TagGen: High-Performance Barcode Generator and Demultiplexer for High-Throughput and Long-Read Sequencing Applications

Tag generation algorithm design

TagGen implements a two-phase algorithm fundamentally different from exhaustive enumeration ( Figure 1).

Figure 1. TagGen overview: barcode design for nanopore sequencing.

(A) Nanopore sequencing errors are dominated by deletions (50%), insertions (25%), and substitutions (25%) at 5–15% total error rates, requiring longer barcodes than short-read platforms. (B) Existing tools enumerate all 4ⁿ sequences exhaustively, causing memory exhaustion at ≥14 bp (268 million sequences). TagGen’s Monte Carlo sampling operates at constant complexity O(k) regardless of length. (C) TagGen’s two-phase generation algorithm: Phase 1 generates ~100,000 random candidates validated against user constraints in parallel; Phase 2 applies greedy diversity selection maximising minimum pairwise Hamming or Levenshtein distance. (D) TagGen’s integrated demultiplexer: a two-stage anchor-free pipeline that assigns reads to source barcodes without requiring flanking adapter sequences. Stage 1 uses k-mer voting to rapidly identify candidate barcodes and estimate their location; Stage 2 performs banded edit-distance alignment to confirm the best match against configurable acceptance criteria. (E) Validation under a nanopore error model demonstrates that 30 bp barcodes (d ≥ 8) maintain 100% correct assignment at 20% error rate, while 10 bp barcodes degrade to 83%. Barcodes ≥14 bp achieve >97% resolution at typical ONT error rates (10–15%). (F) TagGen addresses sample multiplexing, direct RNA sequencing, and spatial transcriptomics through an integrated GUI and CLI pipeline with built-in demultiplexing.

Phase 1: Monte Carlo Candidate Generation. Rather than enumerating all 4ⁿ sequences, TagGen generates random candidates validated in real-time against user-specified constraints (GC content, homopolymer limits, exclude sequences). Generation proceeds in parallel across CPU cores and terminates upon collecting sufficient valid candidates (typically ~1,250 candidates in ~55 ms).

Phase 2: Greedy Diversity Selection. From the candidate pool, TagGen iteratively selects sequences maximizing minimum Hamming or Levenshtein distance to already-selected barcodes, producing a maximally diverse final set (typically 96 barcodes in ~25 ms additional time).

Algorithm pseudocode

The following pseudocode formalizes TagGen’s two-phase approach.

Complexity analysis

The algorithmic approaches differ fundamentally in complexity:

For 14 bp barcodes, this represents a 2,680-fold reduction in operations (100,000 vs. 268 million), translating to milliseconds versus memory exhaustion. Generation time is largely independent of candidate pool size for a fixed target count (Extended Data Figure 1).

Integrated demultiplexer

TagGen includes a built-in demultiplexer (also standalone as taggen-demux) that assigns ONT FASTQ reads to their source barcodes. It is available both as a command-line subcommand (taggen --demux) and as a dedicated tab in the graphical interface.

Each read is processed by a two-stage pipeline. In the first stage, all 8-mers in the read’s search region are looked up against a pre-built index that maps each 8-mer to the set of barcodes containing it, and candidate barcodes are ranked by hit count (k-mer voting). The top candidate’s hit positions are averaged to estimate the barcode location within the read. In the second stage, a banded edit distance computation slides the candidate barcode across a window of (tag length + 15) bp centred on the estimated location, yielding the minimum edit distance to the read. The best-matching barcode is accepted if three criteria are met: (i) the edit distance does not exceed the automatically derived maximum (adaptive rule: tag_length/5 for tags shorter than 20 bp, tag_length/4 for 20–29 bp, tag_length/3 for ≥30 bp; overridable via --max-dist); (ii) the margin between the best and second-best candidate distance is at least 2 (ambiguity guard); and (iii) a confidence score derived from the edit distance exceeds a user-configurable threshold (default 0). Reads failing any criterion are written to an “unassigned” output file, annotated with the reason (no_match, ambiguous, or low_confidence). This conservative strategy ensures that uncertain reads are withheld rather than wrongly assigned.

Distance metrics

Both Hamming distance (substitutions only) and Levenshtein edit distance (substitutions, insertions, deletions) are supported and applied consistently between barcode generation and demultiplexing. Because ONT sequencing produces a characteristic error profile with a significant proportion of insertion and deletion errors – particularly at homopolymer runs – Levenshtein distance is recommended for most ONT workflows. Hamming distance remains available for platforms with predominantly substitution errors or when computational speed is a priority.

Search modes

Two search modes accommodate different library preparation strategies, as follows.

