When a biologist wants to know what a gene does, the fast answer comes from a plastic dish and the trustworthy answer comes from a mouse, and the two routinely disagree. Fast genome-scale screens in cultured cells now feed machine-learning models that predict how a cell responds when a gene is silenced, yet these models are trained and evaluated entirely on dish data, and the field has good evidence that dish-derived regulatory signatures do not survive the move into a living animal. At the same time, the gold-standard alternative — breeding mice in which a gene is deleted only in one cell type — is so slow and costly that it is limited to one or two genes per study, which is nowhere near enough data to teach or test a predictive model. This proposal argues that the two problems solve each other, if one additional confound is removed: the stress caused by how the perturbation is delivered.
The delivery confound is now measurable. A recent study showed that fusogenic liposomes — particles decorated with peptides that merge directly with the cell's outer membrane, dumping their cargo into the cytosol without ever entering the endosomal pathway — transfer mRNA into cells and into living zebrafish brains with more than 80% fewer dead cells and roughly 400-fold less interleukin-6 induction than standard lipofection, and with 35% of telencephalon cells transfected in vivo versus 5.3% for lipofection (Smuggling on the Nanoscale—Fusogenic Liposomes Enable Efficient RNA-Transfer with Negligible Immune Response In Vitro and In Vivo). That study delivered only a fluorescent reporter and read out bulk cytokine PCR. The core claim here is that the same fusogenic platform can carry real genome-editing machinery — Cre recombinase mRNA or Cas9 ribonucleoproteins — into the same cell type in a dish and inside a mouse, producing matched single-cell perturbation pairs on both sides of the translation gap, and that these pairs are exactly the calibration data needed to build and honestly evaluate a model that predicts in vivo multi-omic knockout responses from in vitro ones.
Core Claim
The in vitro–in vivo gap for perturbation prediction decomposes into at least three separable causes: the inflammatory and cytotoxic footprint of the delivery method itself, the missing chromatin-regulatory layer in most models, and genuine microenvironmental context shift. No existing dataset or model isolates these, because every large perturbation resource is dish-only, bulk or RNA-only, and delivered by endosomal uptake (SCALE:Scalable Conditional Atlas-Level Endpoint transport for virtual cell perturbation prediction, Elucidating the Design Space of Generative Models for Single-Cell Perturbation Prediction, Cell-Level Virtual Screening), while every in vivo knockout study is low-throughput and bulk-profiled (Induction of Thoracic Aortic Remodeling by Endothelial-Specific Deletion of MicroRNA-21 in Mice, MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation, Clec7a Worsens Long‐Term Outcomes after Ischemic Stroke by Aggravating Microglia‐Mediated Synapse Elimination). By using membrane-fusing delivery to apply the same edit to the same lineage on both sides of the boundary, and by profiling both sides with paired single-cell chromatin accessibility plus transcriptome, we can measure how much of the gap each cause explains, and train a transfer model on the residual. If the matched-pair dataset shows the gap mostly disappears once delivery stress and regulatory layer are fixed, that reframes the entire modeling agenda; if it does not, we finally have the labeled data to learn what context shift actually looks like.
Research Program
The physical ingredient comes from peptide biophysics. Fusogenic delivery works through shallow amphipathic helices that insert a few ångströms into the bilayer's outer leaflet, the same geometry mapped in atomic detail for the synaptic-vesicle protein α-synuclein, whose N-terminal anchor buries about 8 Å into the hydrocarbon core at a 12-degree tilt (Structural Ensembles of Membrane-bound α-Synuclein Reveal the Molecular Determinants of Synaptic Vesicle Affinity), and whose binding affinity can be tuned without changing the bound structure through N-terminal acetylation, an entropic pre-shaping of the free peptide (The N-terminal Acetylation of α-Synuclein Changes the Affinity for Lipid Membranes but not the Structural Properties of the Bound State). Continuum and coarse-grained simulations put the energetic cost of a monolayer–bilayer fusion stalk at roughly 25–33 $k_B T$, below the 44 $k_B T$ of full bilayer–bilayer fusion, meaning the fusion pathway used for delivery sits in a regime of controlled, partial membrane remodeling rather than lysis (A Model of Lipid Monolayer–Bilayer Fusion of Lipid Droplets and Peroxisomes). One real risk must be engineered around, not ignored: membrane-anchored amphipathic helices expose their unbound hydrophobic segments to solvent, which for α-synuclein's aggregation-prone NAC region accelerates fibril nucleation in crowded environments (Structural Ensembles of Membrane-bound α-Synuclein...). We therefore work with short lipid-targeting helices of the CIDEA type, whose C-terminal helix (residues 163–180) binds phosphatidic acid to destabilize the monolayer locally (The brown adipocyte protein CIDEA promotes lipid droplet fusion via a phosphatidic acid-binding amphipathic helix), rather than any α-synuclein-derived sequence, and we track aggregation markers as a safety readout.
