overlaps described in the above point, each barcode contains a 30nt sequence that was selected as those 30nt sequences that were most accurately predicted by Chimera-seq. That is the defining main feature. An illustration of the barcode design is shown in figure 14. Table 4: Plate 1 sequences Table 5: Plate 2 sequences PBC- / 5Phos / CTC TTTCGTTCTTGAT- 4 AGTCGT / 5Phos / CTCAAGCACGTG- 134 AAGCACGT TAATGAAAACA- 4 ACGCCG GATTTTCGTTCTTGATTAATGAAAACA- Table 6: Plate 3 sequences PBC- / 5Phos / CGAA GGTGGGGCA- 5 TA- / 5Phos / CGAAACATCGGC- 215 ACATCGGC- GCGGGAACTT- 2 GTAG CACGGTGGGGCAGCGGGAACTT- PBC- / 5Phos / CGAA ATTCAATGTA- 5 TA- / 5Phos / CGAAACATCGGCCACATTCAA- PBC- / 5Phos / CGAA CTGAGCGTCTGA- 5 TA- / 5Phos / CGAAACATCGGCCAC- 258 ACATCGGC- GATGTTAGTAT- 6 GTAG CTGAGCGTCTGAGATGTTAGTATTAGT- Preannealing of custom RT adapters Custom RT adaptors (IDT) were annealed using the following conditions: custom oligo A and - In other experiments, we used a different adapter design. A single oligonucleotide, referred to as “capture oligonucleotide.” or CPT oligo. This oligonucleotide comprises a “5´sequence” and a “3´sequence” linked. The 5´ sequence used (TTGACGCGTCTCGGA SEQ ID No 307) can basepair to 5’ end of all 96 barcodes within a set. Three capture oligonucleotides with different 3´ sequences were tested: one for poly-A–tailed RNA where the capture oligo comprises a poly-T tail (5’-TTTTTTTTTT-3’ SEQ ID No 599) in its 5’ end, one targeting 18S human rRNA (sequence to anneal to 5’ of the barcodes: 5’- TTGACGCGTCTCGGA-3’ and sequence to anneal to 3’ end of the barcodes: 5’- GACAAACCCT-3’ SEQ ID No 597), and one targeting 28S human rRNA (sequence to anneal to 5’ of the barcodes: 5’- TTGACGCGTCTCGGA-3’ and sequence to annealed to 3’ end of the barcodes: 5’-TAATGATCCTT-3’ SEQ ID No 598). Single-cell compartmentalization of human lymphocytic cells To obtain compartmentalized single-cells 1.5ml of confluent K562 cells (80-90% confluency), were centrifuged at RT (300g) for 5min. Following removal of the supernatant, cells were resuspended in 1ml of complete RPMI 1640 culture media. To quantify the cells, 10μl of cell suspension were mixed with 10μl of Trypan Blue solution (Thermo Fisher Scientific, 15250061) and cell counts and viability were assessed using the Countess Automated Cell Counter (Invitrogen). Next, cells were centrifuged again at RT (300g) for 5min and resuspended in complete RPMI 1640 culture medium to a final concentration of 3x10^6 / ml cells. Next, a volume containing 200.000 cells is transferred to a fresh 1.5ml DNA LowBind tube and filled to 36.5μl using 1X PBS. Next, the semi-permeable capsule mix is generated following the manufacturer’s instructions, and loaded into a SPC Generation Chip placed on a FLUXTM machine (DGPM02323108001_FLUX_SPC_Generator_UG_V6.3.pdf). Next, SPC recovery, emulsion breaking, and buffer exchange was performed according to the manufacturer’s instruction. Subsequently approximately 100μl of SPCs were retrieved. Live-dead staining to evaluate encapsulation efficiency To perform live-dead staining of encapsulated K562 cells, 2μl of SPC solution were diluted 10- fold in WB2 (10mM Tris-HCl pH 7.6, 0.1% Triton X-100 in ddH20), and supplemented with 0.2μl SYTO-13 (S7575, Thermo Fisher Scientific), and 0.25μl Propidium Iodide (P3566, Thermo Fisher Scientific). Samples were incubated for 30min in a cell incubator (37°C) and subsequently rinsed twice with 1X PBS containing 1X Wash Additive from the SPC Innovator kit (80μm ,Atrandi Biosciences). Next samples were transferred onto a coverslip and imaged using a spinning disk confocal microscope. For unbiased quantification of the encapsulation efficiency, a grid was placed on the imaging plane and a total of 20 images were taken. Next, the amount of encapsulated