A total of 61 K. pneumoniae isolates were collected between April 2018 and November 2023 from 13 departments at the University Hospital Brno (Brno, CZE). For DNA extraction, the DNeasy PowerSoil Pro Kit (Qiagen, Venlo, NL) was used, and DNA purity was measured using a NanoDrop spectrophotometer (Thermo Fisher Scientific, Waltham, MA, USA). The DNA concentration was checked using the Qubit 3.0 Fluorometer (Thermo Fisher Scientific, Wilmington, DE, USA). The sequencing library was prepared using the Rapid Barcoding Kit 96 V14 (Oxford Nanopore Technologies, Oxford, UK). Sequencing was performed using the PromethION 2 Solo sequencing platform (Oxford Nanopore Technologies, Oxford, UK) with R10.4.1 flow cells, and the obtained data were stored in POD5 format.

The signal dataset, comprising squiggles with and without target resistance genes, was prepared as follows. First, POD5 files were basecalled (using a super-accurate model) and demultiplexed with Dorado, integrated into MinKNOW (v23.11.5, nanoporetech.com/community). For each barcode, BAM files were converted to SAM format using Samtools (Danecek et al., 2021). Next, SAM files, together with POD5 files and FASTA files containing query genes, were processed using NANOSLAST (Jakubickova et al., 2025), which identifies squiggles containing target genes and their precise positions within the signal. Only squiggles containing the gene of interest that met empirically defined thresholds for minimum match length and alignment score, which reflect each gene's allelic variability, were stored. The thresholds range between 80%–90% of the maximum alignment score and 80–95% of the length of the searched gene (see Supplementary Table S1). Consequently, CSV files were generated containing nanopore signals with the genes of interest and the indices of their start and end points within the signals.

In total, 26,956 nanopore signals containing genes of interest were obtained (for detailed information, see Supplementary Table S2). The signal lengths ranged from 5,552 to 1.1 million samples, with average and median lengths of 120,590 and 95,361 samples, respectively. Nanopore signals exceeding 300,000 samples occurred only 1,397 times, accounting for 5.2% of the dataset. The mode signal length was 63,736 samples. The signal length histogram is shown in Figure 1. The length of the selected genes (in the signal form) was, on average, 13,341 samples, with a median of 10,812 samples. As shown in Figure 1, the OqxB gene has the longest signal segments, whereas the fosA gene has the shortest.

The obtained data were divided into training, validation, and test sets in a ratio of approximately 80:15:5, with the split performed at the isolate level (i.e., all signals from a given isolate were assigned to a single set). For batch learning, the training signals were randomly fragmented into segments using the data loader. Each segment was then assigned to one of the 11 classes (10 target genes plus a class with no gene of interest) based on its annotated gene content. Within each batch, segments containing gene information were randomly selected, ranging from 1 to the batch size (BS), and the batch was then supplemented with segments lacking any of the genes of interest. In addition, a database of signals without any genes of interest was created for the testing phase, corresponding to five percent of the total number of signals in the database.

Given the nature of the processed signals described in Section 2.2, a hybrid network architecture was proposed, combining a convolutional neural network (CNN) subsampling followed by a transformer encoder block (see Figure 2B). The CNN encoder extracts relevant features and substantially downsamples the signals (see Table 1). The resulting feature representation is then processed by the transformer encoder, which analyses the signal and assigns it to one of 11 classes via a fully connected classifier.

The CNN-based extractor comprises a series of extraction blocks, each comprising two convolutional layers with a 3 × 3 kernel size, followed by a ReLU activation function. The number of blocks varies for different network variants (see Table 1). The ablation study of the number of features and layers in the encoder part was simplified; thus, three model variants were designed and evaluated: Low, Middle, and High (see Table 1). These variants differ in the degree of signal downsampling and the number of extracted features. The CNN encoder produces a sub-sampled signal whose reduced length corresponds to the maximum number of tokens in the transformer block, with each token length determined by the number of extracted features.

