Basecalling Algorithms for Nanopore Sequencing: A Comprehensive Review

Introduction

Nanopore sequencing has transformed genomic research by enabling the direct, real-time analysis of single nucleic acid molecules. Unlike short-read technologies that rely on sequencing by synthesis, nanopore sequencing measures changes in ionic current as a polynucleotide passes through a biological or solid-state nanopore. The conversion of these raw electrical signals into nucleotide sequences, termed basecalling, is a critical computational step that defines the accuracy, throughput, and utility of the technology. This review provides an exhaustive examination of basecalling algorithms, from early hidden Markov model (HMM) approaches to contemporary deep learning architectures, with a focus on applications in veterinary medicine and molecular diagnostics. The content is organized to cover the biophysical basis of signal acquisition, algorithmic evolution, specialized tasks such as RNA modification detection, benchmarking efforts, and integration with downstream analytical workflows relevant to pathogen surveillance, antimicrobial resistance profiling, and host-pathogen interaction studies.

The scope of this article is confined to computational methods and does not address commercial instrumentation details. All references are drawn from peer-reviewed sources to ensure academic rigor [1, 2, 3].

Biophysical Basis of Nanopore Signal Acquisition

The fundamental principle of nanopore sequencing involves a voltage-driven translocation of a single-stranded DNA or RNA molecule through a nanometer-scale pore embedded in a membrane. As the molecule passes, it occludes the pore, causing characteristic disruptions in the ionic current. The magnitude and duration of these disruptions are influenced by the specific combination of nucleotides (k-mer) occupying the pore's sensing region, typically spanning 4–6 nucleotides. The resulting current trace, often termed a "squiggle," is a continuous time series of measurements sampled at rates on the order of 4 kHz. Key factors affecting signal quality include translocation speed, temperature, pore dimensions, and the presence of modifications such as methylation or pseudouridylation [4, 59].

Initial signal processing requires segmentation, where the continuous raw signal is partitioned into discrete events, each corresponding to a single k-mer translocation step. Segmentation algorithms identify step-like transitions in the current trace, and their accuracy directly impacts basecalling performance. A recent comprehensive cross-species comparison of segmentation tools, Dynamont, evaluated multiple algorithms on diverse taxa and highlighted that segmentation quality varies significantly with sequence context and base composition [3]. Other work has focused on overlapping raw signals using hash-based seeding mechanisms (Rawsamble) to improve alignment and consensus without prior basecalling [1].

Evolution of Basecalling Methodologies

Hidden Markov Model and Viterbi-Based Approaches

Early basecalling relied on HMMs that modeled the relationship between k-mer states and observed current levels. The Viterbi algorithm was applied to find the most likely sequence of hidden states (k-mers) given the observed signal [63]. These models required precomputed k-mer models generated from training data with known sequences. The HMM approach was computationally efficient but suffered from limited accuracy, particularly in homopolymer regions and at low signal-to-noise ratios. Subsequent improvements used time-varying cross-membrane voltages to increase translocation control and reduce systematic errors [58]. Despite their limitations, HMM-based methods laid the foundation for modern probabilistic frameworks and remain relevant for certain edge deployments where deep learning models are impractical [5].

Deep Learning Revolution

The introduction of deep neural networks marked a paradigm shift in basecalling accuracy. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs), particularly long short-term memory (LSTM) networks, were adapted to process segments of raw signal and predict nucleotide sequences. Chiron was among the first end-to-end deep learning basecallers, translating raw signal directly into sequence without explicit event detection [60]. Later architectures incorporated attention mechanisms, such as the self-attention basecaller SACall, which improved context capture over long-range dependencies [6]. A comprehensive benchmark of deep learning models for nanopore basecalling demonstrated that architectures combining CNNs with bidirectional LSTMs consistently outperformed pure HMM methods, although performance varied across sequence contexts [7].