and empty SPCs was quantified using custom R scripts. Cell lysis and removal of genomic DNA from compartmentalized human lymphocytic cells To release cellular RNA the lymphocytic cells were lysed. To this end, 100μl of SPCs were washed twice with 1ml of lysis buffer (1X RLT buffer from Quiagen #79216, and 40mM DTT) for 1min at RT (1000g), including incubations of 1min in between spin steps. Next, SPCs were washed five times with 1ml wash buffer supplemented with proteinase K (10mM Tris-HCl pH=8, 0.1% Triton X-100, 1mM EDTA,0.33mg / ml ProteinaseK in ddH2O). To remove residual Proteinase K the SPC were washed ten times with WB1 (10mM Tris-HCl pH=8, 0.1% Triton X-100, 1mM EDTA in ddH2O), and three times with 1ml WBRI (0mM Tris-HCl pH=8, 0.1% Triton X-100, 1mM DTT, 0.5U / μl RNAseOUT in ddH2O). To remove genomic DNA, 500μl of SPCs in WBRI were combined with 10X Turbo DNAseI Buffer (Thermo Fisher Scientific, AM2238) and 5μl of Turbo DNAseI. The reaction was mixed by gentle pipetting and incubated at 37°C for 15min. Following the addition of 56μl EDTA (50mM), the enzyme was heat inactivated at 75C for five minutes. Finally the SPCs were washed thrice with 0.75ml of WBRI. Single-cell direct RNA-sequencing library preparation and RNA release Next, ligation of sequence-specific 18s / 28s RT adapters was performed on compartmentalized cells. To this end, 50μl of SPCs were added to 20μl of 5X NEBNext Quick Ligation Reaction Buffer (NEB, B6058S), 10μl of concentrated T4 DNA Ligase (NEB, M0202T), 1,5μl of RNAse Inhibitor Murine (NEB, M0314S), and 6.45μl of pre-annealed 18s and 28s RT adapters (1.4μM each). The reaction was mixed carefully by pipetting and incubated for 15min at RT. Next, the reverse transcription mix was added consisting of 12.5μl dNTPs (NEB, N0447S), 50μl of 5X SSIV Buffer (Thermo Fisher Scientific, 18090010), 25μl of 0.1M DTT (Thermo Fisher Scientific, 707265ML), and 162.5μl of ddH2O. After a short incubation of 5min at RT, 12.5μl of SSIV (Thermo Fisher Scientific, 18090010) was added to the reaction mix and the reaction incubated at 50°C for 15min followed by heat inactivation at 70°C for 1.5min. Following barcode ligation RNA was released from the semi-permeable capsules. To this end SPCs were first centrifuged for 1min at RT (500g), and the supernatant removed. Next, SPCs were washed thrice with 100μl WB2 (10mM Tris-HCl pH=7.6, 0.1% Triton X-100, in ddH2O). Following the final wash approximately 50μl of SPC suspension was mixed with 1μl of Release Reagent and incubated at RT for 5 min or until the cloudy suspension became clear. Finally, 50μl of ddH20 were added to the solution and the RNA cleaned up using 0.8X RNAse-free AMPure XP beads (Beckman Coulter, ). Beads were incubated with the RNA for 5min followed by two wash steps with 200μl of freshly-prepared 70% EtOH. After air-drying the pellet for 1min, following the final wash step, RNA was eluted in 20μl of ddH2O. RLA ligation and direct RNA-sequencing The final sequencing library was generated by taking 20μl of eluted RNA from the previous step and adding 8μl of 5X NEBNext Quick Ligation Reaction Buffer, 6μl of RNA adaptor (RLA), 3μl of ddH2O, and 3μl of concentrated T4 DNA Ligase. The reaction was incubated at RT for 15min, followed by a clean-up step using 1X RNAse-free AMPure XP beads. Beads were incubated with the RNA for 5min followed by two wash steps with 100μl of wash buffer (WSB). After air-drying the pellet for 1min, following the final wash step, RNA was eluted in 12μl of elution buffer (EB). The eluted library was added to 37.5μl of sequencing buffer (SB) and 25.5μl of library solution (LIS). Following priming of a RNA flowcell (FLO-MIN004RA), the library was loaded and sequenced using a Promethion P2 solo with MinKNOW acquisition software version >24.06.12.

