New high accuracy diagnostics for avian Aspergillus fumigatus infection using Nanopore methylation sequencing of host cell-free DNA and machine learning prediction

Drag, M. H.; Hvilsom, C.; Poulsen, L. L.; Jensen, H. E.; Tahas, S. A.; Leineweber, C.; Cray, C.; Bertelsen, M. F.; Bojesen, A. M.

doi:10.1101/2025.04.11.648151

bioRxiv · 10.1101/2025.04.11.648151

New high accuracy diagnostics for avian Aspergillus fumigatus infection using Nanopore methylation sequencing of host cell-free DNA and machine learning prediction

Drag, M. H.·Hvilsom, C.·Poulsen, L. L.·Jensen, H. E.·Tahas, S. A.·Leineweber, C.·Cray, C.·Bertelsen, M. F.·Bojesen, A. M.

View original source ↗Export metadata ↓

Abstract

Avian aspergillosis is a detrimental fungal infection affecting wild and domestic birds yet sensitive antemortem diagnostics for early clinical infections are lacking. Here we present new diagnostics for Aspergillus fumigatus (Af) infection developed from cell-free DNA (cfDNA) methylation markers. Broiler chickens were experimentally infected with either Af, a non-Af agent (Escherichia coli or Gallibacterium anatis) or assigned as controls. Oxford Nanopore (ONT) sequencing was performed on serum cfDNA (n = 124), and machine learning (ML) models were trained on infection-specific markers. Three tests were developed: A High Accuracy test for best performance (sensitivity: 100%, specificity: 89.2%) and robustness (ROC-AUC: 0.92) as well as Fast- and In situ tests for rapid turnaround and methylation PCR. Diagnostic accuracies were 92.3%, 82.7%, and 73.1%, respectively. In conclusion, new tests using on ML- and host cfDNA methylation markers demonstrated high diagnostic performance comparable to microbial cfDNA (mcfDNA) tests but without concern for environmental contamination. Key highlightsO_LIWe present three new high accuracy diagnostic tests for Aspergillus fumigatus infection in chickens that use methylation markers from serum cell-free DNA (cfDNA). C_LIO_LIDifferentially methylated cfDNA regions (DMRs) were detected by Oxford Nanopore sequencing (ONT) in experimentally infected chickens and used as markers to train machine learning (ML) models for development of three diagnostic tests. C_LIO_LIThe highest accuracy was found with 83 markers of 10 kilobases (KB) using the glmnet algorithm for the ML model, which classified 92.3% blinded samples correctly. C_LIO_LIA Fast test designed for cheap <1h sequencing using adaptive sampling could correctly classify 82.7% samples with 22 markers using a random forest (rf) model. C_LIO_LIAn In situ test with only four markers, envisioned for use in a simple methylation-specific PCR (MSP-PCR) assay, could correctly classify 73.1% blinded samples. C_LIO_LIReference values with associated probabilities of infection were calculated for each of the three tests and are presented for further evaluation. C_LI

Source connections

Published article ↗

Explore related subjects

microbiology Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Drag, M. H., Hvilsom, C., Poulsen, L. L., Jensen, H. E., Tahas, S. A., Leineweber, C., Cray, C., Bertelsen, M. F., Bojesen, A. M.. 2025-04-15. New high accuracy diagnostics for avian Aspergillus fumigatus infection using Nanopore methylation sequencing of host cell-free DNA and machine learning prediction. https://doi.org/10.1101/2025.04.11.648151

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

A population-scale landscape of the subgingival microbiome reveals divergent routes to periodontal dysbiosis

Periodontitis is an archetypical mucosal inflammatory disease in which microbiome dysbiosis at the tooth-epithelial interface interacts with host genetic and behavioral risk factors to drive immune-mediated tissue destruction. Although subgingival microbiome compositional shifts are thought to parallel disease severity, microbiome variation at the population-level and its relationship to periodontal clinical phenotypes and disease-modifying factors remain poorly defined. Here, we use unsupervised manifold learning to map the compositional landscape of the subgingival microbiome in 1,355 adults spanning periodontal health to severe periodontitis. We identified eight latent microbiome states organized along a branching continuum from eubiosis to dysbiosis. An intermediate microbial configuration marked ecological destabilization and bifurcation into two distinct periodontitis-associated dysbiotic trajectories, distinguished by links to gingival inflammation and smoking. Although the microbiome trajectories broadly tracked periodontal destruction, a minority of individuals showed discordant microbiome-clinical phenotypes, with some individuals with periodontitis retaining otherwise eubiotic microbiomes enriched for low-abundance pathobionts, while some cases of health or mild disease had highly dysbiotic communities, suggesting distinct host susceptibility. Together, these findings define a population-scale ecological landscape of the subgingival microbiome, reveal divergent trajectories to periodontal dysbiosis, and highlight heterogeneity in the relationship between microbial community structure and clinical disease expression.

