Freezing the Junction: Click-Tagged Cross-Linking Recovers Intercellular Protein Complexes from Intact Tissue

Co-immunoprecipitation (Co-IP) is the standard way to prove that two proteins physically bind: an antibody against a chosen "bait" protein fishes it, and anything stuck to it, out of ground-up tissue, and mass spectrometry names the partners. Grinding the tissue first costs us two things at once. Weak, transient contacts fall apart during the dilution and washing steps. And all spatial information is destroyed — a Co-IP from a tumor lysate cannot tell you whether a captured complex sat inside one cell or bridged two touching cells, in the tumor core or the peritumoral stroma. Yet much of tissue regulation is exactly between cells: contact-dependent receptor–ligand pairs like PD-1/PD-L1 and Notch–Delta are physically severed the moment tissue is dissociated (Shao et al.), and dissociation itself strips membrane-tethered structures such as tunneling nanotubes that transfer organelles between resident cells (Cytoskeletal remodeling promotes tunneling nanotube formation and drives cardiac resident cell mitochondrial transfer in sepsis). The intercellular half of the physical interactome is, in practice, invisible to the gold-standard method for measuring physical interactions.

We propose to freeze protein complexes chemically while the tissue is still intact, then enrich them afterwards with strain-promoted azide–alkyne cycloaddition (SPAAC) — a copper-free "click" reaction in which a strained ring-shaped alkyne snaps onto an azide group under mild, living-system-compatible conditions. The key precedent is SHVIP, which cross-linked and click-enriched protein complexes inside intact cells and recovered 739 high-confidence interactions in infected fibroblast culture (Structural host-virus interactome profiling of intact infected cells). But SHVIP works only in 2D monolayers, and its selectivity depends on the biological quirk that the virus shuts off host protein synthesis. We port the freeze-then-grind logic to solid tissue, replace viral shutoff with cell-type-targeted chemical tags, and register every recovered interaction back to its tissue niche.

Core Claim

The proposal claims three things can be combined into one workflow that no existing method delivers. First, a membrane-impermeable cross-linker applied to fresh tissue sections joins only those proteins in physical contact at cell surfaces, before any dissociation or lysis — this is the ingredient that saves transient contacts. Second, one cell type's surface proteins carry azide tags, delivered either by an azide-armed antibody against a cell-type-specific marker or by metabolic feeding of azido-monosaccharides that decorate surface glycoproteins, the same glycan-targeting logic used by the self-assembled PROTAC system to reach O-GalNAcylated and O-GlcNAcylated proteins in living cells (Self-assembled PROTACs enable glycoproteins degradation in the living cells). Third, after dissociation, SPAAC ligates a biotin handle onto the azide-tagged complexes for pulldown, and mass spectrometry names every cross-linked partner. The decisive design element is a second, orthogonal tag (a tetrazine handle) placed on a neighboring cell type: a recovered complex carrying one partner from each handle is, by construction, an intercellular (trans) interaction between two specific cell types in contact — something no Co-IP, and no spatial method, can currently certify.

What exists nearby, and how it differs: SHVIP does in-cell cross-linking and click enrichment but only in homogeneous cell culture with no tissue coordinates and viral-shutoff-dependent selectivity (Structural host-virus interactome profiling of intact infected cells). Spatial proteomics models see location, not binding: KRONOS reads multiplexed fluorescence images but two proteins in the same optical patch may be biochemically unassociated, and it is limited to fixed antibody panels (A Foundation Model for Spatial Proteomics); STProtein predicts protein abundance at spatial spots, with no capacity to detect complex formation (STProtein: predicting spatial protein expression from multi-omics data); MELON integrates spatial RNA, ATAC, and antibody-tag abundance, again with no physical-interaction readout (MELON: Multimodal Learning Framework for Spatial Multimodal Omics Data Integration). HuRI mapped ~53,000 binary human interactions by testing pairs in yeast nuclei and layering them onto bulk tissue RNA — spatially blind and missing the mammalian post-translational modifications and transient membrane contacts we care about (A reference map of the human protein interactome).

Research Program

The chemistry phase must solve one known hard problem: SPAAC reagents have never been characterized for uniform diffusion into dense 3D tissue. Second-order rate constants of roughly $10^{-2}$ to $1\ \mathrm{M^{-1}s^{-1}}$ demand micromolar concentrations and long incubations; bulky cyclooctynes like DBCO are hydrophobic and bind nonspecifically; and strained alkynes are slowly quenched by cellular thiols. We will use the machine-learning models that predict ring-strain energies and reaction rates for strain-promoted reactions to shortlist faster, less hydrophobic cyclooctynes (Fast and Accurate Ring Strain Energy Predictions with Machine Learning and Application in Strain-Promoted Reactions), guided by the mechanistic orthogonality analysis of strain-promoted reactions (Origin of Orthogonality of Strain‐Promoted Click Reactions), and then measure labeling uniformity directly as a depth gradient of fluorescent cyclooctyne across cryosections. That SPAAC preserves the integrity of large complex proteins is already established: a biparatopic HER2-targeting antibody–drug conjugate built by copper-free click conjugation retained full specificity and stability (A biparatopic HER2-targeting ADC constructed via site-specific glycan conjugation exhibits superior stability, safety, and efficacy).

