Genetic Anchors for Translating Fungal Metabolite Activity from Culture to Animal

Fungi that live in extreme or symbiotic niches — deep-sea sediments, ascidian tissues, marine sponges — produce small molecules called secondary metabolites: compounds not needed for growth, but evolved as chemical weapons and signals. These molecules are a historically rich source of drugs, and databases now catalog tens of thousands of them; the Comprehensive Marine Natural Products Database alone holds 31,561 distinct entities with 72,343 recorded bioactivities mapped to 2,652 molecular targets CMNPD. Yet almost none of this chemistry is ever tested properly in a living animal. Activity is measured in static cell culture, and the repeated pattern is that a compound kills cells or blocks an enzyme powerfully in a dish, then fails in vivo because it never reaches the right cells, is cleared too fast, or works through a cell type the assay never contained Matrine induces apoptosis of human multiple myeloma cells via activation of the mitochondrial pathway, Curcumol Induces G1 Phase Arrest in SK-Hep-1 Cells by Targeting SKP2-Mediated p27 Degradation, Geissoschizine methyl ether, a corynanthean-type indole alkaloid from Uncaria rhynchophylla as a potential acetylcholinesterase inhibitor.

This proposal argues that the missing piece is not another assay, but ground truth about where in the body a compound must act. Conditional gene knockout — using the Cre-Lox system, where a target gene is deleted only in one chosen cell type in a living mouse — provides exactly that ground truth, and single-cell RNA sequencing (scRNA-seq), which reads the transcriptome of each cell individually, provides the resolution to compare drug action between dish and animal. We propose to pair all three: build a panel of conditional knockouts of a fungal metabolite's putative targets, profile drug-treated tissues at single-cell resolution in vitro and in vivo, and use the knockout signatures as "genetic anchors" that tell single-cell perturbation models which cell type a compound is actually engaging in vivo.

Core Claim

The in vitro-to-in vivo translation gap for a polypharmacological natural product can be decomposed into a per-cell-type, per-target quantity, and conditional knockouts are the instrument that measures it. Concretely: if a compound's benefit in vivo requires inhibiting target $X$ in cell type $C$, then deleting $X$ specifically in $C$ should both contribute to the disease and abolish the compound's benefit — an epistasis test that culture assays cannot run. We further claim that single-cell perturbation models, which today are trained exclusively on in vitro data from synthetic drugs in immortalized lines, will fail on in vivo responses to fungal metabolites, and that conditioning them on knockout-derived per-lineage signatures closes a measurable part of that gap.

This claim is grounded in three documented failures. First, computational perturbation models are fragile under distribution shift: PertEval-scFM benchmarked five single-cell foundation models and found they failed to beat simple multi-layer perceptrons or GEARS at zero-shot perturbation prediction, with GEARS reaching an AUSPC of roughly 0.008 while the best foundation model (scGPT) trailed at about 0.045 — barely better than predicting the population mean. Second, none of these models has ever seen a natural product: the largest chemical perturbation resource, Tahoe-100M, covers 100 million cells but only 379 synthetic drugs across 50 cancer lines scPPDM, and the CRISPR benchmarks (Norman 2019 with 105 single-gene and 131 double-gene perturbations; Replogle 2022 with over 340,000 cells) are all 2D immortalized lines GEARS. Third, the closest existing study that does connect a natural product to human scRNA-seq and an in vivo mouse model — morusin in aortic dissection, which identified 84 overlapping targets and seven hub genes and validated efficacy in a BAPN mouse model Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection — performed no conditional knockout and no scRNA-seq of treated animals, so it cannot say whether the compound works on smooth muscle cells, fibroblasts, or infiltrating immune cells. The only study pairing conditional knockout with in vivo drug rescue, endothelial Dicer1 deletion with enoxacin in angiosarcoma Canonical microRNA loss drives tumor development implicating therapeutic efficacy of enoxacin in angiosarcoma, used a synthetic drug and bulk endpoints. No study combines a fungal metabolite, conditional knockout epistasis, and treatment scRNA-seq of living tissue.