The experimental program has three phases. First, quantify the delivery confound: deliver Cre mRNA to floxed-reporter cells in culture side by side through fusogenic liposomes and lipofection, and profile both plus untreated controls with paired single-cell accessibility and transcriptome sequencing, following the scCAT-seq principle of physically splitting each cell into nuclear and cytoplasmic fractions for joint measurement, which yields on the order of 210,000 chromatin fragments and 4.6 million RNA reads per cell (Deconvolution of single-cell multi-omics layers reveals regulatory heterogeneity). The prediction, from the fusogenic-liposome study's 400-fold cytokine reduction, is that fusion-delivered cells are chromatin-indistinguishable from untreated cells while lipofected cells carry a measurable inflammatory accessibility footprint. Second, generate matched pairs: choose three lineages with published conditional-knockout ground truth — endothelium (miR-21), macrophages (MED1), and microglia (Clec7a) — and use fusogenic delivery of Cas9 ribonucleoproteins against a panel of roughly 30 genes, including the published targets, applied to primary cultures of each lineage in vitro and to the same lineage in vivo in mice, profiling every condition with paired multi-omics. Partial in vivo transfection (35% in the zebrafish telencephalon data) is handled by gating on edit-bearing cells, and the untransfected neighbors serve as an internal within-animal control. Third, build the transfer model: conditioned on the in vitro paired response and a chromatin reference for the target lineage (the adult mouse brain atlas of 2.3 million nuclei provides exactly this kind of cell-type-specific accessibility background for neural lineages; Single-cell analysis of chromatin accessibility in the adult mouse brain), predict the in vivo paired response. Architecturally we follow two lessons from the file: discrete tokenization of counts, which cut distribution-distance metrics by roughly 3–20× relative to continuous latents (Elucidating the Design Space...), and a context-adaptive readout, the ingredient whose removal dropped DeepSADR's average clinical AUC from 0.856 to 0.662 (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction). Evaluation is on genes and lineages held out entirely from training, plus retrodiction of published knockout phenotypes.
Baselines
The naive reference point is zero-shot carryover: take the in vitro measured or predicted response for a held-out gene and use it, unchanged, as the in vivo prediction. This "do nothing" baseline establishes how large the raw gap is in our matched data, and it has a precedent number worth keeping in mind: when an accessibility model trained on the mouse brain atlas was transferred zero-shot to human brain, it retained an AUROC of 0.75 (Single-cell analysis of chromatin accessibility...) — naive cross-context transfer degrades but does not collapse, so our bar must be meaningfully above naive, not merely nonzero.
The strongest existing competitor is the DeepSADR architecture, which to date is the only published model in the file that explicitly learns the dish-to-patient domain shift, reaching a clinical Sorafenib response AUC of 0.957 versus 0.811 for the best baseline (DeepSADR). We will reimplement it faithfully — fragment–pathway bipartite graphs, supervised graph autoencoder, Set-Transformer adaptive readout — retrained on our matched dataset with the same compute, data budget, and hyperparameter search as our model, and evaluated on the same held-out genes. DeepSADR is the strongest competitor precisely because its own ablation shows the adaptive readout is doing the transfer work; our claim is that its bulk, 1,426-gene pathway compression discards the single-cell and chromatin structure where the transfer signal lives, and only a head-to-head retrain can settle that. Beating it means higher differentially-expressed-gene overlap and retrodiction sign accuracy on held-out genes at matched budget.
Two further baselines sharpen the comparison. An in vitro-only perturbation engine of the SCALE type, which improved perturbation-correlation by 12.02% and DE overlap by 10.66% over its predecessor on the Tahoe-100M atlas (SCALE), is trained on our dish data and then evaluated in vivo without any transfer component, isolating how much of the gap is pure domain shift rather than prediction error. And a network-rewiring baseline of the CellVS-Net type, which maps perturbations to gene–gene dependency graphs and achieved a Hits@1 of 0.1250 on drug–disease retrieval against 0.0357 for molecular fingerprints and 0.0179 random (Cell-Level Virtual Screening), tests whether the transferable signal lives in correlation-structure rewiring rather than expression levels — a genuine unresolved tension in the literature, since that paper showed even oracle measured expression underperforms network rewiring on retrieval. Finally, two ablations isolate our key ingredients: the full model trained RNA-only (no accessibility input) isolates the multi-omic contribution, and the full model trained on lipofection-derived pairs instead of fusion-derived pairs isolates the delivery-confound correction.
Why not just…?
Why not just use standard lipofection or viral vectors to deliver the editing machinery? The head-to-head evidence is direct: in the fusogenic-liposome study, lipofection caused more than 80% additional cell death, roughly 400-fold higher interleukin-6 induction, and in vivo produced inflammation exceeding bacterial LPS while transfecting 5.3% of telencephalon cells versus 35% for fusion (Smuggling on the Nanoscale). For this proposal the point is subtler than efficacy: lipofection-stressed cells would contaminate the very confound we are trying to isolate, making the matched pairs useless as calibration data. Our first experiment measures this directly, and if fusion delivery also leaves a chromatin footprint, we will report it and the calibration logic fails honestly.