End mode (default): the demultiplexer searches a window at each end of the read and its reverse complement, corresponding to the standard dual-ended library structure [barcode][adapter][insert] [RC (adapter)][RC (barcode)]. This is the appropriate mode for ligation-based ONT DNA libraries.

Full mode: the entire read is scanned for the barcode, intended for protocols where the barcode may appear at an internal position (e.g., direct RNA sequencing with a mid-read adapter ligation, or custom in-read tagging schemes). In full mode, users can specify one or more position masks to restrict the search to the expected barcode region, reducing false positive k-mer seeds from flanking sequence. Masks may be expressed as absolute positions (start:end in bp), end-relative offsets (5p:N for the first N bp from the 5′ end; 3p:N for the last N bp from the 3′ end), or as read-length fractions (f_start:f_end,resolved per read). Multiple zones are merged and the union is searched.

Read trimming and output

Three trimming modes are available: none (reads written unmodified), ends (barcode trimmed from the matched end), and all (barcodes trimmed from both ends). FASTQ headers of assigned reads are annotated with the barcode name, edit distance, and confidence score. Assigned reads are written to per-sample subdirectories; unassigned reads to a separate subdirectory. Gzip compressed input is supported and processed in streaming batches to maintain constant memory usage. A per-sample statistics TSV file records the number of reads assigned per barcode, expected read count, and mean Q-score, enabling rapid quality assessment. In the GUI, these statistics are displayed as four interactive charts: barcode match position distribution, per sample assignment counts, edit distance and confidence distributions, and per-sample Q-score profiles.

POD5 co-demultiplexing

When raw POD5 signal files are present alongside FASTQ reads, TagGen can invoke the pod5 subset tool to partition signal files by sample, enabling downstream signal-level analyses (e.g., modified base calling) on a per-sample basis.

Implementation

TagGen is implemented in D (dlang.org), a systems programming language combining high-level expressiveness with native code performance. The application is compiled using the LDC2 compiler (LLVM-based D compiler) with release optimizations enabled.

The software architecture comprises six primary modules ( Figure 1):

Operation

Minimum System Requirements:

Installation: Pre-compiled binaries are available for Linux and Windows from the GitHub releases page. Linux users may also compile from source using the DUB package manager:

dub build --compiler = ldc2 --build = release

GUI Workflow:

Demultiplexing workflow (Demultiplex tab):

Command-Line Usage:

TagGen provides a comprehensive CLI for automated pipelines. Barcode generation and demultiplexing examples are shown below:

# Generate 96 barcodes with default parameters
taggen -n 96 -l 14 -d 4 -o my_barcodes

# Generate with custom GC constraints and exclude sequences
taggen -n 96 -l 16 --minGc 40 --maxGc 60 -f adapters.fasta -v

# Generate using JSON configuration file
taggen -c experiment_config.json

# Generate large pool with matrix output
taggen -n 384 -l 20 -d 6 --metric levenshtein -m --heatmap -v

# Demultiplex reads in end mode (standard dual-ended libraries)
taggen --demux --tags my_barcodes.fasta --reads reads.fastq --mode end --trim-mode ends

# Demultiplex with full-mode search and 5′ position mask
taggen --demux --tags my_barcodes.fasta --reads reads.fastq --mode full --position-mask 5p:60

# Demultiplex mid-read barcodes with fractional mask and strict threshold
taggen --demux --tags my_barcodes.fasta --reads reads.fastq --mode full --position-mask 0.05:0.20 --max-dist 4

Full CLI options are available via taggen –help.

Nanopore error simulation

To validate barcode resilience under realistic long-read conditions, we implemented a nanopore error simulator based on published characterizations.^5,6,18 Deletions: 50% of errors (most frequent in nanopore sequencing). Insertions: 25% of errors (random nucleotide insertion). Substitutions: 25% of errors (random nucleotide replacement). We tested error rates from 5% (high-quality R10.4 data) to 25% (stress test beyond typical conditions), encompassing typical ONT performance (10–15%) and degraded sample scenarios (20%). For each condition, 1,000 reads were simulated per barcode, with identification via minimum Levenshtein distance to the original barcode set.

Benchmarking

Performance comparisons were conducted on a workstation with a 12-core AMD processor running Arch Linux. TagGen (v1.2.0) barcode generation runtime was compared against DNABarcodes (v1.20.0) from Bioconductor.¹⁰ Both tools were configured to generate 96 barcodes with equivalent constraint parameters (GC content 25–75%, minimum Hamming distance as specified). Timing measurements excluded I/O operations to focus on algorithmic performance. Each TagGen configuration was tested in triplicate; reported values are means ( Table 3).