Deep Learning Architectures for Basecalling

Convolutional and Recurrent Networks

Most production-grade basecallers employ a hybrid CNN-RNN architecture. The CNN layers extract local features from the raw signal or segmented events, while the RNN layers model temporal dependencies. For example, MSRCall uses a multi-scale deep neural network that processes signal at multiple resolutions to capture both fine and coarse features [8]. Similarly, Lokatt integrates an explicit duration HMM with a residual LSTM network, providing a hybrid approach that combines the strengths of probabilistic duration modeling with deep learning [9]. The RUBICON framework provides a systematic design space exploration for efficient deep learning basecallers, enabling trade-offs between accuracy and computational resource usage [10].

Transformer and Attention-Based Models

Transformer architectures, originally developed for natural language processing, have been adapted for basecalling due to their ability to model long-range dependencies through self-attention. SACall uses a self-attention mechanism to assign importance to different parts of the input signal [6]. BaseNet is a transformer-based toolkit for nanopore signal decoding that achieves competitive accuracy with lower computational overhead compared to LSTM-based models [11]. GCRTcall extends the transformer architecture with gated convolution and relative position embedding, specifically for RNA basecalling, and uses joint loss training to improve both accuracy and robustness [12]. The trend toward transformers suggests they will become the dominant architecture as hardware acceleration improves.

Sequence-to-Sequence and Pair Consensus Decoding

Basecalling can be framed as a sequence-to-sequence problem where the input is the raw signal and the output is the nucleotide sequence. Some approaches process both raw and event-level data jointly to leverage complementary information [13]. Pair consensus decoding improves neural network basecaller accuracy by comparing two independently basecalled reads from the same molecule and resolving discrepancies [14]. This technique is particularly useful for polishing consensus genomes of microbial pathogens, including those relevant to veterinary medicine.

Specialized Basecalling Tasks

Detection of DNA Base Modifications

Nanopore sequencing inherently detects base modifications because modified nucleotides alter the ionic current profile uniquely. Early methods for methylation detection relied on comparing observed signals to unmodified reference models. Modern approaches use deep learning to classify modifications directly from raw signal. DeepMP is a deep learning tool for detecting DNA base modifications, including 5-methylcytosine and 6-methyladenine [15]. A multi-scale neural network approach identifies DNA methylation types and methylated base positions in bacterial genomes [16]. The tool Snappy enables fast identification of DNA methylation motifs from nanopore reads [17]. For veterinary applications, detecting epigenetic markers in bacterial pathogens such as Mycoplasma bovis can inform virulence regulation and host adaptation. (See Mycoplasma bovis in Feedlot Cattle.) Methods that adapt basecalling models for modification detection via incremental learning and anomaly detection allow continuous model improvement without full retraining [18].

RNA Basecalling and Modification Detection

Direct RNA sequencing bypasses reverse transcription and enables direct detection of RNA modifications. However, RNA basecalling faces additional challenges due to the lower stability of RNA molecules and the presence of abundant modifications that distort the signal. Advances in nanopore direct RNA sequencing have enabled detection of N6-methyladenosine (m6A), pseudouridine, and 5-methylcytosine. EpiNano and nanoDoc2 use basecalling error profiles and signal features to detect m6A modifications [19, 20]. Accurate detection of m6A from native RNA sequences was demonstrated using a combination of signal-level and basecalling error features [55]. Quantitative profiling of pseudouridylation dynamics is also possible with nanopore sequencing [21]. The genomic language model approach mitigates chimera artifacts in direct RNA sequencing, improving the reliability of transcriptome-level analyses [2].

A semisupervised learning framework, m6Anet, detects m6A modifications from nanopore direct RNA sequencing by leveraging both labeled and unlabeled data [22]. Another approach uses raw current signals and basecalling errors jointly to detect RNA m6A [23]. The tool MoDorado extends modification detection to tRNA by repurposing modification callers originally designed for other RNA types [24]. These methods are highly relevant for studying RNA viruses of veterinary importance, such as Avian Influenza and Feline Coronavirus.

Poly(A) Tail Length Inference

Direct RNA sequencing also allows measurement of poly(A) tail lengths, which regulate mRNA stability and translation efficiency. A benchmark using synthetic RNAs demonstrated that poly(A) tail length inference from nanopore data is accurate but requires careful calibration, especially for long tails [25]. Misestimation of tail length can occur due to signal compression and basecalling biases.