The computational phase registers every mass-spectrometry-validated interaction to its tissue niche. Adjacent sections of the same tissue block are imaged by multiplexed fluorescence, quality-gated by the KRONOS artifact detector (mean precision 0.98 versus 0.74 for the marker-expression baseline, A Foundation Model for Spatial Proteomics), and cell types are annotated with STELLAR, which reached 93% accuracy on Barrett's esophagus CODEX data and annotated 2.6 million cells across 64 HuBMAP sections (Annotation of Spatially Resolved Single-cell Data with STELLAR). A spatial graph-attention model in the spirit of STELLAR and MELON's neighborhood module is then trained on our labeled interaction–niche pairs to predict which tissue niche hosts a given intercellular pair. Crucially, we also run the tissue-level version of Squires' "roommate rule" test: they showed physical-interaction connectivity predicts subcellular organelle residence at roughly 70% accuracy with overwhelming statistical significance ($p = 10^{-404}$), but that co-localization alone is a "faint signal" wildly insufficient to predict physical binding (A Method of Integrating Spatial Proteomics and Protein-Protein Interaction Network Data). We ask the direct analogue: how many of our chemically frozen trans pairs would a co-localization-based predictor recover?

Baselines

The naive reference point is standard tissue Co-IP-MS: lyse first, no cross-linking, antibody pulldown, mass spectrometry. It is what the field does today, so any value we add is measured against it; it cannot, by construction, attribute an interaction to a cell pair or a niche. The strongest existing competitor is SHVIP's pipeline adapted as fairly as possible: dissociate the same tissue into a single-cell suspension, then apply the same cross-linker, the same azide tagging, the same SPAAC enrichment, the same LC-MS/MS time and false-discovery-rate thresholds. SHVIP reported 739 high-confidence PPIs (441 virus–host) supported by 6,194 cross-links in HSV-1-infected fibroblast culture (Structural host-virus interactome profiling of intact infected cells); no reported result found on Lacuna for any lysis-first or dissociation-first method recovering intercellular complexes from intact tissue, so the head-to-head numbers must be generated in this project. Beating it means recovering a higher fraction of a pre-registered panel of known juxtacrine pairs (PD-1/PD-L1, Notch–Delta, integrin–ECM in tumor/immune tissue) and producing correct cis/trans and cell-type attribution, which the dissociated baseline cannot produce at all because the contacts are already destroyed. The third baseline is the proximity ligation assay (Foster et al.; Fu), which detects two proteins within 40 nm in intact fixed tissue: it is the intact-tissue gold standard, and we run it as an independent oracle on every pair we recover, and on the pre-registered panel as the best targeted method. Two ablations isolate the proposal's key ingredients: cross-linking after dissociation instead of before (isolates the freeze-in-tissue step), and uniform tagging without cell-type selectivity (isolates the attribution step).

Why not just…?

Why not just microdissect regions and Co-IP each one? Laser-capture microdissection preserves region-of-origin, which is a real strength, but it still lyses and mixes every cell inside the captured region, so per-pair intercellular attribution is impossible and post-lysis recombination within the region is uncontrolled — our mixing control below quantifies exactly this failure mode. We will nonetheless compare niche-level assignment against an LCM-style regional baseline, because if regional resolution answers the biological question, the simpler method should win and we would report that honestly.

Why not just do proximity ligation or multiplexed imaging of everything? The proximity ligation assay requires a pre-selected, pair-matched antibody duo for every candidate interaction — it is confirmation, not discovery. Multiplexed imaging of the KRONOS kind sees two proteins in the same ~200 nm optical patch and is capped at fixed antibody panels (175 markers in KRONOS's pretraining corpus, A Foundation Model for Spatial Proteomics); co-localization is not binding. Imaging is therefore our validation layer, not our discovery engine.

Why not just compute it — overlay HuRI on spatial maps? This is the most serious alternative because it is nearly free. Two published results argue it falls short: Squires' faint-signal asymmetry shows co-localization is necessary but wildly insufficient for predicting physical binding (A Method of Integrating Spatial Proteomics and Protein-Protein Interaction Network Data), and HuRI's ~53,000 binary pairs were scored heterologously in yeast nuclei against bulk tissue RNA, missing post-translational-modification-dependent and transient membrane complexes (A reference map of the human protein interactome). We will settle it empirically: if only a small fraction of our validated trans pairs are absent from HuRI, the computational shortcut is a genuine threat to the proposal's value, and we will say so in the results.

Proposed Experiments and Decisive Checks

Experiment one measures SPAAC tissue penetration: fluorescent cyclooctyne incubated into fresh tissue sections, quantified as a labeling-depth gradient, comparing machine-learning-shortlisted cyclooctynes against DBCO, with a no-azide control to price in hydrophobic nonspecific background. Experiment two is the head-to-head on the pre-registered panel of known juxtacrine pairs in tumor/immune tissue: recovery fraction for in-tissue freezing versus dissociate-then-crosslink versus standard lysate Co-IP, all on the same tissue, same mass-spec time, same thresholds. Experiment three is the discovery run on tonsil or lymph node — cohorts covered by the spatial-proteomics benchmarks above — registered to STELLAR-annotated CODEX maps, reporting the number of validated trans pairs per niche and the fraction absent from HuRI. Experiment four is the post-lysis recombination control: azide-tagged cells of one type mixed with untagged cells of another after dissociation, then cross-linked, counting spurious trans calls — a direct measurement of the compartmentalization paradox for intercellular complexes. Experiment five evaluates the niche-prediction model on held-out validated pairs against a co-localization-only predictor, with an abundance-only reference in the style of STProtein (which reported RMSE 0.95 on the Mouse Spleen SPOTS dataset, STProtein: predicting spatial protein expression from multi-omics data) to show what abundance models can and cannot contribute. The decisive check for the whole proposal is experiment two: if freezing in tissue recovers no more known intercellular pairs than the dissociated SHVIP baseline, the central ingredient fails, and the honest conclusion would be that dissociation artifacts do not matter for this class of complexes.