Research Program

We choose aortic dissection as the test bed because the infrastructure exists on every side: human single-cell data from diseased tissue (GSE213740, with the hub genes MAPK8, MAPK1, PTGS2, and MAPK14 upregulated in fibroblasts, B cells, and T cells Integrated Network Pharmacology...), a reproducible BAPN mouse model, and a demonstrated in vivo endpoint set (ultrasonographic aortic dilation, dissection incidence, mortality). The program has four phases.

Phase one is candidate triage. We computationally map CMNPD's 72,343 recorded bioactivities against the dissection hub targets, prioritizing fungal-derived scaffolds — Aspergillus and allied marine taxa — with defined three-dimensional structures (79.9% of CMNPD records have generated 3D conformers CMNPD). Precedents like the versiquinazolines, coral-associated Aspergillus versicolor alkaloids with rare spiro-γ-lactone rings that selectively inhibit thioredoxin reductase Cheng et al., and the emestrin-type epipolythiodioxopiperazines from Aspergillus nidulans that act through PI3K/AKT and mitochondrial apoptosis Emestrin-type epipolythiodioxopiperazines..., show that redox- and MAPK-adjacent fungal chemistry exists and is structurally unlike anything in Tahoe-100M. Because niche fungi often yield only milligrams of product, we will use producer-side genetics to raise supply: deleting a dominant competing pathway, as the phoE polyketide synthase deletion that unlocked silent monoterpene pathways in ascidian-derived Diaporthe sp. SYSU-MS4722 Deletion of a polyketide synthase phoE..., to redirect flux toward our chosen scaffolds. Four to six compounds advance.

Phase two builds the in vitro single-cell perturbation atlas. Primary mouse vascular smooth muscle cells, fibroblasts, macrophages, and endothelial cells are treated with dose-response series of each compound and profiled by scRNA-seq, following the experimental guidelines established by the scPerturb harmonization effort — at least 300 cells per perturbation and about 1,000 UMI counts per cell to reliably separate a perturbation's effect from noise scPerturb: Harmonized Single-Cell Perturbation Data.

Phase three is the knockout epistasis panel, the core new instrument. We generate floxed alleles of the hub targets and cross them to lineage-specific Cre drivers: Mapk14 in myeloid cells, Mapk8 in fibroblasts, Ptgs2 in vascular smooth muscle, following the design logic of studies that proved cell autonomy this way — macrophage-specific MED1 deletion showing myeloid ROS cell-autonomously drives intimal hyperplasia Zhang et al., and sensory-neuron-specific MyD88 deletion showing neuron-autonomous signaling converts acute inflammation into chronic pain Liu et al.. Each knockout line and wild-type controls receive BAPN plus candidate compound or vehicle, scored on the same endpoints as the morusin study. A knockout that abolishes drug benefit while leaving others intact identifies the obligatory site of action. Crucially, because fungal metabolites are polypharmacological — morusin hits 84 disease-relevant targets across seven hubs — a single knockout may not abolish efficacy; if no single deletion does, we test combined deletions, turning the acknowledged tension between single-gene knockout precision and multi-target drugs Tian et al. into a measured property rather than a confound.

Phase four profiles aortic tissue from every cohort in phase three by scRNA-seq and builds the anchor-conditioned model. We define the per-lineage translation gap as

$$D(c,d) = 1 - \rho\big(\, s^{\text{vitro}}_{c,d},\; s^{\text{vivo}}_{c,d} \,\big),$$

where $\rho$ is the Spearman correlation between the differential-expression signature of compound $d$ in cell type $c$ measured in culture ($s^{\text{vitro}}$) and in living tissue ($s^{\text{vivo}}$). The model then predicts in vivo signatures by conditioning on two inputs the field has never combined: the in vitro perturbation signature and the knockout anchor — the scRNA-seq signature of deleting the compound's putative target in that lineage, in the style of GEARS, which gained 40% higher precision on genetic-interaction subtypes and a 53% improvement on zero-shot combinatorial prediction by conditioning on gene co-expression and Gene Ontology graphs GEARS. Chemical encoding uses the 3D conformers from CMNPD rather than flat Morgan fingerprints, since non-planar macrocyclic and epipolythiodioxopiperazine scaffolds are precisely what 2D fingerprints flatten away.

