Rapid and accurate demultiplexing of direct RNA nanopore sequencing data with SeqTagger

Abstract

Nanopore direct RNA sequencing (DRS) enables direct measurement of RNA molecules, including their native RNA modifications, without prior conversion to cDNA. However, commercial methods for molecular barcoding of multiple DRS samples are lacking, and community-driven efforts, such as DeePlexiCon, are not compatible with newer RNA chemistry flowcells and the latest generation of graphics processing units (GPUs). To overcome these limitations, we introduce SeqTagger, a rapid and robust method that can demultiplex DRS data sets with 99% precision and 95% recall. We demonstrate the applicability of SeqTagger in both RNA002/R9.4 and RNA004/RNA chemistries and show its robust performance both for long and short RNA libraries, including custom libraries that do not contain standard poly(A) tails, such as Nano-tRNAseq libraries. Finally, we demonstrate that increasing the multiplexing up to 96 barcodes yields highly accurate demultiplexing models. SeqTagger can be executed in a standalone manner or through the MasterOfPores NextFlow workflow. The availability of an efficient and simple multiplexing strategy improves the cost-effectiveness of this technology and facilitates the analysis of low-input biological samples.

Nanopore sequencing technologies have revolutionized our ability to study the transcriptome, by enabling direct sequencing of the native RNA molecules, offering insights into gene expression profiles while retaining RNA modification and poly(A) tail length information (Garalde et al. 2018). Direct RNA sequencing (DRS) offers several advantages over traditional sequencing methods such as obviating the need for polymerase chain reaction (PCR) amplification and cDNA synthesis steps, thus preserving the native state of RNA molecules and providing a less biased and comprehensive view of the transcriptome (Workman et al. 2020).

However, key challenges persist, such as the lack of commercial barcoding kits and demultiplexing algorithms specific to DRS data, hampering the broader use of this technology. To overcome these limitations, community-driven efforts have been made to enable demultiplexing of DRS data, such as PorePlex or DeePlexiCon (Smith et al. 2020). DeePlexiCon relies on the conversion of signal information to images, which are then classified using 2D convolutional neural networks (2D CNNs) into their corresponding barcodes. This is a computationally expensive task in the bioinformatic processing of DRS data (Cozzuto et al. 2023). In addition, the user has to decide whether high demultiplexing accuracy or recovery takes precedence, with DeePlexiCon being able to reach 99% accuracy sacrificing ∼40% of total reads or 92% accuracy losing 7% of reads (Smith et al. 2020).

Here, we reasoned that developing a novel demultiplexing software that relies on basecalling the reverse transcription adapter (RTA), used during DRS library preparation, should yield superior performance compared to existing tools. We rigorously benchmarked this novel tool, which we named SeqTagger, against the de-facto standard DeePlexiCon, to determine both improvements in model performance (recall and precision) and system performance (computation time and CPU usage). The applicability of SeqTagger was tested across sequencing devices (MinION and PromethION), sequencing chemistries (SQK-RNA002 and SQK-RNA004), RNA biotypes (poly(A)⁺ and tRNA), and for extended sets of barcodes (up to 96 barcodes).

Results

Demultiplexing DRS libraries through direct basecalling of the DNA barcode

To date, efforts for demultiplexing DRS data sets have relied on the transformation of signal intensities to images, followed by the classification of resulting 2D images, as performed by the software DeePlexiCon (Smith et al. 2020). While this approach leads to robust and accurate demultiplexing statistics, it is computationally expensive, limited to four barcodes, and does not support newer RNA004 kit chemistries.

We hypothesized that training a DNA basecaller that could directly basecall a DNA barcode sequence present in the RTA, which is ligated to the native RNA molecules during library preparation, should lead to improved performance of the demultiplexing algorithm (Fig. 1A). In particular, the use of a DNA basecaller would allow skipping the most computationally expensive step of demultiplexing, i.e., the signal transformation step (Smith et al. 2020), and would in principle be flexible toward classifying an unlimited number of barcodes with minimal losses in demultiplexing accuracy, thus allowing an increase in the number of samples that can be loaded in a single flowcell.

Figure 1.

Schematic overview of the DRS and demultiplexing workflow. (A) Overview of the barcoded DRS workflow in which the standard RTA is replaced with a barcode-containing adapter. Following adapter ligation and reverse transcription, the RNA ligation adapter (RLA), containing the helicase enzyme, is ligated, making the library sequencing-ready. (B) Overview of the demultiplexing workflow performed by SeqTagger. The algorithm first segments the raw current intensity signal by identifying the poly(A)-tail signal to extract the barcode-containing reverse transcription adapter (RTA). Following signal normalization, the DNA sequence is basecalled and aligned to a set of reference barcodes. Finally, a filtering step is applied based on the median base quality (baseQ) to remove misassigned barcode sequences.

