Electronic probes (e-probes) were designed for each of the nine viruses using the MiFi^® probe developer pipeline, following the methods described by Espindola and Cardwell (2021). For MeV, e-probes were designed for a generic target, genotypes B3 and D8, and vaccine strains Edmonston (MeV-Edwt) and Moraten (MeV-Mor).

For each target virus, all available genome sequences belonging to the target taxon (inclusivity panel) were retrieved from NCBI and concatenated into a single multi-FASTA file. Closely related viruses and other relevant taxa expected to occur in similar sample matrices were compiled into a separate multi-FASTA file representing the near-neighbor (exclusivity) panel to ensure specificity during probe design (Supplementary Tables 1, 2).

The MiFi^® e-probe software¹ generated raw e-probes by scanning target genomes for unique nucleotide regions absent from the near-neighbor dataset. E-probe length was selected based on sequence range recommended for viruses (20–60 nt) (Espindola and Cardwell, 2021). In this study, most e-probes were generated at 40 nt, whereas measles virus vaccine strains required shorter probes (20 nt for MeV-Edwt and 30 nt for MeV-Mor) to resolve closely related sequences. The total number of e-probes varied depending upon degree of sequence homology/differentiation with the nearest strains.

Candidate e-probes were initially selected based on sequence uniqueness and alignment characteristics, including percent identity and query coverage against near neighbors. Raw e-probes were subsequently subjected to a multi-step curation process to ensure high specificity and minimize the risk of false positives in downstream diagnostic applications.

Each e-probe set was queried against the NCBI GenBank nucleotide (nt) database using BLASTn to evaluate potential cross-reactivity. E-probes exhibiting ≥ 90% identity and ≥ 90% query coverage to non-target taxa, defined as organisms outside the intended virus NCBI taxonomy ID and its sub taxa, were classified as potentially cross-reactive and eliminated using a custom taxonomy-aware filtering script that parses BLASTn tabular output generated against the NCBI GenBank nucleotide (nt) database and evaluates each alignment based on percent identity, query coverage, and associated taxonomy identifiers (staxids). E-probes producing any high-confidence BLASTn hits to non-target taxonomic identifiers were excluded. For e-probes aligning to multiple taxa, retention was permitted only when all significant taxonomic assignments belonged to the target lineage; probes with any non-target taxonomic assignments were removed. This taxonomy-specific filtering step ensured retention of only highly specific e-probes uniquely associated with their respective viral targets and minimized the risk of false-positive detection.

The final curated e-probes were aligned against the full set of selected pathogen genomes to confirm exclusive association with their respective viral targets and to verify the absence of significant cross-alignment to other pathogens included in the study. This verification was performed using Geneious Prime software, providing an additional quality-control step to support accurate and reliable pathogen detection.

Robust diagnostic workflows require internal controls to ensure assay integrity, particularly in HTS data where results depend on quantitative comparisons between positive and negative signals (Espindola and Cardwell, 2021; Wang et al., 2022). However, in highly specific e-probe sets, a challenge arises when negative controls exhibit zero signal variance. This can compromise downstream analytical interpretation and prevent the calculation of detection thresholds (p-values). To mitigate this, internal control (exogenous DNA or RNA) e-probes are introduced to generate a stable background signal in true-negative samples, ensuring that statistical metrics and detection criteria can be applied consistently across all datasets.

In this study, we implemented a dual internal control strategy, using host-specific internal control e-probes for human and animal samples. For use in human-derived clinical samples, an internal control e-probe ACTB_NM001101_40nt was designed from a highly conserved region of the Homo sapiens beta-actin (ACTB) gene (RefSeq accession NM_001101). This gene is abundantly expressed and stably present across diverse human tissues, making it a reliable endogenous control for clinical metagenomic workflows.

For use in animal samples, including bovine, canine, and feline specimens, a separate internal control e-probe, ACTB_NM173979_40nt, was designed from a conserved region of the Bos taurus ACTB gene (RefSeq accession NM_173979). This region was selected for its strong sequence conservation across multiple mammalian species, providing compatibility with different animal hosts while avoiding cross-reactivity with viral sequences.

To evaluate the performance and specificity of each internal control, we analyzed representative virus-free datasets: a publicly available human metagenomic dataset from the NCBI Sequence Read Archive (SRA accession SRR32080769) and an animal-derived negative dataset (SRA accession SRR33223349). Each internal control e-probe was integrated into the MiFi^® software alongside pathogen-specific e-probe databases, and background signal was quantified using p-values from e-probe alignment statistics.