The practical sequencing utility of TagGen-generated barcodes was evaluated by simulating Oxford Nanopore reads with Badread v0.4.1¹⁹ and measuring demultiplexing accuracy against ground truth. Reads were simulated using the nanopore2023 error model across read identity levels of 80–95% (corresponding to approximately 5–20% per-base error rates) spanning the range from early ONT chemistries to most recent high-accuracy data. Two demultiplexers were applied to the same simulated reads: taggen-demux (integrated, using Levenshtein distance) and minibar v0.25 (an established third-party ONT demultiplexer).²⁰ Results are reported as accuracy (percentage of reads correctly assigned to their source barcode) and error rate (reads assigned to a wrong barcode); reads failing matching criteria are classified as unassigned rather than wrong.

Several other tools in this space were not included in the comparison. Qcat,¹⁶ ONT’s open-source demultiplexer, is designed exclusively for ONT’s predefined native barcode kits and does not accept arbitrary user-defined sequences; similarly, the demultiplexing module integrated into ONT’s Dorado basecaller²¹ is restricted to native barcode catalogues. Porechop¹⁷ has been discontinued and does not support user-defined barcodes. PRO¹⁵ is publicly available but a direct comparison remains impractical: PRO optimises for maximum barcode set size at a fixed length rather than accepting user-defined target counts and length ranges, and offers no GC content filtering, homopolymer constraints, or custom exclude sequences. These different design goals make a like-for-like accuracy comparison uninformative.

Demultiplexing accuracy was evaluated across 154 parameter combinations in five systematic test phases using the same Badread simulation setup described above. Tag sets were generated with TagGen using a pool of 10,000 candidates, GC content 30–70%, maximum homopolymer 3 bp, and error tolerance t = 2. A hundred reads were simulated per barcode per run; ground-truth assignments were extracted from Badread read headers. Demultiplexing was performed with taggen-demux using an adaptive acceptance threshold (autoMaxDist: tag_length/5 for tags shorter than 20 bp, tag_length/4 for 20–29 bp, and tag_length/3 for tags of 30 bp or longer) unless otherwise noted. Three accuracy metrics were recorded for each condition: the fraction of reads correctly assigned to the originating sample (accuracy), the fraction assigned to an incorrect sample (misassignment error), and the fraction rejected as unassigned.

TagGen dramatically outperforms exhaustive enumeration

We benchmarked TagGen against DNABarcodes across configurations spanning short-read (8–12 bp) and long-read (14–30 bp) applications ( Figure 2, Table 3).

Figure 2. Performance comparison between TagGen and DNABarcodes.

Generation times (log scale) across barcode lengths 8–30 bp. DNABarcodes (orange) fails at ≥14 bp due to memory exhaustion when attempting to enumerate 268+ million sequences. TagGen (blue) succeeds at all lengths with sub-second generation times for typical applications (14–20 bp).

All benchmarks: 96 target barcodes, Hamming distance, GC 30-70%. Wall-clock times averaged over 3 replicates, Linux. DNABarcodes memory exhaustion occurs when attempting to enumerate 4ⁿ sequences (268 million at 14 bp).

At comparable configurations, TagGen is able to generate longer barcodes (14–30 bp range) with generation times below 50 ms for all tested lengths (Extended Data Figure 3), while the exhaustive enumeration strategy of DNABarcodes fails entirely due to memory requirements exceeding available RAM when attempting to store 268+ million candidate sequences.

Using a literature-based error model on the tags directly, we tested barcode resolution across lengths and error rates, showing that the generated tags are robust to most error rates, in particular longer barcodes ( Table 4).

Values indicate percentage of reads correctly assigned to original barcode. Error model: 50% deletions, 25% insertions, 25% substitutions based on published ONT characterizations. N = 1,000 simulated reads per barcode per condition.

The results reveal a striking pattern. At typical nanopore error rates (10–15%), traditional 10 bp barcodes show degradation (91–96% correct), while barcodes ≥14 bp maintain >97% resolution. Under challenging conditions (20% error) – representing degraded RNA or difficult sequences – the contrast is dramatic: 10 bp barcodes achieve only 83.3% correct assignment (16.7% misassignment), while 30 bp barcodes maintain 100% correct assignment. Even at the extreme 25% error stress test, 30 bp barcodes maintain 99.6% accuracy while 10 bp barcodes fall to 74.7%.

