Per-Cell Pathogen Burden from Discarded Single-Cell Reads

When a disease has no known cause, clinicians turn to sequencing — but the two sequencing cultures never talk to each other. Bulk metagenomic sequencing can find unknown microbes, yet it homogenizes the tissue: more than 99% of its reads come from the human genome, so the sample must be chemically stripped of host material, destroying all information about which cells were infected and how they responded. Single-cell RNA sequencing keeps that cellular context, but its standard pipelines align every read to the human reference genome and silently discard everything that does not match — precisely the reads that might contain an occult virus, bacterium, or reactivated endogenous retrovirus. So today, a single-cell study of a cryptogenic disease sees the damage but never the suspect, and a metagenomic study of the same tissue sees neither the suspect's cellular address nor the local immune reaction.

The proposal here is to close that gap with a computational pipeline — no new wet-lab chemistry — that rescues the discarded non-host reads from droplet-based single-cell data, carries their cell barcodes with them, aligns them against pan-pathogen and endogenous retroviral indexes, and returns a per-cell pathogen burden jointly with the host cell's own transcriptional state. This is aimed squarely at diseases of etiological uncertainty, where the central question is not just "which agent?" but "is the damaged tissue directly infected, or is it an uninfected bystander destroyed by an immune reaction that started elsewhere?" — the exact puzzle posed by the 2022 outbreak of severe acute hepatitis of unknown etiology in children, where adenovirus 41F was detected in roughly 75% of cases in blood and stool yet liver biopsies contained no intact virions Diagnosis, treatment, and prevention of severe acute hepatitis of unknown etiology in children.

Core Claim

The central claim is that cell compartmentalization is itself a host-depletion technology, and that single-cell libraries already contain enough non-host signal to detect and localize occult agents, provided three things are done that current pipelines never do. First, the alignment stage must emit unaligned reads together with their cell barcodes and unique molecular identifiers instead of deleting them — the workflow described in Single Cell RNA-Seq Identifies Cell Subpopulations Contributing to Idiopathic Pulmonary Fibrosis in Humans and Deep sequencing reveals cell-type-specific patterns of single-cell transcriptome variation filters these reads out at step one. Second, ambient RNA — free mRNA from lysed cells that floats into droplets indiscriminately — must be modeled statistically, because a virion lysed in suspension will scatter viral reads across many droplets and manufacture an apparently widespread infection; the debris-modeling approach of DIEM, which achieved 95.7% nuclear droplets in contaminated frozen adipose tissue versus 87.9% for the standard EmptyDrops filter, supplies the machinery. Third, the host cell's own state — interferon-stimulated gene expression, stress response, viral-sensing transcripts — should be used as a prior that tells the search where to concentrate, since Antagonism between viral infection and innate immunity at the single-cell level showed that infected and interferon-producing cells are largely disjoint single-cell fates, meaning the immune signal and the pathogen signal together triangulate the true infected cells.

The payoff is a quantity that neither existing culture can produce: for every cell, both its pathogen burden and its host response program, which jointly distinguish direct pathogen lysis (pathogen-high, damage in the same cells) from bystander immunopathology (pathogen-negative damaged tissue adjacent, in expression space, to a rare infected reservoir). This distinction is exactly what the multi-hit hepatitis hypothesis needs and what bulk testing cannot deliver Acute hepatitis in children: A wide clinical spectrum.

Research Program

The first workstream builds the pipeline, which I will describe concretely. Alignment of the raw FASTQ files proceeds against the human reference as usual, but the unaligned fraction is retained with barcode and UMI tags intact. After subtracting human, ribosomal, and known contaminant sequences, the residual reads are aligned against a composite index of viral, bacterial, and fungal RefSeq genomes plus a curated set of human endogenous retroviral consensus sequences — the latter motivated by Perron et al., who found HERV-W envelope protein in the serum of 73% of multiple sclerosis patients using bulk assays on peripheral blood mononuclear cells, but could never say which cell subpopulation produces it. Reads that match the index are then assigned back to droplets.