We should note that the need for training a new DNA basecalling model—rather than using existing pretrained DNA basecalling models—arises from the fact that the DNA barcode is sequenced using the “RNA” chemistry, which exhibits several key differences to the DNA chemistry, preventing the use existing DNA basecalling models: (1) RNA is sequenced from its 3′ end (in 3′ > 5′ direction), while DNA is sequenced from its 5′ end (in 5′ > 3′ direction); (2) the RNA helicase is slower than the DNA helicase (70 bp for RNA002 and 130 bp for RNA004, compared to 400–450 bp for DNA); and (3) RNA molecules are sampled with a different frequency than DNA molecules (3–4 kHz for RNA, compared to 4–5 kHz for DNA, depending on the flowcell). Thus, the same DNA barcode will generate different signals when sequenced with DNA and RNA kits, making existing DNA basecalling models unusable for basecalling of DNA barcodes in DRS data sets.

To demultiplex reads, SeqTagger performs several consecutive steps on each read: (1) signal segmentation (trimming of barcode signal) and normalization, (2) decoding of the signal into the sequence space (basecalling) followed by mapping to the reference, and (3) filtering out potential misassigned barcodes using per-read median base quality information (Fig. 1B). To train the basecalling model, we chose the latest CTC–CRF model architecture that is achieving over 99% basecalling accuracy for DNA basecalling (https://nanoporetech.com/platform/accuracy/). Training was performed using the Bonito software (https://github.com/nanoporetech/bonito/) (see Methods), and trained models (listed in Supplemental Table S1) were independently validated on PromethION DRS data sets, both for RNA002 and RNA004 chemistries (see Supplemental Table S2 for full list of DRS data sets used).

Systematic benchmarking of DRS demultiplexing software

We first trained SeqTagger on DRS data sequenced with R9.4/RNA002 chemistry, using four different barcodes that were embedded within the reverse transcription adapter (RTA) (Fig. 1A,B, see also Methods). Notably, barcode sequences were kept identical to those used to train the DeePlexiCon model, to allow for direct comparison between SeqTagger and DeePlexiCon (see Supplemental Table S1; Methods).

To test the performance of the trained models, we generated three data sets of 100,000 reads each from an independently sequenced sample, in which each barcode had been ligated to a different in vitro transcribed (IVT) RNA (Supplemental Table S2). In this design, the ground-truth is known thus allowing us to determine the classification metrics of the trained model (see Fig. 2A; Supplemental Table S3). We should note that this independently sequenced run was not used for training/validation or of either demultiplexing tool. To facilitate the monitoring of resources, and ensure that the computational resources used by each of the software were directly comparable, the data were analyzed using MasterOfPores (Di Tommaso et al. 2017; Cozzuto et al. 2023), a nextflow workflow for the analysis of DRS data, with the option “SeqTagger” (model b04_RNA002), “DeePlexiCon” (model resnet20-final.h5), or “no demultiplexing” (ground-truth) (see Methods).

Figure 2.

Comprehensive benchmarking of SeqTagger performance. (A) Schematic overview of the workflow used for comparative analysis of DRS demultiplexing software: DeePlexiCon (purple) and SeqTagger (green). (B) Barplots depicting the demultiplexing precision and recall achieved with SeqTagger default settings (baseQ > 50), DeePlexiCon high-recovery settings (-s 0.5), and DeePlexiCon high accuracy settings (-s 0.9), on the same three data sets described in A. Bars represent the mean (also indicated by the numeric value to the right of each bar) with error bars showing ±1 standard deviation. Dots represent individual replicates. (C, top) Barplot depicting the computation time of SeqTagger and DeePlexiCon, on the benchmarking data sets. Bars represent the mean value with error bars indicating ±1 standard deviation. Dots represent individual replicates. Statistical significance was determined using a two-sided t-test (ns): P > 0.05, (*) P ≤ 0.05, (**) P ≤ 0.01, (***) P ≤ 0.001. (Bottom) Barplot representing the absolute contribution of individual preprocessing steps to the total computation time (Rep-1). (D) Confusion matrices (left), receiver operating characteristic (ROC) curves (middle), and Precision–Recall curves (right) on independent test data generated with RNA002 and RNA004 kit chemistries. Data were analyzed with SeqTagger model b04_RNA002 (upper panels) and b04_RNA004 (bottom panels), respectively. (AUC) Area under the curve, (AP) average precision.

We then compared SeqTagger results (using default settings, i.e., baseQ-cutoff ≥ 50) to those obtained with DeePlexiCon, either using settings for high recovery (-s 0.5) or high accuracy (-s 0.9). Samples were processed using GPU computing nodes running CUDA10, as CUDA11 is not supported by DeePlexiCon, and thus would not be directly comparable to SeqTagger. Of note, SeqTagger can be executed both in GPU computing nodes running CUDA10 and CUDA11. Our analysis revealed that SeqTagger reached 99% precision on the independent test data sets, whereas DeePlexiCon achieved 93% and 97% precision with the high-recovery and high-accuracy settings, respectively (Fig. 2B, top panel; see also Supplemental Table S3). In addition, SeqTagger achieved a recall of 95%, whereas DeePlexiCon's high-recovery mode reached 89%, followed by the high accuracy mode with 75% recall (Fig. 2B, bottom panel; see also Supplemental Table S3).