Benchmarking and Error Profiles

Systematic evaluation of basecaller performance is essential for both developers and end-users. A landmark study compared neural network basecallers and found that accuracy improved dramatically over HMM-based methods, but systematic errors remained in homopolymer regions and at the ends of reads [57]. Benchmarking of deep learning variant callers on bacterial nanopore data revealed that the choice of basecaller and downstream variant caller significantly affects variant detection accuracy [26]. For metagenomic applications, simulation tools like Meta-NanoSim allow comparison of basecaller performance on complex microbial communities [27].

Error correction algorithms such as NanoReviser use deep learning to correct basecalling errors post hoc [28]. Others have developed n-polymer realigners to improve pileup-based variant calling by realigning reads in repetitive regions [29]. The evaluation of polishing tools for microbial genome assembly showed that different polishing strategies (e.g., using short reads or deep learning-based polishing) can substantially improve consensus accuracy [30].

A comparative analysis of segmentation algorithms and their impact on basecalling highlighted that even minor changes in segmentation parameters can propagate errors into final sequences [56]. The study Dynamont provided a comprehensive cross-species comparison of segmentation tools, emphasizing the need for taxon-specific calibration [3].

Integration with Downstream Analysis

Real-Time Basecalling and Selective Sequencing

One of the unique advantages of nanopore sequencing is the ability to perform real-time analysis, enabling selective sequencing where only molecules of interest are fully sequenced. UNCALLED enables real-time mapping of raw electrical signals to a reference genome, allowing targeted sequencing by rejecting off-target molecules [31]. This approach has been applied to sequence specific pathogens from mixed communities, such as Escherichia coli in poultry samples (see Escherichia coli in Chickens and Poultry Products). The tool ReadUntil has been further developed with a simulation framework (SimReadUntil) to benchmark selective sequencing algorithms [32]. Real-time mapping of raw signals is also possible with methods that avoid full basecalling, reducing latency for time-sensitive applications [33].

Basecalling for Clinical and Field Diagnostics

For veterinary diagnostics, rapid and accurate basecalling is critical for identifying pathogens, detecting antimicrobial resistance markers, and performing outbreak investigations. The ability to perform basecalling on portable devices (edge computing) has been explored using FPGA acceleration [53] and by designing lightweight models that fit within resource constraints [10]. A study using a taxon-specific basecaller for Mycoplasma bovis demonstrated that custom basecalling models can improve assembly quality for low-GC-content pathogens [34]. The tool MysteryMaster addresses the problem of extracting usable reads from poorly barcoded samples, which is common in field-collected veterinary specimens [35]. Similarly, Porechop_ABI detects and trims unknown adapters from nanopore reads, improving downstream analysis quality [36].

Simulation Tools for Algorithm Development

Simulation tools are indispensable for developing and benchmarking basecalling algorithms. NanoSim generates synthetic nanopore reads based on statistical characterization of real data [62]. More recent simulators use feed-forward transformers to model raw signal generation end-to-end [37]. Tunable parameters allow testing of algorithm robustness to noise, translocation speed, and modification presence [38]. These simulators enable rigorous evaluation of basecallers without the expense of real sequencing runs.

Future Directions

The field is moving toward higher accuracy through integration of multiple signal modalities, such as concurrent measurement of ionic current and optical or electrochemical signals. The development of non-canonical basecalling for heavily modified DNA has been proposed for data encryption and molecular barcoding, which could have applications in sample tracking for large veterinary surveillance studies [39]. Extraction of metadata from basecalled files allows restoration of flow cell type and basecaller configuration, enabling reproducibility audits [40].

A key gap is the lack of standardized benchmarks for RNA basecalling and modification detection, especially for viral RNA genomes. The community resource SquiDBase provides raw nanopore data from microbes to facilitate algorithm development, but more diverse datasets from veterinary-relevant pathogens are needed [41]. The lightweight k-mer model generation from basecaller move tables offers new ways to accelerate downstream analyses without repeated basecalling [42].

