Where and When the Lung Digests the Virus: A Spatiotemporal Enrichment Atlas of Lysosomal Rerouting

Animal cells have a waste-disposal system: the lysosome, an organelle full of enzymes that shred proteins and particles. During infection, host cells can deliberately divert viral components into this disposal system — a defense called lysosomal rerouting. It works. Overexpressing the trafficking factor RAB11FIP5 in cultured cells diverts a Kaposi's sarcoma-associated herpesvirus protein to lysosomes and yields a roughly 10-fold reduction in released virions (Host RAB11FIP5 protein inhibits the release of Kaposi's sarcoma-associated herpesvirus particles by promoting lysosomal degradation of ORF45); the neuronal isoform Bif-1c unblocks the autophagy pathway — the cell's other route to the lysosome — and suppresses rabies virus in N2a cells (Bif-1c Attenuates Viral Proliferation by Regulating Autophagic Flux Blockade Induced by the Rabies Virus CVS-11 Strain in N2a Cells); and engineered lysosomes delivered to hamster lungs clear Omicron with a greater-than-3-log drop in lung viral RNA (Lysosomal "TRAP": a neotype modality for clearance of viruses and variants). But every one of these demonstrations lives in flat cultures of a single cell type or in homogenized whole organs. Nobody knows where in a real tissue, or at which hour of infection, this disposal defense switches on — or whether some tissue microenvironments fail to mount it and instead become virus-exporting regions.

The obstacle is methodological as much as experimental. Infection time-courses that do exist grind the tissue into a homogenate before sequencing, destroying geography; the rat burn-wound chromatin atlas is the canonical cautionary example, profiling 31,503 cells across nine time points but losing the physical relationship between damaged and healthy zones (A scATAC-seq atlas of stasis zone in rat skin burn injury wound process). Spatial transcriptomics — sequencing at thousands of fixed positions across a tissue slice — is available but static: Starfysh maps pathway-active tumor-immune hubs in single biopsies with no notion of time (Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs). And standard functional enrichment — testing whether a preset gene list is overrepresented among differentially expressed genes — assumes pathways are static, non-spatial, and thresholded, so it cannot say that autophagy is active here but not there, or now but not six hours ago.

Core Claim

We claim two things. First, a biological result: in influenza-infected mouse lung, lysosomal-rerouting activity is not diffuse. It concentrates in identifiable microenvironments, changes direction between early and late infection (from host-directed clearance toward virus-directed hijacking of the same disposal machinery), and its local strength predicts whether that region traps or exports virus. Second, a methodological result: answering this requires, and motivates, a new enrichment statistic — a module-level activity score computed at each measured tissue position and each time point, corrected for which cell types are present, and tested for spatial clustering rather than mere membership in a differential list. No existing tool does this; the gap is explicitly open, since the lysosomal-rerouting literature never measures in situ, and the spatial-enrichment literature never measures over time (Starfysh; Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST).

Research Program

The program has three parts: curating the gene module, building the statistic, and generating the atlas.

The module. We define two signed gene sets from the lysosomal-rerouting literature rather than from a generic database sweep. The clearance set contains host factors that push viral material toward destruction: RAB11FIP5 and the vesicle-trafficking GTPases the host upregulates during norovirus infection, Rab5A and Rab6A (Proteomic analysis of murine norovirus-infected Raw 264.7 cells reveals the induction of Ras GTPases), the autophagy-flux machinery measured by LC3-II and p62 turnover (Bif-1c), and lysosomal hydrolases. The hijack set contains the opposite program: viral exploitation of the ubiquitin-proteasome system to degrade host defenses, exemplified by STING degradation during deltacoronavirus infection (Porcine deltacoronavirus nonstructural protein 2 inhibits type I and III IFN production by targeting STING for degradation) and by HIV accessory proteins degrading PSGL-1 and APOBEC3G (P-Selectin Glycoprotein Ligand 1: A Potential HIV-1 Therapeutic Target; Allogeneic stimulation of the anti-viral APOBEC3G in human CD4+ T cells and prevention of SHIV infectivity in macaques immunized with HLA antigens). Distinguishing the two directions matters because they use overlapping machinery with opposite consequences, and standard Gene Ontology enrichment conflates them.

