Separating and selectively reading RNA samples using nanopore sequencing with barcode-based sorting

We aimed to establish a robust, low-cost, and accurate barcoding methodology to enable dRNA-seq multiplexing and adaptive sampling experiments. We chose to place the barcode sequences within the customizable RTA adapter due to its compatibility with the dRNA-seq library preparation workflow. Because the DNA RTA adapter cannot be read by the RNA-specific basecaller, we implemented signal-based demultiplexing.

Dynamic Time Warping (DTW) is an algorithm used for comparing time-series with temporal distortions. This makes it ideally suited for analyzing raw nanopore signals with their intrinsic variations in dwell-time^9–12. To apply this algorithm to dRNA-seq multiplexing, we integrated a Dynamic Time Warping Distance (DTWD) kernel into a Support Vector Machine (SVM) classifier. This DTWD-based kernel function captures the essential spatial and temporal signal information by quantifying how similar an unknown barcode is to known barcode patterns, termed fingerprints.

We next assessed the performance of this design strategy by comparing non-optimized RTA barcodes (WDX-DPC) to four of the 12 optimized barcodes (WDX4) (Supplementary Note 2 and Supplementary Table 9). The use of optimized barcodes improved both accuracy (99% WDX4 vs. 97% WDX-DPC) and yield (96% WDX4 vs. 92.3% WarpDemuX-DPC) while run times remained unaltered (Table 1 and Fig. 2c). This confirmed that maximizing DTWD improves classification performance. When the model was extended to 12 distinct barcode classes (WDX12), accuracy only slightly decreased despite the complexity of distinguishing a larger number of barcode classes (99% WDX4 vs 98% WDX12; weighted average) (Fig. 2d). A decrease in prediction confidence resulted in a slightly higher rate of unclassified barcodes (4.0% to 8.8% at a confidence threshold of 0.5) (Table 1). Remarkably, the WDX12 model consistently outperformed DPC (4 barcodes) across all confidence thresholds in accuracy, yield and run time (Fig. 2e, Table 1, Supplementary Tables 5–7 and Supplementary Tables 10–12), whilst expanding the number of barcode classes three-fold (also see Supplementary Note 3). Importantly, the performance of WarpDemuX only decreased slowly with increasing numbers of barcodes (Fig. 2f). With an optimized set of barcodes, we predict that 24 barcodes can be implemented at an accuracy of 96% with 14% of reads unclassified (confidence cutoff = 0.5) (Fig. 2f).

We next investigated the read length and read identity distributions of the demultiplexed samples (Fig. 3b–g). We observed clear differences in read length distributions as the infection progressed (Fig. 3b). Discrete bands became visible at 24–72 hpi (Fig. 3b and Supplementary Fig. 13), which could be attributed to distinct subgenomic RNAs by TRS-leader sequence annotation (Supplementary Fig. 14). This analysis showed striking differences in the relative abundances of individual subgenomic RNAs (sgRNAs), with the most abundant sgRNA N present at 22.3% (95% CI = 21.9, 22.7; 48 hpi), compared to the least present subgenomic RNA 10, which was present at 2.5 × 10⁻¹% (95% CI = 2.0 × 10⁻¹, 3.0 × 10⁻¹; 48 hpi) (Fig. 3d). The relative composition of sgRNAs, however, remained relatively stable throughout the infection cycle (Fig. 3d). Importantly, the ability to pool RNA from infected cells on a single flowcell allowed us to assess the replication kinetics of these viruses, without running multiple sequencing experiments. For both viruses, SARS-CoV-2 RNA was first detectable at 24 hpi, and reached a maximum at 48 hpi, comprising approximately 50% of all poly-adenylated RNA in the RNA extract, before declining to 30% at 72 hpi (Supplementary Figs. 15, 16). No apparent difference in kinetics between the wildtype and EGFP-expressing virus was observed, indicating insertion of EGFP does not effect viral replication over the time frame tested (Fig. 3c).

