When a doctor draws blood to measure immune activity, they are sampling the wrong place more often than not. Matched-sampling studies show that local immune responses in the lung, brain, gut, and tumors can run in the opposite direction from what peripheral blood shows: systemically delivered steroids suppress inflammatory genes in blood while lung cells keep interferon signaling fully active, and memory T cells parked in tissue predict melanoma survival while their circulating cousins predict nothing. In parallel, a separate community has built machine-learning models that predict how a gene knockdown reshapes a cell's transcriptome, and protein biochemists can measure which proteins physically bind each other using co-immunoprecipitation, an assay where an antibody fishing out one protein drags down its binding partners. These three threads have never been joined, and the join matters: existing knockdown predictors see only messenger RNA levels, so they are blind to a failure mode where silencing one gene dissolves a multi-protein machine without touching the mRNA of its partners — and they assume the same intracellular wiring in every cell, which compartmentalized immunity data flatly contradicts.
I propose to build and test a perturbation predictor whose operators are conditioned on both a tissue-compartment context and a physical protein-interaction graph, so that a single knockdown yields two distinct outputs: the familiar messenger RNA shift, and a map of which protein complexes are predicted to fall apart. The same knockdown should produce different rewiring in a tissue-resident immune cell than in its circulating counterpart. The computational claims are testable on existing Perturb-seq benchmarks; the central biological claim — that some knockdown effects live entirely at the protein-complex level and are compartment-specific — is testable with small interfering RNA (siRNA, a transient gene-silencing reagent) followed by co-immunoprecipitation and proximity assays in matched resident-versus-circulating primary cells.
Core Claim
Three findings in the literature set up the claim. First, perturbation simulators work only at the transcript level: GEARS predicts combinatorial knockout outcomes from gene co-expression and ontology graphs, DYNAMO composes per-gene perturbation operators as learned embeddings, and scDEBART pretrains on 6.28 million differential-expression profiles; none of them carries a protein-interaction layer, so a knockdown that destabilizes a protein complex without changing partner mRNA is invisible to all of them. Second, PINNACLE shows protein interactions are compartment-dependent: it contextualizes a global interactome across 156 cell types and 24 tissues from the Tabula Sapiens atlas, beating context-free baselines on target nomination, but it is purely observational — it cannot simulate any perturbation. Third, DYNAMO's best cross-cell-type transfer (correlation 0.576 from K562 leukemia cells to RPE1 retinal cells) relies on a frozen network topology, yet compartmentalized immunity data show the same pathway operating in opposite directions in lung versus blood of the same patient. My claim is that conditioning perturbation operators on compartment-specific protein-interaction structure closes both gaps at once: it adds the protein-complex failure mode that mRNA-only models miss, and it replaces the frozen-wiring assumption with wiring that varies by compartment.
Research Program
The model, which I will call a compartment-conditioned perturbation operator, extends the operator-composition idea of DYNAMO with two changes. First, each per-gene operator is conditioned on a compartment embedding drawn from PINNACLE's contextualized protein representations, so the operator for knocking down gene $g$ in an alveolar macrophage differs from the operator for the same gene in a circulating monocyte. Second, alongside the messenger RNA delta, the model predicts an edge-level destabilization score for every protein interaction involving the knocked-down protein's products:
$$q_{uv} = \sigma\!\left(\mathrm{MLP}\big(z_g \,\oplus\, e_c \,\oplus\, d\big)\right)$$
where $z_g$ is the learned embedding of the silenced gene, $e_c$ is the compartment embedding, $d$ is the knockdown efficacy (the fraction of protein depleted, since siRNA rarely achieves full depletion), and $q_{uv}$ is the probability that the physical interaction between proteins $u$ and $v$ is lost. Training uses the same public Perturb-seq corpora the existing models use — the Norman combinatorial screen and the Replogle K562 and RPE1 screens — plus differential-expression pretraining on the CELLxGENE corpus following scDEBART, with PINNACLE's Tabula Sapiens-derived graphs supplying the interaction structure. Because the training perturbations are CRISPR-based while our validation uses graded siRNA, knockdown efficacy enters as an explicit input and partial depletion is treated as interpolation along the dose axis, not as noise.
