In most cases of plasmid construction, the correct (i.e. expected) sequence is known because the process starts with making a blueprint or a map of the plasmid. In addition, typically only plasmids confirmed to have been largely constructed correctly by rapid and inexpensive restriction enzyme digestion or similar laboratory assays are submitted for sequencing analysis. These pieces of a priori information – ‘knowledge of correct, expected sequence of the plasmid’ and ‘certainty that construction largely proceeded correctly’ – can be used to classify the mixed sequence reads and to improve the quality of base calling. Therefore, it should be possible to attain comparable accuracy using a smaller number of reads than that provided in a typical commercial nanopore sequencing sample, enabling barcode-free multiplexing.

The outline of our procedure is as follows: (1) pre-survey, (2) sample submission, and (3) post-analysis (Figure 1). The pre-survey step determines the optimal combination of plasmids suitable for mixing prior to submitting for nanopore sequencing. If the sequences of two plasmids are similar, it becomes more difficult to classify reads from each plasmid a posteriori even with the presence of prior information. Therefore, these kinds of plasmids should be submitted as separate samples. For example, multiple colonies from the same plasmid construction effort cannot be submitted together, because they are expected to have the same sequence (i.e. they share an identical blueprint/map). Therefore, our pre-survey algorithm examines the blueprints of the plasmids and generates recommended groupings as outputs, so that similar ones do not fall into the same group.

One key variable that our pre-survey does not explicitly determine is the maximum number of plasmids that may be safely grouped together. Because read quality, assessed by examining quality score distributions, is typically stable, the main variable affecting multiplexing ability is the overall coverage per plasmid. Having too many plasmids in one group will result in too few reads per plasmid and lower quality scores for the final consensus sequence, but this shortcoming can be compensated if the nanopore flow cell produces higher coverage. However, the coverage provided by commercial sequencing services such as Plasmidsaurus varies from the order of 10²–10³ and is difficult to predict because each nanopore flow cell has different properties. Therefore, it is not possible to determine the maximum number of plasmids that can be mixed in the pre-survey process. Instead, we project that the theoretical minimum number of reads that is required for the reliable consensus calculation is 30 reads per plasmid (discussed in detail later in ‘Maximum number of plasmids that can be mixed’ section). In practice, we typically obtain high-quality results by mixing up to six plasmids when submitting samples for sequencing at the Plasmidsaurus service certified by Oxford Nanopore Technologies, which currently typically provides ~200 or more reads per sample, with read length and quality score distribution properties shown in later figures.

Based on the grouping determined in the pre-survey step, the plasmids classified into the same group are then mixed at equal concentrations. If plasmid concentrations are not equal, the coverage of the plasmids with lower concentrations decreases, affecting the reliability of the results. Following mixing of plasmids according to the grouping, samples are submitted for nanopore sequencing according to specific vendor instructions.

After obtaining results from each nanopore sequencing run, deconvolution is then performed as a post-analysis to obtain consensus sequences for each plasmid. The post-analysis uses the following files as inputs: blueprints of the mixed plasmids (reference sequences) and an output FASTQ file of nanopore sequencing results containing base calling and quality scores of each read (query sequences). The algorithm is divided into three steps: (1) classification of reads, (2) alignment of the classified reads, and (3) calculation of the consensus sequence and quality score (Figure 1). In each step, reference sequences are used as prior information to increase accuracy and quality. The outputs of the post-analysis are FASTQ files containing consensus sequences and quality scores. Two types of FASTQ files are produced: FASTQ files that use arbitrarily set prior probability of error during the plasmid construction at the last step of the post-analysis and those that do not. Results without prior probability are calculated based on an unbiased analysis, whereas results with prior probability are statistically biased toward the blueprint of the plasmid map based on the prior probability (i.e. an arbitrarily set error rate during PCR, ligation, or assembly). The latter option can be considered as analogous to the case where the peak shape of a base within a Sanger sequencing chromatogram is not clear enough for the automatic base calling and is subsequently (and typically manually) compared to the blueprint of the plasmid to determine the identity of the base. Apart from two FASTQ files, a summary GIF file is also provided, showing the composition of reads matched, mismatched, or determined to be deletions or insertions at each position of the plasmid. Because post-analysis is designed assuming that there are no significant differences (insertions or deletions) between blueprints and the actual samples, it is recommended to always check this summary GIF file to confirm whether the sample meets that assumption.