Conclusion

Basecaling algorithms for nanopore sequencing have evolved from simple HMMs to sophisticated deep learning architectures, enabling unprecedented accuracy and versatility. These advances have made nanopore sequencing a powerful tool for veterinary diagnostics, allowing real-time detection of pathogens, epigenetic markers, and RNA modifications. Continued algorithmic innovation, combined with robust benchmarking and simulation tools, will further expand the utility of nanopore sequencing in animal health and disease surveillance.

Workflow Diagram

flowchart TD
    A[Raw Ionic Current Signal] --> B[Signal Segmentation]
    B --> C[Feature Extraction]
    C --> D[Basecalling Engine]
    D --> E[Sequence Output]
    D --> F[Error Correction]
    F --> G[Polished Consensus]
    D --> H[Modification Detection]
    H --> I[Epigenetic/Epitranscriptomic Profile]
    D --> J[Real-Time Mapping]
    J --> K[Selective Sequencing]
    K --> L[Targeted Pathogen Enrichment]
    G --> M[Variant Calling & Assembly]
    M --> N[Veterinary Diagnostic Report]

Summary of Key Basecalling Approaches

References

[1] Firtina C, Mordig M, Mustafa H et al. Rawsamble: overlapping raw nanopore signals using a hash-based seeding mechanism. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/41746292/

[2] Li Y, Wang TY, Guo Q et al. Genomic language model mitigates chimera artifacts in nanopore direct RNA sequencing. Nat Commun. URL: https://pubmed.ncbi.nlm.nih.gov/41554734/

[3] Spangenberg J, Siederdissen CHZ, Goettsch W et al. Dynamont: A comprehensive cross-species comparison of ONT segmentation tools. Gigascience. URL: https://pubmed.ncbi.nlm.nih.gov/41553355/

[4] Xie S, Leung AW, Zheng Z et al. Applications and potentials of nanopore sequencing in the (epi)genome and (epi)transcriptome era. Innovation (Camb). URL: https://pubmed.ncbi.nlm.nih.gov/34901902/

[5] Perešíni P, Boža V, Brejová B et al. Nanopore base calling on the edge. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/34314502/

[6] Huang N, Nie F, Ni P et al. SACall: A Neural Network Basecaller for Oxford Nanopore Sequencing Data Based on Self-Attention Mechanism. IEEE/ACM Trans Comput Biol Bioinform. URL: https://pubmed.ncbi.nlm.nih.gov/33211664/

[7] Pagès-Gallego M, de Ridder J. Comprehensive benchmark and architectural analysis of deep learning models for nanopore sequencing basecalling. Genome Biol. URL: https://pubmed.ncbi.nlm.nih.gov/37041647/

[8] Yeh YM, Lu YC. MSRCall: a multi-scale deep neural network to basecall Oxford Nanopore sequences. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/35766808/

[9] Xu X, Bhalla N, Ståhl P et al. Lokatt: a hybrid DNA nanopore basecaller with an explicit duration hidden Markov model and a residual LSTM network. BMC Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/38062356/

[10] Singh G, Alser M, Denolf K et al. RUBICON: a framework for designing efficient deep learning-based genomic basecallers. Genome Biol. URL: https://pubmed.ncbi.nlm.nih.gov/38365730/

[11] Li Q, Sun C, Wang D et al. BaseNet: A transformer-based toolkit for nanopore sequencing signal decoding. Comput Struct Biotechnol J. URL: https://pubmed.ncbi.nlm.nih.gov/39391372/

[12] Li Q, Sun C, Wang D et al. GCRTcall: a transformer based basecaller for nanopore RNA sequencing enhanced by gated convolution and relative position embedding via joint loss training. Front Genet. URL: https://pubmed.ncbi.nlm.nih.gov/39649096/

[13] Napieralski A, Nowak R. Basecalling Using Joint Raw and Event Nanopore Data Sequence-to-Sequence Processing. Sensors (Basel). URL: https://pubmed.ncbi.nlm.nih.gov/35336445/

[14] Silvestre-Ryan J, Holmes I. Pair consensus decoding improves accuracy of neural network basecallers for nanopore sequencing. Genome Biol. URL: https://pubmed.ncbi.nlm.nih.gov/33468205/

[15] Bonet J, Chen M, Dabad M et al. DeepMP: a deep learning tool to detect DNA base modifications on Nanopore sequencing data. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/34718417/