The statistic. Each spatial measurement (a "spot") pools 10–50 cells, so raw expression confounds program activation with cell composition — more macrophages mechanically means more lysosome genes. We therefore subtract the cell-mixture expectation, using fractions estimated by a top-ranked deconvolution method with a matched single-cell reference (A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics). For a module $P$, spot $s$, and time point $t$:

$$a_{s,t}(P) = \frac{1}{|P|}\sum_{g \in P} \frac{Y_{g,s,t} - \sum_k \hat{w}_{k,s}\, X_{g,k}}{\hat{\sigma}_{g,t}}$$

where $Y_{g,s,t}$ is the observed count for gene $g$, $\hat{w}_{k,s}$ the deconvolution-estimated fraction of cell type $k$ at spot $s$, $X_{g,k}$ that cell type's reference expression, and $\hat{\sigma}_{g,t}$ the residual standard deviation across spots at time $t$. The score $a$ is thus activation beyond what cell composition explains. Two tests then ride on it: a spatial test (permutation of $a_{s,t}$ against physical neighbors, asking whether activation is localized or diffuse) and a temporal test (the contrast $\Delta_s = a_{s,t_2} - a_{s,t_1}$). This design deliberately sidesteps the platform-effect problem that limits cross-lab deconvolution (A comprehensive benchmarking) because we test within-slice structure, not absolute fractions.

The atlas. We infect BALB/c mice with the same two H1N1 strains — pandemic BJ501 and mouse-adapted PR8 — profiled at Day 2 and Day 5 in the existing bulk time-course (MicroRNA Expression Profile of Mouse Lung Infected with 2009 Pandemic H1N1 Influenza Virus), adding Day 0 and Day 7. At each point we run spatial transcriptomics on lung, matched single-cell RNA-seq for the deconvolution reference, a parallel bulk aliquot for the do-nothing baseline, immunofluorescence for LC3 puncta and p62 (the flux readouts used in the Bif-1c study) co-registered to the spots, and qPCR separating extracellular from cell-associated virion RNA — the egress readout that detected the 10-fold trapping effect in the RAB11FIP5 study (Host RAB11FIP5...). The single-cell reference also repairs a specific hole in the prior time-course, which predicted but never measured mRNA targets (Wu et al.). As a robustness check outside infection, we run the statistic across the 1,229 paired profiles of HEST-1k to confirm it stays near baseline in uninfected tissues (HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis).

Baselines

The do-nothing reference is exactly the design of the existing influenza time-course: bulk lung homogenates per time point, differential expression, and standard pathway overrepresentation analysis (MicroRNA Expression Profile...). We run it on the parallel bulk aliquots of the very same lungs, so the comparison differs only in spatial resolution. That study's reported outcomes set the expectation: 47 differentially expressed miRNAs total, with PR8 provoking a broader response (43) than BJ501 (29) and TGF-β signaling as the top enriched hub. Beating this baseline means showing that bulk averaging cancels or misassigns the degradation program — for instance, detecting adjacent regions with opposing clearance-versus-hijack scores that sum to nothing in homogenate, as the salt-stress work already showed for organs executing opposite programs in leaves versus roots (Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings).

The strongest competitor is the best static pipeline available today: deconvolve one infected slice with one of the top-ranked methods — CARD, Cell2location, or Tangram, which the 50-dataset benchmark names as the most robust and accurate across the widest range of conditions (A comprehensive benchmarking...) — then score the lysosomal module per estimated cell type with a standard gene-set statistic. The benchmark paper reports rankings but no absolute accuracy numbers for these methods; no reported result found on Lacuna for absolute deconvolution error of CARD or Cell2location, so for the shared infrastructure we anchor to the two numbers that are reported: Starfysh's deconvolution error of RMSE ≈ 0.09 on simulated mixed spots, described as comparable to reference-based tools (Starfysh), and GraphST's median Adjusted Rand Index of 0.60 for spatial domain identification on human dorsolateral prefrontal cortex slices (GraphST). We give the static baseline the same slices, the same single-cell reference, the same compute, and the same tuning budget; our method must match it on deconvolution-level accuracy (RMSE at or below the ≈ 0.09 level on simulated mixtures) and beat it on the task that matters: predicting the co-registered LC3/p62 imaging signal from transcriptomics alone. The static pipeline cannot use time by construction, so if the imaging signal is well predicted only when the temporal contrast term is included, the static baseline loses for a structural reason, not a tuning one.

Two ablations isolate our ingredients. Removing the cell-mixture correction — scoring raw spot expression — tests whether composition correction matters; and dropping the temporal contrast — collapsing all time points — tests whether time matters. A third comparison handles the per-gene alternative: the STAr platform integrates seven spatially-variable-gene methods such as BOOST-GP, BOOST-MI, and BinSpect, plus two rank-aggregation schemes (Spatial Transcriptomics Arena (STAr)); we run those per module gene and aggregate, against our direct module-level test, on simulated data with known module-level ground truth.

Why not just…?