To simulate realistic host—pathogen metagenomic environments for in-silico sensitivity testing, HTS datasets free of target pathogens were selected as background. Specifically, H. sapiens sequencing data (SRA accession SRR32080769) was used to represent human genomic background, while B. taurus whole-genome sequencing data (SRA accession SRX28470509) was used for animal background simulations. Prior to use, both host datasets were analyzed using the full panel of curated e-probe sets to confirm that no false-positive signals were generated, ensuring their suitability as clean backgrounds for downstream mock metagenome construction.

For each pathogen, reference nucleotide sequences were retrieved from GenBank (accession numbers listed in Supplementary Table 3) and used as input for read simulation. Simulated reads were generated using ART Illumina (version 2.5.8, released June 6, 2016) with the built-in HiSeq 2500 profile (-ss HS25), producing 150 bp paired-end reads (–p –l 150) with an average fragment length of 300 bp and a standard deviation of 10 bp (–m 300 –s 10). A total of 10,000 paired-end reads were simulated for each reference sequence.

The simulated pathogen reads were randomly subsampled to create spike-in datasets at defined concentrations, ranging from 1 to 130 reads per million (RPM). For each mock metagenome, the total number of reads was fixed at 1 million. To achieve this, the appropriate number of pathogen reads was combined with host-derived reads from the selected SRA datasets, ensuring that the total read count remained constant while the proportion of pathogen reads varied according to the target concentration.

The spike-in and dataset assembly processes were implemented using a custom Python script, which utilized the pysam, Biopython, argparse, and random libraries for file manipulation, parsing, and reproducible sampling. For each concentration level, 10 independent replicates were generated to enable statistical assessment of detection thresholds.

Each resulting simulated metagenome was uploaded to the MiFi^® detection software² as a “Metagenome” file and analyzed independently using the corresponding curated e-probe sets. Detection was performed using appropriate e-probe library selections, with default parameters for viral targets, specifically an e-value threshold of 1e^–1 and a minimum of 10 sensitive hits, as recommended by the software developers (Espindola and Cardwell, 2021). A sample was considered positive for a given pathogen when the calculated was ≤ 0.05. The limit of detection (LoD) for each e-probe set was defined as the lowest concentration at which at least 9 out of 10 replicates yielded positive detection.

This approach enabled a biologically relevant in-silico evaluation of probe performance under simple metagenomic conditions.

For more complex in-silico validation, publicly available high-throughput sequencing datasets from the NCBI SRA were selected based on confirmed infection with specific target pathogens. Each dataset was uploaded to the MiFi^® software and analyzed with the complete panel of curated e-probe sets. Only the set corresponding to the pathogen confirmed by prior testing was expected to yield a positive result, while all others served as negative controls. Diagnostic outputs, including the calculated and the final classification result (positive or negative), were recorded for each e-probe set. Any e-probe sequence that produced false-positive results in non-target datasets were identified and removed from the final diagnostic sets to improve specificity.

To evaluate the diagnostic sensitivity of MiFi^® virus-specific e-probes, a total of 44 respiratory samples, 16 pediatric human nasal swabs and 28 known-status animal samples (BRSV and CDV; nasal swabs and brain/lung tissue), were processed using a standardized Oxford Nanopore Technology (ONT) RNA sequencing workflow. Clinical samples were provided by the Oklahoma Children’s Hospital at OU Health and processed in a university research laboratory in Oklahoma City, OK, USA and animal samples were processed at Oklahoma Animal Disease Diagnostic Laboratory at Oklahoma State University in Stillwater, OK, USA.

Nasal swab samples were collected and stored in BD universal viral transport (UVT) (BD, Cat. No. 220526). Total RNA was extracted from 200–500 μL of each sample using the PureLink Viral RNA/DNA Mini Kit (Thermo Fisher Scientific, Cat. No. 12280050) according to the manufacturer’s instructions. This included proteinase K digestion, lysis with carrier RNA, and purification via silica spin columns. Elution was performed in 50 μL of RNase-free water. RNA concentration was quantified with the Qubit^® RNA HS Assay Kit (Invitrogen, Cat. No. 10320093).

To reduce host ribosomal background, 1–5 μg of total RNA per sample underwent rRNA depletion using the RiboMinus Eukaryote System v2 (Thermo Fisher Scientific, Cat. No. A15026). Biotinylated probes targeting cytoplasmic and mitochondrial rRNA were hybridized to the RNA and removed via streptavidin-coated magnetic beads. rRNA-depleted RNA was then purified using the system’s magnetic bead clean-up module and eluted in 12 μL of 70°C nuclease-free water.