[16] Chen Z, Ni P, Wang J. Identifying DNA methylation types and methylated base positions from bacteria using nanopore sequencing with multi-scale neural network. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/40658463/

[17] Konanov DN, Krivonos DV, Babenko VV et al. Snappy: fast identification of DNA methylation motifs based on oxford nanopore reads. Bioinform Adv. URL: https://pubmed.ncbi.nlm.nih.gov/41357306/

[18] Wang Z, Fang Y, Liu Z et al. Adapting nanopore sequencing basecalling models for modification detection via incremental learning and anomaly detection. Nat Commun. URL: https://pubmed.ncbi.nlm.nih.gov/39169028/

[19] Liu H, Begik O, Novoa EM. EpiNano: Detection of m(6)A RNA Modifications Using Oxford Nanopore Direct RNA Sequencing. Methods Mol Biol. URL: https://pubmed.ncbi.nlm.nih.gov/34085237/

[20] Ueda H, Dasgupta B, Yu BY. RNA Modification Detection Using Nanopore Direct RNA Sequencing and nanoDoc2. Methods Mol Biol. URL: https://pubmed.ncbi.nlm.nih.gov/36781737/

[21] Begik O, Lucas MC, Pryszcz LP et al. Quantitative profiling of pseudouridylation dynamics in native RNAs with nanopore sequencing. Nat Biotechnol. URL: https://pubmed.ncbi.nlm.nih.gov/33986546/

[22] Teng H, Stoiber M, Bar-Joseph Z et al. Detecting m6A RNA modification from nanopore sequencing using a semisupervised learning framework. Genome Res. URL: https://pubmed.ncbi.nlm.nih.gov/39406497/

[23] Ni P, Xu J, Zhong Z et al. RNA m6A detection using raw current signals and basecalling errors from Nanopore direct RNA sequencing reads. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/38889266/

[24] Rübsam FNM, Liu-Wei W, Sun Y et al. MoDorado: enhanced detection of tRNA modifications in nanopore sequencing by off-label use of modification callers. Nucleic Acids Res. URL: https://pubmed.ncbi.nlm.nih.gov/40829803/

[25] Chang JJ, Yang X, Teng H et al. Using synthetic RNA to benchmark poly(A) length inference from direct RNA sequencing. Gigascience. URL: https://pubmed.ncbi.nlm.nih.gov/40899916/

[26] Hall MB, Wick RR, Judd LM et al. Benchmarking reveals superiority of deep learning variant callers on bacterial nanopore sequence data. Elife. URL: https://pubmed.ncbi.nlm.nih.gov/39388235/

[27] Yang C, Lo T, Nip KM et al. Characterization and simulation of metagenomic nanopore sequencing data with Meta-NanoSim. Gigascience. URL: https://pubmed.ncbi.nlm.nih.gov/36939007/

[28] Wang L, Qu L, Yang L et al. NanoReviser: An Error-Correction Tool for Nanopore Sequencing Based on a Deep Learning Algorithm. Front Genet. URL: https

[29] Dunn T, Blaauw D, Das R et al. nPoRe: n-polymer realigner for improved pileup-based variant calling. BMC Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/36927439/

[30] Lee JY, Kong M, Oh J et al. Comparative evaluation of Nanopore polishing tools for microbial genome assembly and polishing strategies for downstream analysis. Sci Rep. URL: https://pubmed.ncbi.nlm.nih.gov/34671046/

[31] Kovaka S, Fan Y, Ni B et al. Targeted nanopore sequencing by real-time mapping of raw electrical signal with UNCALLED. Nat Biotechnol. URL: https://pubmed.ncbi.nlm.nih.gov/33257863/

[32] Mordig M, Rätsch G, Kahles A. SimReadUntil for benchmarking selective sequencing algorithms on ONT devices. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/38603597/

[33] Zhang H, Li H, Jain C et al. Real-time mapping of nanopore raw signals. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/34252938/

[34] Vereecke N, Bokma J, Haesebrouck F et al. High quality genome assemblies of Mycoplasma bovis using a taxon-specific Bonito basecaller for MinION and Flongle long-read nanopore sequencing. BMC Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/33176691/