Next, we examined the computing time and resources required for demultiplexing. Our analysis demonstrated that SeqTagger was ∼9× faster than DeePlexiCon (Fig. 2C, top panel; Supplemental Fig. S1A; Supplemental Table S4). Notably, this increase in speed has important implications for the preprocessing of DRS data, as the demultiplexing step is typically a major computational bottleneck, taking up ≥50% of the overall computation time (Fig. 2C, bottom panel; Supplemental Fig. S1B; Supplemental Table S4). Similar results were observed when performing the same analysis on a more complex data set of mouse poly(A)-selected material, where DeePlexiCon was found to take ∼40% of the overall computation time while SeqTagger reduced this to 8.5% of the total time required (Supplemental Fig. S1C).

Finally, we tested SeqTagger's classification performance on in vivo data using two runs of poly(A)-tailed, total RNA from Escherichia coli (Delgado-Tejedor et al. 2024). Both runs contained two (out of the four possible) barcodes used in our b04 models, thus allowing us to determine the false positive rate per barcode. This showed that between 0.1% and 0.2% of total demultiplexed reads were incorrectly assigned per barcode using SeqTagger, resulting in an overall precision of ≥99% (Supplemental Fig. S1D; see also Supplemental Table S5). In contrast, DeePlexiCon had higher amounts of incorrectly predicted barcodes, reaching 3%–6% per barcode in high-recovery mode and 1%–3% in high-accuracy mode (Supplemental Fig. S1D). In addition, SeqTagger recovered higher proportions of demultiplexed reads (90%–93% recall) than DeePlexiCon under high-recovery (81%–88% recall) or high-accuracy modes (65%–73% recall), supporting our observations made on in vitro data sets (see Supplemental Table S5).

Taken together, these results demonstrate that demultiplexing DRS data sets by basecalling the DNA part of the RT adapter yields superior performance metrics, while providing a significant increase in speed over the existing tool DeePlexiCon.

SeqTagger is compatible with both RNA002 and RNA004 chemistries

We then examined whether SeqTagger would be able to demultiplex DRS data sets sequenced with the recently upgraded chemistry that Oxford Nanopore Technologies (ONT) has released for sequencing DRS data sets, which uses SQK-RNA004 kits (replacing SQK-RNA002 kits) and “RNA” flowcells (replacing R9.4 flowcells). Notably, demultiplexing options are currently unavailable for this newer chemistry, neither commercially nor from the scientific community, thus making the transition to the newer chemistry highly problematic.

Here, we prepared DRS libraries using the recently released SQK-RNA004 library preparation kit on the same four barcodes (Supplemental Table S2) ligated to IVT products as previously described, and trained a new demultiplexing model using Bonito (b04_RNA004). The performance of the model was assessed using reads from an independent test data set (flowcell not used for training/validation of the model), as previously done for the SQK-RNA002/R9.4 chemistry (Fig. 2D, top panels; Supplemental Fig. S1E; Supplemental Table S3). Our results showed that SeqTagger was able to reach an overall precision above 99% with a recall of 97% on the new RNA004 chemistry (Supplemental Fig. 2D, bottom panels), thus showing improved performance compared to the old R9.4/RNA002 chemistry.

Altogether, our results demonstrate that SeqTagger is a robust DRS demultiplexing algorithm, reaching ≥99% precision on both RNA002 and RNA004 chemistries while recovering 95% and 97% of all reads, respectively. Notably, the increased demultiplexing speed that direct basecalling of the DNA barcode offers (Fig. 2C), makes SeqTagger well-suited for larger-sized data sets that are expected to be generated with RNA004 kits (due to increased helicase speed and likely higher flowcell longevity), thus removing a potential computational bottleneck in DRS data analyses.

SeqTagger can be extended to different RNA biotypes

Commercial DRS library preparation kits were initially designed to sequence poly(A)-tailed RNA (mRNA, lncRNA, etc.) or in vitro polyadenylated RNA molecules. For this reason, many DRS-related software, such as DeePlexiCon, were trained on polyadenylated RNA molecules, and rely on the identification of the poly(A) homopolymeric region to segment the barcode signal.

Recent works have shown that it is possible to sequence very short RNA reads using DRS (e.g., tRNAs) (Thomas et al. 2021; Lucas et al. 2024), opening novel avenues to explore the small RNA (epi)transcriptome at single molecule resolution. To efficiently capture these short RNA molecules, adapted library preparation protocols, such as those used in “Nano-tRNAseq”, are required (Lucas et al. 2024). Notably, Nano-tRNAseq libraries substantially differ from standard DRS protocols, and the resulting reads contain very short poly(A)-tails, which are RNA–DNA hybrids, and thus potentially lead to incorrect demultiplexing due to imprecise segmentation. Indeed, our initial explorations revealed that DeePlexiCon was not able to accurately demultiplex Nano-tRNAseq libraries, reaching 70% precision and 77% recall on Nano-tRNAseq libraries (Supplemental Table S6). A potential solution to this problem would be an extension of the poly(A) stretch found within the splint adapter which is used during the library preparation of Nano-tRNAseq. However, this would reduce the size difference between tRNAs and adapters, negatively affecting the subsequent clean-up steps and leading to substantial adapter contamination.