TagGen demultiplexer is designed for anchor-free tags

To benchmark taggen-demux in a standard dual-ended library context, we compared its performance against minibar (v0.25), a widely used primer-anchored demultiplexer for ONT data. Simulated reads were generated with a dual-ended structure ([barcode [adapter][insert] [RC (adapter)][RC (barcode)]) using the same TagGen-generated Levenshtein barcodes (n = 96, l = 14/20/30 bp, d = 4/8/8) and Badread simulator used for the rest of the benchmarks. Both tools were given the same barcode FASTA; minibar was additionally provided the flanking adapter sequence, which taggen-demux does not use. The two tools differ fundamentally in their matching strategy. Taggen-demux performs an anchor-free alignment: it searches a window at the read end and computes the edit distance between each reference barcode and the raw read sequence. It therefore operates without any knowledge of where the barcode ends and the adapter begins. Minibar uses a two-step primer-anchored approach: it first locates the adapter by alignment, then extracts the bases immediately upstream as the barcode, and finally compares only those extracted bases against the reference set. When sequencing indels shift the barcode region within the read, primer-anchored extraction isolates the barcode cleanly; the anchor-free approach must absorb those indels within its edit-distance threshold. At high read identity (≥90%), minibar assigned more reads correctly than taggen-demux for 20 and 30 bp barcodes (95–100% vs 78–97%), with near-zero misassignment in both cases ( Figure 3). At 80% read identity, where an average 30 bp barcode carries approximately six sequencing errors, minibar still achieved 88% accuracy for 30 bp barcodes, compared to 59% for taggen-demux. For short barcodes (14 bp) at low identity (80–85%), the two tools performed comparably (17–42%), as the adapter itself becomes too corrupted for reliable anchoring and the tight acceptance threshold (2 edit operations) limits both approaches equally. Taggen-demux consistently produced a small misassignment rate (0.05–1.4%), while minibar produced near-zero misassignment across all conditions. The higher misassignment for taggen-demux at l = 20 bp and low identity reflects the more liberal acceptance threshold (edit_dist = 5 out of 20 bp = 25% of tag length) combined with the absence of primer-assisted barcode extraction. These results confirm that when flanking adapter sequences are known and library structure is fixed, primer-anchored tools such as minibar extract barcodes with higher sensitivity. Taggen-demux is not designed to compete in this scenario; its intended contribution lies in settings where primer sequences are unavailable, variable in position, or absent by design – including direct RNA sequencing, spatial transcriptomics capture plates, and custom library structures in which the barcode is not flanked by a consistent adapter. These scenarios, which minibar cannot address, are evaluated in the following sections.

Figure 3. Comparison of taggen-demux and minibar demultiplexers.

Both tools were applied to the same simulated dual-ended ONT reads ([barcode][adapter][insert] [RC (adapter)][RC (barcode)], 100 reads per barcode, n = 96, Badread nanopore2023 error model). taggen-demux operated in end mode without primer information; minibar additionally received the flanking adapter sequence as a positional anchor. (A) Demultiplexing accuracy and (B) misassignment rate as a function of simulated read identity, for three barcode lengths and minimum pairwise Levenshtein distances (l = 14 bp d = 4, l = 20 bp d = 8, l = 30 bp d = 8). Solid lines, taggen-demux; dashed lines, minibar. Minibar achieves higher assignment rates at medium-to-high read identity by leveraging the primer anchor, whereas taggen-demux is designed for library configurations in which no such anchor is available.

Long barcodes provide superior nanopore error tolerance

The key advantage of longer barcodes becomes evident under simulated nanopore errors ( Table 4, Figure 4). End-mode demultiplexing accuracy increased monotonically with both tag length and read identity ( Figure 4A). At 95% read identity – representative of latest ONT chemistry – accuracy reached 84.7%, 92.2%, 93.3%, and 97.5% for tag lengths of 14, 20, 24, and 30 bp, respectively, with near-zero misassignment (<0.03%) in all cases (Extended Data Figure 4). At 85% identity, which corresponds to older R9.4.1 flow cells or low-complexity regions, accuracy ranged from 42.2% (14 bp) to 77.1% (30 bp). Reads that could not be confidently assigned were rejected as unassigned rather than being misassigned, preserving the integrity of all assigned reads. We therefore recommend 30 bp Levenshtein-generated tags for experiments where read quality is expected to be moderate (≤90% identity).

Figure 4. Systematic evaluation of taggen-demux across library configurations and sequencing conditions.