Finally, we performed a poly(A) tail length analysis, as this is known to regulate RNA translation and stability, and is regulated during infection in other Coronaviruses (Fig. 3e–g). The poly(A) tail length of SARS-CoV-2 sgRNA was previously characterized for a single time point of 24 hpi¹⁴, revealing two tail length populations with a low abundance of shorter tails. Using WarpDemuX, we were able to study the dynamics of poly(A) tail length across multiple time points and compare RNA from cells and virions. This revealed a broad tail length distribution for Vero mRNA, which was stable over time, with a median of 70 nt (2.5th-97.5th percentile: 14–227 nt; 48 hpi), while SARS-CoV-2 RNA had significantly shorter poly(A) tail lengths (median 56 nt, 2.5th–97.5th percentile: 23–108 nt; 48 hpi; two-sided, independent t-test, p < 0.001), and exhibited regulation of tail length over the course of infection. Intriguingly, at 24 hpi, > 99% of cellular SARS-CoV-2 RNA was attributed to a population with a median tail length of 56 nt, with an additional population with a shorter poly(A) tail length around 29 nt becoming more abundant over time, increasing to 3% at 48 hpi and 15% at 72 hpi (Fig. 3e and Supplementary Fig. 17). These shorter tails in the cell appear to be limited to sgRNAs (Fig. 3f) and were not observed in virion RNA (Fig. 3g). The increased presence of shorter tail lengths at the later time points could reflect mRNA degradation upon cell death, compared to genomic RNA protected within membrane vesicles. In agreement, when compared to cellular RNA at 72 hpi, supernatant RNA within virions had a longer average tail length (52 nt cellular vs. 58 nt supernatant) (Fig. 3g).

Leveraging WarpDemuX’s classification speed, we implemented ultra-rapid adaptive sampling for dRNA-seq, aiming to make rejection decisions while the poly(A) tail of RNA molecules is still within the pore. We modified MinKNOW software to provide signal data earlier to the WarpDemuX pipeline, which required adaptations to internal read classification logic (Supplementary Note 5 and Supplementary Figs. 18–20). We also developed our own decision logic to begin rejecting reads once specific thresholds are reached. For this, we use dynamic thresholds that update based on the current sequencing distribution and guide the rejection of reads based on real-time metrics such as adapter counts, read counts, and number of bases (Supplementary Note 6). We applied these decisions to 40 channels each on a MinION R9.4.1 flow cell (Fig. 4b). Notably, 81% of classifications finished within 60 ms after detection of a poly(A) tail, and 98% of classifications completed within 200 ms (Fig. 4c and Supplementary Figs. 21–23). Assuming a translocation speed of 70 nt/s, classification occurs well within 14 nt of the start of the poly(A) tail. In practice, due to the latency of the different softwares, 50% rejections are made within 14 nt and 95% within 34 nt, which would be within the poly(A) tail for the majority of reads (Fig. 4d).

To provide a rigorous test of the adaptive sampling pipeline, we turned to our SARS-CoV-2 samples, which were heavily unbalanced due to differing amounts of RNA present in cells and supernatant (Fig. 4b). All barcode balancing strategies led to a greater number of molecules being sampled, due to higher pore turnover (Fig. 4e, Supplementary Figs. 24–26). Importantly, the rejection of high abundance barcodes not only led to a more even barcode coverage (Supplementary Table 13), reducing the Gini coefficient from 0.44 to 0.15, but also to the enrichment of low abundance barcodes by 33–73% (Fig. 4f and Supplementary Fig. 25).

During the development of WarpDemuX, Oxford Nanopore Technologies released SQK-RNA004, a new direct RNA sequencing chemistry featuring an RNA-specific pore and faster motor protein. While the library preparation approach remains consistent between versions, the new RNA-specific pore exhibits different signal characteristics that necessitate adaptation of our model²¹.