The pre-survey and the post-analysis algorithms are available via a ready-to-execute Jupyter Notebook on Google Colab or locally executable scripts on GitHub (Uematsu, 2024). We have also packaged the script and released it on PyPI to facilitate easy installation and integration by other developers. In the Google Colab version, a function is also included to visualize the classified query sequence along with the reference sequence and the consensus sequence. In addition to analysis of plasmids, SAVEMONEY can also be used for linear DNA such as PCR amplicons and long synthetic DNAs by specifying the ‘topology_of_dna’ option, though we mainly discuss its use for plasmid sequencing in this paper.

The pre-survey algorithm determines the combination of plasmids that are appropriate for mixing. Its initial step is a series of pairwise alignments of each plasmid. Although in principle any alignment algorithm can be used, we first chose to detect chunks of sequences that are sufficiently long and identical to save computational resources, and we then performed an alignment against sequences between chunks by the classical Smith-Waterman algorithm to ensure the accuracy by dynamic programming (Smith and Waterman, 1981). The following parameters were used:match_score=1,mismatch_score=−2,open_gap_penalty=3,extend_gap_penalty=1.

Next, the distance between each plasmid pair was calculated as Levenshtein distance, the minimum number of bases that have to be substituted, deleted, or inserted to change one sequence into the other. Using this framework, the alignment of a plasmid to itself will yield a distance of 0. Calculation of the distances between all plasmid pairs yielded a distance matrix of plasmids. Plasmids with smaller distances were then classified into the same cluster, which we note is different from the final output of the grouping. Hierarchical clustering was adopted using the distance matrix obtained in the previous step, with the nearest point algorithm chosen as the method for calculating the distance between newly formed clusters (Figure 2). Clusters were defined according to a user-defined parameter, distance_threshold, which represents the minimum number of bases that must differ between plasmids in the same group. Therefore, all plasmid pairs with distances lower than distance_threshold are classified into the same cluster. In practice, lower distance_threshold values result in allowing plasmids that are more similar to one another to be classified into different clusters, producing fewer total groups of plasmids for sample submission and thus lowering sequencing costs at the expense of a higher risk of errors during the classification step of the post-analysis algorithm.

Finally, plasmids were classified into groups for sample submission such that plasmids from the same cluster do not fall into the same group and that the minimum distance between plasmids within each group is maximized. To obtain optimal results, we reduced this classification problem to a zero-one integer linear programming problem. First, a rank-3 binary tensor T∈{0,1}g×c×p was prepared, where the sizes of its dimensions correspond to the number of groups, clusters, and plasmids, and each element takes a value of either 0 or 1. The following constraints were applied to this tensor:Tgcp={1if Pp∈Cc,0if Pp∉Cc,∑c,pTgcp=Gg,∑pTgcp≤1,

where Pp and Cc indicate plasmid and cluster with the index of p and c, respectively, and Gg represents the number of plasmids in the group with the index of g, which will be described in more detail below. The constraints represent the following: each plasmid exists only once and belongs to a designated cluster (first constraint), each group contains a designated number of plasmids (second constraint), and plasmids from the same cluster cannot belong to the same group (third constraint). Regarding the first constraint, the cluster to which a plasmid belongs is determined by the results of hierarchical clustering and the values of distance_threshold and number_of_groups. For the second constraint, the number of plasmids in each group can be set arbitrarily, as long as ∑gGg matches the total number of plasmids. However, in the implementation of this algorithm, the number of plasmids in each group was automatically determined to be as equal as possible. In addition to these three constraints, constraints of plasmid pairs that belong to different clusters but should not be distributed into the same group, as expressed by Tgc1p1+Tgc2p2≤1, are sequentially added to a ‘forbidden list’, where c1 and c2 represent the cluster indices to which Pp1 and Pp2 belong, respectively. In the algorithm implementation, in addition to the ‘forbidden list’, an ‘exclusion list’ was also used to prevent plasmid pairs from being added to the ‘forbidden list’. Until all plasmid pairs that belong to different clusters were included in one of the two lists, the following steps were repeated:

This process ultimately leads to the final grouping result.