Why not just breed conventional Cre-loxP conditional knockout mice? For any single gene this remains the gold standard, and we do not replace it — we use published Cre-loxP results as retrodiction targets (miR-21 endothelial knockout dropping diastolic pressure from about 87 to about 72 mmHg and raising aortic opening angle from 119.1° to 161.3°; Induction of Thoracic Aortic Remodeling...). But breeding is capped at one or two genes per study, takes months per line, and invites developmental compensation; it cannot produce the dozens of matched perturbation pairs a transfer model needs. In vivo pooled screening exists but remains technically difficult, as the zebrafish mosaic-transgenesis work makes clear (Library transgenesis in zebrafish through delayed site-specific mosaic integration for in vivo pooled screening of transgenes); fusogenic delivery is our proposed middle path between the two throughput regimes, and phase two is essentially a test of whether that middle path works at all.
Why not just profile RNA-seq and skip the chromatin layer? Because the transcriptome is the output, not the engine. scCAT-seq showed that peak-to-gene links from co-measured cells resolve cell states and master regulators that transcriptomic clustering alone cannot, and validate at higher rates against physical chromatin loops (Deconvolution of single-cell multi-omics layers...). Still, this is an empirical question inside the proposal: the RNA-only ablation is exactly the experiment that would show the accessibility layer is not paying for itself, and if it shows no transfer gain we will conclude the cheaper assay suffices for this task.
Why not just fine-tune DeepSADR or CellVS-Net as they are? This is the most serious alternative, and we treat it as such rather than arguing around it. DeepSADR already achieved strong clinical transfer with an adaptive readout, and CellVS-Net's network-rewiring signal beat oracle expression on retrieval. But both are bulk, PCA- or pathway-compressed, and built for drug response rather than genome editing, and CellVS-Net's benchmarks assume every compound is a pure inhibitor equivalent to a knockdown, a label-noise problem its own authors flag (Cell-Level Virtual Screening). Rather than assert our design is better, we retrain both architectures on our matched data under matched budgets; if a DeepSADR-style bulk pathway model matches our paired single-cell model on held-out-gene transfer and retrodiction, the honest conclusion is that the multi-omic single-cell machinery is unnecessary for this task, and the matched dataset will still be the first resource capable of showing that.
Decisive Checks
Three results decide the proposal. First, the confound measurement: fusion-delivered cells must be statistically indistinguishable from untreated cells in both accessibility and transcription at a defined false-discovery threshold, while lipofected cells show the predicted inflammatory programs; if fusion also perturbs chromatin, the clean-perturbation premise is wrong and we say so. Second, the transfer test on genes never seen in training: our full model must beat both the zero-shot carryover baseline and the retrained DeepSADR on DE overlap and on sign accuracy of predicted in vivo responses, with the RNA-only and lipofection-trained ablations quantifying how much of the improvement comes from each ingredient. Third, retrodiction: from in vitro data alone, the model must recover the direction and approximate magnitude of at least the published miR-21 hemodynamic phenotype and the Clec7a effect on infarct volume and behavioral recovery (Clec7a Worsens Long‐Term Outcomes...). For orientation on the computational side, the in vitro predictor will be validated against the Replogle CRISPRi screen conventions used by SCALE and the eVAE study before any transfer is attempted; no reported result found on Lacuna for paired multi-omic in vivo knockout transfer, which is precisely the empty slot this proposal fills.
References
- Smuggling on the Nanoscale—Fusogenic Liposomes Enable Efficient RNA-Transfer with Negligible Immune Response In Vitro and In Vivo
- SCALE:Scalable Conditional Atlas-Level Endpoint transport for virtual cell perturbation prediction
- Elucidating the Design Space of Generative Models for Single-Cell Perturbation Prediction
- Cell-Level Virtual Screening
- Induction of Thoracic Aortic Remodeling by Endothelial-Specific Deletion of MicroRNA-21 in Mice
- MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation
- Clec7a Worsens Long‐Term Outcomes after Ischemic Stroke by Aggravating Microglia‐Mediated Synapse Elimination
- Structural Ensembles of Membrane-bound α-Synuclein Reveal the Molecular Determinants of Synaptic Vesicle Affinity
- The N-terminal Acetylation of α-Synuclein Changes the Affinity for Lipid Membranes but not the Structural Properties of the Bound State
- A Model of Lipid Monolayer–Bilayer Fusion of Lipid Droplets and Peroxisomes
- The brown adipocyte protein CIDEA promotes lipid droplet fusion via a phosphatidic acid-binding amphipathic helix
- Deconvolution of single-cell multi-omics layers reveals regulatory heterogeneity
- Single-cell analysis of chromatin accessibility in the adult mouse brain
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Library transgenesis in zebrafish through delayed site-specific mosaic integration for in vivo pooled screening of transgenes