[35] Khezri A, Branders S, Bellankimath AB et al. MysteryMaster: scraping the bottom of the barrel of barcoded Oxford nanopore reads. BMC Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/41034732/

[36] Bonenfant Q, Noé L, Touzet H. Porechop_ABI: discovering unknown adapters in Oxford Nanopore Technology sequencing reads for downstream trimming. Bioinform Adv. URL: https://pubmed.ncbi.nlm.nih.gov/36698762/

[37] Beslic D, Kucklick M, Engelmann S et al. End-to-end simulation of nanopore sequencing signals with feed-forward transformers. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/39710838/

[38] Gamaarachchi H, Ferguson JM, Samarakoon H et al. Simulation of nanopore sequencing signal data with tunable parameters. Genome Res. URL: https://pubmed.ncbi.nlm.nih.gov/38692839/

[39] Fan Q, Zhao X, Li J et al. De novo non-canonical nanopore basecalling enables private communication using heavily-modified DNA data at single-molecule level. Nat Commun. URL: https://pubmed.ncbi.nlm.nih.gov/40316536/

[40] Mencius J, Chen W, Zheng Y et al. Restoring flowcell type and basecaller configuration from FASTQ files of nanopore sequencing data. Nat Commun. URL: https://pubmed.ncbi.nlm.nih.gov/40316544/

[41] Cuypers WL, Ceylan H, Turcksin E et al. SquiDBase: a community resource of raw nanopore data from microbes. NAR Genom Bioinform. URL: https://pubmed.ncbi.nlm.nih.gov/41524028/

[42] Samarakoon H, Wan YK, Parameswaran S et al. Leveraging basecaller's move table to generate a lightweight k-mer model for nanopore sequencing analysis. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/40085000/

[43] Hammad K, Wu Z, Ghafar-Zadeh E et al. Nanopore-Aware Embedded Detection for Mobile DNA Sequencing: A Viterbi-HMM Design Versus Deep Learning Approaches. Biosensors (Basel). URL: https://pubmed.ncbi.nlm.nih.gov/41002309/

[44] Sun K, Li J, Chen C et al. Advances in nanopore direct RNA sequencing and its impact on biological research. Biotechnol Adv. URL: https://pubmed.ncbi.nlm.nih.gov/40930310/

[45] Shivakumar VS, Ahmed OY, Kovaka S et al. Sigmoni: classification of nanopore signal with a compressed pangenome index. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/38940135/

[46] Liu-Wei W, van der Toorn W, Bohn P et al. Sequencing accuracy and systematic errors of nanopore direct RNA sequencing. BMC Genomics. URL: https://pubmed.ncbi.nlm.nih.gov/38807060/

[47] Sneddon A, Ravindran A, Shanmuganandam S et al. Biochemical-free enrichment or depletion of RNA classes in real-time during direct RNA sequencing with RISER. Nat Commun. URL: https://pubmed.ncbi.nlm.nih.gov/38789440/

[48] Chiou CS, Chen BH, Wang YW et al. Correcting modification-mediated errors in nanopore sequencing by nucleotide demodification and reference-based correction. Commun Biol. URL: https://pubmed.ncbi.nlm.nih.gov/38030695/

[49] Sitarčík J, Vinař T, Brejová B et al. WarpSTR: determining tandem repeat lengths using raw nanopore signals. Bioinformatics. URL: https://pubmed.ncbi.nlm.nih.gov/37326967/

[50] Wan YK, Hendra C, Pratanwanich PN et al. Beyond sequencing: machine learning algorithms extract biology hidden in Nanopore signal data. Trends Genet. URL: https://pubmed.ncbi.nlm.nih.gov/34711425/

[51] Nguyen SH, Cao MD, Coin LJM. Real-time resolution of short-read assembly graph using ONT long reads. PLoS Comput Biol. URL: https://pubmed.ncbi.nlm.nih.gov/33471816/

Disclaimer: This article is for educational and informational purposes only. It is not intended to substitute for professional veterinary advice, diagnosis, treatment, or regulatory guidance. Always consult a licensed veterinarian or qualified specialist regarding animal health, disease diagnosis, and therapeutic decisions.