The statistical core is the burden model. Let $u_{ck}$ be the count of reads in cell $c$ matching agent $k$, and let $g_c$ be that droplet's total read depth. We model

$$u_{ck} \sim \mathrm{Poisson}(\lambda_k\, g_c + \theta_c\, m_k),$$

where $\lambda_k$ is the ambient rate of agent-$k$ reads per unit depth, estimated from empty droplets in the same run (the DIEM insight that empty droplets carry a clean fingerprint of the ambient soup), and $\theta_c \ge 0$ is the cell's true burden, tested against the null $\theta_c = 0$. This one equation does the host-depletion work that Cheng et al. do chemically with benzonase: their differential lysis increased bacterial reads 2.46-fold in human colon biopsies but necessarily destroys the cells; here the depletion is arithmetic, and the cells survive intact to be annotated.

The second workstream is the coupling analysis. For each cell we compute, from the host-genome reads alone, module scores for interferon signaling, apoptosis, and antigen presentation, then test whether pathogen-positive cells cluster in expression space near interferon-high, pathogen-negative cells — the signature of paracrine bystander damage that Grabowski et al. observed as a "winner-takes-all" switch, with only about 10% of viral-protein-positive cells also producing interferon. The third workstream applies the pipeline to real cryptogenic specimens, principally archived liver tissue from the pediatric hepatitis outbreak and reanalysis of existing idiopathic-disease datasets.

Baselines

The do-nothing reference point is the standard single-cell pipeline — Cell Ranger-style alignment to the host genome, EmptyDrops filtering, Seurat clustering and annotation, exactly as run in Zhang et al. and Dueck et al. — which discards unmapped reads at the first step. Its occult-agent detection rate is zero by construction, so it serves as the floor and as the source of the host-side annotations the proposal reuses; any positive detection the pipeline achieves is gain over this reference, and the host-side clustering must reproduce what the standard pipeline produces, as a sanity check that read salvage has not perturbed cell calling.

The strongest existing competitor is bulk metagenomic sequencing with chemical host depletion, the Cheng et al. benzonase differential-lysis protocol High Sensitivity of Shotgun Metagenomic Sequencing in Colon Tissue Biopsy by Host DNA Depletion, followed by strain-level identification with cgMSI cgMSI: pathogen detection within species from nanopore metagenomic sequencing data, which reports greater than 90% sensitivity at 1× coverage (outperforming ORI by roughly 40%), ANI above 0.999 on the ZymoBIOMICS mock community, and 95% specificity in complex gut simulations. This is the strongest competitor because it is the established answer to "find the unknown agent in a host-dominated sample," and its sensitivity per sequencing dollar is hard to beat. The comparison will be run on matched aliquots of the same dissociated specimens under a fixed total read budget, with both methods tuned by the same spike-in calibration set, so neither gets a depth advantage. Beating it means one of two outcomes: detecting an agent in samples where depleted bulk mNGS is negative at equal depth, or — more realistically, and the more important criterion — detecting the agent and returning its cell type, burden distribution, and local host response, which bulk structurally cannot. If bulk matches the pipeline on detection at every dilution and the cell context turns out not to discriminate mechanisms, the proposal's advantage claim fails and I will say so.

Two ablations isolate the proposal's key ingredients. The ambient-subtraction ablation removes the empty-droplet $\lambda_k$ estimation and keeps everything else, testing whether the Poisson correction is what holds the false-positive rate down — the DIEM adipose-tissue evidence (95.7% versus 87.9% droplet purity for EmptyDrops) predicts it is. The host-state-prior ablation searches all droplets uniformly rather than concentrating deep search on interferon-high droplets, testing whether the host program actually improves detection power or merely decorates it. A third, simpler variant — aligning all reads against a concatenated human-plus-pathogen reference with a standard aligner and no ambient model — is included as the "cheapest possible version" ablation, since a reviewer will rightly ask whether anything more is needed.

Why not just…?