To overcome this limitation, we specifically trained a new model to multiplex Nano-tRNAseq data sets (see Fig. 3A). Our results showed that SeqTagger was able to demultiplex up to four barcodes reaching high precision (98.3%) on the validation set (Fig. 3B), making it possible to multiplex Nano-tRNAseq libraries in the same flowcell, thus reducing Nano-tRNAseq library preparation and sequencing costs.

Figure 3.

SeqTagger can be expanded to work on RNA–DNA hybrids (Nano-tRNAseq libraries) and with larger sets of barcodes. (A) Schematic overview of the demultiplexing models supported by SeqTagger. A four barcode (b04) and 96 barcode (b96) model are available. Additionally, a four barcode demuxing model is available for custom Nano-tRNAseq libraries (b04_tRNA). (B) Confusion matrix for the four barcode tRNA model (b04_tRNA) generated on Nano-tRNAseq validation data. Recorded precision is indicated on top. (C) Confusion matrix of the 96 barcode mRNA model (b96_RNA002) generated on the validation data. Recorded precision is indicated on top. Zoomed panels of the 96 × 96 confusion matrix for eight barcodes are shown on the right.

SeqTagger can be extended to accurately demultiplex large barcode sets

Next, we wondered whether SeqTagger could be extended in terms of multiplexing capacity, i.e., by training a model that would predict additional barcodes, and whether demultiplexing performance would be significantly affected by the increased number of barcodes. To examine this, we trained a model containing a total of 96 different barcodes (b96_RNA002), using the same procedure as previously described for four barcodes (Fig. 3A; see also Supplemental Table S1). Our results obtained on the validation data set demonstrated that increasing the barcoding capacity by 24-fold only had a minor effect on model precision (98.8%) enabling simultaneous sequencing of up to 96 samples on a single flowcell (Fig. 3C).

We then examined the ability of SeqTagger to demultiplex in vivo samples using three independent test data sets. To this end, we extracted total RNA from human samples, which we poly(A)-tailed to make them amenable for DRS (see Methods). Next, we prepared three independent libraries barcoding each with one of the barcodes used for training the 96 barcode model (b96_RNA002). As expected, analysis of all three libraries revealed that the average base quality (baseQ) for barcodes present in the library was significantly higher than those absent, suggesting that base quality is an efficient parameter to remove false positive predictions (Fig. 4A; see also Supplemental Fig. S2A). Moreover, applying the default base quality filter used internally by SeqTagger (baseQ > 50), removed the majority of incorrectly assigned barcodes (96%–97%) while maintaining between 93% and 95% of reads assigned to barcodes present in each library (Fig. 4B; see also Supplemental Fig. S2B). Of note, SCBC-25 which showed lower base quality on the independent test data (Fig. 4A, left panel), performed well on an additional independent test run containing this same barcode suggesting that the low performance on this particular test data set was possibly related to the sample/library preparation, rather than the demultiplexing model (Supplemental Fig. S2C).

Figure 4.

Performance of SeqTagger's 96 barcode model on independent test data. (A) Boxplots depicting the base quality (baseQ) per barcode for three independent test runs using SCBC1-30 (left panel), SCBC-31-60 (middle panel), and SCBC-57-96 (right panel). Libraries were prepared using poly(A)-tailed total RNA from human samples (see Methods). Boxes are limited by the lower quartile Q1 (bottom) and upper quartile Q3 (top). Whiskers are defined as 1.5 × interquartile range (IQR) with outliers not shown. (B) Lineplot representing the total number of reads for each barcode (baseQ > 50) for runs shown in A, following demultiplexing. The total number of reads for each run with baseQ > 50 is indicated by n.

Discussion

Direct RNA nanopore sequencing has recently emerged as a transformative platform to characterize the (epi)transcriptome, offering unparalleled advantages in the analysis of RNA molecules, such as the possibility to detect its native RNA modifications (Garalde et al. 2018; Lucas and Novoa 2023). DRS has so far been used to address a wide range of biological questions in different biological systems, such as deciphering the order of intron removal during RNA splicing (Drexler et al. 2020), dissecting RNA modification dynamics upon chemical or biological stress (Begik et al. 2021; Huang et al. 2021; Delgado-Tejedor et al. 2024; Lucas et al. 2024), and elucidating viral RNA dynamics (Viehweger et al. 2019; Kim et al. 2020; Price et al. 2020; Baquero-Pérez et al. 2024), among others.

Standard protocols for DRS require large amounts of poly(A) tailed input material ranging from ≥50 ng (SQK-RNA002) to ≥300 ng (SQK-RNA004). This prohibits the sequencing of low-input samples as underloading of a flowcell leads to a rapid decay in sequencing throughput. Thus, the possibility to sequence up to 96 samples in a single flowcell can reduce per-sample input requirements down to 3–10 ng per sample. Therefore, optimal sequencing libraries can be generated by combining several barcoded low-input samples, albeit at the expense of sharing the overall throughput of a single flowcell.