Because the target viruses lack natural polyadenylation and the ONT library preparation protocol required poly(A) tails, we performed 3’ polyadenylation using E. coli Poly(A) Polymerase (NEB, Cat. No. M0276L) following Oxford Nanopore’s recommended protocol. Briefly, 1–10 μg of rRNA-depleted RNA was incubated with PAP, ATP, and reaction buffer at 37°C for 1 min to append 50–100 adenosine residues. The reaction was stopped with 50 mM EDTA, and the RNA was purified using RNAClean XP beads (Beckman Coulter, Cat. No. A63987), then resuspended in 12 μL of nuclease-free water for downstream use.

Polyadenylated RNA was used as input for the Oxford Nanopore cDNA-PCR Sequencing Kit V14 with barcoding (SQK-PCB114.24), following the manufacturer’s instructions. First, reverse transcription and strand-switching were performed using CRTA adapters with poly(T) overhangs to ensure capture of full-length polyadenylated transcripts. Second-strand synthesis was followed by 18 cycles of PCR amplification using barcoded primers. Libraries were cleaned with AMPure XP beads, quantified using the Qubit^® fluorometer.

Equimolar pooling of barcoded libraries was performed to a final concentration of 50 fmol in 11 μL for adapter ligation. The pooled library was sequenced on a MinION Mk1B device (ONT) using R10.4.1 flow cells (FLO-MIN114). Sequencing was conducted using MinKNOW software, and high accuracy base calling was performed in real time with Dorado, the ONT base calling software.

The Nanopore metatranscriptomic datasets were uploaded to the MiFi^® software, which automatically screened the reads using e-probes selected from the CDC library of e-probes, following the recommended parameters for virus detection (Espindola and Cardwell, 2021). Screening was performed by aligning sequencing reads against the e-probe database. Positive matches indicated the presence of target viral sequences in the dataset. The E-value threshold was set at 1e–1, and the sensitivity level required a minimum of 10 matching reads to report a positive detection.

For confirmation, metatranscriptomic ONT reads were quality-trimmed with fastp. Host reads were removed by aligning to the appropriate host reference genome(s) with minimap2 (preset map-ont) and retaining the unmapped reads. The host-depleted reads were then aligned to the pathogen reference genomes listed in Supplementary Table 3 using minimap2. For each target virus, we quantified the number of mapped reads per reference and the genome breadth of coverage (%), defined as the percentage of reference bases with ≥ 1 × coverage.

Sequence data were analyzed and compared with PCR confirmation previously provided by the diagnostic clinic(s), as well as with the outputs from the MiFi^® software. Diagnostic sensitivity was calculated by comparing these results using the online calculator developed by Groth-Helms et al. (2025) and available through the Diagnostic Assay Validation Network (DAVN) website.³

The MiFi^® e-probe development process produced 13 distinct probe sets tailored to the genomic characteristics and diagnostic requirements of 9 respiratory pathogens (Table 1). For human viruses, most e-probes were 40 nucleotides in length, although the measles virus strain probe sets required a more targeted sensitivity (smaller e-probe lengths) to differentiate closely related vaccine strains. The generic measles virus set, built from all publicly available MeV sequences in GenBank, yielded 2042 e-probes capable of detecting diverse genotypes. In contrast, the vaccine strains Edmonston (MeV-Edwt) and Moraten (MeV-Mor) resulted in much smaller and highly specific probe sets, 2 and 11 e-probes, respectively, due to the low genetic variability within these strains and the limited sequence availability. For example, only one genome was available for MeV-Mor. Six highly similar near neighbors were included in the design, restricting the number of unique regions available for probe development. Additionally, these probes were shorter (20 nt for Edwt and 30 nt for Mor) because the conserved vaccine regions offered few mutations and required high-resolution targeting.

A similar pattern was observed for the IAV probe set, which generated 4,663 e-probes, the highest among all targets, driven by the vast number of sequences available in GenBank (over 1 million entries). This contrast illustrates how the volume of available genomic data strongly influences probe diversity: targets with high sequence representation and variability generate large, inclusive probe sets, whereas narrowly defined, highly conserved targets with minimal data yield fewer, more specific probes, sacrificing sensitivity, thus potentially requiring deeper sequencing to detect.

The internal control e-probe ACTB_NM001101_40nt, derived from the H. sapiens ACTB gene, was validated using a virus-free human metagenomic dataset (Table 2). The probe consistently generated background signal without interfering with pathogen-specific alignments, confirming its suitability for distinguishing true-negative results in human-derived samples.

For animal pathogens, the internal control e-probe ACTB_NM173979_40nt, designed from a conserved region of the B. taurus ACTB gene, was validated using whole-genome sequencing data from bovine samples. The probe similarly produced consistent, background signals across animal-derived datasets, demonstrating reliable performance as an internal control for bovine, canine, and feline samples (Table 2).