An additional ablation study was conducted for the transformer part of the model, focusing on the number of heads and layers, which mainly affect the trade-off between sensitivity and precision. These parameters also had a significant impact on FPR values. The results of this ablation study are shown in Supplementary Figure S1. Based on the achieved evaluation metrics, a configuration with two layers and one head was chosen for the final model.

The network was implemented in Python 3.12 with PyTorch (v2.8). Given the multiclass nature of the classification task, the cross-entropy loss function was chosen, along with the ADAM optimization algorithm, using β₁ = 0.900 and β₂ = 0.999. Based on preliminary experiments, the number of epochs was set to 250, and the best model was saved according to validation performance to prevent both overfitting and underfitting.

An ablation study to determine the optimal learning hyperparameters for each model was conducted using a sophisticated Bayesian optimization method, specifically, the learning rate (LR), batch size (BS), and dropout probability (DP). The resulting parameters were consistent across all models: LR = 0.001, BS = 64, and DP = 0.0. This optimization was performed using the Bayesian Optimization package (Nogueira, 2014) with 20 iterations and 5 initialization points. A 5-fold cross-validation yielded an average classification accuracy of 81.74% (SD 2.53%), indicating low sensitivity to random sampling and suggesting robustness to variations in data balance and variability.

Basic data processing and testing were performed on a workstation equipped with an Intel Core i9 3.70 GHz processor, 64 GB of RAM, a 2 TB M.2 SSD, and an EVGA GeForce RTX 3090 graphics card with 24 GB of GDDR6 memory. For more computationally intensive tasks, such as network training and Bayesian hyperparameter optimization, resources provided by the e-INFRA CZ project (ID:90254) were used.

The network was trained as a multi-class classifier, assigning signal segments to one of 11 predefined classes. For this classification task, a confusion matrix was computed on the validation set, and the ROC analysis for the Middle model is shown in Figure 3 (see Supplementary Figures S2, S3 for the remaining models). During training, the model achieved an average validation accuracy of 87.5%.

The ROC analysis indicates that the OqxB gene (gray curve in Figure 3) is particularly challenging to classify (AUC = 0.80), likely due to its close genomic proximity to OqxA. Another contributing factor may be the length of the OqxB gene segments, which are the longest among all genes, as shown in Figure 1. Furthermore, a substantial portion of these segments exceeds our window length of 40,000 samples, indicating that partial gene fragments occur more frequently in this gene than in others, which may be the primary reason for the lower accuracy.

Nevertheless, the overall performance improves markedly when using a floating window with a defined overlap, which enables prediction boosting for local classification and supports multiple localization tasks.

To address the variability and extreme length of signals, an approach utilizing a floating window (shown in Figure 2A) with a window size of 40,000 samples was proposed. The window size covers most gene lengths and allows for a defined extent. Each segment within the floating window is classified; for every shift, the network outputs class-assignment probabilities for the segment. Before being processed by the network, the signal is extended (padded) at both ends by half the window length to mitigate edge effects. The chosen overlap determines the density of classified positions in the signal, which are then interpolated across all signal samples (an example of such a function is shown in Figure 2C). This approach enables the approximation of the gene location, with detection resolution influenced by both the overlap and the window size. However, precise gene positioning is beyond the scope of this study, which focuses solely on gene identification within reads. This process yields 11 probability values per signal sample, corresponding to the predefined classes. The final decision on gene presence or absence is based on selecting the class with the highest probability value for each signal sample. As part of the ablation study performed during the development of the sliding window approach, the impact of the window overlap percentage on detection performance was evaluated, as shown in Figure 4.

Introducing a floating window with overlapping acts as a form of boosting, enabling multiple classifications of overlapping signal regions. This approach eliminates false detections caused by the possible inclusion of an incomplete gene region in the currently classified segment. The effect of boosting amplifies with increasing window overlap; however, it also increases computational complexity, as the signal is divided into multiple segments. According to the graph in Figure 4, there is an apparent natural increase in sensitivity; however, the probability of false detection of selected genes in similar regions is also increased, thereby reducing precision. The compromise with the highest average success rate seems to be an 80% overlap.