All simulations used Badread (nanopore2023 error model) at four read identity levels (80–95%) representing the range from older R9.4.1 to current R10.4.1/Kit14 chemistry. (A) End-mode demultiplexing accuracy as a function of read identity for four barcode lengths (14, 20, 24, and 30 bp) generated with Levenshtein distance at the maximum feasible minimum distance per length (d = 5, 8, 10, and 12, respectively; n = 96). Longer barcodes provide progressively higher accuracy; 30 bp barcodes reach ~97% at 95% identity. (B) Effect of distance metric during barcode generation: Levenshtein-generated tags (solid) outperform Hamming-generated tags (dashed) at equal nominal minimum distance, with a gap of 3–8 percentage points at 85–90% identity, owing to the higher realised inter-tag edit distance achieved by greedy selection in Levenshtein space (n = 48, end-mode, d = 8 for l = 20 bp and d = 12 for l = 30 bp). (C) Impact of full-mode position masks on demultiplexing accuracy for libraries with a 5′-end barcode (n = 96, l = 30 bp, d = 8, Levenshtein). End mode (reference) and full-mode with 5′ location information achieve very similar results, while full mode without a mask drops about 20% of accuracy. (D) Impact of full-mode position masks on demultiplexing accuracy for libraries with a mid-read barcode (n = 96, l = 30 bp, d = 8, Levenshtein). End mode (reference) achieves near-zero accuracy as expected, confirming correct mode-restriction behaviour. Full-mode search with a fractional mask (0.05:0.20) reaches ~45% accuracy at 95% identity; a fixed 5′ mask (5p:60) and full mode without a mask show intermediate and lower performance, respectively. (E) Scalability: accuracy as a function of barcode set size from n = 48 to n = 384, for dual-ended (circles) and spatial (diamonds) library types at three read identity levels (l = 30 bp, d = 8, Levenshtein). Accuracy varies by less than 2 percentage points across the full range, confirming applicability to high-throughput spatial transcriptomics and population-scale experiments.

Levenshtein distance provides better accuracy

At equal nominal minimum distance (d = 8) and tag length (20 bp), tags generated with Levenshtein distance constraints consistently outperformed Hamming-generated tags across all read identity levels, with a gap that widened at lower identity (+3–8 percentage points at 85–90% identity) ( Figure 4B). This advantage stems from the greedy selection algorithm achieving a higher actual minimum pairwise Levenshtein distance when operating in Levenshtein space: for d = 8 Levenshtein generation, the realised minimum inter-tag edit distance typically reached 12–15 bp, whereas Hamming-designed tags of the same nominal d achieved lower Levenshtein separations due to the Hamming–Levenshtein inequality. We therefore recommend Levenshtein generation for all ONT applications. Hamming-distance results are provided for comparison. At 30 bp and d = 12, both metrics converged at high identity (>95%), but Levenshtein remained superior at the clinically relevant 85–90% range.

Position masking recovers tags in any spatial configuration

For libraries where the barcode is positioned at the 5′ end of the read (spatial transcriptomics configuration, Figure 4C), end-mode demultiplexing was most effective (96.0% at 95% identity), as the tag is at the read terminus and is read with the lowest per-base error rate. Full-mode search with a 5′ position mask (5p:90, restricting the search window to the first 90 bp) matched end-mode performance (97.0% at 95% identity) and provided robustness to reads where the adapter truncated the tag start. Searching the full read without a mask reduced accuracy by 6–8 percentage points due to spurious off-target matches within the mRNA body, with a concomitant rise in misassignment to ~0.8% at 85% identity. This underscores the importance of correctly specifying the position mask to match the barcode location in the read.

For mid-read barcodes (DRS configuration, Figure 4D), end-mode correctly scored near-zero accuracy (~0%) as expected, confirming that the mode restriction works correctly. Full-mode with a 5p:90 mask achieved 98.0% accuracy at 95% identity. The fractional mask (0.05:0.20, covering 5–20% of read length) provided equivalent performance and is more portable across variable-length reads.

TagGen scales to large barcode sets

Despite increased difficulty of choosing from larger tag sets, accuracy was stable from n = 48 to n = 384 barcodes across all identity levels for both dual-ended and spatial library types, with less than 2 percentage points variation ( Figure 4E). At 95% identity, dual-ended end-mode accuracy remained above 97% up to 384 barcodes, and spatial full-mode (5p:90 mask) reached 97% at the same scale. At 85% identity, accuracy declined slightly with increasing n (from ~80% at n = 48 to ~76% at n = 384 for dual-ended), consistent with a greater probability of ambiguous assignment when the tag set is denser. These results confirm that taggen-demux scales to high-throughput spatial transcriptomics or population scale sequencing experiments without sacrificing accuracy.