These results confirm that both host-specific internal controls provide stable, low-level signals that support robust discrimination between true-negative and low-titer samples without cross-reactivity detection events. All pathogen-specific e-probe tests returned negative results when tested against pathogen-free host HTS data, confirming the absence of cross-reactivity and validating the specificity of the internal control strategy (Table 2). The observed p-values ranged from 0.127 to 0.177, reflecting consistent background signal generation without interference in pathogen detection.

The background signal introduced by the internal control provided consistent normalization across the tested datasets, ensuring robust pathogen detection in the absence of confounding factors.

In silico analyses using simulated metagenomes revealed that the theoretical limit of detection (LoD) varied across the curated e-probe sets (Table 3). LoD is defined as the minimum number of target reads detected within a simulated pool of 1 × 10⁶ total reads and are expressed as a percentage of total reads.

Among human viruses, the e-probes targeting IAV, IBV, MuV, and the generic MeV set showed the highest sensitivity, each detecting as few as 0.0010% of total reads (equivalent to 10 reads per 10⁶). This is consistent with higher e-probe numbers per target. By contrast, the MeV-D8 genotype required 0.0100% (100 reads per 10⁶) to reach the same detection threshold, while MeV-B3, HPIV4, and HRSV e-probes had LoDs of 0.0130 (130 reads per 10⁶). The vaccine strain MeV-Edwt was represented by only two e-probes and therefore could not be assigned a definitive LoD under the criteria used in this study.

For animal viruses, the e-probe sets targeting BRSV, CDV, and FeMV exhibited LoDs of 0.0025, 0.0050, and 0.0020% of total reads, respectively (equivalent to 25, 50 and 20 reads per 10⁶) (Table 3).

The in-silico validation results showed that all e-probe sets successfully detected their corresponding target pathogens across the tested datasets on the MiFi^® software (Table 4). All samples were classified as positive, indicating that each e-probe set was able to detect its target pathogen with high confidence. These results support the sensitivity and effectiveness of the designed e-probes across a diverse panel of sequencing datasets. Notably, datasets for HRSV and MeV-D8 included sequences generated by both Oxford Nanopore and Illumina platforms, while all other targets were validated using Illumina-derived datasets only. This further highlights the robustness of the e-probes across different sequencing technologies. As expected, only the e-probe set corresponding to the confirmed pathogen yielded a positive signal, while all non-target e-probe sets remained negative, indicating no detectable cross-reactivity.

A total of 44 human and animal clinical samples (PII data-disaggregated and with known pathology status) were initially selected for the detection of BRSV, CDV, IAV, IBV, and HRSV using Nanopore HTS in combination with the MiFi^® software and virus-specific e-probes. MiFi^® detection of the correct pathogen across all targeted viruses, showed strong concordance with PCR results. MiFi^® matched the performance of PCR in detecting viral pathogens, demonstrating performance comparable to PCR across the targets evaluated (Table 5).

MiFi^® identified several additional positives that were not detected by PCR for BRSV and CDV. While these may reflect low-abundance viral sequences not detected by PCR, the possibility of contamination or read carryover cannot be excluded. All MiFi^® positive detections were supported by statistically significant P-values and HTS RefSeq mapping results, further validating the analytical robustness of the software.

Of the 13 samples that tested positive for BRSV by PCR, MiFi^® successfully detected the virus in all cases, resulting in a complete concordance with PCR (Table 5). This included samples with low viral abundance based on HTS analysis (e.g., 0.9 and 1.0%). In addition, MiFi^® (and mapping to reference analysis) detected BRSV in two PCR-negative samples, suggesting potential false negatives in PCR testing, or alternatively, the possibility of contamination or read carryover during the sequencing process.

MiFi^® detected CDV in all 11 PCR-positive samples, as well as in four PCR-negative samples, yielding a diagnostic sensitivity of 100% (Table 5). These additional detections may indicate false negatives by PCR, or alternatively, contamination or read carryover during the sequencing process. Notably, MiFi^® consistently identified CDV even in samples with low RefSeq mapping percentages (e.g., 9.3 and 9.8%).

The MiFi^® software identified HRSV in all nine PCR-positive samples, with genome coverage percentages from RefSeq mapping ranging from low (8.8%) to complete (100%) (Table 5). No false positives or false negatives were observed, indicating a diagnostic sensitivity of 100%.

MiFi^® detected IAV in five PCR-positive samples and IBV in two PCR-positive samples (Table 5); diagnostic sensitivity was 100% for both viruses.