Baselines

The naive reference point is current practice: rank compounds by in vitro potency in bulk assays and advance the strongest, as the morusin study did with viability and qRT-PCR readouts in stimulated human aortic smooth muscle cells Integrated Network Pharmacology.... It is the strongest competitor in the trivial sense that it is what everyone actually does and it is nearly free. It will be run on the same compound panel with identical dose ranges. Beating it means demonstrating at least one compound whose in vitro rank and in vivo outcome diverge and are correctly predicted — the exact failure mode documented for matrine, curcumol, and geissoschizine methyl ether, all potent in culture with unproven in vivo efficacy Matrine..., Curcumol..., Geissoschizine methyl ether....

The statistical floor is the Mean baseline highlighted by PertEval-scFM, which predicts each cell type's response as the average in vitro response across cell types; scGPT barely beat it with an AUSPC near 0.045 versus roughly 0.008 for GEARS PertEval-scFM. Any model that cannot clear the mean is doing nothing.

The strongest existing competitors are scGen and scPPDM. scGen established that perturbation responses transfer across cell types, studies, and species by simple latent arithmetic, achieving $R^2$ values up to 0.97 — but only between in vitro systems scGen. scPPDM is the best published chemical-response model, outperforming its nearest competitor by 36.11% in Spearman correlation on differentially expressed genes for unseen drugs on Tahoe-100M scPPDM. Both will be run with the same data splits, the same held-out in vivo compounds, and an equal tuning budget, predicting in vivo per-cell-type responses from in vitro signatures. Beating them means higher top-1000 DEG accuracy per cell type on held-out compounds in vivo, plus correct calls on which knockout abolishes efficacy. A distribution-alignment baseline in the spirit of CELLECTION — which mapped in vitro organoids to in vivo fetal brain stages with Spearman correlations of 0.96–0.99 CELLECTION — will also be run, aligning in vitro and in vivo cell distributions by optimal transport before predicting drug responses.

The key ablation removes only the knockout anchors: the identical architecture trained on the same in vitro and in vivo data but without per-lineage knockout signatures as conditioning. This isolates whether the genetic anchors carry the gain or whether any in vivo training data would suffice. A second ablation swaps 3D conformer features for Morgan fingerprints to isolate the chemistry encoding.

Why not just…?

Why not just use bulk RNA-seq with computational deconvolution? Bulk sequencing is an order of magnitude cheaper and deeper, but it averages over cell types, and the averaging is known to destroy the signal this proposal depends on. CELLECTION directly tested this: models on cell-type proportions achieved an AUROC of 0.70 for Alzheimer's classification while single-cell distribution modeling reached 0.87 from glial populations alone, and pseudo-bulk averaging washed out disease-associated microglial sub-states entirely CELLECTION. Our epistasis score is defined per cell type per compound; a deconvoluted bulk signal cannot attribute a signature to the lineage whose knockout we are testing. As a concession to cost, we will compute pseudo-bulk per sorted cell type as a sanity check, but the primary readout must be single-cell.

Why not just do in vitro CRISPR knockouts in cell lines? This is the cheapest way to generate target-perturbation signatures at scale, and it is exactly what GEARS and PerturbedVAE consume. But it cannot answer the question at hand, for two documented reasons. Static cell lines lack the stroma, clearance, and immune crosstalk that determine in vivo outcomes — the core of the translation gap — and in vitro knockouts cannot distinguish cell-autonomous from non-cell-autonomous mechanisms, which is the entire point of the epistasis design; the angiosarcoma work needed endothelial Dicer1 deletion in living mice, not culture, to establish the phenotype and test systemic drug rescue Canonical microRNA loss.... That said, this alternative is a genuine threat to the expensive arm of the proposal, and we will test it head-on: if in vitro CRISPR signatures of the hub targets predict the in vivo epistasis outcomes as well as in vivo knockout anchors do, the in vivo knockout panel can be trimmed in future iterations. We consider that a valuable finding either way, and it is written into phase four as a decisive check rather than argued around.