To date, commercial options for DRS multiplexing are lacking. Consequently, the field has relied on community-based solutions for demultiplexing, such as PorePlex (https://github.com/hyeshik/poreplex) or DeePlexiCon (Smith et al. 2020). The latter, while being accurate and widely adopted by the community (Begik et al. 2021; Gupta et al. 2023; Javaran et al. 2023; Rajan et al. 2023; White et al. 2023), suffers from several limitations: (1) it currently supports only four barcodes; (2) to date, it only supports old sequencing chemistries (RNA002, currently being deprecated); (3) it requires CUDA v10—and consequently will not work with GPUs released after 2019, which use CUDA v11; (4) it does not work with custom reads that do not contain standard polyadenylated reads, such as Nano-tRNAseq reads, and (5) the signal transformation step, is computationally expensive.

To overcome these limitations, we developed a novel approach for demultiplexing DRS reads, which relies on DNA basecalling of the barcode region, referred to as SeqTagger. We find that SeqTagger is a rapid and accurate demultiplexing program that is robust across different sequencing chemistries (RNA002, RNA004), devices (MinION, PromethION), and RNA species (mRNA, tRNA), supporting both FAST5 and POD5 file formats. It achieves very high precision and recall on both smaller (4) and larger (96) barcode sets, thus opening exciting possibilities for the cost-effective sequencing of multiple samples on a single flowcell. To put these results into a broader context, a study published on cDNA demultiplexing on the ONT platform reached similar performance metrics on a 12-barcode set compared to SeqTagger (b96_RNA002) (Wick et al. 2018). This suggests that demultiplexing using SeqTagger for DRS can reach similar (RNA002) or higher (RNA004) classification metrics than current tools for cDNA demultiplexing.

We would like to note that the speed of demultiplexing with SeqTagger is still mostly bound by data access (Supplemental Fig. S1A) with our results reflecting a network file system. We observed that 1 million reads can be demultiplexed in 5 or 12 min when using a local solid state drive (SSD) or hard disk drive (HDD), respectively. Moreover, when data access is not a bottleneck, for example when the read signals are already loaded in system memory, SeqTagger can classify 50 barcodes in 5 ms. This opens exciting possibilities such as real-time barcode classification of DRS reads, similar to what is currently available for DNA nanopore sequencing.

As an alternative approach to DNA barcodes, used both by DeePlexiCon and SeqTagger, one could imagine ligating either an RNA:DNA heteroduplex or an RNA:RNA barcode at the first step of library preparation (instead of DNA:DNA), which could be readily basecalled with existing RNA basecalling models and subsequently demultiplexed. However, there are several major challenges from both a wet- and dry-lab perspective with such an approach. Firstly, this would require a change to the library preparation protocol as ligases exhibit different substrate specificities, which in turn could reduce the overall efficiency of the library preparation (Bullard and Bowater 2006). Additionally, RNA oligonucleotides pose a substantial economic burden. For example, ordering only one strand of the oligonucleotides used in this study as an RNA molecule increases the cost of one RT adapter by 10-fold, making large barcode sets prohibitively expensive. Moreover, RNA molecules are inherently less stable than DNA, which could lead to problems particularly when freeze-thawed multiple times, which is common for RT adapters. From a computational perspective, RNA oligonucleotides would require major changes to both the acquisition software MinKNOW and the basecalling software, as many definitions rely on differences in the current intensity signals obtained from RNA ligated to the DNA adapter.

Methods

Plasmid extraction and linearization

We selected plasmids encoding in vitro transcripts from a T7 promoter that share little sequence similarity (see Supplemental Table S7) to enable unambiguous mapping of sequencing reads. Overnight cultures (LB-Ampicillin, 100 μg/mL) containing E. coli transfected with the plasmids were processed using the Monarch Plasmid DNA Miniprep Kit Protocol (NEB T1010). Subsequently, plasmids were linearized using 25 μL plasmid solution, 5 μL of Enzyme 1, 5 μL of Enzyme 2 (specific enzymes are mentioned in Supplemental Table S7), and 10 μL of CutSmart Buffer in a total volume of 100 μL for 3 h at 37°C. Volumes were topped up to 300 μL using nuclease-free H₂O before clean-up using 1× volume of basic Phenol:Chloroform:Isoamyl Alcohol (Sigma, P3803). Samples were vortexed followed by centrifugation at 16,000 rpm for 5 min. The aqueous phase was transferred to a new tube and supplemented with 0.1× 3 M NaOAc (pH = 5.2), 2.5× EtOh abs., and 2 μL of GlycoBlue (Invitrogen AM9515). After overnight storage at −20°C, precipitated plasmids were collected by centrifugation at 4°C for 30 min, 16,000 rpm. The pellet was washed twice with 70% EtOH and eluted in 30 μL. Concentrations and purity were determined using a NanoDrop One/One^c (Thermo Scientific).

In vitro transcription and poly(A)-tailing

In vitro transcription was performed using the AmpliScribe T7-Flash Transcription Kit (Biosearch Technologies, ASF3507) with 1 μg of each linearized plasmid as input. The reaction was carried out as specified in the manufacturer's instructions with an overnight incubation at 37°C. The plasmid template was digested by adding 2 μL of TURBO DNase (Invitrogen AM2238) followed by incubation at 37°C for 15 min. In vitro transcripts were cleaned up using the RNeasy Mini Kit (Qiagen 74104) and eluted in 60 μL of nuclease-free H₂O. Concentration and purity were determined using a NanoDrop One/One^c (Thermo Scientific). Transcript integrity was determined using a TapeStation 4150 (Agilent). For poly(A)-tailing 2 μL of 10× E. coli Poly(A) Polymerase Reaction Buffer, 2 μL of 10 mM ATP, 0.5 μL SUPERase•In RNase Inhibitor (Invitrogen AM2696) and 1 μL of E. coli Poly(A) Polymerase (NEB M0276S) were added to 7 μg of input in vitro transcript in a total volume of 20 μL. The reaction was carried out for 3 min at 37°C to obtain short poly(A)-tails. Samples were cleaned up using the RNA Clean & Concentrator-5 Kit (Zymo Research R1013) and eluted in 25 μL of nuclease-free H₂O.