Why not just run bulk metagenomic sequencing with host depletion on the same tissue? This is the most serious alternative, and for pure agent identification it is cheaper and probably more sensitive per read; I will not argue otherwise. The material shows its failure modes are exactly the cryptogenic ones. In the pediatric hepatitis outbreak, the suspected agent was detectable in blood and feces but liver tissue was virion-negative, so bulk metagenomics of the damaged organ returns little and says nothing about mechanism Chen et al.; in Rosai-Dorfman disease of the airway, bulk mNGS of histiocytic proliferations failed to detect the suspected viral trigger at all, leaving the cause indeterminate Wu et al.. More fundamentally, bulk averaging cannot distinguish an infected cell from a bystander — the bimodal single-cell dynamics that Grabowski et al. measured, where viral production and interferon production are mutually exclusive in the same cell, collapse in bulk into a misleading average. The dilution head-to-head in the experiments below is designed to settle this honestly: if depleted bulk matches the pipeline at every spike-in dilution, the honest conclusion is that the pipeline's value is confined to mechanism-resolution, not detection.

Why not just align the reads against a combined human-plus-pathogen reference in the existing pipeline? This needs no new method and is the right first ablation. Its failure mode is statistical rather than architectural: without ambient modeling, viral RNA released from one lysed infected cell scatters across hundreds of droplets and inflates apparent prevalence, and the DIEM benchmark shows standard filtering leaves 12.1% contaminated droplets even in a benign tissue DIEM; the burden test against the Poisson null is what converts raw read counts into a per-cell claim. Short reads also multi-map preferentially onto the larger human genome in a concatenated index, biasing the residual in ways the two-stage subtraction avoids. If the simple version matches the full pipeline's calibrated false-positive rate on spike-ins, I will adopt the simple version — the ablation exists precisely to make that check.

Why not just expand targeted PCR panels and serology? Targeted assays are fast and cheap but hypothesis-bound: they detected nothing in the cryptogenic brain abscess that unbiased sequencing resolved as Providencia rettgeri Zhao et al., and serology fails outright in immunocompromised hosts who cannot mount antibody responses, the population where occult infection concentrates Solomon et al.. They also say nothing about endogenous retroviral reactivation, which required bespoke qPCR even in the multiple sclerosis work Perron et al.. Panels remain the right confirmatory step after the pipeline nominates an agent, but not the discovery step.

Why not just use bulk RNA-seq with computational deconvolution or a pre-trained foundation model? Deconvolution assumes the cell-mixture signature is known in advance and averages away rare populations — the clonogenic progenitors Noseda et al. could only see by isolating cells individually. Foundation models generalize poorly across batches in exactly this regime: the scGeneScope benchmark shows zero-shot models losing 5–7% balanced accuracy relative to fit-to-data models like scVI on held-out experimental conditions scGeneScope. Since the proposal's target is rare, possibly unprecedented signal, a fit-to-data statistical model on the droplet counts is the safer design choice.

Proposed Experiments

The first experiment is ground-truth recovery on a known pathogen with a published single-cell atlas. Age-specific nasal epithelial responses to SARS-CoV-2 infection profiled 139,598 cultured airway epithelial cells from children, adults, and older adults, integrated with 577,243 in vivo airway cells from eight public datasets, and showed that pediatric goblet cells mount strong type I and II interferon responses that restrict viral replication to incomplete cycles. The pipeline must recover per-cell SARS-CoV-2 burden from the unmapped fraction and reproduce that age-stratified tropism and the interferon-antagonism pattern; the decisive metric is concordance of per-cell viral burden with the published annotations, and specifically whether the burden-interferon joint distribution shows the mutual exclusion that Grabowski's model predicts (roughly 10% co-expression).

The second experiment is the calibrated head-to-head against depleted bulk metagenomics, using the ZymoBIOMICS mock community (eight bacteria, two yeasts) that Zhu et al. used to validate cgMSI. The mock community is spiked into dissociated cell suspensions at a dilution series spanning copy numbers per cell from many to fractional, and the same aliquots are split between the single-cell pipeline and the Cheng et al. benzonase-depleted bulk protocol, both at fixed total depth. The metrics are per-agent sensitivity and specificity as a function of multiplicity of infection, the false-positive per-cell burden rate attributable to ambient scattering, and the limit of detection in agent copies per thousand cells. The pre-registered success criterion: match cgMSI-level per-genome sensitivity (>90% at 1× effective coverage) at equal depth while additionally returning correct cell-type assignment for burden-positive droplets; the pre-registered failure criterion: bulk out-detecting the pipeline at all dilutions, which would reduce the contribution to the mechanism-resolution analysis only.