Why not just run pre-trained models zero-shot? Because the evidence says they break under exactly the shift we impose. PertEval-scFM showed foundation models fail at zero-shot perturbation prediction, barely beating the mean baseline PertEval-scFM, and performance drops sharply under realistic lineage and drug-class distribution shifts. Our setting is a maximal shift: fungal scaffolds absent from Tahoe-100M's 379 synthetic drugs, primary cells rather than immortalized lines, and living tissue rather than plastic. The Perturbation Suppression Hypothesis identified why large models smooth over the sparse perturbation-responsive signal that dominates only 10% or less of reads What Makes a Representation Good for Single-Cell Perturbation Prediction?. Nevertheless, we will run scGen and scPPDM zero-shot as fair baselines, and if they match the anchor-conditioned model, the claim is falsified and we will report it.

Proposed Experiments and Decisive Checks

The experimental core is a $4 \times 2 \times 2$ design per compound: four cell types (smooth muscle, fibroblast, macrophage, endothelial), two conditions (in vitro culture, in vivo BAPN aorta), and two genotypes (wild-type, relevant knockout), with scRNA-seq on all in vivo cohorts and bulk-equivalent endpoints (aortic dilation, dissection incidence, mortality) matching the morusin study's instruments. No reported result found on Lacuna for quantitative aortic dilation values under morusin or any fungal metabolite in the BAPN model, so our own wild-type-plus-vehicle and morusin-positive-control arms establish the reference numbers. Three checks decide the proposal. First, whether any model trained only on in vitro data predicts in vivo per-cell-type signatures above the mean baseline — we expect not, per PertEval-scFM. Second, whether the anchor-conditioned model beats its anchor-free ablation on top-1000 DEG accuracy per cell type, the same metric scPPDM led by 28.57% on unseen drugs in vitro scPPDM — this is the direct test of whether genetic anchors, the new ingredient, carry real information. Third, whether the knockout panel resolves the polypharmacology question: a single lineage-specific deletion that abolishes efficacy proves a cell-autonomous obligatory target; absence of any single deletion, combined with additivity under double knockouts, quantifies multi-lineage synergy that single-gene knockout logic would have misread as "ineffective." Each result, including the negative ones, converts a vague translational disappointment into a number attached to a cell type and a target — which is what the field currently lacks.

References

  1. CMNPD: a comprehensive marine natural products database towards facilitating drug discovery from the ocean
  2. Matrine induces apoptosis of human multiple myeloma cells via activation of the mitochondrial pathway
  3. Curcumol Induces G1 Phase Arrest in SK-Hep-1 Cells by Targeting SKP2-Mediated p27 Degradation
  4. Geissoschizine methyl ether, a corynanthean-type indole alkaloid fromUncaria rhynchophyllaas a potential acetylcholinesterase inhibitor
  5. PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction
  6. scPPDM: A Diffusion Model for Single-Cell Drug–Response Prediction
  7. Predicting transcriptional outcomes of novel multigene perturbations with GEARS
  8. Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
  9. Canonical microRNA loss drives tumor development implicating therapeutic efficacy of enoxacin in angiosarcoma
  10. Versiquinazolines L–Q, new polycyclic alkaloids from the marine-derived fungus Aspergillus versicolor
  11. Emestrin-type epipolythiodioxopiperazines from Aspergillus nidulans with cytotoxic activities by regulating PI3K/AKT and mitochondrial apoptotic pathways
  12. Deletion of a polyketide synthase phoE induces monoterpenes production in ascidian-derived fungus Diaporthe sp. SYSU-MS4722
  13. scPerturb: Harmonized Single-Cell Perturbation Data
  14. MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation
  15. TLR signaling adaptor protein MyD88 in primary sensory neurons contributes to persistent inflammatory and neuropathic pain and neuroinflammation
  16. Conditional depletion of macrophages ameliorates cholestatic liver injury and fibrosis via lncRNA-H19
  17. Generative modeling and latent space arithmetics predict single-cell perturbation response across cell types, studies and species
  18. Predicting emergent phenotypes from single cell populations using CELLECTION
  19. What Makes a Representation Good for Single-Cell Perturbation Prediction?