Claims

CLAIMS 1. A method of single-cell direct RNA sequencing comprising:

1. Obtaining a single-cell RNA sample, by using at least one capture oligonucleotide and at least two oligonucleotide-DNA barcodes, wherein said RNA sample comprises sequentially concatenated DNA barcodes arranged in multiple splints, by performing one or more rounds of split-pool barcoding, and 2. obtaining a DRS signal therefrom, and 3. demultiplexing the DRS signal obtained from the sequentially concatenated DNA barcodes using a DNA basecaller, wherein the demultiplexing comprising: a. detecting a subset of sequentially concatenated DNA barcode combinations generated in step 1 by split-pool barcoding, b. assigning basecalled DRS signals to the detected subset of barcode combinations.

2. The method according to the preceding claim wherein Step 3(a), detecting a subset of sequentially concatenated DNA barcode combinations generated during step 1(a) comprises; – aligning each basecalled barcode sequence to a reference set of barcode sequences arranged in multiple splints, Optionally, retaining only the best match at each splint when the difference (mindiff) between the best and second-best match is above a threshold, and / or – identifying valid barcode combinations that occur in at least 3% of the expected number of reads per barcode.

3. The method according to any of the preceding claims 1 or 2 wherein Step 3(b), detecting a subset of barcode combinations generated during split-pool processing – aligning each basecalled barcode sequence to a reference sequence while retaining all matches within a predefined maximum edit distance (maxdiff) from the best match, – generating possible barcode chains from the retained matches, restricted to chains observed in step 3 (a), and / or – selecting a final barcode chain for each read based on the highest alignment score, provided the difference (mindiff) between the best and second-best score is above a defined threshold.

4. The method according to any of the preceding claims 1 or 2, wherein Step 3(b), detecting a subset of barcode combinations generated during split-pool processing – aligning each basecalled barcode sequence to a reference, – generating possible barcode chains from the retained matches, restricted to chains observed in step 3 (a), and / or – selecting a final barcode chain for each read based on the highest alignment score.

5. The method according to any of the preceding claims, wherein the capture oligonucleotide comprises a 5´sequence and a 3´ sequence, the 3´sequence linked to the 3´end of the 5´ sequence, being the 5´sequence a sequence that partially or fully anneals with one of the 5´sequence of the at least two oligonucleotide DNA-barcode, preferably SEQ ID No 307 or a variant thereof having at least 90% identity with SEQ ID No 307 and the 3´ sequence a sequence that anneals with an RNA target of the single-cell RNA, preferably selected from a poly A tail, preferably SEQ ID No 599 or a variant thereof having at least 90% identity with SEQ ID No 599; or an 18s rRNA, preferably SEQ ID No 597 or a variant thereof having at least 90% identity with SEQ ID No 597; or an 28S rRNA, preferably SEQ ID No 598 or a variant thereof having at least 90% identity with SEQ ID No 598.