Examples of the pre-survey results against 14 plasmids are displayed in Figure 2, together with the dendrogram used during the step of hierarchical clustering. When a distance_threshold value of 5 is applied without specifying number_of_groups, plasmids were classified into groups of two, because there are two clusters (clusters 3 and 12 in Figure 2a) containing two plasmids, and plasmids within each cluster have to be distributed into different groups. The minimum distance within each group was 6 (P8 and P10 in Group 2 in Figure 2a) under this condition, but the value increased to 10 (P1 and P4 in Group 1 in Figure 2b) by specifying number_of_groups as 3, indicating that grouping is more quality oriented. By contrast, users can determine the grouping by only specifying number_of_groups (Figure 2c). When the value was set to 4, the algorithm automatically determines the minimum distance_threshold, which in this case was 14, to satisfy the provided number_of_groups. The minimum distance within each group was 20 (P10 and P11 in Group 1 in Figure 2c), indicating the grouping was of the highest quality among the three per-surveys. These results show that lower distance_threshold or number_of_groups values lead to reduced sequencing costs. Conversely, plasmids with higher degrees of similarity will be incorporated into the same group under such settings, which may slightly reduce the reliability of the post-analysis. However, we found that distance_threshold values can be set down to 2 based on the assessment of the limitation of post-analysis, which will be discussed later (see ‘Maximum similarity allowed for mixing’ section for details).

To classify reads from each pore, the alignment was first performed against the reference plasmid pi and the query sequence qk from each pore using the same parameters as described in the pre-survey algorithm. Then, the normalized alignment score, api,qk, was calculated by dividing the alignment score by the length of pi. The plasmid to which the query is assigned, pqk, was determined as follows:pqk=argminpi∈P api,qk,

where P indicates the set of reference plasmids that were mixed. However, if apqk,qk was lower than score_threshold, a user-defined value that represents a cutoff for short reads (see below), the read qk was excluded to increase the quality of subsequent post-analysis. Furthermore, the read qk showing the same normalized alignment score against more than one plasmid was also omitted, because such a read does not contain enough information to determine the plasmid from which it originated. Lastly, the read qk more than two times longer than the length of the assigned reference plasmid pqk, and that showed a higher normalized alignment score than 1 was excluded to omit the read from plasmid multimer.

Four examples of the results from four different groups of plasmids are displayed in Figure 3 (Group I) and Figure 4 (Groups II, III, and IV). Group I involves multiplex sequencing of six plasmids with only modest similarity, and Groups II–IV contain two plasmids each with high similarity. (Note that for Figures 2—4, we have elected to name the plasmids similarly [e.g. P1, P2, etc.] for simplicity; however, they are all distinct, i.e., P1 from Figure 2 is different from P1 in Figure 3.) The pre-survey results for Group I are shown in Figure 3a, indicating two clusters of four (P1–P4) and two (P5–P6) similar plasmids. In reality, each cluster contains plasmids that share a common vector, illustrating how plasmids sharing common vector backbones but different inserts are moderately similar but still suitable for multiplexing, as their distances are far greater than 20, which is a very safe, quality-oriented cutoff value shown in Figure 2c. These six plasmids were mixed and analyzed as a single sample by nanopore sequencing. Figure 3b represents the general quality check of the results for each plasmid: distributions of read length and quality scores.

Reads assigned to each reference plasmid showed sharp read length distributions, indicating the accurate classification of the reads. In addition, the quality score distributions of reads assigned to each reference plasmid are very similar, indicating that each plasmid was sequenced with almost the same quality. Note, however, that the number of reads was not identical for each plasmid, which is reflected by the total histogram area of each colored portion in the read length distribution graphs. Scatter plots of normalized alignment score, api,qk, for each reference plasmid pair are shown in Figure 3c, which is useful to adjust score_threshold. From the graphs, it is apparent that if score_threshold is set higher, the classification accuracy would increase, at the expense of lowering the number of reads assigned to each reference plasmid. For typical experiments, we have found 0.5 to be a reasonable value for score_threshold, but users can fine-tune this parameter to make the most out of the acquired data. For example, if the total number of reads was small but each plasmid was sufficiently different, the threshold can be lowered to increase the number of reads that are passed on to the next step of analysis. Conversely, if the total number of reads is large and the plasmids are highly similar with each other, the quality of subsequent analysis can be improved by raising the threshold. Figure 3c shows the well-separated scatter plots of each read against each reference plasmid, indicating successful classification of reads. In this experiment, a sufficient number of reads (90, 154, 57, 165, 73, 107, and 50 each for P1, P2, P3, P4, P5, P6, and unassigned, respectively) were obtained for each plasmid, which ensures reliability in the subsequent calculation of the consensus sequence.