Barcode design

BC-01 to BC-04 correspond to barcodes previously described and used by DeePlexiCon (Smith et al. 2020). They are comprised of a 30 nt long oligoA with a 5′P end, to enable ligation during library preparation and a 49 nt long oligoB which contains a 10 nt poly-d(T) 3′-end required for annealing to the target poly(A)-containing library. SCBC-01 through SCBC-96 were designed to contain regions that are highly distinguishable in the sequence space (Doroschak et al. 2020). They are comprised of a 47 nt long oligoA with a 5′P group to enable ligation during library preparation, and a 66 nt long oligoB which contains a 10 nt poly-d(T) 3′-end required for annealing to the target poly(A)-containing library. See Supplemental Table S1 for details on all oligonucleotide sequences used to build barcoded libraries.

Hybridization of custom RT adapters containing barcode sequences

All DNA oligos were ordered from Integrated DNA Technologies (IDT) and annealed before library preparation using a final concentration of 1.4 μM of each oligonucleotide, 0.01 M Tris-HCl (pH = 7.5), and 0.05 M NaCl in a total volume of 75 μL nuclease-free H₂O. The mixture was incubated at 94°C for 1 min and slowly cooled down (−0.1°C/s) to room temperature. Small aliquots were prepared to prevent repeated freeze/thawing of hybridized adapters. All oligonucleotides used in this work are listed in Supplemental Table S1.

Direct RNA sequencing of training data

RNA002: Library preparation was carried out according to the manufacturer's instructions (direct-rna-sequencing-sqk-rna002-DRS_9080_v2_revR_14Aug2019-minion) with some adjustments to enable barcode adapter ligation, prevent cross-contamination of barcodes, and enable cost-efficient library preparation of multiple barcodes. Adapter ligation was carried out for 15 min at room temperature using 100 ng of input poly(A)-tailed substrate, 0.5 μL of custom barcode adapter (1.4 μM), 1.5 μL of NEBNext Quick Ligation Reaction Buffer (NEB B6058S), 0.5 μL of RNase Inhibitor (NEB M0314S) and 0.75 μL of concentrated T4 DNA Ligase (NEB M0202T) in a total volume of 7.75 μL nuclease-free H₂O. Subsequently, 4 μL of 5× SuperScript IV Buffer, 1 μL of deoxynucleotide mix (dNTPs) (NEB N0447S), 2 μL 0.1 Dithiothreitol, and 1 μL of SuperScript IV Reverse Transcriptase (Invitrogen 18090010) were added and the reaction incubated for 15 min at 50°C before heat inactivation for 2 min at 70°C. Reverse transcribed RNA:DNA hybrids were cleaned up using 1× RNAClean XP beads (Beckman Coulter A63987), washed twice using freshly prepared 70% EtOH, and eluted in 3 μL of nuclease-free H₂O; 2.5 μL of eluate was transferred to a new tube and RMX ligation was carried out individually for each barcode by adding 1 μL of NEBNext Quick Ligation Reaction Buffer (NEB B6058S), 0.75 μL of RMX adapter, and 0.375 μL of T4 DNA Ligase (NEB M0202T) in a total volume of 5 μL nuclease-free H₂O by incubating for 15 min at room temperature. Libraries were cleaned up using 0.8× RNAClean XP beads (Beckman Coulter A63987), and washed twice using 20 μL of wash buffer (WSB) each time. Samples were eluted in 4 μL of elution buffer (EB). Barcoded libraries were pooled at this step and the volume was adjusted to 18.75 μL using nuclease-free H₂O. Each library was mixed with 18.75 μL of RNA running buffer (RRB) and run on a primed R9.4.1 flowcell using a MinION sequencer with MinKNOW acquisition software version v23.11.4. Live base calling and mapping were activated and runs stopped once ≥40,000 mapped reads were reached per barcode. Flowcells were reused multiple times following flushing (see Flowcell nuclease flushing).

RNA004: Library preparation was carried out according to the manufacturer's instructions (direct-rna-sequencing-sqk-rna004-DRS_9195_v4_revB_20Sep2023-promethion). Changes to the above workflow for RNA002 are mentioned, the remaining steps were identical. For adapter ligation, 0.75 μL of RLA was used. Barcoded libraries were pooled and the volume was adjusted to 32 μL using nuclease-free H₂O. Before loading, the libraries were mixed with 100 μL of sequencing buffer (SB) and 68 μL of library solution (LIS). Libraries were run on a primed FLO-PRO004RA flowcell using a PromethION 2 solo sequencer with MinKNOW acquisition software version v23.11.4.