microbiology↗

bioRxiv2026

Rapid and largely reversible shifts in the canine fecal metabolome during dietary change

Diet can rapidly change the fecal metabolome, but less is known about recovery after the original diet is restored. We used untargeted UPLC-MS metabolomics to analyze 72 fecal samples from nine Pumi dogs during an owner-managed switch from dry food to raw food and back to dry food. Diet phase accounted for a large proportion of variation in both ionization modes. More than 13,000 LC-MS features changed at the first sampling point after the switch to raw food, with a similarly large response after return to dry food. Among features significant in both comparisons, more than 99% changed in opposite directions. At the final sampling point, no positive-mode (ESI+) features and only 13 negative-mode (ESI-) features differed from the second dry-food baseline under the same threshold. BARF-associated patterns persisted in analyses excluding individual dogs and in pedigree-adjusted candidate models, although individual feature effects depended on normalization. Putative metabolites from several biochemical classes differed in their response and recovery. The fecal metabolome therefore changed rapidly and returned largely toward baseline, with differences among dogs.

microbiology↗

bioRxiv2026

Taxonomic and functional concordance between full-length ONT 16S and ONT shotgun metagenomics in the canine gut microbiome

Background: Full-length Oxford Nanopore Technologies (ONT) 16S rRNA sequencing provides a scalable view of microbial community composition and can support phylogeny-based functional prediction, but it is not equivalent to shotgun metagenomics. We asked which biological conclusions are preserved when the same canine fecal specimens are profiled by full-length ONT 16S and ONT whole-genome shotgun (WGS) sequencing, and how their agreement depends on analytical scale, reference representation and classifier. Methods: Ninety-seven fecal specimens from 51 dogs were profiled with both assays from the same DNA extract. Functional profiles predicted from NanoASV/NanoPredict with PICRUSt2 were compared with WGS-supported KEGG Ortholog (KO) profiles generated by Kadath. Taxonomy was benchmarked in a source-genome-matched RefSeq universe and in a host-specific DogMAG universe using minitax and Kraken2. Agreement was evaluated at whole-profile, feature-abundance, detection, between-sample structure and biological-inference scales. Age-associated transfer was assessed with dog-aware continuous mixed models, grouped signed-score analyses and paired/dog-blocked PERMANOVA. Results: Functional whole-profile concordance was high: median within-sample CLR Spearman correlations ranged from 0.781 to 0.860 across developmental strata, while between-sample functional structure remained significant by Mantel (rho=0.543) and Procrustes (r=0.693; both p=0.001). Feature-wise transfer was substantially weaker (median KO-wise CLR Spearman=0.318). Continuous age-associated KO slopes showed substantial cross-assay concordance (Spearman=0.727; signed-score Spearman=0.753; direction agreement=77.9%), although 1,290/5,258 eligible KOs retained significant assay-by-age interactions. Taxonomically, exact genus/species abundance agreement was much lower than agreement in between-sample ecological structure. Host-specific DogMAG improved species-level median Spearman from 0.261 to 0.656 for minitax SpeciesEstimate and from 0.181 to 0.512 for Kraken2. The classifier effect was independent of reference choice: under both RefSeq and DogMAG, minitax yielded stronger 16S-WGS concordance than Kraken2, with all eight prespecified RefSeq paired genus/species endpoints and all 10 DogMAG primary paired endpoints significant after BH correction. The same ordering extended to developmental inference, with DogMAG genus/species age-slope concordance of 0.795/0.799 for SpeciesEstimate versus 0.693/0.702 for Kraken2. Taxonomic Aitchison PERMANOVA detected age-associated structure in every assay/reference/classifier/rank combination, whereas age-by-assay interactions were consistently significant but small (R2 approximately 1.1 to 2.2%). Stricter NanoASV identity thresholds removed substantial 16S abundance without improving species-level agreement. Conclusions: The extent of cross-assay agreement depends on the level of analysis. Full-length ONT 16S preserves broad functional organization, ecological structure and much of the direction of age-associated change, but exact fine-rank composition, individual-feature abundance and effect magnitude remain assay dependent. Host-specific reference representation substantially narrows the taxonomic gap, and classifier choice exerts an additional independent effect: within the same matched reference set, minitax consistently yields stronger 16S-WGS concordance than Kraken2 across abundance, detection, ecological-distance and developmental-inference endpoints. Full-length ONT 16S is therefore well suited to broad ecological screening and hypothesis generation, whereas WGS remains preferable when conclusions depend on quantitative fine-rank composition, directly supported gene content or precise feature-level effect estimates.

microbiology↗