Acceptance threshold can be tuned to reduce error rates

The autoMaxDist parameter controls the maximum edit distance at which a read is accepted as matching a tag. The default adaptive rule (tag_length/5 for <20 bp, tag_length/4 for 20–29 bp, tag_length/3 for ≥30 bp) maximises read assignment while keeping misassignment below 0.5% for all recommended configurations (end-mode and full-mode with position mask) at read identities ≥85%. Users who require near-zero misassignment – for example in clinical or single-cell applications where sample cross-contamination must be minimised – should override the default with a more conservative threshold using the –max-dist flag. For instance, taggen-demux –max-dist 4 with 30 bp Levenshtein tags (d = 8) reduces the misassignment rate to <0.001% at the cost of leaving approximately 10–15% more reads unassigned at 85% identity (Extended Data Figure 4).

Use cases

The following use cases demonstrate TagGen’s application to common long-read sequencing scenarios. Example input parameters are summarized in Table 5.

Introductory example: Standard ONT Sample Multiplexing

Scenario: A researcher needs to multiplex 96 samples for Oxford Nanopore sequencing using the standard flow cell configuration.

Input Parameters:

Barcode length: 14 bp

Target count: 96

Minimum Hamming distance: 4

GC content: 40–60%

Homopolymer limit: 3

CLI Command: taggen -n 96 -l 14 -d 4 --minGc 40 --maxGc 60 -r 3 -o ont_barcodes

Example Output (first 4 barcodes in ont_barcodes.fasta):

>Tag_001

ACGTACGTACGTAC

>Tag_002

TGCATGCATGCATG

>Tag_003

GATCGATCGATCGA

>Tag_004

CTAGCTAGCTAGCT

Generation time: 54 ms

Demultiplexing command: taggen --demux --tags ont_barcodes.fasta --reads reads.fastq --mode end --trim-mode ends

Expected performance: >97% correct barcode assignment at typical ONT error rates (10–15%).

Use Case 1: High-Error-Tolerance Direct RNA Sequencing

End-mode demultiplexing with custom adapter exclusion

Scenario: Direct RNA sequencing of degraded clinical samples where higher error rates are expected. Maximum error tolerance is required. 48-sample multiplexing experiment using a custom ligation adapter containing an 18 bp recognition sequence.

A common challenge in custom ONT library protocols is that proprietary or in-house adapters introduce short sequences that can cross-hybridise with naively generated barcodes, leading to spurious demultiplexing assignments. TagGen’s –exclude flag accepts a FASTA file of sequences to avoid; the greedy selection algorithm discards any candidate barcode whose edit distance to any excluded sequence falls below the minimum distance threshold, ensuring that the final tag set is orthogonal to the adapter.

For this scenario we recommend 24 bp Levenshtein tags with a minimum pairwise distance of d=10 to provide sufficient error tolerance. At 95% read identity – achievable with current R10.4.1/Kit14 chemistry – end-mode demultiplexing reaches 93.3% accuracy with near-zero misassignment. At 90% identity, accuracy is 83.5%, still suitable for standard discovery scale experiments.

Input Parameters:

Barcode length: 24 bp

Target count: 48

Minimum Levenshtein distance: 10

GC content: 45–55% (tighter for RNA stability)

Homopolymer limit: 2 (stricter for basecalling accuracy)

CLI Command: taggen -n 48 -l 24 -d 10 --metric levenshtein --minGc 45 --–maxGc 55 -r 2 --excludeFile custom_adapter.fasta -o uc1_tags -v

taggen --demux --tags uc1_tags.fasta --reads reads.fastq --mode end --trim-mode all

Generation time: ~30 ms

Expected performance: >98% correct assignment even at 20% error rate.

Use Case 2: Standard ONT Multiplexing with Older Chemistry

Scenario: A user has older-chemistry flow cells and needs to generate 96 barcodes that can still be confidently demultiplexed at high error rates.

Older R9.4.1 flow cells and high-molecular-weight or degraded samples routinely produce reads at 85–90% per-base identity. Under these conditions, strict end-mode demultiplexing is sensitive to read-start quality: partial exonuclease activity or pore clogging can introduce a few extra bases before the barcode, shifting it slightly from position 0. Full-mode search with a 5′ position mask (-position-mask 5p:60) restricts the alignment window to the first 60 bp of each read, tolerating this positional jitter while avoiding spurious matches in the downstream read body.

We recommend 30 bp Levenshtein barcodes (d=8) for this scenario. Benchmarks with simulated 5′-tagged reads show that full-mode with a 5p:60 mask achieves 84% accuracy at 85% read identity and 97% at 95% identity, matching or slightly exceeding end mode across the full identity range and maintaining misassignment below 0.1% in all conditions.