To account for the intrinsic variation in the dwell time, as well as the within-event measurement noise, (Supplementary Fig.), we partition the raw signal of the adapter intosegments, based on the method introduced in ONT’s signal analysis tool Tombo (github.com/nanoporetech/tombo; algorithmic details are included in Supplementary Method). Each segment is subsequently represented by its average signal. To mitigate the risk of data loss due to missed events (Supplementary Fig.), we intentionally over-segment the signal by detecting approximately 10% more segments than expected based on the number of bases in the RTA. We normalize the segmented adapter signal and retain only the last 25 segments, which should contain the barcode signal given its known sequence length and relative location within the RTA. 35a, b 2 35c \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n$$\end{document} n

To correct for the likely imperfect segmentation, we perform subsequent Dynamic Time Warping (DTW) to compute the distance between segmented barcode signals. DTW is a robust solution to align two sequences with temporal variation by compressing or stretching their temporal axis to minimize the cumulative Euclidean distance between corresponding elements (thus correcting for imperfect segmentation).

To compute the Dynamic Time Warping Distance (DTWD) between two segmented signals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{1}$$\end{document}s1 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{2}$$\end{document}s2, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${DDTW}({s}_{1},{s}_{2})$$\end{document}DDTW(s1,s2) we use the dtaidistance package (2.3.11)³⁴. The parameters used are detailed in Supplementary Method 3.

We developed a method to generate sets of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document}k RTA barcodes with distinct signal profiles de novo, as it would be computational intractable to perform an exhaustive search over all possible barcode sets of size k (complexity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\begin{array}{c}{4}^{L}\\ k\end{array}\right)$$\end{document}4Lk for barcodes of length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L$$\end{document}L, e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\approx 2.3\times 1{0}^{121}$$\end{document}≈2.3×10121 for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L=18$$\end{document}L=18 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k=12$$\end{document}k=12). First, we constructed a set of wave-like target signal patterns that exhibited high inter-target DTWD, since we expect such patterns to improve signal segmentation and barcode separability. Then, we searched for barcode sequences whose putative signal, constructed using the ONT k-mer models (github.com/nanoporetech/kmer_models), would closely match these target patterns. Because there are no DNA k-mer models for DNA sequenced in 3’ to 5’ direction, we filter for k-mers we expect have similar signal profiles in both sequencing directions, as detailed in Supplementary Method 1. For the twelve target patterns highlighted in Fig. 2a, we selected the top 50 barcodes per target pattern, yielding a total of 600 barcodes, for which we computed all 600 × 600 pairwise DTWDs. Lastly, within the matrix of 600² pairwise DTWDs, we solved the Maximal Diversity Problem (MDP, below) with k = 12 to propose 12 candidate barcodes that we expect to be accurately identifiable and distinguishable. The MDP is defined as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{cc}{\max }_{\delta (\cdot )} & \mathop{\sum }_{i=1}^{M-1}\mathop{\sum }_{j=i+1}^{M}{DDTW}({s}_{i},{s}_{j})\delta (i)\delta (j)\\ {{\rm{Subject\; to}}} & \mathop{\sum }_{i=1}^{M}\delta (i)=k\\ & \delta (i)=\{0,1\},\, 1\le i\le M\end{array}$$\end{document}maxδ(⋅)∑i=1M−1∑j=i+1MDDTW(si,sj)δ(i)δ(j)Subject to∑i=1Mδ(i)=kδ(i)={0,1},1≤i≤Mwhere M denotes the entire set of barcode candidates (600 in our example), k is the number of distinct barcodes we aim to select, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{i}$$\end{document}si is the segmented signal of barcode i, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${DDTW}({s}_{i},{s}_{j})$$\end{document}DDTW(si,sj) is the DTWD between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{i}$$\end{document}si and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{j}$$\end{document}sj, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta (i)$$\end{document}δ(i) is a dirac delta function indicating whether barcode \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i$$\end{document}i is selected. Detailed descriptions of the steps in the design process are included in Supplementary Method 1.