The results of different combinations of plasmids with high sequence similarity are displayed in Figure 4. The biggest concern in this case is whether each read is accurately classified to each reference plasmid. Assuming the case where two plasmids are mixed, theoretically, the classification of a read to either plasmid is determined solely by the regions where the plasmids differ in sequence. This result occurs because errors in regions where the plasmids have identical sequences have the same effects on the normalized alignment score for both plasmids, i.e., errors in such regions decrease the normalized alignment score in the same manner for both plasmids. Hence, the accurate classification can be achieved for plasmids that differ by even a single base, considering the high accuracy of recent nanopore sequencing. To demonstrate this outcome, we mixed and submitted three sets of two plasmids (set 1: P1–P2, set 2: P3–P4, and set 3: P5–P6) with Levenshtein distances of only 1, 2, and 3 as a single sample (Figure 4a–c). The corresponding results after the classification step of the post-analysis are displayed in Figure 4d–f. Because the plasmids are almost identical, most of the data points are nearly overlapping the y=x line in the scatter plot of the normalized alignment score (Figure 4d–f, top). However, a magnified view clearly shows the deviation of data points from y=x line, indicating that reliable separation of many reads can be achieved (Figure 4d–f, bottom).

Further, we performed a quantitative analysis to estimate the rate of incorrect classification. First, the reads covering the regions where the plasmids differ in sequence were extracted. The numbers of such reads were 239 out of 303, 262 out of 297, and 2212 out of 2359 reads for sets 1, 2, and 3, respectively, and the breakdowns of classification results are shown in Figure 4g–i. The histograms on top also show the number of reads with a normalized alignment score below score_threshold (0.5 in this case) in light gray, but the numbers are all very small, suggesting high sequencing quality. The rotated heatmaps below provide more detailed breakdowns of the reads represented in the histograms above. Using these structured data in the rotated heatmaps, we performed a simple fitting using the least squares method, with the number of reads originating from each plasmid and the nanopore error rate as variables. For simplicity, we approximate that the probability of a nanopore-based base calling error is the same independent of the identity of the base (e.g. when errors occur for true base A, the probability of base calling being T, G, or C is one over 3 for each) and that there are no deletion or insertion errors in base calling. The summary of the fitting results is shown in Figure 4j–l. The upper and lower rotated heatmaps in Figure 4j–l represent the breakdown of reads originating from the plasmid with the lower and higher numbering, respectively. These fitting results allowed us to calculate the ratio of reads classified under a reference plasmid that differs from their actual origin, which was represented as ‘incorrect assignment rate’ in Table 1. The values indicate that plasmids that differ even by a single base (set 1) could be classified to a sufficient confidence level with around 0.03–1.4% incorrect classification. The ratio further goes down when Levenshtein distance was increased to 2 (set 2) or 3 (set 3), transforming the base calling error ratio of 1.55% and 2.13% into the incorrect classification ratio of 0.0074–0.0089% and 0.0143–0.0158%, respectively. These results indicate that the nanopore base calling error negligibly affects the classification step, especially for plasmid pairs differing by two or more bases. The results also suggest that the error should not affect the subsequent analysis to obtain consensus base calling. However, we recommend mixing plasmids that differ by at least two bases, i.e., the distance_threshold value should be at least 2, due to the following reasons. First, although this low percentage of incorrect assignment is unlikely to affect the consensus base calling even with plasmids with 1 base difference, it does affect the consensus quality score, and our algorithm does not make any correction for it. Second, although the average error rate of current nanopore technology is low, it can be increased to up to ~40% for a few specific sequences, such as Dam methylation, and this number could result in incorrect consensus base calling. Third, if only one base differs, that base is the only piece of information that can correctly classify reads, indicating that an unexpected mutation in that position during the plasmid construction would ruin the classification. Therefore, we set the minimum distance_threshold value to 2 in the pre-survey, and we recommend setting it to a higher value when possible.

After the classification step, each read is aligned against the corresponding reference sequence, and then a final post-analysis step is executed to obtain the consensus sequence and quality score. Here, the aligned query sequences, quality scores of each read, and prior information are combined using Bayesian analysis, similar to previously reported methods to detect SNPs (Li et al., 2009). When generating consensus sequences in SAVEMONEY, two types of prior information are used: (1) the error rate during plasmid construction (i.e. an arbitrarily set error rate during PCR, ligation, or assembly), and (2) the characteristics of the nanopore reads, i.e., error rate and distribution of quality score for each base. For example, assume that the correct base at a specific position in the blueprint of the plasmid is A, and that 10 reads were obtained for the corresponding part, of which eight had base calls of A and the other two had G. In this case, the following cases can be considered:

By using prior information, each of the above probabilities can be calculated. It is reasonable to adopt the case with the highest probability among them and use it as the consensus base call. Note that the possibility of deletion and insertion is not considered in the above example to make the example simple, but they are implemented in the actual script.