Input Parameters:

Barcode length: 30 bp

Target count: 96

Minimum Levenshtein distance: 8

Exclude sequences: Custom FASTA file containing adapter sequences

Exclude File (adapters.fasta):

>adapter_forward

AGATCGGAAGAGCACACGTCT

>adapter_reverse

AGATCGGAAGAGCGTCGTGTA

>custom_primer

TCGTCGGCAGCGTCAGATGTG

CLI Command: taggen -n 96 -l 30 -d 8 --metric levenshtein -f custom_adapter.fasta -o uc2_tags -v

taggen --demux --tags uc2_tags.fasta --reads reads.fastq --mode full --position-mask 5p:60 --trim-mode all

Generation time: ~95 ms

Expected performance: At 95% identity, full-mode with 5p:60 mask achieves 97% accuracy with near-zero misassignment; at 85% identity, accuracy is 84%, suitable for older chemistry experiments.

Use Case 3: Single-Cell RNA-seq with Spatial Barcodes (Zero Misassignment)

Scenario: Designing location barcodes for a custom spatial transcriptomics array with 384 capture spots, requiring error correction capability for imaging-based readout combined with nanopore sequencing.

In single-cell RNA-seq experiments where each barcode labels an individual patient sample or cell population, even a fraction of a percent of cross-sample contamination confounds downstream differential expression analysis. Here the barcode is embedded within a synthetic RNA spike-in added to each sample: during sequencing the tag appears mid-read, flanked by transcript sequence on both sides, at a position corresponding to roughly 5–20% of read length.

The fractional position mask (--position-mask 0.05:0.20) directs the search to this window regardless of read length, making the parameter portable across cells with varying transcript sizes. To achieve near-zero misassignment, we override the default acceptance threshold with --max-dist 4: at 90% read identity, this reduces misassignment from <0.1% (default) to <0.001%, at the cost of leaving approximately 10–15% more reads unassigned – a worthwhile trade when sample integrity is paramount.

Input Parameters: Barcode length: 30 bp

Target count: 384

Minimum Levenshtein distance: 8

GC content: 40–60%

Homopolymer limit: 3

CLI Command: taggen -n 384 -l 30 -d 8 --metric levenshtein --heatmap -o uc3_tags -v

taggen --demux --tags uc3_tags.fasta --reads reads.fastq \

--mode full --position-mask 0.05:0.20 \

--max-dist 4 --trim-mode all

Generation time: ~1.4 s. Expected performance: With –max-dist 4, misassignment is reduced to <0.001% at 90% read identity; approximately 10–15% of reads are left unassigned, which is acceptable when sample integrity is paramount.

Filling a critical gap for long-read sequencing

The adoption of nanopore sequencing for transcriptomics, metagenomics, and clinical applications has outpaced the development of supporting bioinformatic tools. While barcode generators designed for Illumina sequencing work well at 8–12 bp, they cannot scale to the lengths required for error-tolerant nanopore barcoding.

TagGen (see Figure 5 for the graphical and command line user interface snapshots) addresses this gap through algorithmic innovation. By replacing O(4ⁿ) enumeration with O(k) sampling, we transform an exponentially scaling problem into a constant-time operation. The practical consequence is that TagGen succeeds – in milliseconds – at lengths where existing tools fail entirely. TagGen was designed to address a specific and underserved scenario in long-read sequencing: multiplexing experiments in which barcodes are not flanked by known primer or adapter sequences. In standard short-read or kit-based ONT workflows, demultiplexers such as minibar²⁰ or Dorado²¹ leverage flanking adapter sequences as positional anchors, first locating the adapter and then extracting the adjacent barcode for comparison. This strategy works well when the library structure is fixed and the adapter sequence is intact. It fails – or cannot be applied at all – in custom library preparations, direct RNA sequencing, spatial transcriptomics capture plates, and synthetic RNA spike-in approaches, where either no consistent flanking sequence exists or the barcode can appear at variable positions within the read.

Figure 5. TagGen graphical user interface and command-line interface.

The composite screenshot shows four views of the TagGen application. Top left: the Generate tab, where users configure barcode parameters (tag length, count, minimum distance, GC content bounds, homopolymer limit, and exclude sequences) and preview generated barcodes with their core sequences, lengths, and GC content. Top right: the interactive pairwise distance heatmap displaying minimum Hamming or Levenshtein distances between all generated barcodes, enabling visual quality verification; generated barcodes can be exported in FASTA or TSV format. Bottom left: the Demultiplex tab, where users load barcode and FASTQ files, select search mode (end or full), configure position masks and acceptance thresholds, and view demultiplexing results. Centre: the command-line interface showing equivalent generation and demultiplexing commands with their options, demonstrating the full CLI workflow.