The third experiment is reanalysis of existing idiopathic-disease data, beginning with the IPF cohort of Zhang et al. (five end-stage IPF lungs, five healthy donors), mining its unmapped reads for endogenous retroviral and microbial signal concentrated in the pro-fibrotic myoCAF_FAP myofibroblasts and macrophage_SPP1 macrophages the study identified. This arm is explicitly exploratory: the paper reports proportional shifts qualitatively, and no reported result found on Lacuna for numerical per-cell abundances in that cohort, so no ground-truth numeric target exists; the deliverable is a bound on how much non-host signal standard pipelines of record have been discarding, which itself has not been measured anywhere in the literature this file contains.

The fourth experiment is the clinical application that motivated the proposal: single-nucleus profiling of archived frozen liver tissue from pediatric cryptogenic hepatitis cases, where the tissue is virion-negative in bulk Chen et al.. Frozen tissue mandates the ambient machinery — the DIEM adipose benchmark is the relevant precedent — and the decisive readout is binary in spirit: either a rare burden-positive cell population (say, sinusoidal endothelial or Kupffer-lineage cells) sits at the epicenter of the interferon-high neighborhood, supporting a focal-reservoir mechanism, or the damaged parenchyma is uniformly burden-negative with clonally expanded cytotoxic T cells nearby, supporting pure bystander immunopathology. Either outcome is publishable and clinically actionable; no bulk assay currently on the market can distinguish the two.

The overall risk profile is honest: the pipeline could work technically and still find nothing in the clinical tissue, because some cryptogenic diseases may have no nucleic-acid trace in the organ at the time of biopsy. That would itself be an informative negative — it would argue for hit-and-run immunopathology, the leading hypothesis for the pediatric hepatitis outbreak, and would redirect etiological search toward the priming compartments rather than the damaged organ. The spike-in and atlas-recovery experiments, however, are fully within our control, and they are the ones that decide whether the method is sound.

References

  1. Diagnosis, treatment, and prevention of severe acute hepatitis of unknown etiology in children
  2. Single Cell <scp>RNA</scp>‐Seq Identifies Cell Subpopulations Contributing to Idiopathic Pulmonary Fibrosis in Humans
  3. Deep sequencing reveals cell-type-specific patterns of single-cell transcriptome variation
  4. Enhancing droplet-based single-nucleus RNA-seq resolution using the semi-supervised machine learning classifier DIEM
  5. Antagonism between viral infection and innate immunity at the single-cell level
  6. Acute hepatitis in children: A wide clinical spectrum
  7. Human endogenous retrovirus type W envelope expression in blood and brain cells provides new insights into multiple sclerosis disease
  8. High Sensitivity of Shotgun Metagenomic Sequencing in Colon Tissue Biopsy by Host DNA Depletion
  9. cgMSI: pathogen detection within species from nanopore metagenomic sequencing data
  10. Primary diffuse Rosai-Dorfman disease in central airway: a case report and literature review
  11. Case report: Cryptogenic giant brain abscess caused by Providencia rettgeri mimicking stroke and tumor in a patient with impaired immunity
  12. Fatal Case of Chronic Jamestown Canyon Virus Encephalitis Diagnosed by Metagenomic Sequencing in Patient Receiving Rituximab
  13. PDGFRα demarcates the cardiogenic clonogenic Sca1+ stem/progenitor cell in adult murine myocardium
  14. scGeneScope: A Treatment-Matched Single Cell Imaging and Transcriptomics Dataset and Benchmark for Treatment Response Modeling
  15. Age-specific nasal epithelial responses to SARS-CoV-2 infection