The wet-lab arm turns the model's unique output into a falsifiable prediction. The melanoma tissue-atlas study gives a concrete entry point: it defines a CD4+ tissue-resident memory program (the CD69-high, KLF2/S1PR1-low state) whose members differentiate into cytotoxic effectors in tissue but not in blood. From the model I will nominate the residency regulators whose knockdown is predicted to dissolve the largest set of protein interactions in resident-like cells while sparing those same interactions in circulating cells. Primary human T cells will be split into resident-like and circulating conditions, silenced with siRNA at two or three doses, and measured. Because classical co-immunoprecipitation requires roughly $10^6$–$10^7$ cells per pull-down — prohibitive for rare primary resident cells — the assay plan is tiered: targeted co-immunoprecipitation with Western blot detection where biomass allows, following the bait-prey validation logic used in the IQGAP3–GLI1 lung-cancer study, and proximity ligation assays, which read out protein-pair proximity in intact single cells, for the low-input and spatially resolved measurements.
Proposed Experiments
The first experiment is the standard benchmark suite, run fairly. On the Norman combinatorial benchmark, GEARS reports 40% higher precision than prior approaches on four genetic-interaction subtypes, and DYNAMO reports 2.7-times-better prediction of unseen double perturbations than GEARS; on Replogle K562 unseen genes, DYNAMO reports a zero-shot correlation of $\rho_\Delta = 0.283$ (dropping to 0.108 for genes lacking regulatory-network coverage), and scDEBART reports a mean Enrichment Factor of 11.96 on Perturb-seq benchmarks versus 1.74–2.99 for scGPT and GEARS, plus 71.4% top-1 accuracy on reverse knockout identification where prior models were near zero. The compartment-conditioned model must at least match these numbers on the messenger RNA task — the protein layer is an addition, not an excuse for worse mRNA prediction — and the decisive comparison is a stratified split: perturbations whose targets sit in dense protein complexes (hubs) versus those that do not. The hypothesis is that gains concentrate on hub knockdowns, where mRNA-only models systematically underpredict functional damage.
The second experiment tests the frozen-wiring tension head-on. DYNAMO's K562-to-RPE1 transfer correlation of 0.576 is the strongest evidence that a fixed network suffices — but that transfer is between immortalized cell lines, not between physiological compartments of one organism, and the dexamethasone study shows within-patient lung-versus-blood discordance exceeding anything a frozen topology could represent. I will run the analogous transfer between compartment-matched and compartment-mismatched contexts and ask whether conditioning on $e_c$ recovers accuracy where the frozen model degrades. If the frozen model matches the conditioned one here, the compartment-conditioning ingredient is not needed and the proposal's second half fails.
The third experiment is the wet-lab validation. For each nominated knockdown, the model outputs a ranked list of protein interactions predicted to be lost, separately for resident-like and circulating cells. The co-immunoprecipitation and proximity-ligation measurements score these lists by precision at the tested cutoff and by whether the compartment contrast (loss in resident, sparing in circulating) is reproduced. No reported result exists on Lacuna for predicting protein-complex rewiring after knockdown in any cell type, so this experiment establishes its own reference numbers; success is pre-registered as at least half of the top-ranked predicted losses being measurably reduced, with the compartment contrast holding at the same siRNA dose. As a sanity anchor for the siRNA-plus-readout pipeline itself, the RhoA study provides a worked precedent: siRNA silencing of RhoA in LPS-injured lung epithelial cells cut apoptosis from roughly 35% to about 12%, showing that graded silencing with phenotypic readouts behaves as expected in a lung-relevant system.
Baselines
The naive reference point is the zero-change predictor: for a knockdown, predict no messenger RNA shift and no interaction loss. It defines the floor, and it is not trivial — a large fraction of Perturb-seq perturbations produce small deltas, so a model must beat it meaningfully, not marginally. The strongest existing competitor depends on the task. For messenger RNA prediction, it is scDEBART, which reports a mean Enrichment Factor of 11.96 on Perturb-seq benchmarks and 71.4% top-1 accuracy on reverse knockout identification; I will run its public pipeline on the same held-out splits, with the same pretraining corpus access and the same tuning budget, and beating it means a higher Enrichment Factor on the same held-out perturbations with the stratified hub-versus-non-hub breakdown reported alongside. For cross-context transfer, the competitor is DYNAMO, which reports 0.576 for K562-to-RPE1 transfer without retraining and 2.7-times-better double-perturbation prediction than GEARS; beating it means higher transfer correlation under compartment mismatch specifically. For the interaction-loss task, no existing model makes such predictions, so the strongest surrogate baseline is a two-stage pipeline: PINNACLE's static contextual interaction scores re-weighted by the predicted messenger RNA delta, which tests whether a simple product of existing components already captures complex destabilization. If it does, the learned operator adds nothing. Finally, the key ablation is the full model with the compartment-conditioned interactome replaced by the context-free global interactome (BioGRID/HuRI as used in PINNACLE's baseline configuration); PINNACLE's own results motivate this ablation — it outperformed a random-walk baseline in 100% of cell types and GAT in 44.2% at APR@10 for rheumatoid arthritis target nomination, and context-free embeddings gave a PD-1/PD-L1 binding-discrimination score gap of 0.003 versus 0.011 for contextualized ones — and the ablation isolates whether compartment conditioning, rather than architecture or extra parameters, drives any observed gain.