Our benchmarks confirm both sides of this distinction. In a head-to-head comparison on primer-flanked dual-ended reads – the use case for which minibar was designed – minibar achieves higher assignment rates than taggen-demux in anchor-free mode, particularly at longer barcodes and higher read identities (88–100% vs 59–97% at 80–95% identity, 30 bp barcodes). This is expected: primer anchoring is a genuinely more powerful strategy when the primer is present and intact. The comparison therefore does not identify a weakness in TagGen; it demarcates the two tools’ intended operating regimes. Where TagGen adds unique value is the large class of applications in which a primer anchor is absent, unreliable, or structurally variable – precisely the conditions under which minibar and analogous tools cannot be used.

Quantified error tolerance for nanopore applications

Our systematic benchmarking across 154 simulated conditions (Panels A–E, Figure 4) provides the first comprehensive characterisation of anchor-free barcode demultiplexing performance as a function of tag length, minimum distance, error metric, position mask, and barcode set size.

The choice of error metric during tag generation has a direct and substantial effect on demultiplexing accuracy. Tags generated with Levenshtein distance constraints outperform Hamming-generated tags at equal nominal minimum distance, with a gap of 3–8 percentage points at 85–90% read identity. The mechanistic explanation is that the greedy selection algorithm, when operating in Levenshtein space, achieves a higher realised minimum pairwise edit distance between tags than the required minimum (typically 12–15 bp achieved vs 8 bp required for l = 30 bp, n = 96). Tags generated with Hamming d = 8 achieve lower realized Levenshtein separation due to the Hamming–Levenshtein inequality, leaving them more vulnerable to insertion and deletion errors characteristic of nanopore sequencing. We therefore recommend Levenshtein generation for all ONT applications. For standard 5′-tagged libraries, end-mode demultiplexing of 30 bp Levenshtein barcodes (d = 8, n = 96) achieves 96% accuracy at 95% read identity and 78% at 85% identity. Full-mode search with a 5′ position mask (−-position-mask 5p:60) matches or slightly exceeds end-mode performance and additionally tolerates reads in which the barcode is shifted from position 0 due to partial adapter degradation – a common occurrence with older R9.4.1 flow cells or degraded samples. This mode is recommended for high sample-count experiments on older chemistry where read starts are less uniform.

For mid-read barcodes – including tags embedded in synthetic RNA spike-ins for single-cell applications or tags placed internal to amplicons – end-mode demultiplexing correctly assigns near-zero reads, confirming that the mode restriction functions as a safety mechanism: reads are rejected rather than misassigned when the search window does not encompass the barcode position. Full-mode with a fractional position mask (−-position-mask 0.05:0.20) achieves 98% accuracy at 90% identity and is robust across variable read lengths. For zero-misassignment applications such as patient sample tracking in single-cell clinical studies, the –max-dist override reduces misassignment to <0.001% at a cost of approximately 10–15% more unassigned reads. Noise simulation provides the first systematic validation of barcode resilience under realistic nanopore error profiles. The results have direct implications for experimental design. Standard ONT sequencing (10–15% error): Barcodes ≥14 bp with d ≥ 4 achieve >97% resolution. Challenging conditions (20% error): Barcodes ≥20 bp with d ≥ 6 maintain >98% resolution. Direct RNA sequencing with degraded samples: 30 bp barcodes with d ≥ 8 provide near-perfect resolution even at 25% error. These findings suggest that researchers using nanopore sequencing should consider migrating from traditional 10–12 bp barcodes to 14–20 bp barcodes for routine applications, and to 24–30 bp barcodes when sample quality is uncertain.

Enabling new applications

TagGen’s capabilities enable several applications previously impractical due to computational limitations:

Limitations

Several limitations should be considered when using TagGen.

Distance metric: TagGen supports both Hamming and Levenshtein distance for barcode generation, and uses Levenshtein distance for demultiplexing. While the validation results confirm high accuracy under indel-rich nanopore error profiles, future work could explore probability divergence¹⁵ as an alternative metric for selection, which may offer additional gains in barcode separability for high-error profiles.

Optimality: The greedy selection algorithm produces locally optimal but not globally optimal barcode sets. For most practical applications with 96–384 barcodes, this approximation is sufficient, but users requiring mathematically optimal sets for very large barcode counts may need to employ more computationally intensive methods such as integer linear programming.

Memory scaling: While TagGen avoids the exponential memory requirements of exhaustive enumeration, very large candidate pools (>1 million sequences) for extremely long barcodes may still require substantial RAM. The default candidate pool size is tuned for typical applications.