Flowcell nuclease flushing

Preventing carry-over during sequencing runs was paramount for generating high-quality training data. Hence, we adapted the existing flowcell flushing protocol (EXP-WSH004) by preparing a modified flowcell wash mix. To this end, we mixed 20 μL of TURBO DNase (Invitrogen AM2238) with 380 μL of wash diluent (DIL, provided with EXP-WSH004). After loading the mix into the flowcell, the reaction was incubated for 20 min. The remaining steps were identical to the manufacturer's protocol. In our hands, this yielded negligible carryover between runs (0.01%–0.03%).

Sample preparation and direct RNA sequencing of independent test data for b96

Total RNA was isolated using the AllPrep RNA/DNA/miRNA Universal Kit (Qiagen 80224) using the manufacturer's instructions. Frozen lung cryosections were homogenized with the TissueLyser mixer-mill disruptor (2 × 2 min, 25 Hz, Qiagen, Hilden, Germany). The quality of total RNA was assessed with an Agilent 2100 Bioanalyzer and Agilent RNA 6000 Nano Kit (Agilent Technologies, Boeblingen, Germany). Next, total RNA was split into a long (≥200 nt) and short (<200 nt) fraction using the RNeasy MinElute kit (Qiagen 74204). Subsequently, the long RNA fraction was poly(A)-tailed (NEB M0276) in a total of 40 μL. Individual poly(A)-tailed samples were cleaned up using RNAClean XP beads (Beckman A63987).

For DRS library preparation, individual samples (150 ng each) were ligated to preannealed (as described above) barcode-containing RT adapters (SCBCs) using concentrated T4 DNA Ligase (NEB M0202T). Following reverse transcription using SuperScript IV Reverse Transcriptase (Invitrogen 18090010), barcoded samples were pooled before the clean-up RNAClean XP beads (Beckman A63987). For RMX ligation 250 ng of pooled library were used. Subsequent steps followed the standard DRS library preparation protocol for RNA002 (direct-rna-sequencing-sqk-rna002-DRS_9080_v2_revS_14Aug2019-promethion). Libraries were run on a primed R9.4.1 flowcell using a PromethION sequencer with MinKNOW acquisition software version v23.11.4.

Training the DNA basecaller

The training algorithm, Bonito (https://github.com/nanoporetech/bonito/), was used to train the DNA basecaller model. Bonito requires as input signal chunks of fixed length and their corresponding barcode sequences. To know the signal-to-sequence correspondence for training purposes, one can either sequence one barcode per flowcell/run or ligate each barcode to a unique RNA molecule and later assign each read to the barcode by mapping reads to the corresponding references (see Supplemental Table S1 for barcode sequence design).

We used 120,000 reads per barcode to train four-barcode models (b04) and 40,000 reads per barcode for larger barcode sets (b96). Reads were randomly subsampled to reach the required number of reads per barcode. Importantly, we used only the last 3000 (RNA002) or 2000 (RNA004) samples of every barcode signal for training and basecalling. The reads with shorter barcodes were skipped.

The CTC–CRF models were trained with Bonito v0.7.2. We used a window of 31 and a 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). Since we found only marginal differences in their accuracy for barcode basecalling, we decided to use the lightest (and fastest) version with 96 features (fast).

Basecalling of DNA barcode from direct RNA sequencing

The demultiplexing algorithm follows several steps: (1) signal segmentation, (2) normalization, (3) barcode sequence decoding, (4) barcode identification, and (5) quality filtering. First, the signal corresponding to the barcode is identified by t-test statistics in two window sizes (450 and 1000) rolling over the first 30,000 read samples. We define barcode end as the position with the highest t-test score, additionally adding 100 samples of poly(A) tail. Secondly, the barcode signal is normalized using median absolute deviation. Thirdly, the barcode sequence (and corresponding base qualities) are decoded using Bonito v0.7.2 with a custom barcode basecalling model. Subsequently, the sequence is aligned onto barcode reference using minimap2 v2.26 (Li 2018) executed with custom parameters: -k6 -w3 -A1 -B1 -O1 -E1 -c1 -m10 -s13. The best primary alignment for every barcode sequence is reported as the predicted barcode. Finally, only barcodes with median basecall quality above 50 are reported.

Comparative analysis of demultiplexing software

For comparative performance analysis of demultiplexing software, we benchmarked SeqTagger against DeePlexiCon, which was the only DRS demultiplexing tool available as of January 2024, when this experiment was performed. We should note that WarpDemux was not included in our benchmarking as the code was made publicly available later that year (July 2024 for RNA002 chemistry and November 2024 for RNA004 chemistry) (van der Toorn et al. 2024). PorePlex was not included in the comparison as it relies on a set of barcodes different from the ones used in this study, and is not further maintained by the developers.