6. The method according to the preceding claim where the capture oligonucleotide is SEQ ID No 308 or a variant thereof having at least 90% identity with SEQ ID No 308 7. The method according to any of the preceding claims, wherein the oligonucleotide- DNA barcode comprises a 5´sequence, a barcode and a 3´ sequence, the barcode linked to the 3´end of the 5´sequence and the 3´sequence linked to the 3´end of the barcode, being - the 5´sequence a sequence that partially or fully anneals with the 5´ sequence of the capture oligonucleotide, preferably SEQ ID No 301 or a variant thereof having at least 90% identity with SEQ ID No 301, or the 5´sequence a sequence that partially or fully anneals with the 3´ sequence of the oligonucleotide-DNA barcode used in a previous round of the split-pool barcoding, preferably SEQ ID No 302 or a variant thereof having at least 90%identity with SEQ ID No 302, or SEQ ID No 303 or a variant thereof having at least 90% identity with SEQ ID No 303; and - the barcode being one of SEQ ID No 309-596 or a variant thereof having at least 90% identity with SEQ ID No: 309-596; and - the 3´ sequence a sequence that anneals with the 5´ sequence of the oligonucleotide-DNA barcode used in a next round of the split-pool barcoding, preferably SEQ ID No 304 or a variant thereof having at least 90% identity with SEQ ID No:304, or SEQ ID No 305 or a variant thereof having at least 90% identity with SEQ ID No:305, or the 3´ sequence a sequence that anneals with oligonucleotide with a helicase enzyme for RNA sequencing.

8. The method according to any of the preceding claims wherein the threshold value mindiff is equal to or greater than 7.

9. The method according to any of the preceding claims wherein the maximum edit distance maxdiff is equal to or less than 5.

10. The method according to any of the preceding claims wherein the barcode combinations are represented as chains of sequential barcodes over two or more splints, each barcode chain being identified by its component barcodes at each splint position.

11. The method according to any of the preceding claims, wherein the single-cell is coupled to a semi-permeable capsule.

12. A method to train a DNA basecaller for Direct RNA Sequencing (DRS), carried out by a computer, comprising the following steps: a) providing a Direct RNA Sequencing data obtained from a sample in the form of linked RNA::DNA or RNA::cDNA molecules, wherein the DNA or cDNA is complementary to the RNA, b) establishing a correlation of each signal of the Direct RNA Sequencing data to a nucleotide by signal-segmentation to obtain a DRS training dataset, and c) training the DNA basecaller for Direct RNA Sequencing (DRS) using the DRS training datasets of step b) 13. The method according to claim 12, wherein in step a) the RNA of the linked RNA::DNA or RNA::cDNA molecules are obtained from a synthetic or biological sample, preferably a transcriptome, or a biological transcriptome.

14. The method according to any of the preceding claims, 12 or 13, wherein the DRS training dataset has a high sequence complexity, preferably of at least 1000.

15. The method according to any of the preceding claims 12 to 14, wherein the linked RNA::DNA or RNA::cDNA molecules of step a) are linked covalently, preferably from the 5’ end of the cDNA to the 3´ end of the RNA molecule 16. The method according to any of the preceding claims 12 to 15, wherein the cDNA of the linked RNA::cDNA molecule is generated by reverse transcription from the RNA of the RNA::cDNA molecule, preferably linked by ligation with a dedicated ligase or a hairpin ligation, or by using adapter sequences.

17. The method according to any of the preceding claims 12 to 16, wherein step b) comprises the following stages: b.1 basecalling a read with RNA basecalling model b.2 normalization b.3 identifying events in the signal b.4 generation of signal chunks, signal-sequence pairs and using the positions of events normalized by the total number of bases in a reference sequence to predict the sequence of nucleotides for any signal chunk generated 18. The method according to the preceding claim, wherein the sample is RNA::cDNA step b1) comprises establishing the ground-truth sequence of cDNA from complementary RNA strand.

19. The method according to any of the preceding claims 12 to 18, wherein step c) comprises the training of DNA basecaller for DRS using the obtained training dataset of step b), preferably a CTC-CRF DNA basecaller.

20. A DNA basecaller for Direct RNA Sequencing (DRS) obtained by the method of any of the preceding claims 12 to 19.

21. The use of the DNA basecaller for DRS according to the preceding claim, for demultiplexing nucleic acids sequenced by using DRS.

22. The use of the DNA basecaller for DRS according to the preceding claim for basecalling single-cell native RNA sequence, and / or demultiplexing single-cell native RNA sequences.

23. The use according to the preceding claim, wherein the RNA comprises one or more DNA barcodes, concatenated sequential DNA barcodes and / or randomly generated DNA barcodes, wherein the barcodes are DNA oligonucleotides.

24. The method according to any of the preceding claims 1 to 11 wherein the DNA basecaller is trained according to the method disclosed in any of the claims 12 to 19.

Images