To evaluate our barcode design and train a classifier for demultiplexing, we generated labeled RTA signals by synthesizing and sequencing various custom RTA adapters (Supplementary Table) attached to RNA molecules with integrated (in-line) RNA barcodes—serving as ground truth labels. Notably, the in-line RNA barcodes used were integrated into the RNA transcripts using PCR, positioning them at the 5’ end of the poly(A) tail to enhance the accuracy of basecalling and, consequently, the precision of barcode classification, compared to standard RNA barcodes that are ligated to the 3’ end of the poly(A) tail. We randomized the RNA in-line barcode to RTA barcode assignment across replicates to circumvent any potential of bias introduced by the in-line RNA barcode signal. The generated data sets (1–6) are detailed in Supplementary Table. 24 1

The barcoded RNA designated for training the WarpDemuX classifier was synthesized via in vitro transcription. To enrich the diversity of our dataset, we used a variety of RNA templates. Briefly, starting from different DNA templates, 12 PCRs per template were performed with a forward primer containing a T7 RNA polymerase promoter and varying reverse primers to add different RNA barcodes to the 3’ end, followed by a 10nt poly(A) tail (Supplementary Fig.and Supplementary Table,). After purification of PCR products, they were used as templates for in vitro transcription, followed by DNA digest and RNA clean-up. The RNA was then further polyadenylated to ensure no signal spillover from the RNA barcode into the DNA signal. Further details and a visual summary of the protocol are included in Supplementary Method. 36 25 26 4

The direct RNA sequencing library preparation was started by ligating 100 ng of poly-adenylated barcoded RNA onto a custom RTA adapter. Further library preparation was performed according to ONT’s recommendations, with the exception of the use of MarathonRT as a reverse transcription enzyme (see Supplementary Method). After motor protein ligation, 7 µL of the prepared library was loaded on a Flongle FLO-FLG001 flow cell, and reads were acquired until all pores became inactive. For a detailed description of all steps, see Supplementary Method. Reads were basecalled post-acquisition with the direct RNA model of ONTs research basecaller Dorado. See Supplementary Methodfor the full configuration. 5 5 6

We trained a support vector machine (SVM) to classify the sequenced dual-barcoded RNA reads into RTA barcode categories based on their RTA-barcode signal, using a combination of read mapping and RNA barcode classification to infer their ground truth labels.

To determine ground truth labels, reads were first uniquely mapped to RNA reference sequences with LAST³⁵. Then, the RNA barcode region was identified and reads were associated with the RNA barcode that had the smallest Levenshtein distance to them. Reads for which this distance surpassed a defined threshold of 4 were discarded. Based on the specific combinations of RNA and RNA-RTA barcode pairs used during library preparation, the read assignments were used to establish the RTA-barcode ground truth labels required to train the classifier.

Datasets 1 and 2 served as the test sets for the WarpDemuX-DPC and DPC models. For the WarpDemuX models, we designed two distinct test sets: one comprising 600 reads per barcode from datasets 3–5, and another using the entirety of dataset 6, which contains RNA without enzymatic poly-adenylation.

We constructed validation sets for our models, comprising 200 reads per barcode, utilizing data from replicate 2–5 of the DeePlexiCon training data, as detailed in Smith et al.⁷, for the WarpDemuX-DPC model, and data from datasets 3–5 for the WarpDemuX models (WDX4, WDX12).

Training sets, consisting of 400 reads per barcode, were sampled from data from replicate 2–5 of the DeePlexiCon training data⁷ for the WarpDemuX-DPC model, and our datasets 3–5 for the WarpDemuX models (WDX4, WDX12). First, 400 training instances for the noise class were selected based on the robust Z-score (defined below) of the median inter-class distance per read. Identification of noise instances is further detailed in Supplementary Note 1. Then, to ensure the training set captured the full range of the original data distribution despite its limited size, we used greedy approximations of the Maximal- and Minimal Diversity Problem, stratified across barcode classes (implementation is included in Supplementary Note 1), to select 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}×80 instances per barcode that respectively maximized and minimized their mutual distance. The training set was completed with 240 randomly sampled instances per barcode. Supplementary Materials provide further details on training set size determination (Supplementary Note 1).

The robust Z-score is calculated as follows, whererepresents the value under evaluation (median inter-class distance), Median() is the dataset’s median, and MAD is the Median Absolute Deviation: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{i}$$\end{document} X i \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X$$\end{document} X 2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{Robust\; Zscore}}}=\frac{{X}_{i}-{{\rm{Median}}}(X)}{{{\rm{MAD}}}}$$\end{document} Robust Zscore = X i − Median ( ) X MAD 3 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{MAD}}}={{\rm{Median}}}\left(\left|{X}_{i}-{{\rm{Median}}}(X)\right|\right)$$\end{document} MAD Median = X i − Median ( ) X

All sampling processes for validation, training, and test sets were stratified across replicate groups to compensate for variations in sequencing yields, ensuring a balanced representation across all data.