Briefly, we sampled 3 × 100,000 reads from an independent RNA002 run which consisted of IVT-01-04 ligated to BC-01-04 (Fig. 2A; see also Supplemental Table S2), which had not been used to train or validate the models. Each data set was processed using the MasterOfPores (Cozzuto et al. 2020) version 3 nextflow (Di Tommaso et al. 2017) workflow, using either SeqTagger (model: b04_RNA002), DeePlexiCon (Smith et al. 2020) (model: resnet20-final.h5), or no demultiplexing (ground-truth) while keeping the remaining parameters identical. We compared default SeqTagger settings (baseQ ≥ 50, -b 50) with DeePlexiCon settings for high recovery (-s 0.5) and high accuracy (-s 0.9). Demultiplexing of individual samples (100,000 reads) was performed on a single GPU (RTX2080 Ti, CUDA10) and a single CPU with 12 GB of memory allocated (‐‐granularity 25). Samples were basecalled using guppy (v6.0.6, model = rna_r9.4.1_70bps_hac.cfg) and aligned to the reference sequence using minimap2 (v2.17) with -ax map-ont -k14.

Model performance metrics used to build confusion matrices and receiver operating characteristic curves were extracted using custom Python scripts. Computational resources required per sample were extracted from the nextflow report (-with-report, see Supplemental Table S4) and analyzed using custom R scripts (R Core Team 2021).

To compare the computation time of processes executed in the mop_preprocess workflow we used rep-1 of the benchmarking data set and recorded the computation time required by each process (Fig. 2C, bottom; Supplemental Fig. S1B). We realized that these results would underestimate the relative contribution of the mapping step as mapping was performed to a small reference of four sequences. To obtain a more realistic representation of the required computation time per process we sampled 100,000 reads from a poly(A)-selected Mus musculus library which was aligned to the GRCm39 (mm39) reference genome (minimap2, v2.17, -uf -ax splice -k14). In addition, we activated counting, another optional step in the preprocessing pipeline. We obtained comparable results to the benchmarking data set with DeePlexiCon taking up 40% of the overall computation time while SeqTagger reduces the time required to 8.5% (Supplemental Fig. S1C). To determine the performance of SeqTagger on in vivo data sets, we demultiplexed two publicly available sequencing runs (obtained from the NCBI BioProject database [https://www.ncbi.nlm.nih.gov/bioproject/] under accession number PRJEB42568), containing poly(A)-tailed, total RNA extracted from E. coli (Supplemental Fig. S1D; Supplemental Table S5; Delgado-Tejedor et al. 2024). Both runs were aligned to the E. coli rRNA reference (minimap2, v2.17, -ax map-ont -k14) after demultiplexing using either SeqTagger (model: b04_RNA002) or DeePlexiCon (model: resnet20-final.h5) with high-recovery (-s 0.5) or high-accuracy (-s 0.9) settings. Accuracy was determined by counting the number of reads assigned to each barcode in the FASTQ files (Supplemental Table S5). All scripts used to perform this analysis, as well as the reference sequences, can be found at GitHub (see Software availability).

Benchmarking model performance for RNA004, extended barcode models, and tRNA models

To determine model performance on the new RNA004 chemistry, we used an independent test data set in which IVT-01-04 was ligated to BC-01-04, which had not been previously used for model training or validation. Results were processed identically to the process described above (see Comparative analysis of demultiplexing software) with exceptions mentioned next. Demultiplexing was run on a CUDA11-enabled single GPU (NVIDIA RTX A4000) and a single CPU with 12 GB of memory allocated. For basecalling, we used Dorado (v0.5.3, https://github.com/nanoporetech/dorado, model=rna004_130bps_sup@v3.0.1). The extended barcode model (b96_RNA002) was trained on 24 runs each containing four in vitro transcripts ligated to the respective barcode. Confusion matrices were generated based on the validation data set (analysis notebooks available at https://github.com/novoalab). The tRNA model was trained by using four individual Nano-tRNAseq runs performed on human cell lines, each using a different barcode sequence. Confusion matrices were generated based on the validation data set. All runs used to generate training and testing data are listed in Supplemental Table S2.

Software availability

SeqTagger can be executed using a Docker container following the instructions on GitHub (https://github.com/novoalab/SeqTagger) or using the MasterOfPores (version 3.0) nextflow pipeline (MoP3) (https://github.com/biocorecrg/MOP3) (Di Tommaso et al. 2017; Cozzuto et al. 2023). Processing of raw FAST5 files was performed using MasterOfPores (version 3.0). Full documentation on installation, usage, and dependencies related to MoP3 (including running SeqTagger via MoP3) can be found at https://biocorecrg.github.io/MoP3/. All custom R and Python scripts used to generate figures in this manuscript can be found at GitHub (https://github.com/novoalab/SeqTagger) and as Supplemental Code.

Data access

All raw and processed sequencing data generated in this study have been submitted to the European Nucleotide Archive (ENA; https://www.ebi.ac.uk/ena/browser/) under accession numbers PRJEB78482 and PRJEB84026. A summary of all data sets used in this work can be found in Supplemental Table S2.

Competing interest statement

L.P.P., G.D., and E.M.N. have filed patents on the SeqTagger demultiplexing algorithm and method (application EP24382340) and an extension thereof (application EP24383144). E.M.N. has received travel and accommodation expenses to speak at Oxford Nanopore Technologies conferences. G.D. has received travel bursaries from ONT to present his work at conferences. E.M.N. is a member of the Scientific Advisory Board of IMMAGINA Biotech.