Although quality scores of each read from Oxford Nanopore Technologies do not exactly follow Phred scores, they are approaching Phred scores in recent years (Delahaye and Nicolas, 2021; Laver et al., 2015). Therefore, for practical purposes, we elected to consider the quality scores as Phred scores, while accepting a small error in the calculation of consensus quality score to some extent. Based on this assumption, the error rate, Ebasecall=B, when the consensus base calling is B at a specific location, can be calculated as follows using Bayes’ theorem:Ebasecall=B=1−P(B∣D1,D2,⋯)=1−P(B)P(D1,D2,⋯∣B)P(D1,D2,⋯),

where PB represents the probability that the true base is B at the specific location of the plasmid, and Dk represents the data obtained from pore k. The prior probability PB can be arbitrarily set from the error ratio during plasmid construction, which corresponds to the first piece of prior information. Here, the likelihood P(D1,D2,⋯∣B) and the probability of obtaining the data PD1,D2,⋯ can be calculated as follows:P(D1,D2,⋯∣B)=∏kP(Dk∣B),P(D1,D2,⋯)=∑B′P(B′,D1,D2,⋯)=∑B′[P(B′)∏kP(Dk∣B′)].

This conversion is guaranteed under the condition that the data obtained by each pore is independent if the true base is known. Further, the likelihood P(Dk∣B) can be converted as follows:P(Dk∣B)=P(Qk,bk∣B)=P(Qk∣bk,B)P(bk∣B),

where bk and Qk represent the base calling and the quality score obtained from pore k, respectively. Here, P(Qk∣bk,B) and P(bk∣B) can be calculated based on the quality score distribution and the error ratio of the nanopore sequencing, respectively, which corresponds to a second piece of prior information. These characteristics can be obtained from nanopore sequencing data of plasmids with known sequences, or, in practice, with unknown sequences by considering the consensus sequence as a known sequence. Although these characteristics change depending on the types of flow cell and library preparation chemistry, in this paper we proceed based on the characteristics of R10.4.1 flow cells with V14 library preparation chemistry by Oxford Nanopore Technologies that are currently used for sequencing via the Plasmidsaurus service, but the principle should be the same with other long-read sequencers. The representative quality score distribution and the error ratio we obtained by analyzing one plasmid by nanopore sequencing are displayed in Figure 5. Thus, the final consensus base calling, Bconsensus, and the consensus Phred score, Qconsensus, can be calculated as follows:Bconsensus=argminB Ebasecall=B,Qconsensus=−10log10⁡Ebasecall=Bconsensus.

The extent to which similar plasmids can be mixed (i.e. how low the distance_threshold value can be in the pre-survey step) is greatly affected by the resolution of the ‘classification’ in the post-analysis. In this step, the use of prior information is important. Assuming that two plasmids, p1 and p2, with different sequences, are mixed and analyzed by nanopore sequencing. If the data from one read from one pore returns a sequence similar to p1, the following two cases are possible:

Intuitively, case 1 is the correct answer, but a more precise expression is that the probability of case 2 is extremely low if plasmid p1 and p2 are ‘sufficiently different’, so it is safe to consider only the possibility of case 1 in practice. This example implies that the prior information of known plasmid sequences in samples improves the accuracy of the classification of the reads from each pore. It is difficult to perform such a classification process with versatility and high accuracy in the absence of prior information.

To specifically determine what kinds of plasmid pairs are ‘sufficiently different’, detailed analysis was performed. First, the percentage of reads with correct base calling was calculated for each position of plasmids and the distribution was obtained (Figure 6a). More than 98% of positions exhibited a correct base calling rate over 0.9, but a small number (0.1%) of them showed low correct rates below 0.7. Averaged quality score distributions of high (>0.9) and low (<0.7) correct rates are displayed in Figure 6b. In the former case, most reads showed the maximum quality score, whereas in the latter case, the percentage is lower. The rate of correct base calling, incorrect base calling, and base calling of deletion were 67.1%, 19.0%, and 13.9%, respectively. Figure 6c shows the bases that were enriched around 5-mers of positions that showed correct rates lower than 0.7 in the ‘worst-case scenario’ indicated in Figure 6a. Using this ‘worst-case scenario’, the probability of incorrect classification when data from one read is classified to p1 was calculated as follows:P(p2∣D=p1)=P(D=p1∣p2)P(p2)P(D=p1),

where PD=p1 represents the probability that one read was classified as p1, i.e., normalized alignment score of the read was higher for p1 than p2. The denominator PD=p1 can be transformed as follows:P(D=p1)=∑k=1,2P(pk)P(D=p1∣pk).