The RTA barcode classifiers were trained on preprocessed RTA barcode signals. To obtain these RTA barcode signals, we first identify the DNA adapter signal in the raw signal using ADAPTed (our custom tool for DNA adapter signal detection; https://github.com/KleistLab/ADAPTed↗; LLR-based method for RNA002, convolutional classifier method for RNA004). Adapter detection is necessary but also functions as a signal quality filter: ADAPTed processing entails poly(A) detection and various signal characteristics checks. After successfully identifying the DNA adapter signal, we segment, normalize, and trim it, as described in the “Adapter signal processing and alignment” section. For each RTA barcode signal, we generated a feature representation that quantifies how similar the signal is to the RTA barcode signals of reads with known labels. To do this, DTWDs between the signal and all labeled signals in the training data set are computed.

We then use Support Vector Machines (SVMs) for classification by introducing a Dynamic Time Warping Kernel function,, which is defined as:whereandare two preprocessed RTA barcode signal instances, andis a hyper-parameter of the kernel function. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K$$\end{document} K \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{1}$$\end{document} s 1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{2}$$\end{document} s 2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma=1.2$$\end{document} γ = 1.2 4 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K({s}_{1},{s}_{2})=\exp \left(-\gamma \cdot {DDTW}({s}_{1},{s}_{2})\right)$$\end{document} K exp ( ) s 1 s 2 , = − ⋅ γ D D T W ( ) s 1 s 2 ,

During training, the SVM kernel matrix is based on the pairwise DTWDs between the barcode signals in the training data.

We use an SVM implementation based on libsvm³⁶, with a one-vs-one scheme for multi-class support, as implemented in Scikit-Learn (1.3.1)³⁷. Herein, calibrated class membership probability estimates are provided based on an internal 5-fold cross-validation combined with a multiclass extension to Platt scaling³⁸. See Supplementary Method 3 for further details on the model parameterization.

During inference, for each read to be classified,is computed against all barcode signals in the training data, i.e.,, with, wherecontains all segmented barcode signals that belong to the RTA barcodebased on their ground truth label. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K({s}_{1},{s}_{2})$$\end{document} K ( ) s 1 s 2 , \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{2}\in B$$\end{document} s 2 ∈ B \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B=\{{B}_{1},{B}_{2},\ldots {B}_{k}\}$$\end{document} B = { } B 1 B 2 B k , , … \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${B}_{i}$$\end{document} B i \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i$$\end{document} i

Each prediction in our classification system is paired with a confidence score, indicating the reliability of the predicted class label. Reads for which the prediction has a confidence score below a user-defined cutoff are discarded (unclassified). For comparability, we adopt the same method as used in DeePlexiCon and define the confidence score as the interval between the classifier’s probability estimate of the predicted class and the second most probable class.

To prepare the sequencing samples, 5 × 105 Vero E6 TMPRSS2 cells (CVCL_YQ49, a generous gift from S. Pöhlmann; authenticated by the provider) were infected in duplicate with two different recombinant SARS-CoV-2 viruses at MOI = 0.01. Briefly, rSARS-CoV-2 WT and rSARS-CoV-2 GFP inoculum were prepared in DMEM supplemented with 1% FCS³⁹. Before the infection, Vero E6 TMPRSS2 cells were washed once with PBS and incubated with the respective inoculum for 1 h at 37 °C with gently shake every 10 min. The inoculum was removed, and fresh DMEM supplemented with 5% FCS, 100 mg/mL of penicillin, and 100 μ/mL of streptomycin was added to the cells.