Why not just…?
Why not just profile peripheral blood instead of perturbing hard-to-get tissue-resident cells? Because the blood readout is demonstrably the wrong signal for exactly this question. The dexamethasone study found blood transcriptomics showing over 4,000 differentially expressed genes and apparent immune suppression while lung cells maintained steroid-resistant interferon signaling; the melanoma atlas study found circulating memory signatures carried no survival signal while resident signatures strongly predicted it; the liver-failure study found blood CD8+ T cells failed to reflect dysfunction of lymphocytes trapped in the liver; and the Atlanta CSF study documented cytokine storms confined to cerebrospinal fluid and invisible in serum. Blood profiling cannot substitute for a compartment-resolved perturbation experiment — it is the phenomenon being escaped, not a cheap proxy for it.
Why not just trust the in silico models and skip the co-immunoprecipitation work? The computational evidence itself says the models are not trustworthy on this question. GEARS and scGPT reach only Enrichment Factors of 1.74–2.99 on Perturb-seq benchmarks, GEARS reportedly crashed on unseen genes in DYNAMO's evaluation, and DYNAMO's correlation falls to 0.108 for genes without regulatory-network coverage — and none of these models has any representation of physical binding. The IQGAP3 study is the cautionary precedent in reverse: transcriptomic data alone could not establish that IQGAP3 drives GLI1 directly rather than through intermediate steps, and only co-immunoprecipitation settled it. My model's interaction-loss predictions are precisely its most novel and least verified output; without the wet-lab arm the proposal would be an untestable architecture change. This is a real cost, and the tiered assay design (proximity ligation before bulk pull-downs) exists to keep it affordable.
Why not just use CRISPR knockouts instead of siRNA? CRISPR gives complete, permanent deletion, which sounds cleaner, but the Selective Gene Knockdown Using Small Interfering RNAs paper argues permanent deletion is a heavy-handed intervention that is slow and often lethal for essential genes, whereas siRNA offers tunable, transient, dosage-dependent silencing. The dosage axis is not a convenience here — it is the variable my model takes as input ($d$), and graded depletion is what lets us test whether complex dissolution scales with residual protein level, which a binary knockout cannot distinguish. The honest weakness is that the training corpora are CRISPR-based, so the siRNA-CRISPR mismatch is itself a transfer problem; the wet-lab experiment tests it directly, and if predictions only hold at full depletion, the dose-conditioning claim should be abandoned.
Why not just use the global BioGRID/HuRI interactome as the graph? Because compartment-independent interaction maps are known to carry false positives and negatives exactly where this proposal needs fidelity: PINNACLE's authors state plainly that the human interactome they start from is not measured in a cell-type-specific manner, and their contextualization beat context-free baselines across target-nomination tasks. This alternative is fully covered by the planned ablation rather than argued away — if the context-free graph performs identically under compartment mismatch, the simpler graph wins and the compartment conditioning should be dropped.
Decisive Checks
The proposal stands or falls on three pre-registered outcomes. If the hub-versus-non-hub stratification shows no differential gain over scDEBART and DYNAMO, the protein layer adds nothing measurable even in silico. If the context-free ablation matches the conditioned model on compartment-mismatched transfer, the frozen-wiring assumption survives and Gap 4 is empirically refuted rather than open. And if the co-immunoprecipitation and proximity-ligation measurements show no compartment contrast in interaction loss at matched siRNA doses, then the central biological claim — that the same knockdown rewires resident and circulating cells differently at the protein-complex level — is wrong, and the model, however accurate on benchmarks, would be describing a phenomenon that does not exist.