Why not just run bulk time-course RNA-seq and enrich for autophagy terms? This is the cheapest design and it already exists in essence (MicroRNA Expression Profile...). It falls short on evidence already in hand: the salt-stress study found that different tissue compartments run entirely different stress programs, so averaging can hide or invert what is happening locally (Tissue-Specific Transcriptome Analysis...), and the RAB11FIP5 work showed the clearance program acts in specific cells at a specific stage — trapping virions at egress without touching viral DNA replication (Host RAB11FIP5...), a subtlety invisible in homogenate. Our decisive check is direct: we compute what the bulk analysis would report by averaging our spots, and compare it to the imaging-validated spot-level truth on the same lungs.

Why not just deconvolve one infected slice and score the lysosome genes? This is a real threat, and we treat it as our strongest baseline rather than arguing around it. It could plausibly work if module activation is strong and spatially stable. Two facts argue it will not: the influenza time-course showed dramatic temporal escalation between Day 2 and Day 5 in the very tissue we profile (Wu et al.), and autophagy is defined by flux — the turnover of LC3-II and p62 — not by levels, so a static snapshot can report accumulation where the pathway is actually blocked, exactly the error the Bif-1c study was designed to catch (Bif-1c...). The settling experiment is the imaging-prediction test: if the static pipeline predicts the flux readout as well as the temporal model, we will report that the simpler design suffices.

Why not just do a non-spatial single-cell time-course, like the burn atlas? Cheaper, and cell-type-resolved, which may be all the biology there is — if cell identity fully determines module activity, spatial coordinates add nothing. We take this seriously in two ways. First, we need the single-cell time-course anyway as the deconvolution reference, in the manner of stereoscope's probabilistic mapping (Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography), so the alternative is incorporated rather than excluded. Second, our spatial permutation test on the composition-corrected score is precisely the falsification device: if no residual spatial structure survives after accounting for cell types, we will report that geography adds nothing here and the single-cell design is sufficient. The burn atlas itself warns which biology would be lost: dissociation erased the relationship between the stasis zone, necrotic core, and healthy border that determines cell fate (A scATAC-seq atlas...).

Decisive Checks

The method is validated before the biology. We simulate spatial arrays from high-resolution imaging data, as the 50-dataset benchmark does from seqFISH+ and MERFISH (A comprehensive benchmarking...), planting module activation in the five spatial patterns of STAr and evaluating detection with AUC, F1, and MCC against the per-gene-plus-aggregation baseline; the STAr summary reports that BOOST-GP, BOOST-MI, and BinSpect generally perform well but no reported result found on Lacuna for their numeric AUC, F1, or MCC values, so those numbers come from our own runs. As a secondary prediction test we hold out module genes and predict their spatial pattern from the remaining genes, benchmarking against SpaOTsc's reported average AUC of 0.95 for Zebrafish and 0.88 for Drosophila (Inferring spatial and signaling relationships between cells from single cell transcriptomic data). The biological payoff has a concrete form: spot-level clearance scores should anti-correlate with the local extracellular-to-intracellular virion RNA ratio, and regions predicted as hijack-dominant should show impaired flux markers. If those correlations hold across both strains and all time points, we will have the first map of where and when a tissue decides to digest a virus — and a statistic that others can point at any degradation program, host or pathogen.

References

  1. Host RAB11FIP5 protein inhibits the release of Kaposi’s sarcoma-associated herpesvirus particles by promoting lysosomal degradation of ORF45
  2. Bif-1c Attenuates Viral Proliferation by Regulating Autophagic Flux Blockade Induced by the Rabies Virus CVS-11 Strain in N2a Cells
  3. Lysosomal “TRAP”: a neotype modality for clearance of viruses and variants
  4. A scATAC-seq atlas of stasis zone in rat skin burn injury wound process
  5. Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs
  6. Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST
  7. Proteomic analysis of murine norovirus-infected Raw 264.7 cells reveals the induction of Ras GTPases
  8. Porcine deltacoronavirus nonstructural protein 2 inhibits type I and III IFN production by targeting STING for degradation
  9. P-Selectin Glycoprotein Ligand 1: A Potential HIV-1 Therapeutic Target
  10. P19-20. Allogeneic stimulation of the anti-viral APOBEC3G in human CD4+ T cells and prevention of SHIV infectivity in macaques immunized with HLA antigens
  11. A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics
  12. MicroRNA Expression Profile of Mouse Lung Infected with 2009 Pandemic H1N1 Influenza Virus
  13. HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis
  14. Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings
  15. Spatial Transcriptomics Arena (STAr): an Integrated Platform for Spatial Transcriptomics Methodology Research
  16. Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography
  17. Inferring spatial and signaling relationships between cells from single cell transcriptomic data