Then, cells were harvested at 8 and 24 h post infection. A mock control of uninfected cells was also performed. Total RNA was extracted with Trizol (Invitrogen) using the manufacturer’s recommendations, precipitated by the addition of 0.1 volumes of 3 M NaOAc and 3 volumes of ice-cold 100 % EtOH and incubation at − 20 °C for > 12 h, followed by centrifugation at 16,000 × g for 30 min at 4 °C, a wash with 70 % EtOH and another round of centrifugation before resuspension in 20 µL RNase-free H2O. RNA concentration was quantifed on Nanodrop, and 1 ug of total RNA per sample was used for WarpDemuX-multiplexed direct RNA library preparation as described in Supplementary Method 5 and Supplementary Table 27, with the exception of the usage of the 100 ms chunk size protocol (see Supplementary Note 5). The generated reads were demultiplexed with the WDX 12-barcode model (WDX12) at a confidence threshold of 0.8, and reads were aligned with minimap (see Supplementary Note 4 for details).

We used the Read Until API (see https://github.com/nanoporetech/read_until_api↗) to communicate with the MinKNOW server during sequencing. During sequencing, read data is streamed to Read Until channel caches in fixed-duration chunks, and continually added to as sequencing of the read progresses. WarpDemuX analyzes these caches to detect RTA signals and classify barcodes.

Live demultiplexing begins by identifying the transition from adapter to poly(A) tail in the raw signal, utilizing ADAPTed’s fast method tailored for streaming data, after which the read is processed and classified.

To minimize the time spent sequencing unwanted reads, we developed a custom MinKNOW protocol with a reduced read chunk duration (100 ms; approx. 300 observations per chunk). Details are included in Supplementary Note 6. Instructions to setting up MinKNOW for adaptive sampling using our custom protocol are available at Github [https://github.com/KleistLab/WarpDemuX↗]. Details on the system specifications used for barcode-specific adaptive sampling are included in Supplementary Method 7.

Due to the nature of dRNA-seq sample preparation, not all adapter molecules are ligated to a target molecule, leading to adapter-only reads. These reads, while largely filtered by MinKNOW after sequencing, can skew the real-time WarpDemuX-classified barcode distribution, necessitating correction based on actual sequencer output (pod5 files). WarpDemuX tracks these outputs to maintain an accurate sample distribution and estimate the sequenced base count per sample. The user can determine the aspect to be balanced and the degree of sensitivity applied.

For RNA004 chemistry evaluation, we generated training, validation, calibration, and test datasets from five independent sequencing replicates that were synthesized using our dual-barcoding approach as described in Supplementary Method, with the exception of the usage of the RNA004 motor protein RLA, and a PromethION RNA004 flow cell (FLO-PRO004RA). From replicates 1–3, we randomly selected 2000 reads per barcode (stratified across replicates) for initial data distribution analysis and noise class identification. The training set comprised 400 reads per barcode from this set, with noise class instances selected based on robust Z-scores of median inter-class distances as before. To ensure a comprehensive representation of signal variations in the training set, we employed a stratified sampling strategy combining maximally diverse (80 reads), minimally diverse (80 reads), and random (240 reads) instances per barcode. 5

Validation sets (200 reads per barcode, replicates 1–3) were used to select optimal models for different barcode set sizes (WDX4-WDX12). Two independent test sets were created: test set 1 containing 10,000 reads per barcode from replicates 1–3, and test set 2 combining replicates 4–5 with 10,000–18,000 reads per barcode. A separate calibration set of 2000 reads per barcode from replicates 1–3 enabled confidence score mapping refinement.

To establish reliable accuracy estimates, we developed a barcode-specific calibration procedure that maps model confidence scores to predefined target accuracy levels (95–99.9%). For each barcode, we first calculated the baseline error rate () on the calibration dataset. We then determined confidence score thresholds () that would achieve desired target accuracy levels () by computing the confidence score quantile () from misclassified reads: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{b}$$\end{document} E b \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document} T \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A}_{t}$$\end{document} A t \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q$$\end{document} q 5 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q=1-(100-{A}_{t})/{E}_{b}$$\end{document} q = − / 1 ( ) 100 − A t E b

Confidence scores above T were retained to achieve the target accuracy level. This calibration was performed independently for each barcode to account for barcode-specific characteristics and error patterns.

Further information on research design is available in thelinked to this article. Nature Portfolio Reporting Summary