The possible inclusion of only part of a gene in a floating window is a natural effect that can affect the success of identification. The graph in Supplementary Figure S4 shows the dependence of the average identification sensitivity (calculated across all ten genes) on the proportion of the gene covered by the floating window, revealing an exponential trend. As the model may rely on different informative regions for different genes and cases, it is therefore unsurprising that the sensitivity reaches its maximum only when the entire gene is contained within the floating window. When only half of the gene is present in the window, the sensitivity decreases to 18%. Problems may thus arise when a gene lies at the end of a nanopore signal and is therefore not fully captured. However, since each strain is sequenced at an average depth of ~100 × , other reads typically contain the full gene, effectively mitigating this limitation.

ARGs may occur multiple times within a bacterial genome, and similar genes often cluster in close proximity, complicating detection. Therefore, classifying entire read alone is insufficient; the floating window approach offers a new perspective on gene detection, even for multiple occurrences within a single read (signal). For each floating window position, probabilities for all gene classes, including the no gene category, are computed, and the class with the highest probability is selected as a final decision. Figure 6 illustrates examples of four signals with multiple occurrences of ARGs. In Figure 6C, a false negative case for the blaOXA gene (blue region) is observed, where the predicted absence probability exceeded the presence probability, despite BLAST confirming the gene's presence. This is despite the fact that, according to BLAST detection, the gene is present. Conversely, in the next signal (Figure 6D), OqxA shows a high predicted probability despite local alignment failing to detect the gene, with nucleotide sequence identity at only approximately 50%. No such false detections occur in the High model, suggesting that the false detections may be caused by information loss at higher subsampling rates introduced by the encoder or by lower network capacity, making it more difficult to identify slight changes in the signal pattern.

The time complexity of the selected model was evaluated on a test dataset under CPU and GPU configurations, with and without parallelization. Figure 7 illustrates how time complexity varies with the overlap parameter in the floating window approach. GPU batch processing offers substantial parallelization benefits, reducing computation time to approximately 6 h per million reads. Scaling to a larger number of classes (up to 1,000 gene types) is expected to minimally impact computational complexity, thanks to the network's architecture. Thus, the proposed model shows strong potential for detecting additional resistance-related genes, provided that the annotated dataset is expanded and the model is retrained.

Nanopore read processing traditionally relies on basecalling, where raw signals are converted into nucleotide sequences before downstream analysis. This conversion introduces substantial delays in the analytic workflow, as shown in Figure 8, which depicts cumulative read processing over time. Figure 8 reflects data from the sequencing run described in Section 2.1. Sequencing and basecalling statistics were extracted from MinKNOW logs, while NanoResFormer performance metrics were derived from computational time analysis (Section 3.3).

As shown in Figure 8, after 48 h of sequencing, only 20% of reads have been basecalled. Moreover, on the basecalled data, other tools capable of ARG identification need to be applied, further extending the time required to obtain crucial results. In contrast, our basecalling-free approach enables real-time direct analysis of raw signals, including AMR gene detection during sequencing. Within the first six h of sequencing, over one million reads can be screened with >92% sensitivity. The full sequencing run can be processed in approximately 96 h, less than half the time required for basecalling (216 h).

While basecalling remains necessary for some downstream applications, rapid ARGs identification via a basecalling-free strategy offers clear clinical benefits. Using Bernoulli distribution-based calculations and considering the model's sensitivity, the minimum number n of reads required to ensure that at least one of the resistant genes is included in the random selection with the specified level of confidence has been estimated. Using the lowest observed gene frequency (p_min = 0.0298%, Supplementary Table S2) and the formula n = ln(1 − 0.999)/ln(1−p_min), it has been calculated that 24,586 reads are required to achieve 99.9% confidence in detecting at least one target gene. Given NanoResFormer's processing speed (Section 3.3), the first resistant gene is likely to be detected within 9.3 min. If no gene is found during this time, the absence can be inferred. If detected, processing an additional 805,545 reads (~5.2 hours) ensures 99.9% confidence that all resistance genes present are identified.