Therefore, the plot of P(p2∣D=p1) can be calculated and the results are displayed in Figure 6d, which varies the number of bases that are different between the plasmids.

When p1 and p2 differ by only one base, the percentage of incorrect classification was estimated to be 39% in this rarely occurring (0.1% frequency) ‘worst-case scenario’. When the sequence differs by two bases, this percentage drops to 22%. This value is sufficiently smaller than 50% and can be considered to have no effect on the consensus sequence. Also, this percentage is calculated assuming that the two bases are both the ‘worst-case scenario’, i.e., the 0.1% case shown in Figure 6a. That means the chance of this happening for two bases that differ in p1 and p2 is 10⁻⁶, which we consider to be small enough to not be a cause for concern. Thus, in most cases, the percentage of incorrect classification will be much lower even if there are only two base differences between p1 and p2. In fact, the percentage of misclassified plasmids was estimated as less than 0.3% at the highest for a two-nucleotide difference from experimental data in Figure 4e–h. Therefore, we ascertain that a distance_threshold value of 2 or more during the pre-survey is sufficient for reliable post-analysis.

The question of how many plasmids can be safely mixed together can be partially replaced by the question of how many reads are required at minimum to obtain a reliable consensus sequence. To answer this question, detailed analysis was performed again using the ‘worst-case scenario’. We simulated nanopore base calling according to quality score distribution when the correct rate was less than 0.7 (Figure 6b, top panel). For a simulation with 20 reads, as an example, 20 sets of quality scores and read types (match, mismatch, deletion, omitted) were generated according to the distribution, and the consensus base calling and quality scores were calculated using SAVEMONEY. Sampling of 10,000 events was performed in each condition over 1–40 reads to calculate the probability of correct/incorrect consensus base calling (Figure 6e). The consensus was calculated in three ways: (1) without prior probability for the unbiased base calling (blue), (2) with prior probability to incorporate an arbitrarily set plasmid construction error rate (orange), and (3) with wrong prior probability for testing error detection (green). The first two conditions correspond to the two outputs from SAVEMONEY, and the last condition indicates the sensitivity of ‘with prior probability’ analysis to detect errors when the reference sequence differs from the actual plasmid.

The results show that the ratio of incorrect base calling decreases exponentially (i.e. linearly on the logarithmic plot), and all three lines reach an error rate of less than 0.01 after more than 30 reads. This analysis shows that, even in the ‘worst-case scenario’, which occurs with a frequency of ~0.1% (Figure 6a), the probability of incorrect base calling is less than 0.1%, indicating that the net probability is ~10⁻⁶. This value is small enough to be negligible even considering the size of usual plasmids (on the order of 10⁴ bases). Therefore, when samples are submitted to a nanopore sequencing service such as Plasmidsaurus, which typically provides a minimum number of ~200 reads per sample, we have found that mixing of up to six plasmids is routinely possible. These conditions bring down the current sequencing cost per plasmid to lower than a single Sanger sequencing run. Further savings and efficiency come from the costs associated with multiple Sanger sequencing runs for inserts longer than 1 kb and re-analysis of occasional failed Sanger sequencing runs. For operators of long-read sequencing, such analysis may enable conservation of expensive reagents. Overall, we expect that the principles underlying our computational approach will accelerate widespread adoption of whole-plasmid sequencing.

All plasmids were either purchased or constructed by Gibson Assembly (Gibson et al., 2009) or cut and paste cloning techniques. Restriction enzyme digestion tests were performed for those constructed before submitting to sequencing. All sequencing was performed by the Plasmidsaurus service certified by Oxford Nanopore Sequencing Technology. Sequencing data were obtained using V14 chemistry on PromethION with an R10.4.1 flow cell, followed by base calling using a high accuracy (HAC) model.

All analyses were performed by using Google Colab or local environment (Python 3.10.0). The following packages were used: BioPython (Chapman and Chang, 2000), kpLogo (Wu and Bartel, 2017), Numpy (Harris et al., 2020), Pandas (pandas-dev, 2024), parasail (Daily, 2016), Pillow (python-pillow, 2025), PuLP (Peschiera and Mitchell, 2024), Spoa (Vaser et al., 2017), tqdm (da Costa-Luis, 2019), Scipy (Virtanen et al., 2020), and SnapGene Reader (Luo, 2018).