Due to the lack of similar published algorithms for the same task, our proposed approach is primarily compared with NanoGeneNet (Nykrynova et al., 2022), which, as mentioned in the Introduction, uses a combination of LSTM and CNN. This model was originally evaluated for the classification of seven MLST genes in nanopore signals, achieving a sensitivity of 92.0%. Although our model was designed to detect a different set of genes (10 resistance genes), it achieved very similar results for the same task (i.e., gene identification) on comparable nanopore signals generated using a newer sequencing chemistry (R10 instead of R9), with a sensitivity of 92.6%.

When comparing inference times, transformer-based models demonstrate a clear advantage thanks to their parallelization capabilities and efficient batch processing. The experiment shown in Figure 8, where both tools were evaluated under identical conditions, i.e., on the same hardware and nanopore signal testing database, confirmed this advantage, as NanoGeneNet was up to three times slower at identifying nanopore signals. This performance gap highlights transformer models' potential and their suitability for the real-time signal processing tasks addressed in this work.

As noted by the authors of NanoGeneNet (Nykrynova et al., 2022), training LSTM-based models is challenging and not entirely trivial, often requiring sequential training to mitigate the vanishing and exploding gradient problem. In contrast, transformers enable single-shot training, which makes it feasible to release, in future work, a user-friendly tool that enables users to retrain the model on new or custom nanopore signals.

The final area of comparison is the ability to detect multiple gene occurrences within a single signal. NanoGeneNet does not support this functionality because it uses global signal classification, albeit with localization. In our proposed tool, the implementation of a floating-window approach enables local classification, thereby detecting multiple gene occurrences within a single read.

Although this study demonstrates the feasibility of transformer-based resistance gene detection from squiggles, several limitations must be addressed before clinical implementation. The current model was trained on a limited set of resistance genes, restricting its clinical applicability. Expanding the model to cover the comprehensive resistance gene panel required for clinical diagnostics poses significant challenges. Adding more gene classes directly to the existing architecture may be constrained by memory limitations and could reduce classification accuracy due to increased model complexity. A hierarchical modeling approach could be implemented to address this limitation through a two-tier classification system: broad-spectrum models would identify the presence of resistance gene families, followed by specialized models that resolve specific allelic variants when required. This strategy optimizes computational efficiency by avoiding simultaneous searches for all possible alleles and focusing only on those detected during the initial screening, thus making real-time implementation more feasible.

Another limitation may arise during the preparation of training data. Current labeling relies on BLAST-based alignment, which can introduce systematic bias. Specifically, highly variable alleles may evade detection yet remain recognizable to the transformer model, potentially causing false-positive classifications. Furthermore, these tools may not reliably identify newly emerged or significantly mutated variants of resistance genes that the model could still detect.

A further challenge lies in the diversity of sequencing data used to train the model. Training datasets should ideally span multiple library chemistries, flowcell types, and even different sequencing platforms to enhance robustness and generalization. However, creating such a diverse dataset is challenging in practice. Nevertheless, the dataset used in this study is based on the latest R10 chemistry, which is currently the most widely used and serves as the de facto standard in Oxford Nanopore Technologies (ONT) workflows.

To handle over-length reads that transformers cannot efficiently process at their original length, a floating-window approach has been proposed. This method enables the processing of variable-length signals, including extremely long ones, and improves detection success rates. Transformers complexity scales quadratically with token (window length) count; therefore, a 40,000-sample floating window was chosen, covering the lengths of all nine genes except OqxB (average length of 40,412 samples). Nevertheless, this gene was included in the study as it represents a borderline case in terms of the length of successful gene identification. Despite its borderline length, OqxB identification achieved 66% sensitivity. However, the transformer network cannot capture the entire signal context, and enlarging the window would significantly increase computational time. Overlapping floating windows partially mitigate this limitation, but it remains a practical concern.